Brief Reports

¹Medical College of Wisconsin Affiliated Hospitals, Milwaukee, Wisconsin. At the time of this study, Dr. Bergl was with the Division of General Internal Medicine, Medical College of Wisconsin, Milwaukee, Wisconsin. ²Medical College of Wisconsin, Milwaukee, Wisconsin.Physical examination (PE) is a core clinical skill in undergraduate medical education.¹ Although the optimal approach to teaching clinical skills is debated, robust preclinical curricula should generally be followed by iterative skill development during clinical rotations.^2,3

The internal medicine rotation represents a critical time to enhance PE skills. Diagnostic decision making and PE are highly prioritized competencies for the internal medicine clerkship,⁴ and students will likely utilize many core examination skills^1,2 during this time. Bedside teaching of PE during the internal medicine service also provides an opportunity for students to receive feedback based on direct observation,⁵ a sine qua non of competency-based assessment.

Unfortunately, current internal medicine training environments limit opportunities for workplace-based instruction in PE. Recent studies suggest diminishing time spent on bedside patient care and teaching, with computer-based “indirect patient care” dominating much of the clinical workday of internal medicine services.^6-8 However, the literature does not delineate how often medical students are enhancing their PE skills during clinical rotations or describe how the educational environment may influence PE teaching.

We aimed to describe the content and context of PE instruction during the internal medicine clerkship workflow. Specifically, we sought to explore what strategies physician team members used to teach PE to students. We also sought to describe factors in the inpatient learning environment that might explain why physical examination (PE) instruction occurs infrequently.

We conducted a prospective mixed-methods study using time motion analysis, checklists on clinical teaching, and daily open-ended observations written by a trained observer from June through August 2015 at a single academic medical center. Subjects were recruited from internal medicine teaching teams and were allowed to opt out. Teaching teams had 2 formats: (1) traditional team with an attending physician (hospitalist or general internist), a senior resident, 2 interns, a fourth-year medical student, and 2 third-year students or (2) hospitalist team in which a third-year student works directly with a hospitalist and advanced practitioner. The proposal was submitted to the Medical College of Wisconsin Institutional Review Board and deemed exempt from further review.

All observations were carried out by a single investigator (A.T.), who was a second-year medical student at the time. To train this observer and to pilot the data collection instruments, our lead investigator (P.B.) directly supervised our observer on 4 separate occasions, totaling over 12 hours of mentored co-observation. Immediately after each training session, both investigators (A.T. and P.B.) debriefed to compare notes, to review checklists on recorded observations, and to discuss areas of uncertainty. During the training period, formal metrics of agreement (eg, kappa coefficients) were not gathered, as data collection instruments were still being refined.

Observation periods were centered on third-year medical students and their interactions with patients and members of the teaching team. Observed activities included pre-rounding, teaching rounds with the attending physician, and new patient admissions during call days. Observations generally occurred between the hours of 7 AM and 6 PM, and we limited periods of observation to 3 consecutive hours to minimize observer fatigue. Observation periods were selected to maximize the number of subjects and teams observed, to adequately capture pre-rounding and new admissions activities, and to account for variations in rounding styles throughout the call cycle. Teams were excluded if a member of the study team was an attending physician on the clinical team or if any member of the patient care team had opted out of the study.

Data were collected on paper checklists that included idealized bedside teaching activities around PE. Teaching activities were identified through a review of relevant literature^9,10 and were further informed by our senior investigator’s own experience with faculty development in this area¹¹ and team members’ attendance at bedside teaching workshops. At the end of each day, our observer also wrote brief observations that summarized factors affecting bedside teaching of PE. Checklist data were transferred to an Excel file (Microsoft), and written observations were imported into NVivo 10 (QRS International, Melbourne, Australia) for coding and analysis.

Checklist data were analyzed using simple descriptive statistics. We compared time spent on various types of rounding using ANOVA, and we used a Student two-tailed t-test to compare the amount of time students spent examining patients on pre-rounds versus new admissions. To ascertain differences in the frequency of PE teaching activities by location, we used chi-squared tests. Statistical analysis was performed using embedded statistics functions in Microsoft Excel. A P value of <.05 was used as the cut-off for significance.

We analyzed the written observations using conventional qualitative content analysis. Two investigators (A.T. and P.B.) reviewed the written comments and used open coding to devise a preliminary inductive coding scheme. Codes were refined iteratively, and a schema of categories and nodes was outlined in a codebook that was periodically reviewed by the entire research team. The coding investigators met regularly to ensure consistency in coding, and a third team member remained available to reconcile significant disagreements in code definitions.

Eighty-one subjects participated in the study: 21 were attending physicians, 12 residents, 21 interns, 11 senior medical students, and 26 junior medical students. We observed 16 distinct inpatient teaching teams and 329 unique patient-related events (discussions and/or patient-clinician encounters), with most events being observed during attending rounds (269/329, or 82%). There were 123 encounters at the bedside, averaging 7 minutes; 43 encounters occurred in the hallway, averaging 8 minutes each; and 163 encounters occurred in a workroom and averaged 7 minutes per patient discussion. We also observed 28 student-patient encounters during pre-round activities and 30 student-patient encounters during new admissions.

Teaching and Direct Observation

During 28 pre-rounding encounters, students usually examined the patient (26 out of 28 instances, 93%) but were observed only 4 times doing so (out of 26 instances, or 15%). During 30 new patient admissions, students examined 27 patients (90%) and had their PE observed 6 times (out of 27 instances, or 22%). There were no significant differences in frequency of these activities (P > .05, chi-squared) between pre-rounds or new admissions.

Observations on Teaching Strategies

In the written observations, we categorized various methods being used to teach PE. Bedside teaching of PE most often involved teachers simply describing or discussing physical findings (42 mentions in observations) or verifying a student’s reported findings (15 mentions). Teachers were also observed to use bedside teaching to contextualize findings (13 mentions), such as relating the quality of bowel sounds to the patient’s constipation or to discuss expected pupillary light reflexes in a neurologically intact patient. Less commonly, attending physicians narrated steps in their PE technique (9 mentions). Students were infrequently encouraged to practice a specific PE skill again (7 mentions) or allowed to re-examine and reconsider their initial interpretations (5 mentions).

This observational study of clinical teaching on internal medicine teaching services demonstrates that PE teaching is most likely to occur during bedside rounding. However, even in bedside encounters, most PE instruction is limited to physician team members pointing out significant findings. Although physical findings were mentioned for the majority of patients seen on rounds, attending physicians infrequently verified students’ or residents’ findings, demonstrated technique, or incorporated PE into clinical decision making. We witnessed an alarming dearth of direct observation of students and almost no real-time feedback in performing and teaching PE. Thus, students rarely had opportunities to engage in higher-order learning activities related to PE on the internal medicine rotation.

We posit that the learning environment influenced PE instruction on the internal medicine rotation. To optimize inpatient teaching of PE, attending physicians need to consider the factors we identified in Table 2. Such teaching may be effective with a more limited number of participants and without distraction from technology. Time constraints are one of the major perceived barriers to bedside teaching of PE, and our data support this concern, as teams spent an average of only 7 minutes on each bedside encounter. However, many of the strategies observed to be used in real-time PE instruction, such as validating the learners’ findings or examining patients as a team, naturally fit into clinical routines and generally do not require extra thought or preparation.

One of the key strengths of our study is the use of direct observation of students and their teachers. This study is unique in its exclusive focus on PE and its description of factors affecting PE teaching activities on an internal medicine service. This observational, descriptive study also has obvious limitations. The study was conducted at a single institution during a limited time period. Moreover, the study period June through August, which was chosen based on our observer’s availability, includes the transition to a new academic year (July 1, 2015) when medical students and residents were becoming acclimated to their new roles. Additionally, the data were collected by a single researcher, and observer bias may affect the results of qualitative analysis of journal entries.

In conclusion, this study highlights the infrequency of applied PE skills in the daily clinical and educational workflow of internal medicine teaching teams. These findings may reflect a more widespread problem in clinical education, and replication of our findings at other teaching centers could galvanize faculty development around bedside PE teaching.

Disclosures

Dr. Bergl has nothing to disclose. Ms. Taylor reports grant support from the Cohen Endowment for Medical Student Research at the Medical College of Wisconsin during the conduct of the study. Mrs. Klumb, Ms. Quirk, Dr. Muntz, and Dr. Fletcher have nothing to disclose.

Funding

This work was funded in part by the Cohen Endowment for Medical Student Research at the Medical College of Wisconsin.

¹Medical College of Wisconsin Affiliated Hospitals, Milwaukee, Wisconsin. At the time of this study, Dr. Bergl was with the Division of General Internal Medicine, Medical College of Wisconsin, Milwaukee, Wisconsin. ²Medical College of Wisconsin, Milwaukee, Wisconsin.Physical examination (PE) is a core clinical skill in undergraduate medical education.¹ Although the optimal approach to teaching clinical skills is debated, robust preclinical curricula should generally be followed by iterative skill development during clinical rotations.^2,3

The internal medicine rotation represents a critical time to enhance PE skills. Diagnostic decision making and PE are highly prioritized competencies for the internal medicine clerkship,⁴ and students will likely utilize many core examination skills^1,2 during this time. Bedside teaching of PE during the internal medicine service also provides an opportunity for students to receive feedback based on direct observation,⁵ a sine qua non of competency-based assessment.

Unfortunately, current internal medicine training environments limit opportunities for workplace-based instruction in PE. Recent studies suggest diminishing time spent on bedside patient care and teaching, with computer-based “indirect patient care” dominating much of the clinical workday of internal medicine services.^6-8 However, the literature does not delineate how often medical students are enhancing their PE skills during clinical rotations or describe how the educational environment may influence PE teaching.

We aimed to describe the content and context of PE instruction during the internal medicine clerkship workflow. Specifically, we sought to explore what strategies physician team members used to teach PE to students. We also sought to describe factors in the inpatient learning environment that might explain why physical examination (PE) instruction occurs infrequently.

We conducted a prospective mixed-methods study using time motion analysis, checklists on clinical teaching, and daily open-ended observations written by a trained observer from June through August 2015 at a single academic medical center. Subjects were recruited from internal medicine teaching teams and were allowed to opt out. Teaching teams had 2 formats: (1) traditional team with an attending physician (hospitalist or general internist), a senior resident, 2 interns, a fourth-year medical student, and 2 third-year students or (2) hospitalist team in which a third-year student works directly with a hospitalist and advanced practitioner. The proposal was submitted to the Medical College of Wisconsin Institutional Review Board and deemed exempt from further review.

All observations were carried out by a single investigator (A.T.), who was a second-year medical student at the time. To train this observer and to pilot the data collection instruments, our lead investigator (P.B.) directly supervised our observer on 4 separate occasions, totaling over 12 hours of mentored co-observation. Immediately after each training session, both investigators (A.T. and P.B.) debriefed to compare notes, to review checklists on recorded observations, and to discuss areas of uncertainty. During the training period, formal metrics of agreement (eg, kappa coefficients) were not gathered, as data collection instruments were still being refined.

Observation periods were centered on third-year medical students and their interactions with patients and members of the teaching team. Observed activities included pre-rounding, teaching rounds with the attending physician, and new patient admissions during call days. Observations generally occurred between the hours of 7 AM and 6 PM, and we limited periods of observation to 3 consecutive hours to minimize observer fatigue. Observation periods were selected to maximize the number of subjects and teams observed, to adequately capture pre-rounding and new admissions activities, and to account for variations in rounding styles throughout the call cycle. Teams were excluded if a member of the study team was an attending physician on the clinical team or if any member of the patient care team had opted out of the study.

Data were collected on paper checklists that included idealized bedside teaching activities around PE. Teaching activities were identified through a review of relevant literature^9,10 and were further informed by our senior investigator’s own experience with faculty development in this area¹¹ and team members’ attendance at bedside teaching workshops. At the end of each day, our observer also wrote brief observations that summarized factors affecting bedside teaching of PE. Checklist data were transferred to an Excel file (Microsoft), and written observations were imported into NVivo 10 (QRS International, Melbourne, Australia) for coding and analysis.

Checklist data were analyzed using simple descriptive statistics. We compared time spent on various types of rounding using ANOVA, and we used a Student two-tailed t-test to compare the amount of time students spent examining patients on pre-rounds versus new admissions. To ascertain differences in the frequency of PE teaching activities by location, we used chi-squared tests. Statistical analysis was performed using embedded statistics functions in Microsoft Excel. A P value of <.05 was used as the cut-off for significance.

We analyzed the written observations using conventional qualitative content analysis. Two investigators (A.T. and P.B.) reviewed the written comments and used open coding to devise a preliminary inductive coding scheme. Codes were refined iteratively, and a schema of categories and nodes was outlined in a codebook that was periodically reviewed by the entire research team. The coding investigators met regularly to ensure consistency in coding, and a third team member remained available to reconcile significant disagreements in code definitions.

Eighty-one subjects participated in the study: 21 were attending physicians, 12 residents, 21 interns, 11 senior medical students, and 26 junior medical students. We observed 16 distinct inpatient teaching teams and 329 unique patient-related events (discussions and/or patient-clinician encounters), with most events being observed during attending rounds (269/329, or 82%). There were 123 encounters at the bedside, averaging 7 minutes; 43 encounters occurred in the hallway, averaging 8 minutes each; and 163 encounters occurred in a workroom and averaged 7 minutes per patient discussion. We also observed 28 student-patient encounters during pre-round activities and 30 student-patient encounters during new admissions.

Teaching and Direct Observation

During 28 pre-rounding encounters, students usually examined the patient (26 out of 28 instances, 93%) but were observed only 4 times doing so (out of 26 instances, or 15%). During 30 new patient admissions, students examined 27 patients (90%) and had their PE observed 6 times (out of 27 instances, or 22%). There were no significant differences in frequency of these activities (P > .05, chi-squared) between pre-rounds or new admissions.

Observations on Teaching Strategies

In the written observations, we categorized various methods being used to teach PE. Bedside teaching of PE most often involved teachers simply describing or discussing physical findings (42 mentions in observations) or verifying a student’s reported findings (15 mentions). Teachers were also observed to use bedside teaching to contextualize findings (13 mentions), such as relating the quality of bowel sounds to the patient’s constipation or to discuss expected pupillary light reflexes in a neurologically intact patient. Less commonly, attending physicians narrated steps in their PE technique (9 mentions). Students were infrequently encouraged to practice a specific PE skill again (7 mentions) or allowed to re-examine and reconsider their initial interpretations (5 mentions).

This observational study of clinical teaching on internal medicine teaching services demonstrates that PE teaching is most likely to occur during bedside rounding. However, even in bedside encounters, most PE instruction is limited to physician team members pointing out significant findings. Although physical findings were mentioned for the majority of patients seen on rounds, attending physicians infrequently verified students’ or residents’ findings, demonstrated technique, or incorporated PE into clinical decision making. We witnessed an alarming dearth of direct observation of students and almost no real-time feedback in performing and teaching PE. Thus, students rarely had opportunities to engage in higher-order learning activities related to PE on the internal medicine rotation.

We posit that the learning environment influenced PE instruction on the internal medicine rotation. To optimize inpatient teaching of PE, attending physicians need to consider the factors we identified in Table 2. Such teaching may be effective with a more limited number of participants and without distraction from technology. Time constraints are one of the major perceived barriers to bedside teaching of PE, and our data support this concern, as teams spent an average of only 7 minutes on each bedside encounter. However, many of the strategies observed to be used in real-time PE instruction, such as validating the learners’ findings or examining patients as a team, naturally fit into clinical routines and generally do not require extra thought or preparation.

One of the key strengths of our study is the use of direct observation of students and their teachers. This study is unique in its exclusive focus on PE and its description of factors affecting PE teaching activities on an internal medicine service. This observational, descriptive study also has obvious limitations. The study was conducted at a single institution during a limited time period. Moreover, the study period June through August, which was chosen based on our observer’s availability, includes the transition to a new academic year (July 1, 2015) when medical students and residents were becoming acclimated to their new roles. Additionally, the data were collected by a single researcher, and observer bias may affect the results of qualitative analysis of journal entries.

In conclusion, this study highlights the infrequency of applied PE skills in the daily clinical and educational workflow of internal medicine teaching teams. These findings may reflect a more widespread problem in clinical education, and replication of our findings at other teaching centers could galvanize faculty development around bedside PE teaching.

Disclosures

Dr. Bergl has nothing to disclose. Ms. Taylor reports grant support from the Cohen Endowment for Medical Student Research at the Medical College of Wisconsin during the conduct of the study. Mrs. Klumb, Ms. Quirk, Dr. Muntz, and Dr. Fletcher have nothing to disclose.

Funding

This work was funded in part by the Cohen Endowment for Medical Student Research at the Medical College of Wisconsin.

¹Medical College of Wisconsin Affiliated Hospitals, Milwaukee, Wisconsin. At the time of this study, Dr. Bergl was with the Division of General Internal Medicine, Medical College of Wisconsin, Milwaukee, Wisconsin. ²Medical College of Wisconsin, Milwaukee, Wisconsin.Physical examination (PE) is a core clinical skill in undergraduate medical education.¹ Although the optimal approach to teaching clinical skills is debated, robust preclinical curricula should generally be followed by iterative skill development during clinical rotations.^2,3

The internal medicine rotation represents a critical time to enhance PE skills. Diagnostic decision making and PE are highly prioritized competencies for the internal medicine clerkship,⁴ and students will likely utilize many core examination skills^1,2 during this time. Bedside teaching of PE during the internal medicine service also provides an opportunity for students to receive feedback based on direct observation,⁵ a sine qua non of competency-based assessment.

Unfortunately, current internal medicine training environments limit opportunities for workplace-based instruction in PE. Recent studies suggest diminishing time spent on bedside patient care and teaching, with computer-based “indirect patient care” dominating much of the clinical workday of internal medicine services.^6-8 However, the literature does not delineate how often medical students are enhancing their PE skills during clinical rotations or describe how the educational environment may influence PE teaching.

We aimed to describe the content and context of PE instruction during the internal medicine clerkship workflow. Specifically, we sought to explore what strategies physician team members used to teach PE to students. We also sought to describe factors in the inpatient learning environment that might explain why physical examination (PE) instruction occurs infrequently.

We conducted a prospective mixed-methods study using time motion analysis, checklists on clinical teaching, and daily open-ended observations written by a trained observer from June through August 2015 at a single academic medical center. Subjects were recruited from internal medicine teaching teams and were allowed to opt out. Teaching teams had 2 formats: (1) traditional team with an attending physician (hospitalist or general internist), a senior resident, 2 interns, a fourth-year medical student, and 2 third-year students or (2) hospitalist team in which a third-year student works directly with a hospitalist and advanced practitioner. The proposal was submitted to the Medical College of Wisconsin Institutional Review Board and deemed exempt from further review.

All observations were carried out by a single investigator (A.T.), who was a second-year medical student at the time. To train this observer and to pilot the data collection instruments, our lead investigator (P.B.) directly supervised our observer on 4 separate occasions, totaling over 12 hours of mentored co-observation. Immediately after each training session, both investigators (A.T. and P.B.) debriefed to compare notes, to review checklists on recorded observations, and to discuss areas of uncertainty. During the training period, formal metrics of agreement (eg, kappa coefficients) were not gathered, as data collection instruments were still being refined.

Observation periods were centered on third-year medical students and their interactions with patients and members of the teaching team. Observed activities included pre-rounding, teaching rounds with the attending physician, and new patient admissions during call days. Observations generally occurred between the hours of 7 AM and 6 PM, and we limited periods of observation to 3 consecutive hours to minimize observer fatigue. Observation periods were selected to maximize the number of subjects and teams observed, to adequately capture pre-rounding and new admissions activities, and to account for variations in rounding styles throughout the call cycle. Teams were excluded if a member of the study team was an attending physician on the clinical team or if any member of the patient care team had opted out of the study.

Data were collected on paper checklists that included idealized bedside teaching activities around PE. Teaching activities were identified through a review of relevant literature^9,10 and were further informed by our senior investigator’s own experience with faculty development in this area¹¹ and team members’ attendance at bedside teaching workshops. At the end of each day, our observer also wrote brief observations that summarized factors affecting bedside teaching of PE. Checklist data were transferred to an Excel file (Microsoft), and written observations were imported into NVivo 10 (QRS International, Melbourne, Australia) for coding and analysis.

Checklist data were analyzed using simple descriptive statistics. We compared time spent on various types of rounding using ANOVA, and we used a Student two-tailed t-test to compare the amount of time students spent examining patients on pre-rounds versus new admissions. To ascertain differences in the frequency of PE teaching activities by location, we used chi-squared tests. Statistical analysis was performed using embedded statistics functions in Microsoft Excel. A P value of <.05 was used as the cut-off for significance.

We analyzed the written observations using conventional qualitative content analysis. Two investigators (A.T. and P.B.) reviewed the written comments and used open coding to devise a preliminary inductive coding scheme. Codes were refined iteratively, and a schema of categories and nodes was outlined in a codebook that was periodically reviewed by the entire research team. The coding investigators met regularly to ensure consistency in coding, and a third team member remained available to reconcile significant disagreements in code definitions.

Eighty-one subjects participated in the study: 21 were attending physicians, 12 residents, 21 interns, 11 senior medical students, and 26 junior medical students. We observed 16 distinct inpatient teaching teams and 329 unique patient-related events (discussions and/or patient-clinician encounters), with most events being observed during attending rounds (269/329, or 82%). There were 123 encounters at the bedside, averaging 7 minutes; 43 encounters occurred in the hallway, averaging 8 minutes each; and 163 encounters occurred in a workroom and averaged 7 minutes per patient discussion. We also observed 28 student-patient encounters during pre-round activities and 30 student-patient encounters during new admissions.

Teaching and Direct Observation

During 28 pre-rounding encounters, students usually examined the patient (26 out of 28 instances, 93%) but were observed only 4 times doing so (out of 26 instances, or 15%). During 30 new patient admissions, students examined 27 patients (90%) and had their PE observed 6 times (out of 27 instances, or 22%). There were no significant differences in frequency of these activities (P > .05, chi-squared) between pre-rounds or new admissions.

Observations on Teaching Strategies

In the written observations, we categorized various methods being used to teach PE. Bedside teaching of PE most often involved teachers simply describing or discussing physical findings (42 mentions in observations) or verifying a student’s reported findings (15 mentions). Teachers were also observed to use bedside teaching to contextualize findings (13 mentions), such as relating the quality of bowel sounds to the patient’s constipation or to discuss expected pupillary light reflexes in a neurologically intact patient. Less commonly, attending physicians narrated steps in their PE technique (9 mentions). Students were infrequently encouraged to practice a specific PE skill again (7 mentions) or allowed to re-examine and reconsider their initial interpretations (5 mentions).

This observational study of clinical teaching on internal medicine teaching services demonstrates that PE teaching is most likely to occur during bedside rounding. However, even in bedside encounters, most PE instruction is limited to physician team members pointing out significant findings. Although physical findings were mentioned for the majority of patients seen on rounds, attending physicians infrequently verified students’ or residents’ findings, demonstrated technique, or incorporated PE into clinical decision making. We witnessed an alarming dearth of direct observation of students and almost no real-time feedback in performing and teaching PE. Thus, students rarely had opportunities to engage in higher-order learning activities related to PE on the internal medicine rotation.

We posit that the learning environment influenced PE instruction on the internal medicine rotation. To optimize inpatient teaching of PE, attending physicians need to consider the factors we identified in Table 2. Such teaching may be effective with a more limited number of participants and without distraction from technology. Time constraints are one of the major perceived barriers to bedside teaching of PE, and our data support this concern, as teams spent an average of only 7 minutes on each bedside encounter. However, many of the strategies observed to be used in real-time PE instruction, such as validating the learners’ findings or examining patients as a team, naturally fit into clinical routines and generally do not require extra thought or preparation.

One of the key strengths of our study is the use of direct observation of students and their teachers. This study is unique in its exclusive focus on PE and its description of factors affecting PE teaching activities on an internal medicine service. This observational, descriptive study also has obvious limitations. The study was conducted at a single institution during a limited time period. Moreover, the study period June through August, which was chosen based on our observer’s availability, includes the transition to a new academic year (July 1, 2015) when medical students and residents were becoming acclimated to their new roles. Additionally, the data were collected by a single researcher, and observer bias may affect the results of qualitative analysis of journal entries.

In conclusion, this study highlights the infrequency of applied PE skills in the daily clinical and educational workflow of internal medicine teaching teams. These findings may reflect a more widespread problem in clinical education, and replication of our findings at other teaching centers could galvanize faculty development around bedside PE teaching.

Disclosures

Dr. Bergl has nothing to disclose. Ms. Taylor reports grant support from the Cohen Endowment for Medical Student Research at the Medical College of Wisconsin during the conduct of the study. Mrs. Klumb, Ms. Quirk, Dr. Muntz, and Dr. Fletcher have nothing to disclose.

Funding

This work was funded in part by the Cohen Endowment for Medical Student Research at the Medical College of Wisconsin.

Alarm fatigue is a patient safety hazard in hospitals¹ that occurs when exposure to high rates of alarms leads clinicians to ignore or delay their responses to the alarms.^2,3 To date, most studies of physiologic monitor alarms in hospitalized children have used data from single institutions and often only a few units within each institution.⁴ These limited studies have found that alarms in pediatric units are rarely actionable.² They have also shown that physiologic monitor alarms occur frequently in children’s hospitals and that alarm rates can vary widely within a single institution,⁵ but the extent of variation between children’s hospitals is unknown. In this study, we aimed to describe and compare physiologic monitor alarm characteristics and the proportion of patients monitored in the inpatient units of 5 children’s hospitals.

We performed a cross-sectional study using a point-prevalence design of physiologic monitor alarms and monitoring during a 24-hour period at 5 large, freestanding tertiary-care children’s hospitals. At the time of the study, each hospital had an alarm management committee in place and was working to address alarm fatigue. Each hospital’s institutional review board reviewed and approved the study.

We collected 24 consecutive hours of data from the inpatient units of each hospital between March 24, 2015, and May 1, 2015. Each hospital selected the data collection date within that window based on the availability of staff to perform data collection.⁶ We excluded emergency departments, procedural areas, and inpatient psychiatry and rehabilitation units. By using existing central alarm-collection software that interfaced with bedside physiologic monitors, we collected data on audible alarms generated for apnea, arrhythmia, low and high oxygen saturation, heart rate, respiratory rate, blood pressure, and exhaled carbon dioxide. Bedside alarm systems and alarm collection software differed between centers; therefore, alarm types that were not consistently collected at every institution (eg, alarms for electrode and device malfunction, ventilators, intracranial and central venous pressure monitors, and temperatures probes) were excluded. To estimate alarm rates and to account for fluctuations in hospital census throughout the day,⁷ we collected census (to calculate the number of alarms per patient day) and the number of monitored patients (to calculate the number of alarms per monitored-patient day, including only monitored patients in the denominator) on each unit at 3 time points, 8 hours apart. Patients were considered continuously monitored if they had presence of a waveform and data for pulse oximetry, respiratory rate, and/or heart rate at the time of data collection. We then determined the rate of alarms by unit type—medical-surgical unit (MSU), neonatal intensive care unit (NICU), or pediatric intensive care unit (PICU)—and the alarm types. Based on prior literature demonstrating up to 95% of alarms contributed by a minority of patients on a single unit,⁸ we also calculated the percentage of alarms contributed by beds in the highest quartile of alarms. We also assessed the percentage of patients monitored by unit type. The Supplementary Appendix shows the alarm parameter thresholds in use at the time of the study.

Physiologic monitor alarm rates and the proportion of patients monitored varied widely between unit types and among the tertiary-care children’s hospitals in our study. We found that among MSUs, the hospital with the lowest proportion of beds monitored had the highest alarm rate, with over triple the rate seen at the hospital with the lowest alarm rate. Regardless of unit type, a small subgroup of patients at each hospital contributed a disproportionate share of alarms. These findings are concerning because of the patient morbidity and mortality associated with alarm fatigue¹ and the studies suggesting that higher alarm rates may lead to delays in response to potentially critical alarms.²

We previously described alarm rates at a single children’s hospital and found that alarm rates were high both in and outside of the ICU areas.⁵ This study supports those findings and goes further to show that alarm rates on some MSUs approached rates seen in the ICU areas at other centers.⁴ However, our results should be considered in the context of several limitations. First, the 5 study hospitals utilized different bedside monitors, equipment, and software to collect alarm data. It is possible that this impacted how alarms were counted, though there were no technical specifications to suggest that results should have been biased in a specific way. Second, our data did not reflect alarm validity (ie, whether an alarm accurately reflected the physiologic state of the patient) or factors outside of the number of patients monitored—such as practices around ICU admission and transfer as well as monitor practices such as lead changes, the type of leads employed, and the degree to which alarm parameter thresholds could be customized, which may have also affected alarm rates. Finally, we excluded alarm types that were not consistently collected at all hospitals. We were also unable to capture alarms from other alarm-generating devices, including ventilators and infusion pumps, which have also been identified as sources of alarm-related safety issues in hospitals.^9-11 This suggests that the alarm rates reported here underestimate the total number of audible alarms experienced by staff and by hospitalized patients and families.

While our data collection was limited in scope, the striking differences in alarm rates between hospitals and between similar units in the same hospitals suggest that unit- and hospital-level factors—including default alarm parameter threshold settings, types of monitors used, and monitoring practices such as the degree to which alarm parameters are customized to the patient’s physiologic state—likely contribute to the variability. It is also important to note that while there were clear outlier hospitals, no single hospital had the lowest alarm rate across all unit types. And while we found that a small number of patients contributed disproportionately to alarms, monitoring fewer patients overall was not consistently associated with lower alarm rates. While it is difficult to draw conclusions based on a limited study, these findings suggest that solutions to meaningfully lower alarm rates may be multifaceted. Standardization of care in multiple areas of medicine has shown the potential to decrease unnecessary utilization of testing and therapies while maintaining good patient outcomes.^12-15 Our findings suggest that the concept of positive deviance,¹⁶ by which some organizations produce better outcomes than others despite similar limitations, may help identify successful alarm reduction strategies for further testing. Larger quantitative studies of alarm rates and ethnographic or qualitative studies of monitoring practices may reveal practices and policies that are associated with lower alarm rates with similar or improved monitoring outcomes.

We found wide variability in physiologic monitor alarm rates and the proportion of patients monitored across 5 children’s hospitals. Because alarm fatigue remains a pressing patient safety concern, further study of the features of high-performing (low-alarm) hospital systems may help identify barriers and facilitators of safe, effective monitoring and develop targeted interventions to reduce alarms.

The authors thank Melinda Egan, Matt MacMurchy, and Shannon Stemler for their assistance with data collection.

Disclosure

Dr. Bonafide is supported by the National Heart, Lung, and Blood Institute of the National Institutes of Health under Award Number K23HL116427. Dr. Brady is supported by the Agency for Healthcare Research and Quality under Award Number K08HS23827. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the Agency for Healthcare Research and Quality. There was no external funding obtained for this study. The authors have no conflicts of interest to disclose.

Alarm fatigue is a patient safety hazard in hospitals¹ that occurs when exposure to high rates of alarms leads clinicians to ignore or delay their responses to the alarms.^2,3 To date, most studies of physiologic monitor alarms in hospitalized children have used data from single institutions and often only a few units within each institution.⁴ These limited studies have found that alarms in pediatric units are rarely actionable.² They have also shown that physiologic monitor alarms occur frequently in children’s hospitals and that alarm rates can vary widely within a single institution,⁵ but the extent of variation between children’s hospitals is unknown. In this study, we aimed to describe and compare physiologic monitor alarm characteristics and the proportion of patients monitored in the inpatient units of 5 children’s hospitals.

We performed a cross-sectional study using a point-prevalence design of physiologic monitor alarms and monitoring during a 24-hour period at 5 large, freestanding tertiary-care children’s hospitals. At the time of the study, each hospital had an alarm management committee in place and was working to address alarm fatigue. Each hospital’s institutional review board reviewed and approved the study.

We collected 24 consecutive hours of data from the inpatient units of each hospital between March 24, 2015, and May 1, 2015. Each hospital selected the data collection date within that window based on the availability of staff to perform data collection.⁶ We excluded emergency departments, procedural areas, and inpatient psychiatry and rehabilitation units. By using existing central alarm-collection software that interfaced with bedside physiologic monitors, we collected data on audible alarms generated for apnea, arrhythmia, low and high oxygen saturation, heart rate, respiratory rate, blood pressure, and exhaled carbon dioxide. Bedside alarm systems and alarm collection software differed between centers; therefore, alarm types that were not consistently collected at every institution (eg, alarms for electrode and device malfunction, ventilators, intracranial and central venous pressure monitors, and temperatures probes) were excluded. To estimate alarm rates and to account for fluctuations in hospital census throughout the day,⁷ we collected census (to calculate the number of alarms per patient day) and the number of monitored patients (to calculate the number of alarms per monitored-patient day, including only monitored patients in the denominator) on each unit at 3 time points, 8 hours apart. Patients were considered continuously monitored if they had presence of a waveform and data for pulse oximetry, respiratory rate, and/or heart rate at the time of data collection. We then determined the rate of alarms by unit type—medical-surgical unit (MSU), neonatal intensive care unit (NICU), or pediatric intensive care unit (PICU)—and the alarm types. Based on prior literature demonstrating up to 95% of alarms contributed by a minority of patients on a single unit,⁸ we also calculated the percentage of alarms contributed by beds in the highest quartile of alarms. We also assessed the percentage of patients monitored by unit type. The Supplementary Appendix shows the alarm parameter thresholds in use at the time of the study.

Physiologic monitor alarm rates and the proportion of patients monitored varied widely between unit types and among the tertiary-care children’s hospitals in our study. We found that among MSUs, the hospital with the lowest proportion of beds monitored had the highest alarm rate, with over triple the rate seen at the hospital with the lowest alarm rate. Regardless of unit type, a small subgroup of patients at each hospital contributed a disproportionate share of alarms. These findings are concerning because of the patient morbidity and mortality associated with alarm fatigue¹ and the studies suggesting that higher alarm rates may lead to delays in response to potentially critical alarms.²

We previously described alarm rates at a single children’s hospital and found that alarm rates were high both in and outside of the ICU areas.⁵ This study supports those findings and goes further to show that alarm rates on some MSUs approached rates seen in the ICU areas at other centers.⁴ However, our results should be considered in the context of several limitations. First, the 5 study hospitals utilized different bedside monitors, equipment, and software to collect alarm data. It is possible that this impacted how alarms were counted, though there were no technical specifications to suggest that results should have been biased in a specific way. Second, our data did not reflect alarm validity (ie, whether an alarm accurately reflected the physiologic state of the patient) or factors outside of the number of patients monitored—such as practices around ICU admission and transfer as well as monitor practices such as lead changes, the type of leads employed, and the degree to which alarm parameter thresholds could be customized, which may have also affected alarm rates. Finally, we excluded alarm types that were not consistently collected at all hospitals. We were also unable to capture alarms from other alarm-generating devices, including ventilators and infusion pumps, which have also been identified as sources of alarm-related safety issues in hospitals.^9-11 This suggests that the alarm rates reported here underestimate the total number of audible alarms experienced by staff and by hospitalized patients and families.

While our data collection was limited in scope, the striking differences in alarm rates between hospitals and between similar units in the same hospitals suggest that unit- and hospital-level factors—including default alarm parameter threshold settings, types of monitors used, and monitoring practices such as the degree to which alarm parameters are customized to the patient’s physiologic state—likely contribute to the variability. It is also important to note that while there were clear outlier hospitals, no single hospital had the lowest alarm rate across all unit types. And while we found that a small number of patients contributed disproportionately to alarms, monitoring fewer patients overall was not consistently associated with lower alarm rates. While it is difficult to draw conclusions based on a limited study, these findings suggest that solutions to meaningfully lower alarm rates may be multifaceted. Standardization of care in multiple areas of medicine has shown the potential to decrease unnecessary utilization of testing and therapies while maintaining good patient outcomes.^12-15 Our findings suggest that the concept of positive deviance,¹⁶ by which some organizations produce better outcomes than others despite similar limitations, may help identify successful alarm reduction strategies for further testing. Larger quantitative studies of alarm rates and ethnographic or qualitative studies of monitoring practices may reveal practices and policies that are associated with lower alarm rates with similar or improved monitoring outcomes.

We found wide variability in physiologic monitor alarm rates and the proportion of patients monitored across 5 children’s hospitals. Because alarm fatigue remains a pressing patient safety concern, further study of the features of high-performing (low-alarm) hospital systems may help identify barriers and facilitators of safe, effective monitoring and develop targeted interventions to reduce alarms.

The authors thank Melinda Egan, Matt MacMurchy, and Shannon Stemler for their assistance with data collection.

Disclosure

Dr. Bonafide is supported by the National Heart, Lung, and Blood Institute of the National Institutes of Health under Award Number K23HL116427. Dr. Brady is supported by the Agency for Healthcare Research and Quality under Award Number K08HS23827. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the Agency for Healthcare Research and Quality. There was no external funding obtained for this study. The authors have no conflicts of interest to disclose.

Alarm fatigue is a patient safety hazard in hospitals¹ that occurs when exposure to high rates of alarms leads clinicians to ignore or delay their responses to the alarms.^2,3 To date, most studies of physiologic monitor alarms in hospitalized children have used data from single institutions and often only a few units within each institution.⁴ These limited studies have found that alarms in pediatric units are rarely actionable.² They have also shown that physiologic monitor alarms occur frequently in children’s hospitals and that alarm rates can vary widely within a single institution,⁵ but the extent of variation between children’s hospitals is unknown. In this study, we aimed to describe and compare physiologic monitor alarm characteristics and the proportion of patients monitored in the inpatient units of 5 children’s hospitals.

We performed a cross-sectional study using a point-prevalence design of physiologic monitor alarms and monitoring during a 24-hour period at 5 large, freestanding tertiary-care children’s hospitals. At the time of the study, each hospital had an alarm management committee in place and was working to address alarm fatigue. Each hospital’s institutional review board reviewed and approved the study.

We collected 24 consecutive hours of data from the inpatient units of each hospital between March 24, 2015, and May 1, 2015. Each hospital selected the data collection date within that window based on the availability of staff to perform data collection.⁶ We excluded emergency departments, procedural areas, and inpatient psychiatry and rehabilitation units. By using existing central alarm-collection software that interfaced with bedside physiologic monitors, we collected data on audible alarms generated for apnea, arrhythmia, low and high oxygen saturation, heart rate, respiratory rate, blood pressure, and exhaled carbon dioxide. Bedside alarm systems and alarm collection software differed between centers; therefore, alarm types that were not consistently collected at every institution (eg, alarms for electrode and device malfunction, ventilators, intracranial and central venous pressure monitors, and temperatures probes) were excluded. To estimate alarm rates and to account for fluctuations in hospital census throughout the day,⁷ we collected census (to calculate the number of alarms per patient day) and the number of monitored patients (to calculate the number of alarms per monitored-patient day, including only monitored patients in the denominator) on each unit at 3 time points, 8 hours apart. Patients were considered continuously monitored if they had presence of a waveform and data for pulse oximetry, respiratory rate, and/or heart rate at the time of data collection. We then determined the rate of alarms by unit type—medical-surgical unit (MSU), neonatal intensive care unit (NICU), or pediatric intensive care unit (PICU)—and the alarm types. Based on prior literature demonstrating up to 95% of alarms contributed by a minority of patients on a single unit,⁸ we also calculated the percentage of alarms contributed by beds in the highest quartile of alarms. We also assessed the percentage of patients monitored by unit type. The Supplementary Appendix shows the alarm parameter thresholds in use at the time of the study.

Physiologic monitor alarm rates and the proportion of patients monitored varied widely between unit types and among the tertiary-care children’s hospitals in our study. We found that among MSUs, the hospital with the lowest proportion of beds monitored had the highest alarm rate, with over triple the rate seen at the hospital with the lowest alarm rate. Regardless of unit type, a small subgroup of patients at each hospital contributed a disproportionate share of alarms. These findings are concerning because of the patient morbidity and mortality associated with alarm fatigue¹ and the studies suggesting that higher alarm rates may lead to delays in response to potentially critical alarms.²

We previously described alarm rates at a single children’s hospital and found that alarm rates were high both in and outside of the ICU areas.⁵ This study supports those findings and goes further to show that alarm rates on some MSUs approached rates seen in the ICU areas at other centers.⁴ However, our results should be considered in the context of several limitations. First, the 5 study hospitals utilized different bedside monitors, equipment, and software to collect alarm data. It is possible that this impacted how alarms were counted, though there were no technical specifications to suggest that results should have been biased in a specific way. Second, our data did not reflect alarm validity (ie, whether an alarm accurately reflected the physiologic state of the patient) or factors outside of the number of patients monitored—such as practices around ICU admission and transfer as well as monitor practices such as lead changes, the type of leads employed, and the degree to which alarm parameter thresholds could be customized, which may have also affected alarm rates. Finally, we excluded alarm types that were not consistently collected at all hospitals. We were also unable to capture alarms from other alarm-generating devices, including ventilators and infusion pumps, which have also been identified as sources of alarm-related safety issues in hospitals.^9-11 This suggests that the alarm rates reported here underestimate the total number of audible alarms experienced by staff and by hospitalized patients and families.

While our data collection was limited in scope, the striking differences in alarm rates between hospitals and between similar units in the same hospitals suggest that unit- and hospital-level factors—including default alarm parameter threshold settings, types of monitors used, and monitoring practices such as the degree to which alarm parameters are customized to the patient’s physiologic state—likely contribute to the variability. It is also important to note that while there were clear outlier hospitals, no single hospital had the lowest alarm rate across all unit types. And while we found that a small number of patients contributed disproportionately to alarms, monitoring fewer patients overall was not consistently associated with lower alarm rates. While it is difficult to draw conclusions based on a limited study, these findings suggest that solutions to meaningfully lower alarm rates may be multifaceted. Standardization of care in multiple areas of medicine has shown the potential to decrease unnecessary utilization of testing and therapies while maintaining good patient outcomes.^12-15 Our findings suggest that the concept of positive deviance,¹⁶ by which some organizations produce better outcomes than others despite similar limitations, may help identify successful alarm reduction strategies for further testing. Larger quantitative studies of alarm rates and ethnographic or qualitative studies of monitoring practices may reveal practices and policies that are associated with lower alarm rates with similar or improved monitoring outcomes.

We found wide variability in physiologic monitor alarm rates and the proportion of patients monitored across 5 children’s hospitals. Because alarm fatigue remains a pressing patient safety concern, further study of the features of high-performing (low-alarm) hospital systems may help identify barriers and facilitators of safe, effective monitoring and develop targeted interventions to reduce alarms.

The authors thank Melinda Egan, Matt MacMurchy, and Shannon Stemler for their assistance with data collection.

Disclosure

Dr. Bonafide is supported by the National Heart, Lung, and Blood Institute of the National Institutes of Health under Award Number K23HL116427. Dr. Brady is supported by the Agency for Healthcare Research and Quality under Award Number K08HS23827. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the Agency for Healthcare Research and Quality. There was no external funding obtained for this study. The authors have no conflicts of interest to disclose.

While patients may be newly exposed to opioids during medical and surgical hospitalization and the prescription of opioids at discharge is common,^1-5 prescribers of opioids at discharge may not intend to initiate long-term opioid (LTO) use. By understanding the frequency of progression to LTO use, hospitalists can better balance postdischarge pain treatment and the risk for unintended LTO initiation.

Estimates of LTO use rates following hospital discharge in selected populations^1,2,4-6 have varied depending on the population studied and the method of defining LTO use.⁷ Rates of LTO use following incident opioid prescription have not been directly compared at medical versus surgical discharge or compared with initiation in the ambulatory setting. We present the rates of LTO use following incident opioid exposure at surgical discharge and medical discharge and identify the factors associated with LTO use following surgical and medical discharge.

Data Sources

Veterans Health Administration (VHA) data were obtained through the Austin Information Technology Center for fiscal years (FYs) 2003 through 2012 (Austin, Texas). Decision support system national data extracts were used to identify prescription-dispensing events, and inpatient and outpatient medical SAS data sets were used to identify diagnostic codes. The study was approved by the University of Iowa Institutional Review Board and the Iowa City Veterans Affairs (VA) Health Care System Research and Development Committee.

Patients

We included all patients with an outpatient opioid prescription during FY 2011 that was preceded by a 1-year opioid-free period.⁷ Patients with broadly accepted indications for LTO use (eg, metastatic cancer, palliative care, or opioid-dependence treatment) were excluded.⁷

Opioid Exposure

We included all outpatient prescription fills for noninjectable dosage forms of butorphanol, fentanyl, hydrocodone, hydromorphone, levorphanol, meperidine, methadone, morphine, oxycodone, oxymorphone, pentazocine, and tramadol. Consistent with the Centers for Disease Control and Prevention and VA/Department of Defense guidelines, LTO use was defined conceptually as regular use for >90 days. Operationalizing this definition to pharmacy refill data was established by using a cabinet supply methodology,⁷ which allows for the construction of episodes of continuous medication therapy by estimating the medication supply available to a patient for each day during a defined period based on the pattern of observed refills. LTO use was defined as an episode of continuous opioid supply for >90 days and beginning within 30 days of the initial prescription. While some studies have defined LTO use based on onset within 1 year following surgery,⁵ the requirement for onset within 30 days of initiation was applied to more strongly tie the association of developing LTO use with the discharge event and minimize various forms of bias that are introduced with extended follow-up periods.

Clinical Characteristics

Patients were classified as being medical discharges, surgical discharges, or outpatient initiators. Patients with an opioid index date within 2 days following discharge were designated based on discharge bed section; additionally, if patients had a surgical bed section during hospitalization, they were assigned as surgical discharges. Demographic, diagnosis, and medication exposure variables that were previously associated with LTO use were selected.^8,9 Substance use disorder, chronic pain, anxiety disorder, and depressive disorder were based on International Classification of Diseases, 9th Revision (ICD-9) codes in the preceding year. The use of concurrent benzodiazepines, skeletal muscle relaxants, and antidepressants were determined at opioid initiation.¹⁰ Rural or urban residence was assigned by using the Rural-Urban Commuting Area Codes system and mapped with the zip code of a veteran’s residence.¹¹

Analysis

Bivariate and multivariable relationships were determined by using logistic regression. The multivariable model considered all pairwise interaction terms between inpatient service (surgery versus medicine) and each of the variables in the model. Statistically significant interaction terms (P < .05) were retained, and all others were omitted from the final model. The main effects for variables that were involved in a significant interaction term were not reported in the final multivariable model; instead, we created fully specified multivariable models for surgery service and medicine service and reported odds ratios (ORs) for the main effects. All analyses were conducted by using SAS version 9.4 (SAS Institute Inc, Cary, North Carolina).

The observation that subsequent LTO use occurs more frequently in discharged medical patients than surgical patients is consistent with the findings of Calcaterra et al.¹ that among patients with no surgery versus surgery during hospitalization, opioid receipt at discharge resulted in a higher adjusted OR (7.24 for no surgery versus 3.40 for surgery) for chronic opioid use at 1 year. One explanation for this finding may be an artifact of cohort selection in the study design: patients with prior opioid use are excluded from the cohort, and prior use may be more common among surgical patients presenting for elective inpatient surgery for painful conditions. Previous work suggests that opioid use preoperatively is a robust predictor of postoperative use, and rates of LTO use are low among patients without preoperative opioid exposure.⁶

Demographic characteristics associated with persistent opioid receipt were similar to those previously reported.^5,8,9 The inclusion of medication classes indicated in the treatment of mental health or pain conditions (ie, antidepressants, benzodiazepines, muscle relaxants, and nonopioid analgesics) resulted in diagnoses based on ICD-9 codes being no longer associated with LTO use. Severity or activity of illness, preferences regarding pharmacologic or nonpharmacologic treatment and undiagnosed or undocumented pain-comorbid conditions may all contribute to this finding. Future work studying opioid-related outcomes should include variables that reflect pharmacologic management of comorbid diagnoses in the cohort development or analytic design.

The strongest risk factors were potentially modifiable: days’ supply, dose, and concurrent medications. The measures of opioid quantity supplied are associated with subsequent ongoing use and are consistent with recent work based on prescription drug–monitoring data in a single state¹⁴ and in a nationally representative sample.¹⁵ That this relationship persists following hospital discharge, a scenario in which LTO use is unlikely to be initiated by a provider (who would be expected to subsequently titrate or monitor therapy), further supports the potential to curtail unintended LTO use through judicious early prescribing decisions.

We assessed only opioids that were supplied through a VA pharmacy, which may lead to the misclassification of patients as opioid naive for inclusion and an underestimation of the rate of opioid use following discharge. It is possible that differences in the rates of non-VA pharmacy use differ in medical and surgical populations in a nonrandom way. This study was performed in a large, integrated health system and may not be generalizable outside the VA system, where more discontinuities between hospital and ambulatory care may exist.

The initiation of LTO use at discharge is more common in veterans who are discharged from medical than surgical hospitalizations, likely reflecting differences in the patient population, pain conditions, and discharge prescribing decisions. While patient characteristics are associated with LTO use, the strongest associations are with increasing index dose and days’ supply; both represent potentially modifiable prescriber behaviors. These findings support policy changes and other efforts to minimize dose and days supplied when short-term use is intended as a means to address the current opioid epidemic.

Acknowledgments

The work reported here was supported by the Department of Veterans Affairs Office of Academic Affiliations and Office of Research and Development (Dr. Mosher and Dr. Hofmeyer), and Health Services Research and Development Service (HSR&D) through the Comprehensive Access and Delivery Research and Evaluation Center (CIN 13-412) and a Career Development Award (CDA 10-017; Dr. Lund).

Disclosures

The authors report no conflict of interest in regard to this study. The authors had full access to and take full responsibility for the integrity of the data. All analyses were conducted by using SAS version 9.2 (SAS Institute Inc, Cary, NC). This manuscript is not under review elsewhere, and there is no prior publication of the manuscript contents. The views expressed in this article are those of the authors and do not necessarily represent the views of the Department of Veterans Affairs. The study was approved by the University of Iowa Institutional Review Board and the Iowa City Healthcare System Research and Development Committee.

While patients may be newly exposed to opioids during medical and surgical hospitalization and the prescription of opioids at discharge is common,^1-5 prescribers of opioids at discharge may not intend to initiate long-term opioid (LTO) use. By understanding the frequency of progression to LTO use, hospitalists can better balance postdischarge pain treatment and the risk for unintended LTO initiation.

Estimates of LTO use rates following hospital discharge in selected populations^1,2,4-6 have varied depending on the population studied and the method of defining LTO use.⁷ Rates of LTO use following incident opioid prescription have not been directly compared at medical versus surgical discharge or compared with initiation in the ambulatory setting. We present the rates of LTO use following incident opioid exposure at surgical discharge and medical discharge and identify the factors associated with LTO use following surgical and medical discharge.

Data Sources

Veterans Health Administration (VHA) data were obtained through the Austin Information Technology Center for fiscal years (FYs) 2003 through 2012 (Austin, Texas). Decision support system national data extracts were used to identify prescription-dispensing events, and inpatient and outpatient medical SAS data sets were used to identify diagnostic codes. The study was approved by the University of Iowa Institutional Review Board and the Iowa City Veterans Affairs (VA) Health Care System Research and Development Committee.

Patients

We included all patients with an outpatient opioid prescription during FY 2011 that was preceded by a 1-year opioid-free period.⁷ Patients with broadly accepted indications for LTO use (eg, metastatic cancer, palliative care, or opioid-dependence treatment) were excluded.⁷

Opioid Exposure

We included all outpatient prescription fills for noninjectable dosage forms of butorphanol, fentanyl, hydrocodone, hydromorphone, levorphanol, meperidine, methadone, morphine, oxycodone, oxymorphone, pentazocine, and tramadol. Consistent with the Centers for Disease Control and Prevention and VA/Department of Defense guidelines, LTO use was defined conceptually as regular use for >90 days. Operationalizing this definition to pharmacy refill data was established by using a cabinet supply methodology,⁷ which allows for the construction of episodes of continuous medication therapy by estimating the medication supply available to a patient for each day during a defined period based on the pattern of observed refills. LTO use was defined as an episode of continuous opioid supply for >90 days and beginning within 30 days of the initial prescription. While some studies have defined LTO use based on onset within 1 year following surgery,⁵ the requirement for onset within 30 days of initiation was applied to more strongly tie the association of developing LTO use with the discharge event and minimize various forms of bias that are introduced with extended follow-up periods.

Clinical Characteristics

Patients were classified as being medical discharges, surgical discharges, or outpatient initiators. Patients with an opioid index date within 2 days following discharge were designated based on discharge bed section; additionally, if patients had a surgical bed section during hospitalization, they were assigned as surgical discharges. Demographic, diagnosis, and medication exposure variables that were previously associated with LTO use were selected.^8,9 Substance use disorder, chronic pain, anxiety disorder, and depressive disorder were based on International Classification of Diseases, 9th Revision (ICD-9) codes in the preceding year. The use of concurrent benzodiazepines, skeletal muscle relaxants, and antidepressants were determined at opioid initiation.¹⁰ Rural or urban residence was assigned by using the Rural-Urban Commuting Area Codes system and mapped with the zip code of a veteran’s residence.¹¹

Analysis

Bivariate and multivariable relationships were determined by using logistic regression. The multivariable model considered all pairwise interaction terms between inpatient service (surgery versus medicine) and each of the variables in the model. Statistically significant interaction terms (P < .05) were retained, and all others were omitted from the final model. The main effects for variables that were involved in a significant interaction term were not reported in the final multivariable model; instead, we created fully specified multivariable models for surgery service and medicine service and reported odds ratios (ORs) for the main effects. All analyses were conducted by using SAS version 9.4 (SAS Institute Inc, Cary, North Carolina).

The observation that subsequent LTO use occurs more frequently in discharged medical patients than surgical patients is consistent with the findings of Calcaterra et al.¹ that among patients with no surgery versus surgery during hospitalization, opioid receipt at discharge resulted in a higher adjusted OR (7.24 for no surgery versus 3.40 for surgery) for chronic opioid use at 1 year. One explanation for this finding may be an artifact of cohort selection in the study design: patients with prior opioid use are excluded from the cohort, and prior use may be more common among surgical patients presenting for elective inpatient surgery for painful conditions. Previous work suggests that opioid use preoperatively is a robust predictor of postoperative use, and rates of LTO use are low among patients without preoperative opioid exposure.⁶

Demographic characteristics associated with persistent opioid receipt were similar to those previously reported.^5,8,9 The inclusion of medication classes indicated in the treatment of mental health or pain conditions (ie, antidepressants, benzodiazepines, muscle relaxants, and nonopioid analgesics) resulted in diagnoses based on ICD-9 codes being no longer associated with LTO use. Severity or activity of illness, preferences regarding pharmacologic or nonpharmacologic treatment and undiagnosed or undocumented pain-comorbid conditions may all contribute to this finding. Future work studying opioid-related outcomes should include variables that reflect pharmacologic management of comorbid diagnoses in the cohort development or analytic design.

The strongest risk factors were potentially modifiable: days’ supply, dose, and concurrent medications. The measures of opioid quantity supplied are associated with subsequent ongoing use and are consistent with recent work based on prescription drug–monitoring data in a single state¹⁴ and in a nationally representative sample.¹⁵ That this relationship persists following hospital discharge, a scenario in which LTO use is unlikely to be initiated by a provider (who would be expected to subsequently titrate or monitor therapy), further supports the potential to curtail unintended LTO use through judicious early prescribing decisions.

We assessed only opioids that were supplied through a VA pharmacy, which may lead to the misclassification of patients as opioid naive for inclusion and an underestimation of the rate of opioid use following discharge. It is possible that differences in the rates of non-VA pharmacy use differ in medical and surgical populations in a nonrandom way. This study was performed in a large, integrated health system and may not be generalizable outside the VA system, where more discontinuities between hospital and ambulatory care may exist.

The initiation of LTO use at discharge is more common in veterans who are discharged from medical than surgical hospitalizations, likely reflecting differences in the patient population, pain conditions, and discharge prescribing decisions. While patient characteristics are associated with LTO use, the strongest associations are with increasing index dose and days’ supply; both represent potentially modifiable prescriber behaviors. These findings support policy changes and other efforts to minimize dose and days supplied when short-term use is intended as a means to address the current opioid epidemic.

Acknowledgments

The work reported here was supported by the Department of Veterans Affairs Office of Academic Affiliations and Office of Research and Development (Dr. Mosher and Dr. Hofmeyer), and Health Services Research and Development Service (HSR&D) through the Comprehensive Access and Delivery Research and Evaluation Center (CIN 13-412) and a Career Development Award (CDA 10-017; Dr. Lund).

Disclosures

The authors report no conflict of interest in regard to this study. The authors had full access to and take full responsibility for the integrity of the data. All analyses were conducted by using SAS version 9.2 (SAS Institute Inc, Cary, NC). This manuscript is not under review elsewhere, and there is no prior publication of the manuscript contents. The views expressed in this article are those of the authors and do not necessarily represent the views of the Department of Veterans Affairs. The study was approved by the University of Iowa Institutional Review Board and the Iowa City Healthcare System Research and Development Committee.

While patients may be newly exposed to opioids during medical and surgical hospitalization and the prescription of opioids at discharge is common,^1-5 prescribers of opioids at discharge may not intend to initiate long-term opioid (LTO) use. By understanding the frequency of progression to LTO use, hospitalists can better balance postdischarge pain treatment and the risk for unintended LTO initiation.

Estimates of LTO use rates following hospital discharge in selected populations^1,2,4-6 have varied depending on the population studied and the method of defining LTO use.⁷ Rates of LTO use following incident opioid prescription have not been directly compared at medical versus surgical discharge or compared with initiation in the ambulatory setting. We present the rates of LTO use following incident opioid exposure at surgical discharge and medical discharge and identify the factors associated with LTO use following surgical and medical discharge.

Data Sources

Veterans Health Administration (VHA) data were obtained through the Austin Information Technology Center for fiscal years (FYs) 2003 through 2012 (Austin, Texas). Decision support system national data extracts were used to identify prescription-dispensing events, and inpatient and outpatient medical SAS data sets were used to identify diagnostic codes. The study was approved by the University of Iowa Institutional Review Board and the Iowa City Veterans Affairs (VA) Health Care System Research and Development Committee.

Patients

We included all patients with an outpatient opioid prescription during FY 2011 that was preceded by a 1-year opioid-free period.⁷ Patients with broadly accepted indications for LTO use (eg, metastatic cancer, palliative care, or opioid-dependence treatment) were excluded.⁷

Opioid Exposure

We included all outpatient prescription fills for noninjectable dosage forms of butorphanol, fentanyl, hydrocodone, hydromorphone, levorphanol, meperidine, methadone, morphine, oxycodone, oxymorphone, pentazocine, and tramadol. Consistent with the Centers for Disease Control and Prevention and VA/Department of Defense guidelines, LTO use was defined conceptually as regular use for >90 days. Operationalizing this definition to pharmacy refill data was established by using a cabinet supply methodology,⁷ which allows for the construction of episodes of continuous medication therapy by estimating the medication supply available to a patient for each day during a defined period based on the pattern of observed refills. LTO use was defined as an episode of continuous opioid supply for >90 days and beginning within 30 days of the initial prescription. While some studies have defined LTO use based on onset within 1 year following surgery,⁵ the requirement for onset within 30 days of initiation was applied to more strongly tie the association of developing LTO use with the discharge event and minimize various forms of bias that are introduced with extended follow-up periods.

Clinical Characteristics

Patients were classified as being medical discharges, surgical discharges, or outpatient initiators. Patients with an opioid index date within 2 days following discharge were designated based on discharge bed section; additionally, if patients had a surgical bed section during hospitalization, they were assigned as surgical discharges. Demographic, diagnosis, and medication exposure variables that were previously associated with LTO use were selected.^8,9 Substance use disorder, chronic pain, anxiety disorder, and depressive disorder were based on International Classification of Diseases, 9th Revision (ICD-9) codes in the preceding year. The use of concurrent benzodiazepines, skeletal muscle relaxants, and antidepressants were determined at opioid initiation.¹⁰ Rural or urban residence was assigned by using the Rural-Urban Commuting Area Codes system and mapped with the zip code of a veteran’s residence.¹¹

Analysis

Bivariate and multivariable relationships were determined by using logistic regression. The multivariable model considered all pairwise interaction terms between inpatient service (surgery versus medicine) and each of the variables in the model. Statistically significant interaction terms (P < .05) were retained, and all others were omitted from the final model. The main effects for variables that were involved in a significant interaction term were not reported in the final multivariable model; instead, we created fully specified multivariable models for surgery service and medicine service and reported odds ratios (ORs) for the main effects. All analyses were conducted by using SAS version 9.4 (SAS Institute Inc, Cary, North Carolina).

The observation that subsequent LTO use occurs more frequently in discharged medical patients than surgical patients is consistent with the findings of Calcaterra et al.¹ that among patients with no surgery versus surgery during hospitalization, opioid receipt at discharge resulted in a higher adjusted OR (7.24 for no surgery versus 3.40 for surgery) for chronic opioid use at 1 year. One explanation for this finding may be an artifact of cohort selection in the study design: patients with prior opioid use are excluded from the cohort, and prior use may be more common among surgical patients presenting for elective inpatient surgery for painful conditions. Previous work suggests that opioid use preoperatively is a robust predictor of postoperative use, and rates of LTO use are low among patients without preoperative opioid exposure.⁶

Demographic characteristics associated with persistent opioid receipt were similar to those previously reported.^5,8,9 The inclusion of medication classes indicated in the treatment of mental health or pain conditions (ie, antidepressants, benzodiazepines, muscle relaxants, and nonopioid analgesics) resulted in diagnoses based on ICD-9 codes being no longer associated with LTO use. Severity or activity of illness, preferences regarding pharmacologic or nonpharmacologic treatment and undiagnosed or undocumented pain-comorbid conditions may all contribute to this finding. Future work studying opioid-related outcomes should include variables that reflect pharmacologic management of comorbid diagnoses in the cohort development or analytic design.

The strongest risk factors were potentially modifiable: days’ supply, dose, and concurrent medications. The measures of opioid quantity supplied are associated with subsequent ongoing use and are consistent with recent work based on prescription drug–monitoring data in a single state¹⁴ and in a nationally representative sample.¹⁵ That this relationship persists following hospital discharge, a scenario in which LTO use is unlikely to be initiated by a provider (who would be expected to subsequently titrate or monitor therapy), further supports the potential to curtail unintended LTO use through judicious early prescribing decisions.

We assessed only opioids that were supplied through a VA pharmacy, which may lead to the misclassification of patients as opioid naive for inclusion and an underestimation of the rate of opioid use following discharge. It is possible that differences in the rates of non-VA pharmacy use differ in medical and surgical populations in a nonrandom way. This study was performed in a large, integrated health system and may not be generalizable outside the VA system, where more discontinuities between hospital and ambulatory care may exist.

The initiation of LTO use at discharge is more common in veterans who are discharged from medical than surgical hospitalizations, likely reflecting differences in the patient population, pain conditions, and discharge prescribing decisions. While patient characteristics are associated with LTO use, the strongest associations are with increasing index dose and days’ supply; both represent potentially modifiable prescriber behaviors. These findings support policy changes and other efforts to minimize dose and days supplied when short-term use is intended as a means to address the current opioid epidemic.

Acknowledgments

The work reported here was supported by the Department of Veterans Affairs Office of Academic Affiliations and Office of Research and Development (Dr. Mosher and Dr. Hofmeyer), and Health Services Research and Development Service (HSR&D) through the Comprehensive Access and Delivery Research and Evaluation Center (CIN 13-412) and a Career Development Award (CDA 10-017; Dr. Lund).

Disclosures

The authors report no conflict of interest in regard to this study. The authors had full access to and take full responsibility for the integrity of the data. All analyses were conducted by using SAS version 9.2 (SAS Institute Inc, Cary, NC). This manuscript is not under review elsewhere, and there is no prior publication of the manuscript contents. The views expressed in this article are those of the authors and do not necessarily represent the views of the Department of Veterans Affairs. The study was approved by the University of Iowa Institutional Review Board and the Iowa City Healthcare System Research and Development Committee.

Physicians face growing pressure to reduce their use of “low value” care—services that provide either little to no benefit, little benefit relative to cost, or outsized potential harm compared to benefit. One emerging policy solution for deterring such services is to financially penalize physicians who prescribe them.^1,2

Physicians’ willingness to support such policies may depend on who they believe benefits from reductions in low-value care. In previous studies of cancer screening, the more that primary care physicians felt that the money saved from cost-containment efforts went to insurance company profits rather than to patients, the less willing they were to use less expensive cancer screening approaches.³

Similarly, physicians may be more likely to support financial penalty policies if they perceive that the benefits from reducing low-value care accrue to patients (eg, lower out-of-pocket costs) rather than insurers or hospitals (eg, profits and salaries of their leaders). If present, such perceptions could inform incentive design. We explored the hypothesis that support of financial penalties for low-value care would be associated with where physicians thought the money goes.

Study Sample

By using a panel of internists maintained by the American College of Physicians, we conducted a randomized, web-based survey among 484 physicians who were either internal medicine residents or internal medicine physicians practicing hospital medicine.

Survey Instrument

Respondents used a 5-point scale (“strongly disagree” to “strongly agree”) to indicate their agreement with a policy that financially penalizes physicians for prescribing services that provide few benefits to patients. Respondents were asked to simultaneously consider the following hospital medicine services, deemed to be low value based on medical evidence and consensus guidelines⁴: (1) placing, and leaving in, urinary catheters for urine output monitoring in noncritically ill patients, (2) ordering continuous telemetry monitoring for nonintensive care unit patients without a protocol governing continuation, and (3) prescribing stress ulcer prophylaxis for medical patients not at a high risk for gastrointestinal complications. Policy support was defined as “somewhat” or “strongly” agreeing with the policy. As part of another study of this physician cohort, this question varied in how the harm of low-value services was framed: either as harm to patients, to society, or to hospitals and insurers as institutions. Respondent characteristics were balanced across survey versions, and for the current analysis, we pooled responses across all versions.

All other questions in the survey, described in detail elsewhere,⁵ were identical for all respondents. For this analysis, we focused on a question that asked physicians to assume that reducing these services saves money without harming the quality of care and to rate on a 4-point scale (“none” to “a lot”) how much of the money saved would ultimately go to the following 6 nonmutually exclusive areas: (a) other healthcare services for patients, (b) reduced charges to patients’ employers or insurers, (c) reduced out-of-pocket costs for patients, (d) salaries and bonuses for physicians, (e) salaries and profits for insurance companies and their leaders, and (f) salaries and profits for hospitals and/or health systems and their leaders.

Based on the positive correlation identified between the first 4 items (a to d) and negative correlation with the other 2 items (e and f), we reverse-coded the latter 2 and summed all 6 into a single-outcome scale, effectively representing the degree to which the money saved from reducing low-value services accrues generally to patients or physicians instead of to hospitals, insurance companies, and their leaders. The Cronbach alpha for the scale was 0.74, indicating acceptable reliability. Based on scale responses, we dichotomized respondents at the median into those who believe that the money saved from reducing low-value services would accrue as benefits to patients or physicians and those who believe benefits accrue to insurance companies or hospitals and/or health systems and their leaders. The protocol was exempted by the University of Pennsylvania Institutional Review Board.

Statistical Analysis

We used a χ² test and multivariable logistic regression analysis to evaluate the association between policy support and physician beliefs about who benefits from reductions in low-value care. A χ² test and a Kruskal-Wallis test were also used to evaluate the association between other respondent characteristics and beliefs about who benefits from reductions in low-value care. Analyses were performed by using Stata version 14.1 (StataCorp, College Station, TX). Tests of significance were 2-tailed at an alpha of .05.

Compared with nonrespondents, the 187 physicians who responded (39% response rate) were more likely to be female (30% vs 26%, P = 0.001), older (mean age 41 vs 36 years old, P < 0.001), and practicing clinicians rather than internal medicine residents (87% vs 69%, P < 0.001). Twenty-one percent reported that their personal compensation was tied to cost incentives.

Overall, respondents believed that more of any money saved from reducing low-value services would go to profits and leadership salaries for insurance companies and hospitals and/or health systems rather than to patients (panel A of Figure). Few respondents felt that the money saved would ultimately go toward physician compensation.

Despite ongoing efforts to highlight how reducing low-value care benefits patients, physicians in our sample did not believe that much of the money saved would benefit patients.

This result may reflect that while some care patterns are considered low value because they provide little benefit at a high cost, others yield potential harm, regardless of cost. For example, limiting stress ulcer prophylaxis largely aims to avoid clinical harm (eg, adverse drug effects and nosocomial infections). Limiting telemetric monitoring largely aims to reduce costly care that provides only limited benefit. Therefore, the nature of potential benefit to patients is very different—improved clinical outcomes in the former and potential cost savings in the latter. Future studies could separately assess physician attitudes about these 2 different definitions of low-value services.

Our study also demonstrates that the more physicians believe that much of any money saved goes to the profits and salaries of insurance companies, hospitals and/or health systems, and their leaders rather than to patients, the less likely they are to support policies financially penalizing physicians for prescribing low-value services.

Our study does not address why physicians have the beliefs that they have, but a likely explanation, at least in part, is that financial flows in healthcare are complex and tangled. Indeed, a clear understanding of who actually benefits is so hard to determine that these stated beliefs may really derive from views of power or justice rather than from some understanding of funds flow. Whether or not ideological attitudes underlie these expressed beliefs, policymakers and healthcare institutions might be advised to increase transparency about how cost savings are realized and whom they benefit.

Our analysis has limitations. Although it provides insight into where physicians believe relative amounts of money saved go with respect to 6 common options, the study did not include an exhaustive list of possibilities. The response rate also limits the representativeness of our results. Additionally, the study design prevents conclusions about causality; we cannot determine whether the belief that savings go to insurance companies and their executives is what reduces physicians’ enthusiasm for penalties, whether the causal association is in the opposite direction, or whether the 2 factors are linked in another way.

Nonetheless, our findings are consistent with a sense of healthcare justice in which physicians support penalties imposed on themselves only if the resulting benefits accrue to patients rather than to corporate or organizational interests. Effective physician penalties will likely need to address the belief that insurers and provider organizations stand to gain more than patients when low-value care services are reduced.

Disclosure 

Drs. Liao, Schapira, Mitra, and Weissman have no conflicts to disclose. Dr. Navathe serves as advisor to Navvis and Company, Navigant Inc., Lynx Medical, Indegene Inc., and Sutherland Global Services and receives an honorarium from Elsevier Press, none of which have relationship to this manuscript. Dr. Asch is a partner and partial owner of VAL Health, which has no relationship to this manuscript.

Funding

This work was supported by The Leonard Davis Institute of Health Economics at the University of Pennsylvania, which had no role in the study design, data collection, analysis, or interpretation of results.

Physicians face growing pressure to reduce their use of “low value” care—services that provide either little to no benefit, little benefit relative to cost, or outsized potential harm compared to benefit. One emerging policy solution for deterring such services is to financially penalize physicians who prescribe them.^1,2

Physicians’ willingness to support such policies may depend on who they believe benefits from reductions in low-value care. In previous studies of cancer screening, the more that primary care physicians felt that the money saved from cost-containment efforts went to insurance company profits rather than to patients, the less willing they were to use less expensive cancer screening approaches.³

Similarly, physicians may be more likely to support financial penalty policies if they perceive that the benefits from reducing low-value care accrue to patients (eg, lower out-of-pocket costs) rather than insurers or hospitals (eg, profits and salaries of their leaders). If present, such perceptions could inform incentive design. We explored the hypothesis that support of financial penalties for low-value care would be associated with where physicians thought the money goes.

Study Sample

By using a panel of internists maintained by the American College of Physicians, we conducted a randomized, web-based survey among 484 physicians who were either internal medicine residents or internal medicine physicians practicing hospital medicine.

Survey Instrument

Respondents used a 5-point scale (“strongly disagree” to “strongly agree”) to indicate their agreement with a policy that financially penalizes physicians for prescribing services that provide few benefits to patients. Respondents were asked to simultaneously consider the following hospital medicine services, deemed to be low value based on medical evidence and consensus guidelines⁴: (1) placing, and leaving in, urinary catheters for urine output monitoring in noncritically ill patients, (2) ordering continuous telemetry monitoring for nonintensive care unit patients without a protocol governing continuation, and (3) prescribing stress ulcer prophylaxis for medical patients not at a high risk for gastrointestinal complications. Policy support was defined as “somewhat” or “strongly” agreeing with the policy. As part of another study of this physician cohort, this question varied in how the harm of low-value services was framed: either as harm to patients, to society, or to hospitals and insurers as institutions. Respondent characteristics were balanced across survey versions, and for the current analysis, we pooled responses across all versions.

All other questions in the survey, described in detail elsewhere,⁵ were identical for all respondents. For this analysis, we focused on a question that asked physicians to assume that reducing these services saves money without harming the quality of care and to rate on a 4-point scale (“none” to “a lot”) how much of the money saved would ultimately go to the following 6 nonmutually exclusive areas: (a) other healthcare services for patients, (b) reduced charges to patients’ employers or insurers, (c) reduced out-of-pocket costs for patients, (d) salaries and bonuses for physicians, (e) salaries and profits for insurance companies and their leaders, and (f) salaries and profits for hospitals and/or health systems and their leaders.

Based on the positive correlation identified between the first 4 items (a to d) and negative correlation with the other 2 items (e and f), we reverse-coded the latter 2 and summed all 6 into a single-outcome scale, effectively representing the degree to which the money saved from reducing low-value services accrues generally to patients or physicians instead of to hospitals, insurance companies, and their leaders. The Cronbach alpha for the scale was 0.74, indicating acceptable reliability. Based on scale responses, we dichotomized respondents at the median into those who believe that the money saved from reducing low-value services would accrue as benefits to patients or physicians and those who believe benefits accrue to insurance companies or hospitals and/or health systems and their leaders. The protocol was exempted by the University of Pennsylvania Institutional Review Board.

Statistical Analysis

We used a χ² test and multivariable logistic regression analysis to evaluate the association between policy support and physician beliefs about who benefits from reductions in low-value care. A χ² test and a Kruskal-Wallis test were also used to evaluate the association between other respondent characteristics and beliefs about who benefits from reductions in low-value care. Analyses were performed by using Stata version 14.1 (StataCorp, College Station, TX). Tests of significance were 2-tailed at an alpha of .05.

Compared with nonrespondents, the 187 physicians who responded (39% response rate) were more likely to be female (30% vs 26%, P = 0.001), older (mean age 41 vs 36 years old, P < 0.001), and practicing clinicians rather than internal medicine residents (87% vs 69%, P < 0.001). Twenty-one percent reported that their personal compensation was tied to cost incentives.

Overall, respondents believed that more of any money saved from reducing low-value services would go to profits and leadership salaries for insurance companies and hospitals and/or health systems rather than to patients (panel A of Figure). Few respondents felt that the money saved would ultimately go toward physician compensation.

Despite ongoing efforts to highlight how reducing low-value care benefits patients, physicians in our sample did not believe that much of the money saved would benefit patients.

This result may reflect that while some care patterns are considered low value because they provide little benefit at a high cost, others yield potential harm, regardless of cost. For example, limiting stress ulcer prophylaxis largely aims to avoid clinical harm (eg, adverse drug effects and nosocomial infections). Limiting telemetric monitoring largely aims to reduce costly care that provides only limited benefit. Therefore, the nature of potential benefit to patients is very different—improved clinical outcomes in the former and potential cost savings in the latter. Future studies could separately assess physician attitudes about these 2 different definitions of low-value services.

Our study also demonstrates that the more physicians believe that much of any money saved goes to the profits and salaries of insurance companies, hospitals and/or health systems, and their leaders rather than to patients, the less likely they are to support policies financially penalizing physicians for prescribing low-value services.

Our study does not address why physicians have the beliefs that they have, but a likely explanation, at least in part, is that financial flows in healthcare are complex and tangled. Indeed, a clear understanding of who actually benefits is so hard to determine that these stated beliefs may really derive from views of power or justice rather than from some understanding of funds flow. Whether or not ideological attitudes underlie these expressed beliefs, policymakers and healthcare institutions might be advised to increase transparency about how cost savings are realized and whom they benefit.

Our analysis has limitations. Although it provides insight into where physicians believe relative amounts of money saved go with respect to 6 common options, the study did not include an exhaustive list of possibilities. The response rate also limits the representativeness of our results. Additionally, the study design prevents conclusions about causality; we cannot determine whether the belief that savings go to insurance companies and their executives is what reduces physicians’ enthusiasm for penalties, whether the causal association is in the opposite direction, or whether the 2 factors are linked in another way.

Nonetheless, our findings are consistent with a sense of healthcare justice in which physicians support penalties imposed on themselves only if the resulting benefits accrue to patients rather than to corporate or organizational interests. Effective physician penalties will likely need to address the belief that insurers and provider organizations stand to gain more than patients when low-value care services are reduced.

Disclosure 

Drs. Liao, Schapira, Mitra, and Weissman have no conflicts to disclose. Dr. Navathe serves as advisor to Navvis and Company, Navigant Inc., Lynx Medical, Indegene Inc., and Sutherland Global Services and receives an honorarium from Elsevier Press, none of which have relationship to this manuscript. Dr. Asch is a partner and partial owner of VAL Health, which has no relationship to this manuscript.

Funding

This work was supported by The Leonard Davis Institute of Health Economics at the University of Pennsylvania, which had no role in the study design, data collection, analysis, or interpretation of results.

Physicians face growing pressure to reduce their use of “low value” care—services that provide either little to no benefit, little benefit relative to cost, or outsized potential harm compared to benefit. One emerging policy solution for deterring such services is to financially penalize physicians who prescribe them.^1,2

Physicians’ willingness to support such policies may depend on who they believe benefits from reductions in low-value care. In previous studies of cancer screening, the more that primary care physicians felt that the money saved from cost-containment efforts went to insurance company profits rather than to patients, the less willing they were to use less expensive cancer screening approaches.³

Similarly, physicians may be more likely to support financial penalty policies if they perceive that the benefits from reducing low-value care accrue to patients (eg, lower out-of-pocket costs) rather than insurers or hospitals (eg, profits and salaries of their leaders). If present, such perceptions could inform incentive design. We explored the hypothesis that support of financial penalties for low-value care would be associated with where physicians thought the money goes.

Study Sample

By using a panel of internists maintained by the American College of Physicians, we conducted a randomized, web-based survey among 484 physicians who were either internal medicine residents or internal medicine physicians practicing hospital medicine.

Survey Instrument

Respondents used a 5-point scale (“strongly disagree” to “strongly agree”) to indicate their agreement with a policy that financially penalizes physicians for prescribing services that provide few benefits to patients. Respondents were asked to simultaneously consider the following hospital medicine services, deemed to be low value based on medical evidence and consensus guidelines⁴: (1) placing, and leaving in, urinary catheters for urine output monitoring in noncritically ill patients, (2) ordering continuous telemetry monitoring for nonintensive care unit patients without a protocol governing continuation, and (3) prescribing stress ulcer prophylaxis for medical patients not at a high risk for gastrointestinal complications. Policy support was defined as “somewhat” or “strongly” agreeing with the policy. As part of another study of this physician cohort, this question varied in how the harm of low-value services was framed: either as harm to patients, to society, or to hospitals and insurers as institutions. Respondent characteristics were balanced across survey versions, and for the current analysis, we pooled responses across all versions.

All other questions in the survey, described in detail elsewhere,⁵ were identical for all respondents. For this analysis, we focused on a question that asked physicians to assume that reducing these services saves money without harming the quality of care and to rate on a 4-point scale (“none” to “a lot”) how much of the money saved would ultimately go to the following 6 nonmutually exclusive areas: (a) other healthcare services for patients, (b) reduced charges to patients’ employers or insurers, (c) reduced out-of-pocket costs for patients, (d) salaries and bonuses for physicians, (e) salaries and profits for insurance companies and their leaders, and (f) salaries and profits for hospitals and/or health systems and their leaders.

Based on the positive correlation identified between the first 4 items (a to d) and negative correlation with the other 2 items (e and f), we reverse-coded the latter 2 and summed all 6 into a single-outcome scale, effectively representing the degree to which the money saved from reducing low-value services accrues generally to patients or physicians instead of to hospitals, insurance companies, and their leaders. The Cronbach alpha for the scale was 0.74, indicating acceptable reliability. Based on scale responses, we dichotomized respondents at the median into those who believe that the money saved from reducing low-value services would accrue as benefits to patients or physicians and those who believe benefits accrue to insurance companies or hospitals and/or health systems and their leaders. The protocol was exempted by the University of Pennsylvania Institutional Review Board.

Statistical Analysis

We used a χ² test and multivariable logistic regression analysis to evaluate the association between policy support and physician beliefs about who benefits from reductions in low-value care. A χ² test and a Kruskal-Wallis test were also used to evaluate the association between other respondent characteristics and beliefs about who benefits from reductions in low-value care. Analyses were performed by using Stata version 14.1 (StataCorp, College Station, TX). Tests of significance were 2-tailed at an alpha of .05.

Compared with nonrespondents, the 187 physicians who responded (39% response rate) were more likely to be female (30% vs 26%, P = 0.001), older (mean age 41 vs 36 years old, P < 0.001), and practicing clinicians rather than internal medicine residents (87% vs 69%, P < 0.001). Twenty-one percent reported that their personal compensation was tied to cost incentives.

Overall, respondents believed that more of any money saved from reducing low-value services would go to profits and leadership salaries for insurance companies and hospitals and/or health systems rather than to patients (panel A of Figure). Few respondents felt that the money saved would ultimately go toward physician compensation.

Despite ongoing efforts to highlight how reducing low-value care benefits patients, physicians in our sample did not believe that much of the money saved would benefit patients.

This result may reflect that while some care patterns are considered low value because they provide little benefit at a high cost, others yield potential harm, regardless of cost. For example, limiting stress ulcer prophylaxis largely aims to avoid clinical harm (eg, adverse drug effects and nosocomial infections). Limiting telemetric monitoring largely aims to reduce costly care that provides only limited benefit. Therefore, the nature of potential benefit to patients is very different—improved clinical outcomes in the former and potential cost savings in the latter. Future studies could separately assess physician attitudes about these 2 different definitions of low-value services.

Our study also demonstrates that the more physicians believe that much of any money saved goes to the profits and salaries of insurance companies, hospitals and/or health systems, and their leaders rather than to patients, the less likely they are to support policies financially penalizing physicians for prescribing low-value services.

Our study does not address why physicians have the beliefs that they have, but a likely explanation, at least in part, is that financial flows in healthcare are complex and tangled. Indeed, a clear understanding of who actually benefits is so hard to determine that these stated beliefs may really derive from views of power or justice rather than from some understanding of funds flow. Whether or not ideological attitudes underlie these expressed beliefs, policymakers and healthcare institutions might be advised to increase transparency about how cost savings are realized and whom they benefit.

Our analysis has limitations. Although it provides insight into where physicians believe relative amounts of money saved go with respect to 6 common options, the study did not include an exhaustive list of possibilities. The response rate also limits the representativeness of our results. Additionally, the study design prevents conclusions about causality; we cannot determine whether the belief that savings go to insurance companies and their executives is what reduces physicians’ enthusiasm for penalties, whether the causal association is in the opposite direction, or whether the 2 factors are linked in another way.

Nonetheless, our findings are consistent with a sense of healthcare justice in which physicians support penalties imposed on themselves only if the resulting benefits accrue to patients rather than to corporate or organizational interests. Effective physician penalties will likely need to address the belief that insurers and provider organizations stand to gain more than patients when low-value care services are reduced.

Disclosure 

Drs. Liao, Schapira, Mitra, and Weissman have no conflicts to disclose. Dr. Navathe serves as advisor to Navvis and Company, Navigant Inc., Lynx Medical, Indegene Inc., and Sutherland Global Services and receives an honorarium from Elsevier Press, none of which have relationship to this manuscript. Dr. Asch is a partner and partial owner of VAL Health, which has no relationship to this manuscript.

Funding

This work was supported by The Leonard Davis Institute of Health Economics at the University of Pennsylvania, which had no role in the study design, data collection, analysis, or interpretation of results.

Reducing low-value care—services for which there is little to no benefit, little benefit relative to cost, or outsized potential harm compared with benefit—is an essential step toward maintaining or improving quality while lowering cost. Unfortunately, low-value services persist widelydespite professional consensus, guidelines, and national campaigns aimed to reduce them.^1-3 In turn, policy makers are beginning to consider financially penalizing physicians in order to deter low-value services.^4,5 Physician support for such penalties remains unknown. In this study, we used a randomized survey experiment to evaluate how the framing of harms from low-value care—in terms of those to patients, healthcare institutions, or society—influenced physician support of financial penalties for low-value care services.

Study Sample

By using a stratified random sample maintained by the American College of Physicians, we conducted a web-based survey among 484 physicians who were either internal medicine residents or internists practicing hospital medicine.

Instrument Design and Administration

Our study focused on 3 low-value services relevant to inpatient medicine: (1) placing, and leaving in, urinary catheters for urine output monitoring in noncritically ill patients; (2) ordering continuous telemetry monitoring for nonintensive care unit (non-ICU) patients without a protocol governing continuation; and (3) prescribing stress ulcer prophylaxis for medical patients not at a high risk for gastrointestinal (GI) complications. Although the nature and trade-offs between costs, harms, and benefits vary by individual service, all 3 are promulgated through the Choosing Wisely^® guidelines as low value based on existing data and professional consensus from the Society of Hospital Medicine.⁶

To evaluate intended behavior related to these 3 low-value services, respondents were first presented with 3 clinical vignettes focused on the care of patients hospitalized for pneumonia, congestive heart failure, and alcohol withdrawal, which were selected to reflect common inpatient medicine scenarios. Respondents were asked to use a 4-point scale (very likely to very unlikely) to estimate how likely they were to recommend various tests or treatments, including the low-value services noted above. Respondents who were “somewhat unlikely” and “very unlikely” to recommend low-value services were considered concordant with low-value care guidelines.

Following the vignettes, respondents then used a 5-point scale (strongly agree to strongly disagree) to indicate their agreement with a policy that financially penalizes physicians for prescribing each service. Support was defined as “somewhat or strongly” agreeing with the policy. Respondents were randomized to receive 1 of 3 versions of this question (supplementary Appendix).

All versions stated that, “According to research and expert opinion, certain aspects of inpatient care provide little benefit to patients” and listed the 3 low-value services noted above. The “patient harm” version also described the harm of low-value care as costs to patients and risk for clinical harms and complications. The “societal harm” version described the harms as costs to society and utilization of limited healthcare resources. The “institutional harm” version described harms as costs to hospitals and insurers.

Other survey items were adapted from existing literature^7-9 and evaluated respondent beliefs about the effectiveness of physician incentives in improving the value of care, as well as the appropriateness of including cost considerations in clinical decision-making.

The instrument was pilot tested among study team members and several independent internists affiliated with the University of Pennsylvania. After incorporating feedback into the final instrument, the web-based survey was distributed to eligible physicians via e-mail. Responses were anonymous and respondents received a $15 gift card for participation. The protocol was reviewed and deemed exempt by the University of Pennsylvania Institutional Review Board.

Statistical Analysis

Respondent characteristics (sociodemographic, intended clinical behavior, and cost control attitudes) were described by using percentages for categorical variables and medians and interquartile ranges for continuous variables. Balance in respondent characteristics across survey versions was evaluated using χ² and Kruskal-Wallis tests. Multivariable logistic regression, adjusted for characteristics in the Table, was used to evaluate the association between survey version and policy support. All tests of significance were 2-tailed with significance level alpha = 0.05. Analyses were performed using STATA version 14.1 (StataCorp LLC, College Station, TX, http://www.stata.com).

Of 484 eligible respondents, 187 (39%) completed the survey. Compared with nonrespondents, respondents were more likely to be female (30% vs 26%, P = 0.001), older (mean age 41 vs 36 years, P < 0.001), and practicing clinicians rather than internal medicine residents (87% vs 69%, P < 0.001). Physician characteristics were similar across the 3 survey versions (Table). Most respondents agreed that financial incentives for individual physicians is an effective way to improve the value of healthcare (73.3%) and that physicians should consider the costs of a test or treatment to society when making clinical decisions for patients (79.1%). The majority also felt that clinicians have a duty to offer a test or treatment to a patient if it has any chance of helping them (70.1%) and that it is inappropriate for anyone beyond the clinician and patient to decide if a test or treatment is “worth the cost” (63.6%).

To our knowledge, this study is the first to prospectively evaluate physician support of financial penalties for low-value services relevant to hospital medicine. It has 2 main findings.

First, although overall policy support was relatively low (39.6%), it varied significantly on the basis of how the harms of low-value care were framed. Support was highest in the “societal harm” version, suggesting that emphasizing these harms may increase acceptability of financial penalties among physicians and contribute to the larger effort to decrease low-value care in hospital settings. The comparatively low support for the “patient harm” version is somewhat surprising but may reflect variation in the nature of harm, benefit, and cost trade-offs for individual low-value services, as noted above, and physician belief that some low-value services do not in fact produce significant clinical harms.

For example, whereas evidence demonstrates that stress ulcer prophylaxis in non-ICU patients can harm patients through nosocomial infections and adverse drug effects,^10,11 the clinical harms of telemetry are less obvious. Telemetry’s low value derives more from its high cost relative to benefit, rather than its potential for clinical harm.⁶ The many paths to “low value” underscore the need to examine attitudes and uptake toward these services separately and may explain the wide range in concordance between intended clinical behavior and low-value care guidelines (11.8% to 78.6%).

Reinforcing policies could more effectively deter low-value care. For example, multiple forces, including Medicare payment reform and national accreditation policies,^12,13 have converged to discourage low-value use of urinary catheters in hospitalized patients. In contrast, there has been little reinforcement beyond consensus guidelines to reduce low-value use of telemetric monitoring. Given questions about whether consensus methods alone can deter low-value care beyond obvious “low hanging fruit,”¹⁴ policy makers could coordinate policies to accelerate progress within other priority areas.

Broad policies should also be paired with local initiatives to influence physician behavior. For example, health systems have begun successfully leveraging the electronic medical record and utilizing behavioral economics principles to design interventions to reduce inappropriate overuse of antibiotics for upper respiratory infections in primary care clinics.¹⁵ Organizations are also redesigning care processes in response to resource utilization imperatives under ongoing value-based care payment reform. Care redesign and behavioral interventions embedded at the point of care can both help deter low-value services in inpatient settings.

Study limitations include a relatively low response rate, which limits generalizability. However, all 3 randomized groups were similar on measured characteristics, and experimental randomization reduces the nonresponse bias concerns accompanying descriptive surveys. Additionally, although we evaluated intended clinical behavior in a national sample, our results may not reflect actual behavior among all physicians practicing hospital medicine. Future work could include assessments of actual or self-reported practices or examine additional factors, including site, years of practice, knowledge about guidelines, and other possible determinants of guideline-concordant behaviors.

Despite these limitations, our study provides important early evidence about physician support of financial penalties for low-value care relevant to hospital medicine. As policy makers design and organizational leaders implement financial incentive policies, this information can help increase their acceptability among physicians and more effectively reduce low-value care within hospitals.

Disclosure

Drs. Liao, Schapira, Mitra, and Weissman have no conflicts to disclose. Dr. Navathe serves as advisor to Navvis and Company, Navigant Inc, Lynx Medical, Indegene Inc, and Sutherland Global Services and receives an honorarium from Elsevier Press, none of which have relationship to this manuscript. Dr. Asch is a partner and part owner of VAL Health, which has no relationship to this manuscript.

Funding

This work was supported by The Leonard Davis Institute of Health Economics at the University of Pennsylvania, which had no role in the study design, data collection, analysis, or interpretation of results.

Reducing low-value care—services for which there is little to no benefit, little benefit relative to cost, or outsized potential harm compared with benefit—is an essential step toward maintaining or improving quality while lowering cost. Unfortunately, low-value services persist widelydespite professional consensus, guidelines, and national campaigns aimed to reduce them.^1-3 In turn, policy makers are beginning to consider financially penalizing physicians in order to deter low-value services.^4,5 Physician support for such penalties remains unknown. In this study, we used a randomized survey experiment to evaluate how the framing of harms from low-value care—in terms of those to patients, healthcare institutions, or society—influenced physician support of financial penalties for low-value care services.

Study Sample

By using a stratified random sample maintained by the American College of Physicians, we conducted a web-based survey among 484 physicians who were either internal medicine residents or internists practicing hospital medicine.

Instrument Design and Administration

Our study focused on 3 low-value services relevant to inpatient medicine: (1) placing, and leaving in, urinary catheters for urine output monitoring in noncritically ill patients; (2) ordering continuous telemetry monitoring for nonintensive care unit (non-ICU) patients without a protocol governing continuation; and (3) prescribing stress ulcer prophylaxis for medical patients not at a high risk for gastrointestinal (GI) complications. Although the nature and trade-offs between costs, harms, and benefits vary by individual service, all 3 are promulgated through the Choosing Wisely^® guidelines as low value based on existing data and professional consensus from the Society of Hospital Medicine.⁶

To evaluate intended behavior related to these 3 low-value services, respondents were first presented with 3 clinical vignettes focused on the care of patients hospitalized for pneumonia, congestive heart failure, and alcohol withdrawal, which were selected to reflect common inpatient medicine scenarios. Respondents were asked to use a 4-point scale (very likely to very unlikely) to estimate how likely they were to recommend various tests or treatments, including the low-value services noted above. Respondents who were “somewhat unlikely” and “very unlikely” to recommend low-value services were considered concordant with low-value care guidelines.

Following the vignettes, respondents then used a 5-point scale (strongly agree to strongly disagree) to indicate their agreement with a policy that financially penalizes physicians for prescribing each service. Support was defined as “somewhat or strongly” agreeing with the policy. Respondents were randomized to receive 1 of 3 versions of this question (supplementary Appendix).

All versions stated that, “According to research and expert opinion, certain aspects of inpatient care provide little benefit to patients” and listed the 3 low-value services noted above. The “patient harm” version also described the harm of low-value care as costs to patients and risk for clinical harms and complications. The “societal harm” version described the harms as costs to society and utilization of limited healthcare resources. The “institutional harm” version described harms as costs to hospitals and insurers.

Other survey items were adapted from existing literature^7-9 and evaluated respondent beliefs about the effectiveness of physician incentives in improving the value of care, as well as the appropriateness of including cost considerations in clinical decision-making.

The instrument was pilot tested among study team members and several independent internists affiliated with the University of Pennsylvania. After incorporating feedback into the final instrument, the web-based survey was distributed to eligible physicians via e-mail. Responses were anonymous and respondents received a $15 gift card for participation. The protocol was reviewed and deemed exempt by the University of Pennsylvania Institutional Review Board.

Statistical Analysis

Respondent characteristics (sociodemographic, intended clinical behavior, and cost control attitudes) were described by using percentages for categorical variables and medians and interquartile ranges for continuous variables. Balance in respondent characteristics across survey versions was evaluated using χ² and Kruskal-Wallis tests. Multivariable logistic regression, adjusted for characteristics in the Table, was used to evaluate the association between survey version and policy support. All tests of significance were 2-tailed with significance level alpha = 0.05. Analyses were performed using STATA version 14.1 (StataCorp LLC, College Station, TX, http://www.stata.com).

Of 484 eligible respondents, 187 (39%) completed the survey. Compared with nonrespondents, respondents were more likely to be female (30% vs 26%, P = 0.001), older (mean age 41 vs 36 years, P < 0.001), and practicing clinicians rather than internal medicine residents (87% vs 69%, P < 0.001). Physician characteristics were similar across the 3 survey versions (Table). Most respondents agreed that financial incentives for individual physicians is an effective way to improve the value of healthcare (73.3%) and that physicians should consider the costs of a test or treatment to society when making clinical decisions for patients (79.1%). The majority also felt that clinicians have a duty to offer a test or treatment to a patient if it has any chance of helping them (70.1%) and that it is inappropriate for anyone beyond the clinician and patient to decide if a test or treatment is “worth the cost” (63.6%).

To our knowledge, this study is the first to prospectively evaluate physician support of financial penalties for low-value services relevant to hospital medicine. It has 2 main findings.

First, although overall policy support was relatively low (39.6%), it varied significantly on the basis of how the harms of low-value care were framed. Support was highest in the “societal harm” version, suggesting that emphasizing these harms may increase acceptability of financial penalties among physicians and contribute to the larger effort to decrease low-value care in hospital settings. The comparatively low support for the “patient harm” version is somewhat surprising but may reflect variation in the nature of harm, benefit, and cost trade-offs for individual low-value services, as noted above, and physician belief that some low-value services do not in fact produce significant clinical harms.

For example, whereas evidence demonstrates that stress ulcer prophylaxis in non-ICU patients can harm patients through nosocomial infections and adverse drug effects,^10,11 the clinical harms of telemetry are less obvious. Telemetry’s low value derives more from its high cost relative to benefit, rather than its potential for clinical harm.⁶ The many paths to “low value” underscore the need to examine attitudes and uptake toward these services separately and may explain the wide range in concordance between intended clinical behavior and low-value care guidelines (11.8% to 78.6%).

Reinforcing policies could more effectively deter low-value care. For example, multiple forces, including Medicare payment reform and national accreditation policies,^12,13 have converged to discourage low-value use of urinary catheters in hospitalized patients. In contrast, there has been little reinforcement beyond consensus guidelines to reduce low-value use of telemetric monitoring. Given questions about whether consensus methods alone can deter low-value care beyond obvious “low hanging fruit,”¹⁴ policy makers could coordinate policies to accelerate progress within other priority areas.

Broad policies should also be paired with local initiatives to influence physician behavior. For example, health systems have begun successfully leveraging the electronic medical record and utilizing behavioral economics principles to design interventions to reduce inappropriate overuse of antibiotics for upper respiratory infections in primary care clinics.¹⁵ Organizations are also redesigning care processes in response to resource utilization imperatives under ongoing value-based care payment reform. Care redesign and behavioral interventions embedded at the point of care can both help deter low-value services in inpatient settings.

Study limitations include a relatively low response rate, which limits generalizability. However, all 3 randomized groups were similar on measured characteristics, and experimental randomization reduces the nonresponse bias concerns accompanying descriptive surveys. Additionally, although we evaluated intended clinical behavior in a national sample, our results may not reflect actual behavior among all physicians practicing hospital medicine. Future work could include assessments of actual or self-reported practices or examine additional factors, including site, years of practice, knowledge about guidelines, and other possible determinants of guideline-concordant behaviors.

Despite these limitations, our study provides important early evidence about physician support of financial penalties for low-value care relevant to hospital medicine. As policy makers design and organizational leaders implement financial incentive policies, this information can help increase their acceptability among physicians and more effectively reduce low-value care within hospitals.

Disclosure

Drs. Liao, Schapira, Mitra, and Weissman have no conflicts to disclose. Dr. Navathe serves as advisor to Navvis and Company, Navigant Inc, Lynx Medical, Indegene Inc, and Sutherland Global Services and receives an honorarium from Elsevier Press, none of which have relationship to this manuscript. Dr. Asch is a partner and part owner of VAL Health, which has no relationship to this manuscript.

Funding

This work was supported by The Leonard Davis Institute of Health Economics at the University of Pennsylvania, which had no role in the study design, data collection, analysis, or interpretation of results.

Reducing low-value care—services for which there is little to no benefit, little benefit relative to cost, or outsized potential harm compared with benefit—is an essential step toward maintaining or improving quality while lowering cost. Unfortunately, low-value services persist widelydespite professional consensus, guidelines, and national campaigns aimed to reduce them.^1-3 In turn, policy makers are beginning to consider financially penalizing physicians in order to deter low-value services.^4,5 Physician support for such penalties remains unknown. In this study, we used a randomized survey experiment to evaluate how the framing of harms from low-value care—in terms of those to patients, healthcare institutions, or society—influenced physician support of financial penalties for low-value care services.

Study Sample

By using a stratified random sample maintained by the American College of Physicians, we conducted a web-based survey among 484 physicians who were either internal medicine residents or internists practicing hospital medicine.

Instrument Design and Administration

Our study focused on 3 low-value services relevant to inpatient medicine: (1) placing, and leaving in, urinary catheters for urine output monitoring in noncritically ill patients; (2) ordering continuous telemetry monitoring for nonintensive care unit (non-ICU) patients without a protocol governing continuation; and (3) prescribing stress ulcer prophylaxis for medical patients not at a high risk for gastrointestinal (GI) complications. Although the nature and trade-offs between costs, harms, and benefits vary by individual service, all 3 are promulgated through the Choosing Wisely^® guidelines as low value based on existing data and professional consensus from the Society of Hospital Medicine.⁶

To evaluate intended behavior related to these 3 low-value services, respondents were first presented with 3 clinical vignettes focused on the care of patients hospitalized for pneumonia, congestive heart failure, and alcohol withdrawal, which were selected to reflect common inpatient medicine scenarios. Respondents were asked to use a 4-point scale (very likely to very unlikely) to estimate how likely they were to recommend various tests or treatments, including the low-value services noted above. Respondents who were “somewhat unlikely” and “very unlikely” to recommend low-value services were considered concordant with low-value care guidelines.

Following the vignettes, respondents then used a 5-point scale (strongly agree to strongly disagree) to indicate their agreement with a policy that financially penalizes physicians for prescribing each service. Support was defined as “somewhat or strongly” agreeing with the policy. Respondents were randomized to receive 1 of 3 versions of this question (supplementary Appendix).

All versions stated that, “According to research and expert opinion, certain aspects of inpatient care provide little benefit to patients” and listed the 3 low-value services noted above. The “patient harm” version also described the harm of low-value care as costs to patients and risk for clinical harms and complications. The “societal harm” version described the harms as costs to society and utilization of limited healthcare resources. The “institutional harm” version described harms as costs to hospitals and insurers.

Other survey items were adapted from existing literature^7-9 and evaluated respondent beliefs about the effectiveness of physician incentives in improving the value of care, as well as the appropriateness of including cost considerations in clinical decision-making.

The instrument was pilot tested among study team members and several independent internists affiliated with the University of Pennsylvania. After incorporating feedback into the final instrument, the web-based survey was distributed to eligible physicians via e-mail. Responses were anonymous and respondents received a $15 gift card for participation. The protocol was reviewed and deemed exempt by the University of Pennsylvania Institutional Review Board.

Statistical Analysis

Respondent characteristics (sociodemographic, intended clinical behavior, and cost control attitudes) were described by using percentages for categorical variables and medians and interquartile ranges for continuous variables. Balance in respondent characteristics across survey versions was evaluated using χ² and Kruskal-Wallis tests. Multivariable logistic regression, adjusted for characteristics in the Table, was used to evaluate the association between survey version and policy support. All tests of significance were 2-tailed with significance level alpha = 0.05. Analyses were performed using STATA version 14.1 (StataCorp LLC, College Station, TX, http://www.stata.com).

Of 484 eligible respondents, 187 (39%) completed the survey. Compared with nonrespondents, respondents were more likely to be female (30% vs 26%, P = 0.001), older (mean age 41 vs 36 years, P < 0.001), and practicing clinicians rather than internal medicine residents (87% vs 69%, P < 0.001). Physician characteristics were similar across the 3 survey versions (Table). Most respondents agreed that financial incentives for individual physicians is an effective way to improve the value of healthcare (73.3%) and that physicians should consider the costs of a test or treatment to society when making clinical decisions for patients (79.1%). The majority also felt that clinicians have a duty to offer a test or treatment to a patient if it has any chance of helping them (70.1%) and that it is inappropriate for anyone beyond the clinician and patient to decide if a test or treatment is “worth the cost” (63.6%).

To our knowledge, this study is the first to prospectively evaluate physician support of financial penalties for low-value services relevant to hospital medicine. It has 2 main findings.

First, although overall policy support was relatively low (39.6%), it varied significantly on the basis of how the harms of low-value care were framed. Support was highest in the “societal harm” version, suggesting that emphasizing these harms may increase acceptability of financial penalties among physicians and contribute to the larger effort to decrease low-value care in hospital settings. The comparatively low support for the “patient harm” version is somewhat surprising but may reflect variation in the nature of harm, benefit, and cost trade-offs for individual low-value services, as noted above, and physician belief that some low-value services do not in fact produce significant clinical harms.

For example, whereas evidence demonstrates that stress ulcer prophylaxis in non-ICU patients can harm patients through nosocomial infections and adverse drug effects,^10,11 the clinical harms of telemetry are less obvious. Telemetry’s low value derives more from its high cost relative to benefit, rather than its potential for clinical harm.⁶ The many paths to “low value” underscore the need to examine attitudes and uptake toward these services separately and may explain the wide range in concordance between intended clinical behavior and low-value care guidelines (11.8% to 78.6%).

Reinforcing policies could more effectively deter low-value care. For example, multiple forces, including Medicare payment reform and national accreditation policies,^12,13 have converged to discourage low-value use of urinary catheters in hospitalized patients. In contrast, there has been little reinforcement beyond consensus guidelines to reduce low-value use of telemetric monitoring. Given questions about whether consensus methods alone can deter low-value care beyond obvious “low hanging fruit,”¹⁴ policy makers could coordinate policies to accelerate progress within other priority areas.

Broad policies should also be paired with local initiatives to influence physician behavior. For example, health systems have begun successfully leveraging the electronic medical record and utilizing behavioral economics principles to design interventions to reduce inappropriate overuse of antibiotics for upper respiratory infections in primary care clinics.¹⁵ Organizations are also redesigning care processes in response to resource utilization imperatives under ongoing value-based care payment reform. Care redesign and behavioral interventions embedded at the point of care can both help deter low-value services in inpatient settings.

Study limitations include a relatively low response rate, which limits generalizability. However, all 3 randomized groups were similar on measured characteristics, and experimental randomization reduces the nonresponse bias concerns accompanying descriptive surveys. Additionally, although we evaluated intended clinical behavior in a national sample, our results may not reflect actual behavior among all physicians practicing hospital medicine. Future work could include assessments of actual or self-reported practices or examine additional factors, including site, years of practice, knowledge about guidelines, and other possible determinants of guideline-concordant behaviors.

Despite these limitations, our study provides important early evidence about physician support of financial penalties for low-value care relevant to hospital medicine. As policy makers design and organizational leaders implement financial incentive policies, this information can help increase their acceptability among physicians and more effectively reduce low-value care within hospitals.

Disclosure

Drs. Liao, Schapira, Mitra, and Weissman have no conflicts to disclose. Dr. Navathe serves as advisor to Navvis and Company, Navigant Inc, Lynx Medical, Indegene Inc, and Sutherland Global Services and receives an honorarium from Elsevier Press, none of which have relationship to this manuscript. Dr. Asch is a partner and part owner of VAL Health, which has no relationship to this manuscript.

Funding

This work was supported by The Leonard Davis Institute of Health Economics at the University of Pennsylvania, which had no role in the study design, data collection, analysis, or interpretation of results.

As part of the Choosing Wisely^® campaign, the Society of Hospital Medicine recommends against performing “repetitive complete blood count [CBC] and chemistry testing in the face of clinical and lab stability.”¹ This recommendation stems from a body of research that shows that frequent or excessive phlebotomy can have negative consequences, including iatrogenic anemia (termed hospital-acquired anemia), which may necessitate blood transfusion.² The downstream effects of potentially unnecessary testing, including the evaluation of false-positive results, must also be considered. Additional important effects include patient discomfort and disruption of sleep and unproductive work by hospital staff, including nurses, phlebotomists, and laboratory technicians.

Though interventions to reduce unnecessary daily labs have been previously evaluated, there are no studies that focus on decreasing lab testing on patients deemed clinically stable and close to discharge. This is in part due to the absence of clear criteria or guidelines to define clinical stability in the context of lab utilization.

We therefore aimed to implement a multifaceted, patient-centered initiative—the Necessity of Labs Assessed Bedside (NO LABS)—that focused on reducing lab testing in patients at 24 to 48 hours before discharge. We targeted the 24 to 48-hour period before the anticipated date of discharge, as this may be a period of greater stability and provide an opportunity to identify and decrease unnecessary testing.

The study took place at Mount Sinai Hospital, which is an 1174-bed tertiary care teaching hospital in New York City. We targeted 2 inpatient medicine units where virtually all patients are assigned to a hospitalist rotating for a 2- to 4-week period, for the period of July 1, 2015, to July 31, 2016. These units employed bedside interdisciplinary rounds (IDR) attended by the hospitalist, social worker, case manager, nurse, nurse manager, and medical director. Bedside IDR focuses on the daily plan and patient safety by utilizing a scripted format.³ Our multifaceted intervention included prompting the hospitalist physician during bedside IDR, education of the clinicians, and regular data review for the hospitalists and unit staff.

As described by Dunn et al.,³ the IDR script included the following: a review of the plan of care by the hospitalist, identifying a patient’s personal goals for the day, a brief update of discharge planning (as appropriate), and a safety assessment performed by the nurse (identifying Foley catheters, falls risk, etc). We incorporated an inquiry into the daily IDR script identifying clinically stable patients for discharge in the next 24 to 48 hours (based on physician judgment), followed by a prompt to the hospitalist to discontinue labs when appropriate. The unit medical director and nurse manager were both tasked with prompting the hospitalist at the bedside. Our hospital utilizes computerized physician order entry. Lab orders were then discontinued by the clinician during rounds using a computer on wheels (or after rounds when one was not available). The hospitalist, unit medical director, and nurse manager were reminded about the project through weekly e-mails and in-person communication.

To assess whether the prompt was being incorporated consistently, an observer was added to rounds beginning in the second month of the project. The observer was present at least 3 times a week for the subsequent 3 months of the project. Our intervention also included education geared towards hospitalists, including a brief presentation on reducing unnecessary lab testing during a monthly hospitalist faculty meeting (the first and sixth month of the intervention). The group’s data on laboratory testing within the 24 to 48 hours prior to discharge were also presented at these monthly meetings (beginning 2 months into the intervention and monthly thereafter). Lastly, we provided the unit staff with unit-level metrics, biweekly for the first 3 months and every 2 to 3 months thereafter.

We extracted electronic medical record (EMR) data on lab utilization for patients on the 2 hospitalist units for the intervention period. Baseline data were obtained from July 1, 2014, to June 30, 2015. Patients with a length of stay (LOS) ≤7 days (75th percentile) were included; on these units, longer stays were considered more likely to have complex social issues delaying discharge and thus less likely to require laboratory testing. We tracked ordering for 4 common lab tests: basic metabolic panel, CBC, CBC with differential, and the comprehensive metabolic panel. The primary outcome was the monthly percentage of patients for whom testing was ordered in the 24 and 48 hours preceding discharge. A secondary outcome was testing at 72 hours preceding discharge to identify any potential compensatory (increased) testing the evening prior. We applied a quasi-experimental interrupted time series design with a segmented regression analysis to estimate changes before and after our intervention, expressed in acute changes (change in intercept) and over time (changes in trend) while adjusting for preintervention trends. All analyses were performed with SAS v9.4 statistical software (SAS Institute, Cary, NC). Our project was deemed a quality improvement project and thus an IRB submission was not required.

There were 1579 discharges in the preintervention period and 1308 discharges in the postintervention period. The average age of the patient population was similar in the baseline and postintervention groups (61.5 vs 59.3 years; P = 0.400), and there was no difference in the mean LOS before and after implementation (3.67 vs 3.68 days; P = 0.817).

Our structured, multifaceted approach effectively reduced daily lab testing in the 24 to 48 hours prior to discharge. Bedside IDR provided a unique opportunity to effectively communicate to the patient about necessary (or unnecessary) testing. Moreover, given the complexity of identifying clinical stability, our strategy focused on the onset of discharge planning, a more easily discernible and less obtrusive focal point to promote the discontinuation of lab testing.

Though the nature of bundled interventions can make it difficult to identify which intervention is most effective, we believe that all interventions were effective in different capacities during various phases in the intervention period. We believe that the decrease in lab testing in the 24 to 48 hours preceding discharge was primarily driven by the new rounding structure. This is evident in the significant decrease seen in the first few months of the intervention period. Six months into the intervention, we begin to see a decrease at 72 hours prior to discharge. Additionally, we see a decrease in the mean number of labs per patient day over the entire hospitalization period. We attribute these results to a gradual shift in the culture in our division as a direct consequence of educational sessions and individual feedback provided during this time.

To our knowledge, this is the first study to use anticipated discharge as a correlate for clinical stability and therefore as an opportunity to prompt discontinuation of laboratory testing. Other studies evaluated interventions targeting the EMR and the ease with which providers can order recurring labs. These include restricting recurring orders in the EMR,⁴ a robust education and awareness campaign targeting house staff,⁵ and other multifaceted approaches to decreasing lab utilization,⁶ all of which have shown promising results. While these approaches show varying degrees of success, ours is unique in its focus on the period prior to discharge. In addition, the intervention can be readily implemented in settings that utilize scripted IDR. It also brings high-value decision-making to the bedside by informing the patient that in the setting of presumed clinical stability, no additional tests are warranted.

Our study has several limitations. First, while interdisciplinary discharge rounds are widely implemented,^7,8 our rounds occur at the bedside and employ a script, potentially limiting generalizability. The structured prompting may be feasible during structured IDR in a standard conference room setting, though we did not assess this model. Second, bedside rounds only included patients who were able to participate. Rounding on patients unable to participate, such as patients with delirium with agitation, was done outside the patient room rather than at the bedside. A modified script was used in these instances (absent questions addressed to the patient), allowing for the prompt to be incorporated. These patients were included in the analysis. Lastly, as previously stated, we cannot clearly identify which intervention (the prompt, education, or feedback) most effectively led to a sustained decrease in lab ordering.

Our structured, multifaceted intervention reduced laboratory testing during the last 48 hours of admission. Hospitals that aim to decrease potentially unnecessary lab testing should consider implementing a bundle, including a prompt at a uniform and structured point during the hospitalization of patients who are expected to be discharged within 24 to 48 hours, clinician education, an audit, and feedback.

Disclosure

 All authors report no conflicts of interest to disclose.

As part of the Choosing Wisely^® campaign, the Society of Hospital Medicine recommends against performing “repetitive complete blood count [CBC] and chemistry testing in the face of clinical and lab stability.”¹ This recommendation stems from a body of research that shows that frequent or excessive phlebotomy can have negative consequences, including iatrogenic anemia (termed hospital-acquired anemia), which may necessitate blood transfusion.² The downstream effects of potentially unnecessary testing, including the evaluation of false-positive results, must also be considered. Additional important effects include patient discomfort and disruption of sleep and unproductive work by hospital staff, including nurses, phlebotomists, and laboratory technicians.

Though interventions to reduce unnecessary daily labs have been previously evaluated, there are no studies that focus on decreasing lab testing on patients deemed clinically stable and close to discharge. This is in part due to the absence of clear criteria or guidelines to define clinical stability in the context of lab utilization.

We therefore aimed to implement a multifaceted, patient-centered initiative—the Necessity of Labs Assessed Bedside (NO LABS)—that focused on reducing lab testing in patients at 24 to 48 hours before discharge. We targeted the 24 to 48-hour period before the anticipated date of discharge, as this may be a period of greater stability and provide an opportunity to identify and decrease unnecessary testing.

The study took place at Mount Sinai Hospital, which is an 1174-bed tertiary care teaching hospital in New York City. We targeted 2 inpatient medicine units where virtually all patients are assigned to a hospitalist rotating for a 2- to 4-week period, for the period of July 1, 2015, to July 31, 2016. These units employed bedside interdisciplinary rounds (IDR) attended by the hospitalist, social worker, case manager, nurse, nurse manager, and medical director. Bedside IDR focuses on the daily plan and patient safety by utilizing a scripted format.³ Our multifaceted intervention included prompting the hospitalist physician during bedside IDR, education of the clinicians, and regular data review for the hospitalists and unit staff.

As described by Dunn et al.,³ the IDR script included the following: a review of the plan of care by the hospitalist, identifying a patient’s personal goals for the day, a brief update of discharge planning (as appropriate), and a safety assessment performed by the nurse (identifying Foley catheters, falls risk, etc). We incorporated an inquiry into the daily IDR script identifying clinically stable patients for discharge in the next 24 to 48 hours (based on physician judgment), followed by a prompt to the hospitalist to discontinue labs when appropriate. The unit medical director and nurse manager were both tasked with prompting the hospitalist at the bedside. Our hospital utilizes computerized physician order entry. Lab orders were then discontinued by the clinician during rounds using a computer on wheels (or after rounds when one was not available). The hospitalist, unit medical director, and nurse manager were reminded about the project through weekly e-mails and in-person communication.

To assess whether the prompt was being incorporated consistently, an observer was added to rounds beginning in the second month of the project. The observer was present at least 3 times a week for the subsequent 3 months of the project. Our intervention also included education geared towards hospitalists, including a brief presentation on reducing unnecessary lab testing during a monthly hospitalist faculty meeting (the first and sixth month of the intervention). The group’s data on laboratory testing within the 24 to 48 hours prior to discharge were also presented at these monthly meetings (beginning 2 months into the intervention and monthly thereafter). Lastly, we provided the unit staff with unit-level metrics, biweekly for the first 3 months and every 2 to 3 months thereafter.

We extracted electronic medical record (EMR) data on lab utilization for patients on the 2 hospitalist units for the intervention period. Baseline data were obtained from July 1, 2014, to June 30, 2015. Patients with a length of stay (LOS) ≤7 days (75th percentile) were included; on these units, longer stays were considered more likely to have complex social issues delaying discharge and thus less likely to require laboratory testing. We tracked ordering for 4 common lab tests: basic metabolic panel, CBC, CBC with differential, and the comprehensive metabolic panel. The primary outcome was the monthly percentage of patients for whom testing was ordered in the 24 and 48 hours preceding discharge. A secondary outcome was testing at 72 hours preceding discharge to identify any potential compensatory (increased) testing the evening prior. We applied a quasi-experimental interrupted time series design with a segmented regression analysis to estimate changes before and after our intervention, expressed in acute changes (change in intercept) and over time (changes in trend) while adjusting for preintervention trends. All analyses were performed with SAS v9.4 statistical software (SAS Institute, Cary, NC). Our project was deemed a quality improvement project and thus an IRB submission was not required.

There were 1579 discharges in the preintervention period and 1308 discharges in the postintervention period. The average age of the patient population was similar in the baseline and postintervention groups (61.5 vs 59.3 years; P = 0.400), and there was no difference in the mean LOS before and after implementation (3.67 vs 3.68 days; P = 0.817).

Our structured, multifaceted approach effectively reduced daily lab testing in the 24 to 48 hours prior to discharge. Bedside IDR provided a unique opportunity to effectively communicate to the patient about necessary (or unnecessary) testing. Moreover, given the complexity of identifying clinical stability, our strategy focused on the onset of discharge planning, a more easily discernible and less obtrusive focal point to promote the discontinuation of lab testing.

Though the nature of bundled interventions can make it difficult to identify which intervention is most effective, we believe that all interventions were effective in different capacities during various phases in the intervention period. We believe that the decrease in lab testing in the 24 to 48 hours preceding discharge was primarily driven by the new rounding structure. This is evident in the significant decrease seen in the first few months of the intervention period. Six months into the intervention, we begin to see a decrease at 72 hours prior to discharge. Additionally, we see a decrease in the mean number of labs per patient day over the entire hospitalization period. We attribute these results to a gradual shift in the culture in our division as a direct consequence of educational sessions and individual feedback provided during this time.

To our knowledge, this is the first study to use anticipated discharge as a correlate for clinical stability and therefore as an opportunity to prompt discontinuation of laboratory testing. Other studies evaluated interventions targeting the EMR and the ease with which providers can order recurring labs. These include restricting recurring orders in the EMR,⁴ a robust education and awareness campaign targeting house staff,⁵ and other multifaceted approaches to decreasing lab utilization,⁶ all of which have shown promising results. While these approaches show varying degrees of success, ours is unique in its focus on the period prior to discharge. In addition, the intervention can be readily implemented in settings that utilize scripted IDR. It also brings high-value decision-making to the bedside by informing the patient that in the setting of presumed clinical stability, no additional tests are warranted.

Our study has several limitations. First, while interdisciplinary discharge rounds are widely implemented,^7,8 our rounds occur at the bedside and employ a script, potentially limiting generalizability. The structured prompting may be feasible during structured IDR in a standard conference room setting, though we did not assess this model. Second, bedside rounds only included patients who were able to participate. Rounding on patients unable to participate, such as patients with delirium with agitation, was done outside the patient room rather than at the bedside. A modified script was used in these instances (absent questions addressed to the patient), allowing for the prompt to be incorporated. These patients were included in the analysis. Lastly, as previously stated, we cannot clearly identify which intervention (the prompt, education, or feedback) most effectively led to a sustained decrease in lab ordering.

Our structured, multifaceted intervention reduced laboratory testing during the last 48 hours of admission. Hospitals that aim to decrease potentially unnecessary lab testing should consider implementing a bundle, including a prompt at a uniform and structured point during the hospitalization of patients who are expected to be discharged within 24 to 48 hours, clinician education, an audit, and feedback.

Disclosure

 All authors report no conflicts of interest to disclose.

As part of the Choosing Wisely^® campaign, the Society of Hospital Medicine recommends against performing “repetitive complete blood count [CBC] and chemistry testing in the face of clinical and lab stability.”¹ This recommendation stems from a body of research that shows that frequent or excessive phlebotomy can have negative consequences, including iatrogenic anemia (termed hospital-acquired anemia), which may necessitate blood transfusion.² The downstream effects of potentially unnecessary testing, including the evaluation of false-positive results, must also be considered. Additional important effects include patient discomfort and disruption of sleep and unproductive work by hospital staff, including nurses, phlebotomists, and laboratory technicians.

Though interventions to reduce unnecessary daily labs have been previously evaluated, there are no studies that focus on decreasing lab testing on patients deemed clinically stable and close to discharge. This is in part due to the absence of clear criteria or guidelines to define clinical stability in the context of lab utilization.

We therefore aimed to implement a multifaceted, patient-centered initiative—the Necessity of Labs Assessed Bedside (NO LABS)—that focused on reducing lab testing in patients at 24 to 48 hours before discharge. We targeted the 24 to 48-hour period before the anticipated date of discharge, as this may be a period of greater stability and provide an opportunity to identify and decrease unnecessary testing.

The study took place at Mount Sinai Hospital, which is an 1174-bed tertiary care teaching hospital in New York City. We targeted 2 inpatient medicine units where virtually all patients are assigned to a hospitalist rotating for a 2- to 4-week period, for the period of July 1, 2015, to July 31, 2016. These units employed bedside interdisciplinary rounds (IDR) attended by the hospitalist, social worker, case manager, nurse, nurse manager, and medical director. Bedside IDR focuses on the daily plan and patient safety by utilizing a scripted format.³ Our multifaceted intervention included prompting the hospitalist physician during bedside IDR, education of the clinicians, and regular data review for the hospitalists and unit staff.

As described by Dunn et al.,³ the IDR script included the following: a review of the plan of care by the hospitalist, identifying a patient’s personal goals for the day, a brief update of discharge planning (as appropriate), and a safety assessment performed by the nurse (identifying Foley catheters, falls risk, etc). We incorporated an inquiry into the daily IDR script identifying clinically stable patients for discharge in the next 24 to 48 hours (based on physician judgment), followed by a prompt to the hospitalist to discontinue labs when appropriate. The unit medical director and nurse manager were both tasked with prompting the hospitalist at the bedside. Our hospital utilizes computerized physician order entry. Lab orders were then discontinued by the clinician during rounds using a computer on wheels (or after rounds when one was not available). The hospitalist, unit medical director, and nurse manager were reminded about the project through weekly e-mails and in-person communication.

To assess whether the prompt was being incorporated consistently, an observer was added to rounds beginning in the second month of the project. The observer was present at least 3 times a week for the subsequent 3 months of the project. Our intervention also included education geared towards hospitalists, including a brief presentation on reducing unnecessary lab testing during a monthly hospitalist faculty meeting (the first and sixth month of the intervention). The group’s data on laboratory testing within the 24 to 48 hours prior to discharge were also presented at these monthly meetings (beginning 2 months into the intervention and monthly thereafter). Lastly, we provided the unit staff with unit-level metrics, biweekly for the first 3 months and every 2 to 3 months thereafter.

We extracted electronic medical record (EMR) data on lab utilization for patients on the 2 hospitalist units for the intervention period. Baseline data were obtained from July 1, 2014, to June 30, 2015. Patients with a length of stay (LOS) ≤7 days (75th percentile) were included; on these units, longer stays were considered more likely to have complex social issues delaying discharge and thus less likely to require laboratory testing. We tracked ordering for 4 common lab tests: basic metabolic panel, CBC, CBC with differential, and the comprehensive metabolic panel. The primary outcome was the monthly percentage of patients for whom testing was ordered in the 24 and 48 hours preceding discharge. A secondary outcome was testing at 72 hours preceding discharge to identify any potential compensatory (increased) testing the evening prior. We applied a quasi-experimental interrupted time series design with a segmented regression analysis to estimate changes before and after our intervention, expressed in acute changes (change in intercept) and over time (changes in trend) while adjusting for preintervention trends. All analyses were performed with SAS v9.4 statistical software (SAS Institute, Cary, NC). Our project was deemed a quality improvement project and thus an IRB submission was not required.

There were 1579 discharges in the preintervention period and 1308 discharges in the postintervention period. The average age of the patient population was similar in the baseline and postintervention groups (61.5 vs 59.3 years; P = 0.400), and there was no difference in the mean LOS before and after implementation (3.67 vs 3.68 days; P = 0.817).

Our structured, multifaceted approach effectively reduced daily lab testing in the 24 to 48 hours prior to discharge. Bedside IDR provided a unique opportunity to effectively communicate to the patient about necessary (or unnecessary) testing. Moreover, given the complexity of identifying clinical stability, our strategy focused on the onset of discharge planning, a more easily discernible and less obtrusive focal point to promote the discontinuation of lab testing.

Though the nature of bundled interventions can make it difficult to identify which intervention is most effective, we believe that all interventions were effective in different capacities during various phases in the intervention period. We believe that the decrease in lab testing in the 24 to 48 hours preceding discharge was primarily driven by the new rounding structure. This is evident in the significant decrease seen in the first few months of the intervention period. Six months into the intervention, we begin to see a decrease at 72 hours prior to discharge. Additionally, we see a decrease in the mean number of labs per patient day over the entire hospitalization period. We attribute these results to a gradual shift in the culture in our division as a direct consequence of educational sessions and individual feedback provided during this time.

To our knowledge, this is the first study to use anticipated discharge as a correlate for clinical stability and therefore as an opportunity to prompt discontinuation of laboratory testing. Other studies evaluated interventions targeting the EMR and the ease with which providers can order recurring labs. These include restricting recurring orders in the EMR,⁴ a robust education and awareness campaign targeting house staff,⁵ and other multifaceted approaches to decreasing lab utilization,⁶ all of which have shown promising results. While these approaches show varying degrees of success, ours is unique in its focus on the period prior to discharge. In addition, the intervention can be readily implemented in settings that utilize scripted IDR. It also brings high-value decision-making to the bedside by informing the patient that in the setting of presumed clinical stability, no additional tests are warranted.

Our study has several limitations. First, while interdisciplinary discharge rounds are widely implemented,^7,8 our rounds occur at the bedside and employ a script, potentially limiting generalizability. The structured prompting may be feasible during structured IDR in a standard conference room setting, though we did not assess this model. Second, bedside rounds only included patients who were able to participate. Rounding on patients unable to participate, such as patients with delirium with agitation, was done outside the patient room rather than at the bedside. A modified script was used in these instances (absent questions addressed to the patient), allowing for the prompt to be incorporated. These patients were included in the analysis. Lastly, as previously stated, we cannot clearly identify which intervention (the prompt, education, or feedback) most effectively led to a sustained decrease in lab ordering.

Our structured, multifaceted intervention reduced laboratory testing during the last 48 hours of admission. Hospitals that aim to decrease potentially unnecessary lab testing should consider implementing a bundle, including a prompt at a uniform and structured point during the hospitalization of patients who are expected to be discharged within 24 to 48 hours, clinician education, an audit, and feedback.

Disclosure

 All authors report no conflicts of interest to disclose.

Pneumonia is a major cause of hospitalization, mortality, and healthcare cost. ^1,2 The diagnosis involves clinical features plus radiographic evidence of infection. Hospital-acquired pneumonia (HAP) is defined by the Infectious Disease Society of America (IDSA) as a pneumonia that occurs ≥48 hours after admission and is not associated with mechanical ventilation. ³

IDSA recommendations suggest that patients with suspected HAP be treated based on results of noninvasively obtained sputum cultures rather than being treated empirically. ³ This recommendation is graded as weak with low-quality evidence based on a lack of both evidence showing that respiratory cultures improve clinical outcomes and studies examining the yield of noninvasive collection methods. ^4,5 However, resistant pathogens lead to a risk of inadequate empiric therapy, which is associated with increased mortality. ⁶ Culture data may provide an opportunity for escalation or de-escalation of antibiotic coverage. IDSA recommendations for microbiologic sampling are thus aimed at increasing appropriate coverage and minimizing unnecessary antibiotic exposure.

While the yield and clinical utility of sputum culture in community-acquired pneumonia has been studied extensively, data examining the yield of sputum culture in HAP (non–ventilator-associated pneumonia [non-VAP]) are sparse. In 1 small single-center study, researchers demonstrated positive sputum cultures in 17/35 (48.6%) patients with radiographically confirmed cases of HAP, ⁷ while in another study, researchers demonstrated positive sputum cultures in 57/63 (90.5%). ⁸ We aimed to identify the frequency with which sputum cultures positively identify an organism, identify predictors of positive sputum cultures, and characterize the microbiology of sputum cultures in a large cohort of HAP cases.

We conducted a retrospective cohort study of patients admitted to a large academic medical center in Boston, Massachusetts, from January 2007 to July 2013. All patients ≥18 years of age were eligible for inclusion. We excluded outside hospital transfers, those with a length of hospitalization <48 hours, and psychiatric admissions.

The study was approved by the institutional review board at the Beth Israel Deaconess Medical Center and granted a waiver of informed consent. Data were collected from electronic databases and supplemented by chart review.

Hospital-Acquired Pneumonia

We defined HAP as pneumonia occurring at least 48 hours after admission, consistent with American Thoracic Society and IDSA criteria.³ To identify cases, we reviewed the charts of all admissions identified as having a discharge International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) code for bacterial pneumonia (481, 482, 483, 485, 486, 507), indicated as not “present-on-admission.” We validated that the treating clinician had clinically diagnosed pneumonia and initiated antibiotics for this purpose by performing chart review. We reviewed the radiologist interpretation of radiographs surrounding the date of the clinical diagnosis of pneumonia to confirm the presence of a new opacity. Uncertain cases (with respect to either the presence of pneumonia or the timing of the diagnosis) were reviewed by a second member of the study team and, in the case of disagreement, adjudicated by a third member of the study team. Only the first clinically validated HAP per hospitalization was included in the analysis. To focus on HAP rather than VAP, we excluded hospitalizations in which the date of a procedure code for mechanical ventilation preceded the date of pneumonia diagnosis.

Microbiology

In our analysis, we used sputum samples obtained from expectorated or induced samples to evaluate the yield of noninvasive sputum sampling, as recommended by the IDSA. We included sputum samples collected ≥48 hours after admission and within 7 days of the clinical diagnosis of HAP. Sputum samples with >10 epithelial cells per high-power field (hpf) were considered to be contaminated. Among noncontaminated samples, positive sputum cultures were defined as those with a microbiologic diagnosis other than “oral flora,” while those with no growth or growth of oral flora or only yeast were considered to be negative. The hospital’s microbiology laboratory does not routinely provide species identification for Gram-negative rods (GNRs) growing on culture in the presence of growth of ≥3 other colony types. We considered such GNRs (not further speciated) to represent a positive culture result in our analysis given that colonization versus pathogenicity is a clinical distinction and, as such, these results may impact antibiotic choice.

Statistical Analysis

Data were analyzed by using SAS software, version 9.3. We used a 2-sided P value of <0.05 to indicate statistical significance for all comparisons. We used the χ² test and the nonparametric median test for unadjusted comparisons.

To identify predictors of a positive (versus negative or contaminated) sputum culture among patients with HAP, we used a generalized estimating equation model with a Poisson distribution error term, log link, and first-order autoregressive correlation structure to account for multiple sputum specimens per patient. We combined culture negative and contaminated samples to highlight the clinical utility of sputum culture in a real-world setting. Potential predictors chosen based on clinical grounds included all variables listed in Table 1. We defined comorbidities specified in Table 1 via ICD-9-CM secondary diagnosis codes and diagnosis related groups (DRGs) using Healthcare Cost and Utilization Project Comorbidity Software, version 3.7, based on the work of Elixhauser et al.^9,10; dialysis use was defined by an ICD-9-CM procedure code of 39.95; inpatient steroid use was defined by a hospital pharmacy charge for a systemic steroid in the 7 days preceding the sputum sample.

There were 230,635 hospitalizations of patients ≥18 years of age from January 2007 to July 2013. After excluding outside hospital transfers (n = 14,422), hospitalizations <48 hours in duration (n = 59,774), and psychiatric hospitalizations (n = 9887), there were 146,552 hospitalizations in the cohort.

To our knowledge, our study is the first to assess the predictors of positive sputum culture among patients with HAP (non-VAP) who had sputum samples obtained noninvasively, and this study is larger than prior studies in which researchers reported on sputum culture yield in HAP. Sputum samples were obtained in 29.4% cases of clinically diagnosed HAP. Although 87% of specimens obtained were culture-negative or contaminated, 13% yielded a bacterial organism. Although we do not report the antibiotic sensitivity patterns of the isolated organisms, the organisms identified frequently demonstrate antibiotic resistance, highlighting the potential for both antibiotic escalation and de-escalation based on sputum culture. In a multivariable model, presence of chronic lung disease and steroid use in the preceding week were both significantly associated with culture positivity.

The retrospective nature of the study raises the possibility of selection bias from systematic differences between the 29.4% of patients with HAP who had sputum collected and those who did not. Patients with sputum cultures were similar to patients without cultures in most measured characteristics, but we are unable to know what the yield of noninvasive sputum culture would have been had all patients with HAP been sampled. As such, our findings reflect the yield of sputum culture among patients with HAP for whom cultures were successfully obtained. It is not clear why only 29.4% of HAP patients received IDSA guideline-concordant care, but similar rates of culture use are reported elsewhere.⁷ While physician decision-making could have contributed to this finding, it is also possible that many sick, hospitalized patients are simply unable to produce sputum for analysis. In future studies, researchers should examine barriers to guideline-concordant care.

We considered a culture result of GNRs (not further speciated) as positive in our analysis because this result indicates growth of mixed bacterial types, the pathogenicity of which is a clinical determination. Physicians may request speciation and antibiotic sensitivities and, as such, these results have the potential to impact antibiotic choice. Had we considered such cultures to be negative or contaminated, the rate of culture positivity would have been only slightly reduced from 63/478 (13.2%) to 50/478 (10.5%).

The strengths of our study include the chart-based validation of administratively identified cases of pneumonia and a large cohort. There are also limitations. The single-center nature of the study has implications for pretest probability and generalizability. Additionally, in our study, we did not examine outcomes among patients treated empirically versus those treated based on sputum culture results. Finally, our reliance on administrative codes to identify cases of HAP for subsequent validation could have resulted in incomplete capture of HAP cases.

In conclusion, in our study, we provide an estimate of the diagnostic yield of sputum culture in a large cohort with chart-validated HAP, a description of HAP microbiology, and predictors of positive sputum culture. Thirteen percent of patients who had sputum culture testing received a microbiologic diagnosis. Because of the relative ease of obtaining a sputum sample and the microbiologic distribution in our study (representing a mix of commonly drug-resistant pathogens and more typical community-acquired pathogens), we suggest that sputum culture in HAP is a useful diagnostic tool with the potential to inform antibiotic escalation or de-escalation.

Acknowledgments

Dr. Herzig was funded by grant number K23AG042459 from the National Institute on Aging. Dr. Marcantonio was funded by grant number K24AG035075 from the National Institute on Aging. The funding organizations had no involvement in any aspect of the study, including design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript.

Disclosure

No conflicts of interest apply for any of the authors.

Pneumonia is a major cause of hospitalization, mortality, and healthcare cost. ^1,2 The diagnosis involves clinical features plus radiographic evidence of infection. Hospital-acquired pneumonia (HAP) is defined by the Infectious Disease Society of America (IDSA) as a pneumonia that occurs ≥48 hours after admission and is not associated with mechanical ventilation. ³

IDSA recommendations suggest that patients with suspected HAP be treated based on results of noninvasively obtained sputum cultures rather than being treated empirically. ³ This recommendation is graded as weak with low-quality evidence based on a lack of both evidence showing that respiratory cultures improve clinical outcomes and studies examining the yield of noninvasive collection methods. ^4,5 However, resistant pathogens lead to a risk of inadequate empiric therapy, which is associated with increased mortality. ⁶ Culture data may provide an opportunity for escalation or de-escalation of antibiotic coverage. IDSA recommendations for microbiologic sampling are thus aimed at increasing appropriate coverage and minimizing unnecessary antibiotic exposure.

While the yield and clinical utility of sputum culture in community-acquired pneumonia has been studied extensively, data examining the yield of sputum culture in HAP (non–ventilator-associated pneumonia [non-VAP]) are sparse. In 1 small single-center study, researchers demonstrated positive sputum cultures in 17/35 (48.6%) patients with radiographically confirmed cases of HAP, ⁷ while in another study, researchers demonstrated positive sputum cultures in 57/63 (90.5%). ⁸ We aimed to identify the frequency with which sputum cultures positively identify an organism, identify predictors of positive sputum cultures, and characterize the microbiology of sputum cultures in a large cohort of HAP cases.

We conducted a retrospective cohort study of patients admitted to a large academic medical center in Boston, Massachusetts, from January 2007 to July 2013. All patients ≥18 years of age were eligible for inclusion. We excluded outside hospital transfers, those with a length of hospitalization <48 hours, and psychiatric admissions.

The study was approved by the institutional review board at the Beth Israel Deaconess Medical Center and granted a waiver of informed consent. Data were collected from electronic databases and supplemented by chart review.

Hospital-Acquired Pneumonia

We defined HAP as pneumonia occurring at least 48 hours after admission, consistent with American Thoracic Society and IDSA criteria.³ To identify cases, we reviewed the charts of all admissions identified as having a discharge International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) code for bacterial pneumonia (481, 482, 483, 485, 486, 507), indicated as not “present-on-admission.” We validated that the treating clinician had clinically diagnosed pneumonia and initiated antibiotics for this purpose by performing chart review. We reviewed the radiologist interpretation of radiographs surrounding the date of the clinical diagnosis of pneumonia to confirm the presence of a new opacity. Uncertain cases (with respect to either the presence of pneumonia or the timing of the diagnosis) were reviewed by a second member of the study team and, in the case of disagreement, adjudicated by a third member of the study team. Only the first clinically validated HAP per hospitalization was included in the analysis. To focus on HAP rather than VAP, we excluded hospitalizations in which the date of a procedure code for mechanical ventilation preceded the date of pneumonia diagnosis.

Microbiology

In our analysis, we used sputum samples obtained from expectorated or induced samples to evaluate the yield of noninvasive sputum sampling, as recommended by the IDSA. We included sputum samples collected ≥48 hours after admission and within 7 days of the clinical diagnosis of HAP. Sputum samples with >10 epithelial cells per high-power field (hpf) were considered to be contaminated. Among noncontaminated samples, positive sputum cultures were defined as those with a microbiologic diagnosis other than “oral flora,” while those with no growth or growth of oral flora or only yeast were considered to be negative. The hospital’s microbiology laboratory does not routinely provide species identification for Gram-negative rods (GNRs) growing on culture in the presence of growth of ≥3 other colony types. We considered such GNRs (not further speciated) to represent a positive culture result in our analysis given that colonization versus pathogenicity is a clinical distinction and, as such, these results may impact antibiotic choice.

Statistical Analysis

Data were analyzed by using SAS software, version 9.3. We used a 2-sided P value of <0.05 to indicate statistical significance for all comparisons. We used the χ² test and the nonparametric median test for unadjusted comparisons.

To identify predictors of a positive (versus negative or contaminated) sputum culture among patients with HAP, we used a generalized estimating equation model with a Poisson distribution error term, log link, and first-order autoregressive correlation structure to account for multiple sputum specimens per patient. We combined culture negative and contaminated samples to highlight the clinical utility of sputum culture in a real-world setting. Potential predictors chosen based on clinical grounds included all variables listed in Table 1. We defined comorbidities specified in Table 1 via ICD-9-CM secondary diagnosis codes and diagnosis related groups (DRGs) using Healthcare Cost and Utilization Project Comorbidity Software, version 3.7, based on the work of Elixhauser et al.^9,10; dialysis use was defined by an ICD-9-CM procedure code of 39.95; inpatient steroid use was defined by a hospital pharmacy charge for a systemic steroid in the 7 days preceding the sputum sample.

There were 230,635 hospitalizations of patients ≥18 years of age from January 2007 to July 2013. After excluding outside hospital transfers (n = 14,422), hospitalizations <48 hours in duration (n = 59,774), and psychiatric hospitalizations (n = 9887), there were 146,552 hospitalizations in the cohort.

To our knowledge, our study is the first to assess the predictors of positive sputum culture among patients with HAP (non-VAP) who had sputum samples obtained noninvasively, and this study is larger than prior studies in which researchers reported on sputum culture yield in HAP. Sputum samples were obtained in 29.4% cases of clinically diagnosed HAP. Although 87% of specimens obtained were culture-negative or contaminated, 13% yielded a bacterial organism. Although we do not report the antibiotic sensitivity patterns of the isolated organisms, the organisms identified frequently demonstrate antibiotic resistance, highlighting the potential for both antibiotic escalation and de-escalation based on sputum culture. In a multivariable model, presence of chronic lung disease and steroid use in the preceding week were both significantly associated with culture positivity.

The retrospective nature of the study raises the possibility of selection bias from systematic differences between the 29.4% of patients with HAP who had sputum collected and those who did not. Patients with sputum cultures were similar to patients without cultures in most measured characteristics, but we are unable to know what the yield of noninvasive sputum culture would have been had all patients with HAP been sampled. As such, our findings reflect the yield of sputum culture among patients with HAP for whom cultures were successfully obtained. It is not clear why only 29.4% of HAP patients received IDSA guideline-concordant care, but similar rates of culture use are reported elsewhere.⁷ While physician decision-making could have contributed to this finding, it is also possible that many sick, hospitalized patients are simply unable to produce sputum for analysis. In future studies, researchers should examine barriers to guideline-concordant care.

We considered a culture result of GNRs (not further speciated) as positive in our analysis because this result indicates growth of mixed bacterial types, the pathogenicity of which is a clinical determination. Physicians may request speciation and antibiotic sensitivities and, as such, these results have the potential to impact antibiotic choice. Had we considered such cultures to be negative or contaminated, the rate of culture positivity would have been only slightly reduced from 63/478 (13.2%) to 50/478 (10.5%).

The strengths of our study include the chart-based validation of administratively identified cases of pneumonia and a large cohort. There are also limitations. The single-center nature of the study has implications for pretest probability and generalizability. Additionally, in our study, we did not examine outcomes among patients treated empirically versus those treated based on sputum culture results. Finally, our reliance on administrative codes to identify cases of HAP for subsequent validation could have resulted in incomplete capture of HAP cases.

In conclusion, in our study, we provide an estimate of the diagnostic yield of sputum culture in a large cohort with chart-validated HAP, a description of HAP microbiology, and predictors of positive sputum culture. Thirteen percent of patients who had sputum culture testing received a microbiologic diagnosis. Because of the relative ease of obtaining a sputum sample and the microbiologic distribution in our study (representing a mix of commonly drug-resistant pathogens and more typical community-acquired pathogens), we suggest that sputum culture in HAP is a useful diagnostic tool with the potential to inform antibiotic escalation or de-escalation.

Acknowledgments

Dr. Herzig was funded by grant number K23AG042459 from the National Institute on Aging. Dr. Marcantonio was funded by grant number K24AG035075 from the National Institute on Aging. The funding organizations had no involvement in any aspect of the study, including design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript.

Disclosure

No conflicts of interest apply for any of the authors.

Pneumonia is a major cause of hospitalization, mortality, and healthcare cost. ^1,2 The diagnosis involves clinical features plus radiographic evidence of infection. Hospital-acquired pneumonia (HAP) is defined by the Infectious Disease Society of America (IDSA) as a pneumonia that occurs ≥48 hours after admission and is not associated with mechanical ventilation. ³

IDSA recommendations suggest that patients with suspected HAP be treated based on results of noninvasively obtained sputum cultures rather than being treated empirically. ³ This recommendation is graded as weak with low-quality evidence based on a lack of both evidence showing that respiratory cultures improve clinical outcomes and studies examining the yield of noninvasive collection methods. ^4,5 However, resistant pathogens lead to a risk of inadequate empiric therapy, which is associated with increased mortality. ⁶ Culture data may provide an opportunity for escalation or de-escalation of antibiotic coverage. IDSA recommendations for microbiologic sampling are thus aimed at increasing appropriate coverage and minimizing unnecessary antibiotic exposure.

While the yield and clinical utility of sputum culture in community-acquired pneumonia has been studied extensively, data examining the yield of sputum culture in HAP (non–ventilator-associated pneumonia [non-VAP]) are sparse. In 1 small single-center study, researchers demonstrated positive sputum cultures in 17/35 (48.6%) patients with radiographically confirmed cases of HAP, ⁷ while in another study, researchers demonstrated positive sputum cultures in 57/63 (90.5%). ⁸ We aimed to identify the frequency with which sputum cultures positively identify an organism, identify predictors of positive sputum cultures, and characterize the microbiology of sputum cultures in a large cohort of HAP cases.

We conducted a retrospective cohort study of patients admitted to a large academic medical center in Boston, Massachusetts, from January 2007 to July 2013. All patients ≥18 years of age were eligible for inclusion. We excluded outside hospital transfers, those with a length of hospitalization <48 hours, and psychiatric admissions.

The study was approved by the institutional review board at the Beth Israel Deaconess Medical Center and granted a waiver of informed consent. Data were collected from electronic databases and supplemented by chart review.

Hospital-Acquired Pneumonia

We defined HAP as pneumonia occurring at least 48 hours after admission, consistent with American Thoracic Society and IDSA criteria.³ To identify cases, we reviewed the charts of all admissions identified as having a discharge International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) code for bacterial pneumonia (481, 482, 483, 485, 486, 507), indicated as not “present-on-admission.” We validated that the treating clinician had clinically diagnosed pneumonia and initiated antibiotics for this purpose by performing chart review. We reviewed the radiologist interpretation of radiographs surrounding the date of the clinical diagnosis of pneumonia to confirm the presence of a new opacity. Uncertain cases (with respect to either the presence of pneumonia or the timing of the diagnosis) were reviewed by a second member of the study team and, in the case of disagreement, adjudicated by a third member of the study team. Only the first clinically validated HAP per hospitalization was included in the analysis. To focus on HAP rather than VAP, we excluded hospitalizations in which the date of a procedure code for mechanical ventilation preceded the date of pneumonia diagnosis.

Microbiology

In our analysis, we used sputum samples obtained from expectorated or induced samples to evaluate the yield of noninvasive sputum sampling, as recommended by the IDSA. We included sputum samples collected ≥48 hours after admission and within 7 days of the clinical diagnosis of HAP. Sputum samples with >10 epithelial cells per high-power field (hpf) were considered to be contaminated. Among noncontaminated samples, positive sputum cultures were defined as those with a microbiologic diagnosis other than “oral flora,” while those with no growth or growth of oral flora or only yeast were considered to be negative. The hospital’s microbiology laboratory does not routinely provide species identification for Gram-negative rods (GNRs) growing on culture in the presence of growth of ≥3 other colony types. We considered such GNRs (not further speciated) to represent a positive culture result in our analysis given that colonization versus pathogenicity is a clinical distinction and, as such, these results may impact antibiotic choice.

Statistical Analysis

Data were analyzed by using SAS software, version 9.3. We used a 2-sided P value of <0.05 to indicate statistical significance for all comparisons. We used the χ² test and the nonparametric median test for unadjusted comparisons.

To identify predictors of a positive (versus negative or contaminated) sputum culture among patients with HAP, we used a generalized estimating equation model with a Poisson distribution error term, log link, and first-order autoregressive correlation structure to account for multiple sputum specimens per patient. We combined culture negative and contaminated samples to highlight the clinical utility of sputum culture in a real-world setting. Potential predictors chosen based on clinical grounds included all variables listed in Table 1. We defined comorbidities specified in Table 1 via ICD-9-CM secondary diagnosis codes and diagnosis related groups (DRGs) using Healthcare Cost and Utilization Project Comorbidity Software, version 3.7, based on the work of Elixhauser et al.^9,10; dialysis use was defined by an ICD-9-CM procedure code of 39.95; inpatient steroid use was defined by a hospital pharmacy charge for a systemic steroid in the 7 days preceding the sputum sample.

There were 230,635 hospitalizations of patients ≥18 years of age from January 2007 to July 2013. After excluding outside hospital transfers (n = 14,422), hospitalizations <48 hours in duration (n = 59,774), and psychiatric hospitalizations (n = 9887), there were 146,552 hospitalizations in the cohort.

To our knowledge, our study is the first to assess the predictors of positive sputum culture among patients with HAP (non-VAP) who had sputum samples obtained noninvasively, and this study is larger than prior studies in which researchers reported on sputum culture yield in HAP. Sputum samples were obtained in 29.4% cases of clinically diagnosed HAP. Although 87% of specimens obtained were culture-negative or contaminated, 13% yielded a bacterial organism. Although we do not report the antibiotic sensitivity patterns of the isolated organisms, the organisms identified frequently demonstrate antibiotic resistance, highlighting the potential for both antibiotic escalation and de-escalation based on sputum culture. In a multivariable model, presence of chronic lung disease and steroid use in the preceding week were both significantly associated with culture positivity.

The retrospective nature of the study raises the possibility of selection bias from systematic differences between the 29.4% of patients with HAP who had sputum collected and those who did not. Patients with sputum cultures were similar to patients without cultures in most measured characteristics, but we are unable to know what the yield of noninvasive sputum culture would have been had all patients with HAP been sampled. As such, our findings reflect the yield of sputum culture among patients with HAP for whom cultures were successfully obtained. It is not clear why only 29.4% of HAP patients received IDSA guideline-concordant care, but similar rates of culture use are reported elsewhere.⁷ While physician decision-making could have contributed to this finding, it is also possible that many sick, hospitalized patients are simply unable to produce sputum for analysis. In future studies, researchers should examine barriers to guideline-concordant care.

We considered a culture result of GNRs (not further speciated) as positive in our analysis because this result indicates growth of mixed bacterial types, the pathogenicity of which is a clinical determination. Physicians may request speciation and antibiotic sensitivities and, as such, these results have the potential to impact antibiotic choice. Had we considered such cultures to be negative or contaminated, the rate of culture positivity would have been only slightly reduced from 63/478 (13.2%) to 50/478 (10.5%).

The strengths of our study include the chart-based validation of administratively identified cases of pneumonia and a large cohort. There are also limitations. The single-center nature of the study has implications for pretest probability and generalizability. Additionally, in our study, we did not examine outcomes among patients treated empirically versus those treated based on sputum culture results. Finally, our reliance on administrative codes to identify cases of HAP for subsequent validation could have resulted in incomplete capture of HAP cases.

In conclusion, in our study, we provide an estimate of the diagnostic yield of sputum culture in a large cohort with chart-validated HAP, a description of HAP microbiology, and predictors of positive sputum culture. Thirteen percent of patients who had sputum culture testing received a microbiologic diagnosis. Because of the relative ease of obtaining a sputum sample and the microbiologic distribution in our study (representing a mix of commonly drug-resistant pathogens and more typical community-acquired pathogens), we suggest that sputum culture in HAP is a useful diagnostic tool with the potential to inform antibiotic escalation or de-escalation.

Acknowledgments

Dr. Herzig was funded by grant number K23AG042459 from the National Institute on Aging. Dr. Marcantonio was funded by grant number K24AG035075 from the National Institute on Aging. The funding organizations had no involvement in any aspect of the study, including design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript.

Disclosure

No conflicts of interest apply for any of the authors.

Ascites is the most common complication of cirrhosis and often leads to hospitalization. ¹ Paracentesis is recommended for all patients admitted with ascites and cirrhosis. ¹ Additionally, the Society of Hospital Medicine considers the ability to perform paracenteses a core competency for hospitalists. ² Although considered a safe procedure, major bleeding complications occur in 0.2% to 1.7% of paracenteses. ^3-7 Patients with cirrhosis form new abdominal wall vessels because of portal hypertension, and hemoperitoneum from the laceration of these vessels during paracentesis carries a high morbidity and mortality. ^6,8 Ultrasound guidance using a low-frequency ultrasound probe is currently standard practice for paracentesis and has been shown to reduce bleeding complications. ^9-11 However, the use of vascular ultrasound (high-frequency probe) is also recommended to identify blood vessels within the intended needle pathway to reduce bleeding, but no studies have been performed to demonstrate a benefit. ^3,11 This study aimed to evaluate whether this “2-probe technique” reduces paracentesis-related bleeding complications.

The procedure service at Cedars Sinai Medical Center (CSMC) in Los Angeles performs paracentesis regularly with ultrasound guidance. CSMC is a tertiary care, academic medical center with 861 licensed beds. We performed a pre- to postintervention study of consecutive patients (admitted and ambulatory) who underwent paracentesis done by 1 proceduralist (MJA) from the procedure service at CSMC from February 2010 through February 2016. From February 1, 2010, through August 2011, paracenteses were performed using only low-frequency, phased array ultrasound probes (preintervention group). From September 1, 2011, through February 2016, a 2-probe technique was used, whereby ultrasound interrogation of the abdomen using a low-frequency, phased array probe (to identify ascites) was supplemented with a second scan using a high-frequency, linear probe to identify vasculature within the planned needle path (postintervention group). As a standard part of quality assurance, CSMC documented all paracentesis-related complications from procedures performed by their center. Northwestern University investigators (JHB, EC, JF) independently evaluated these data to look at bleeding complications before and after the implementation of the 2-probe technique. The CSMC and Northwestern University institutional review boards approved this study.

Procedure Protocol

Each patient’s primary team or outpatient physician requested a consultation for paracentesis from the CSMC procedure service. All patient evaluations began with an abdominal ultrasound using the low-frequency probe to determine the presence of ascites and a potential window of access to the fluid. After September 1, 2011, the CSMC procedure service implemented the 2-probe technique to also evaluate the abdominal wall for the presence of vessels. Color flow Doppler ultrasound further helped to differentiate blood vessels as necessary. The optimal window was then marked on the abdominal wall, and the paracentesis was performed. Per the routine of the CSMC procedure service, antiplatelet or anticoagulant medications were not held for paracenteses.

Measurement

All data were collected prospectively at the time of the procedure, including the volume of fluid removed, the number of needle passes required, and whether the patient was on antiplatelet or anticoagulant medications (including warfarin, direct oral anticoagulants, thrombin inhibitors, heparin, or low molecular weight heparins). Patients were followed for complications for up to 24 hours after the procedure or until a clinical question of a complication was reconciled. Minor bleeding was defined as new serosanguinous fluid on repeat paracentesis not associated with hemodynamic changes, local bruising or bleeding at the site, or abdominal wall hematoma. Major bleeding was defined by the development of hemodynamic instability or by reaccumulation of fluid on ultrasound within 24 hours postparacentesis and one of the following: an associated hemoglobin drop of greater than 2 g/dl, blood seen on repeat paracentesis, blood density fluid on a computed tomography scan, or the lack of an alternative explanation. All data were recorded in a handheld database (HanDbase; DDH Software, Wellington, FL).

A query of the electronic medical record was performed to obtain patient demographics and relevant clinical information, including age, sex, body mass index, International Normalized Ratio (INR), partial thromboplastin time (PTT), platelet counts (10³/uL, hematocrit (%) and creatinine (mg/dl). Our query for laboratory data retrieved the closest laboratory entry up to 48 hours before the procedure.

Statistical Analysis

We used a χ² test, Student t test, or Kruskal-Wallis test to compare demographic and clinical characteristics of procedure patients between the 2 study groups (pre- and postintervention). Major and minor bleeding were compared between the 2 groups using the χ² test.¹² We used the χ² test instead of the Fisher’s exact test for several reasons. The usual rule is that the Fisher’s exact test is necessary when 1 or more expected outcome values are less than 5. However, McDonald argues that the χ² test should be used with large sample sizes (more than 1000) in lieu of the outcome-value-of-5 rule.¹² The Fisher’s exact test also assumes that the row and column totals are fixed. However, the outcomes in our study were not fixed because any patient could have a bleeding complication during each procedure. When row and column totals are not fixed, only 5% of the time will a P value be less than 0.05, and the Fisher’s exact test is too conservative.¹² We performed all statistical analyses using IBM SPSS Statistics Version 22 (IBM Corp, Armonk, NY).

Patient demographic and clinical information can be found in the Table. The proceduralist (MJA) performed a total of 5777 paracenteses (1000 preintervention, 4777 postintervention) on 1639 patients. Four hundred eighty-nine (10.2%) vascular anomalies were identified within the intended needle path in the postintervention group (Figure). More patients in the preintervention group were on aspirin (93 [9.3%] vs 230 [4.8%]; P < 0.001) and therapeutic intravenous anticoagulants (33 [3.3%] vs 89 [1.9%]; P = 0.004), while more patients in the postintervention group were on both an antiplatelet and oral anticoagulant (1 [0.1%] vs 38 [0.8%]; P = 0.015) and subcutaneous prophylactic anticoagulants (184 [18.4%] vs 1120 [23.4%]; P = 0.001) at the time of the procedure. There were no other differences between groups with antiplatelet or anticoagulant drugs. We found no difference in minor bleeding between pre- and postintervention groups. Major bleeding was lower after the 2-probe technique was implemented (3 [0.3%] vs 4 [0.08%]; P = 0.07). There were no between-group differences in INR, PTT, or platelet counts among major bleeders. One patient in the postintervention group had hemodynamic instability and dropped his hemoglobin by 3.8 g/dl at 7 hours after the procedure. This was unexplained, as the patient had no abdominal symptoms or findings on examination. The patient received several liters of fluid before ultimately dying, and the primary team considered sepsis as a possible cause, but no postmortem examination was performed. This was the only death attributed to a major bleeding complication. We included this patient in our analysis because the cause of his demise was not completely clear. However, excluding this patient would change the results from a trend to a statistically significant difference between groups (3 [0.3%] vs 3 [0.06%]; P = 0.03).

To our knowledge, we report the largest series of paracentesis prospectively evaluated for bleeding complications, and this is the first study to evaluate whether adding a vascular ultrasound (high-frequency probe) avoids major bleeding. In our series, up to 10% of patients had abnormal vessels seen with a vascular ultrasound that were within the original intended trajectory path of the needle. These vessels were also likely present yet invisible when ultrasound-guided paracentesis using only the standard, low-frequency probe was being performed. It is unknown whether these vessels are routinely traversed with the needle, nicked, or narrowly avoided during paracenteses performed using only a low-frequency probe.

Procedure-related bleeding may not be completely avoidable, despite using the vascular probe. Some authors have suggested that the mechanism of bleeding is more related to the rapid reduction in intraperitoneal pressure, which increases the gradient across vessel walls, resulting in rupture and bleeding.⁶ However, in our series, using vascular ultrasound also reduced major bleeding to numbers lower than those historically reported in the literature (0.2%).^3-4 Our preintervention number needed to harm was 333 procedures to cause 1 major bleed, compared to 1250 (or 1666 using the 3-patient bleeding analysis) in the postintervention group. In 2008, 150,000 Medicare beneficiaries underwent paracentesis.¹³ Using our study analysis, if vascular ultrasound was used on these patients, up to 360 major bleeds may have been prevented, along with a corresponding reduction in unnecessary morbidity and mortality.

Our study has several limitations. First, it was limited to 1 center with 1 very experienced proceduralist. Although it is possible that the reduction in major bleeding may have been due to the increasing experience of the proceduralist over time, we do not think that this is likely because he had already performed thousands of paracenteses over 9 years before the start of our study. Second, major bleeding was rare and therefore precluded a multivariate analysis to control for temporal trends that might have occurred in our pre- to poststudy design. Statistically significant demographic and clinical variable differences between groups were likely not clinically meaningful. Although more patients were on intravenous anticoagulants in the preintervention group, coagulopathy or low platelets do not increase the bleeding risk during paracenteses,^1,8 and there was no clinical difference in INR, PTT, or platelets between groups (Table). Third, it is possible that unmeasured characteristics contributed to more patient complications in the preintervention group. Finally, we were unable to evaluate length of stay and mortality differences between groups that might have been attributable to the procedure because of the low number of major bleeding complications and the inability to perform a multivariate analysis.

Our results suggest that using the 2-probe technique to predetermine the needle path before performing paracentesis might prevent major bleeding. Based on our findings, we believe that the addition of a vascular ultrasound during paracentesis should be considered by all hospitalists.

Acknowledgments

The authors acknowledge Drs. Douglas Vaughan and Kevin O’Leary for their support and encouragement of this work. They would also like to thank the Cedars-Sinai Enterprise Information Systems Department for assistance with their data query.

Disclosure

The authors have no relevant financial disclosures or conflicts of interest to report.

Ascites is the most common complication of cirrhosis and often leads to hospitalization. ¹ Paracentesis is recommended for all patients admitted with ascites and cirrhosis. ¹ Additionally, the Society of Hospital Medicine considers the ability to perform paracenteses a core competency for hospitalists. ² Although considered a safe procedure, major bleeding complications occur in 0.2% to 1.7% of paracenteses. ^3-7 Patients with cirrhosis form new abdominal wall vessels because of portal hypertension, and hemoperitoneum from the laceration of these vessels during paracentesis carries a high morbidity and mortality. ^6,8 Ultrasound guidance using a low-frequency ultrasound probe is currently standard practice for paracentesis and has been shown to reduce bleeding complications. ^9-11 However, the use of vascular ultrasound (high-frequency probe) is also recommended to identify blood vessels within the intended needle pathway to reduce bleeding, but no studies have been performed to demonstrate a benefit. ^3,11 This study aimed to evaluate whether this “2-probe technique” reduces paracentesis-related bleeding complications.

The procedure service at Cedars Sinai Medical Center (CSMC) in Los Angeles performs paracentesis regularly with ultrasound guidance. CSMC is a tertiary care, academic medical center with 861 licensed beds. We performed a pre- to postintervention study of consecutive patients (admitted and ambulatory) who underwent paracentesis done by 1 proceduralist (MJA) from the procedure service at CSMC from February 2010 through February 2016. From February 1, 2010, through August 2011, paracenteses were performed using only low-frequency, phased array ultrasound probes (preintervention group). From September 1, 2011, through February 2016, a 2-probe technique was used, whereby ultrasound interrogation of the abdomen using a low-frequency, phased array probe (to identify ascites) was supplemented with a second scan using a high-frequency, linear probe to identify vasculature within the planned needle path (postintervention group). As a standard part of quality assurance, CSMC documented all paracentesis-related complications from procedures performed by their center. Northwestern University investigators (JHB, EC, JF) independently evaluated these data to look at bleeding complications before and after the implementation of the 2-probe technique. The CSMC and Northwestern University institutional review boards approved this study.

Procedure Protocol

Each patient’s primary team or outpatient physician requested a consultation for paracentesis from the CSMC procedure service. All patient evaluations began with an abdominal ultrasound using the low-frequency probe to determine the presence of ascites and a potential window of access to the fluid. After September 1, 2011, the CSMC procedure service implemented the 2-probe technique to also evaluate the abdominal wall for the presence of vessels. Color flow Doppler ultrasound further helped to differentiate blood vessels as necessary. The optimal window was then marked on the abdominal wall, and the paracentesis was performed. Per the routine of the CSMC procedure service, antiplatelet or anticoagulant medications were not held for paracenteses.

Measurement

All data were collected prospectively at the time of the procedure, including the volume of fluid removed, the number of needle passes required, and whether the patient was on antiplatelet or anticoagulant medications (including warfarin, direct oral anticoagulants, thrombin inhibitors, heparin, or low molecular weight heparins). Patients were followed for complications for up to 24 hours after the procedure or until a clinical question of a complication was reconciled. Minor bleeding was defined as new serosanguinous fluid on repeat paracentesis not associated with hemodynamic changes, local bruising or bleeding at the site, or abdominal wall hematoma. Major bleeding was defined by the development of hemodynamic instability or by reaccumulation of fluid on ultrasound within 24 hours postparacentesis and one of the following: an associated hemoglobin drop of greater than 2 g/dl, blood seen on repeat paracentesis, blood density fluid on a computed tomography scan, or the lack of an alternative explanation. All data were recorded in a handheld database (HanDbase; DDH Software, Wellington, FL).

A query of the electronic medical record was performed to obtain patient demographics and relevant clinical information, including age, sex, body mass index, International Normalized Ratio (INR), partial thromboplastin time (PTT), platelet counts (10³/uL, hematocrit (%) and creatinine (mg/dl). Our query for laboratory data retrieved the closest laboratory entry up to 48 hours before the procedure.

Statistical Analysis

We used a χ² test, Student t test, or Kruskal-Wallis test to compare demographic and clinical characteristics of procedure patients between the 2 study groups (pre- and postintervention). Major and minor bleeding were compared between the 2 groups using the χ² test.¹² We used the χ² test instead of the Fisher’s exact test for several reasons. The usual rule is that the Fisher’s exact test is necessary when 1 or more expected outcome values are less than 5. However, McDonald argues that the χ² test should be used with large sample sizes (more than 1000) in lieu of the outcome-value-of-5 rule.¹² The Fisher’s exact test also assumes that the row and column totals are fixed. However, the outcomes in our study were not fixed because any patient could have a bleeding complication during each procedure. When row and column totals are not fixed, only 5% of the time will a P value be less than 0.05, and the Fisher’s exact test is too conservative.¹² We performed all statistical analyses using IBM SPSS Statistics Version 22 (IBM Corp, Armonk, NY).

Patient demographic and clinical information can be found in the Table. The proceduralist (MJA) performed a total of 5777 paracenteses (1000 preintervention, 4777 postintervention) on 1639 patients. Four hundred eighty-nine (10.2%) vascular anomalies were identified within the intended needle path in the postintervention group (Figure). More patients in the preintervention group were on aspirin (93 [9.3%] vs 230 [4.8%]; P < 0.001) and therapeutic intravenous anticoagulants (33 [3.3%] vs 89 [1.9%]; P = 0.004), while more patients in the postintervention group were on both an antiplatelet and oral anticoagulant (1 [0.1%] vs 38 [0.8%]; P = 0.015) and subcutaneous prophylactic anticoagulants (184 [18.4%] vs 1120 [23.4%]; P = 0.001) at the time of the procedure. There were no other differences between groups with antiplatelet or anticoagulant drugs. We found no difference in minor bleeding between pre- and postintervention groups. Major bleeding was lower after the 2-probe technique was implemented (3 [0.3%] vs 4 [0.08%]; P = 0.07). There were no between-group differences in INR, PTT, or platelet counts among major bleeders. One patient in the postintervention group had hemodynamic instability and dropped his hemoglobin by 3.8 g/dl at 7 hours after the procedure. This was unexplained, as the patient had no abdominal symptoms or findings on examination. The patient received several liters of fluid before ultimately dying, and the primary team considered sepsis as a possible cause, but no postmortem examination was performed. This was the only death attributed to a major bleeding complication. We included this patient in our analysis because the cause of his demise was not completely clear. However, excluding this patient would change the results from a trend to a statistically significant difference between groups (3 [0.3%] vs 3 [0.06%]; P = 0.03).

To our knowledge, we report the largest series of paracentesis prospectively evaluated for bleeding complications, and this is the first study to evaluate whether adding a vascular ultrasound (high-frequency probe) avoids major bleeding. In our series, up to 10% of patients had abnormal vessels seen with a vascular ultrasound that were within the original intended trajectory path of the needle. These vessels were also likely present yet invisible when ultrasound-guided paracentesis using only the standard, low-frequency probe was being performed. It is unknown whether these vessels are routinely traversed with the needle, nicked, or narrowly avoided during paracenteses performed using only a low-frequency probe.

Procedure-related bleeding may not be completely avoidable, despite using the vascular probe. Some authors have suggested that the mechanism of bleeding is more related to the rapid reduction in intraperitoneal pressure, which increases the gradient across vessel walls, resulting in rupture and bleeding.⁶ However, in our series, using vascular ultrasound also reduced major bleeding to numbers lower than those historically reported in the literature (0.2%).^3-4 Our preintervention number needed to harm was 333 procedures to cause 1 major bleed, compared to 1250 (or 1666 using the 3-patient bleeding analysis) in the postintervention group. In 2008, 150,000 Medicare beneficiaries underwent paracentesis.¹³ Using our study analysis, if vascular ultrasound was used on these patients, up to 360 major bleeds may have been prevented, along with a corresponding reduction in unnecessary morbidity and mortality.

Our study has several limitations. First, it was limited to 1 center with 1 very experienced proceduralist. Although it is possible that the reduction in major bleeding may have been due to the increasing experience of the proceduralist over time, we do not think that this is likely because he had already performed thousands of paracenteses over 9 years before the start of our study. Second, major bleeding was rare and therefore precluded a multivariate analysis to control for temporal trends that might have occurred in our pre- to poststudy design. Statistically significant demographic and clinical variable differences between groups were likely not clinically meaningful. Although more patients were on intravenous anticoagulants in the preintervention group, coagulopathy or low platelets do not increase the bleeding risk during paracenteses,^1,8 and there was no clinical difference in INR, PTT, or platelets between groups (Table). Third, it is possible that unmeasured characteristics contributed to more patient complications in the preintervention group. Finally, we were unable to evaluate length of stay and mortality differences between groups that might have been attributable to the procedure because of the low number of major bleeding complications and the inability to perform a multivariate analysis.

Our results suggest that using the 2-probe technique to predetermine the needle path before performing paracentesis might prevent major bleeding. Based on our findings, we believe that the addition of a vascular ultrasound during paracentesis should be considered by all hospitalists.

Acknowledgments

The authors acknowledge Drs. Douglas Vaughan and Kevin O’Leary for their support and encouragement of this work. They would also like to thank the Cedars-Sinai Enterprise Information Systems Department for assistance with their data query.

Disclosure

The authors have no relevant financial disclosures or conflicts of interest to report.

Ascites is the most common complication of cirrhosis and often leads to hospitalization. ¹ Paracentesis is recommended for all patients admitted with ascites and cirrhosis. ¹ Additionally, the Society of Hospital Medicine considers the ability to perform paracenteses a core competency for hospitalists. ² Although considered a safe procedure, major bleeding complications occur in 0.2% to 1.7% of paracenteses. ^3-7 Patients with cirrhosis form new abdominal wall vessels because of portal hypertension, and hemoperitoneum from the laceration of these vessels during paracentesis carries a high morbidity and mortality. ^6,8 Ultrasound guidance using a low-frequency ultrasound probe is currently standard practice for paracentesis and has been shown to reduce bleeding complications. ^9-11 However, the use of vascular ultrasound (high-frequency probe) is also recommended to identify blood vessels within the intended needle pathway to reduce bleeding, but no studies have been performed to demonstrate a benefit. ^3,11 This study aimed to evaluate whether this “2-probe technique” reduces paracentesis-related bleeding complications.

The procedure service at Cedars Sinai Medical Center (CSMC) in Los Angeles performs paracentesis regularly with ultrasound guidance. CSMC is a tertiary care, academic medical center with 861 licensed beds. We performed a pre- to postintervention study of consecutive patients (admitted and ambulatory) who underwent paracentesis done by 1 proceduralist (MJA) from the procedure service at CSMC from February 2010 through February 2016. From February 1, 2010, through August 2011, paracenteses were performed using only low-frequency, phased array ultrasound probes (preintervention group). From September 1, 2011, through February 2016, a 2-probe technique was used, whereby ultrasound interrogation of the abdomen using a low-frequency, phased array probe (to identify ascites) was supplemented with a second scan using a high-frequency, linear probe to identify vasculature within the planned needle path (postintervention group). As a standard part of quality assurance, CSMC documented all paracentesis-related complications from procedures performed by their center. Northwestern University investigators (JHB, EC, JF) independently evaluated these data to look at bleeding complications before and after the implementation of the 2-probe technique. The CSMC and Northwestern University institutional review boards approved this study.

Procedure Protocol

Each patient’s primary team or outpatient physician requested a consultation for paracentesis from the CSMC procedure service. All patient evaluations began with an abdominal ultrasound using the low-frequency probe to determine the presence of ascites and a potential window of access to the fluid. After September 1, 2011, the CSMC procedure service implemented the 2-probe technique to also evaluate the abdominal wall for the presence of vessels. Color flow Doppler ultrasound further helped to differentiate blood vessels as necessary. The optimal window was then marked on the abdominal wall, and the paracentesis was performed. Per the routine of the CSMC procedure service, antiplatelet or anticoagulant medications were not held for paracenteses.

Measurement

All data were collected prospectively at the time of the procedure, including the volume of fluid removed, the number of needle passes required, and whether the patient was on antiplatelet or anticoagulant medications (including warfarin, direct oral anticoagulants, thrombin inhibitors, heparin, or low molecular weight heparins). Patients were followed for complications for up to 24 hours after the procedure or until a clinical question of a complication was reconciled. Minor bleeding was defined as new serosanguinous fluid on repeat paracentesis not associated with hemodynamic changes, local bruising or bleeding at the site, or abdominal wall hematoma. Major bleeding was defined by the development of hemodynamic instability or by reaccumulation of fluid on ultrasound within 24 hours postparacentesis and one of the following: an associated hemoglobin drop of greater than 2 g/dl, blood seen on repeat paracentesis, blood density fluid on a computed tomography scan, or the lack of an alternative explanation. All data were recorded in a handheld database (HanDbase; DDH Software, Wellington, FL).

A query of the electronic medical record was performed to obtain patient demographics and relevant clinical information, including age, sex, body mass index, International Normalized Ratio (INR), partial thromboplastin time (PTT), platelet counts (10³/uL, hematocrit (%) and creatinine (mg/dl). Our query for laboratory data retrieved the closest laboratory entry up to 48 hours before the procedure.

Statistical Analysis

We used a χ² test, Student t test, or Kruskal-Wallis test to compare demographic and clinical characteristics of procedure patients between the 2 study groups (pre- and postintervention). Major and minor bleeding were compared between the 2 groups using the χ² test.¹² We used the χ² test instead of the Fisher’s exact test for several reasons. The usual rule is that the Fisher’s exact test is necessary when 1 or more expected outcome values are less than 5. However, McDonald argues that the χ² test should be used with large sample sizes (more than 1000) in lieu of the outcome-value-of-5 rule.¹² The Fisher’s exact test also assumes that the row and column totals are fixed. However, the outcomes in our study were not fixed because any patient could have a bleeding complication during each procedure. When row and column totals are not fixed, only 5% of the time will a P value be less than 0.05, and the Fisher’s exact test is too conservative.¹² We performed all statistical analyses using IBM SPSS Statistics Version 22 (IBM Corp, Armonk, NY).

Patient demographic and clinical information can be found in the Table. The proceduralist (MJA) performed a total of 5777 paracenteses (1000 preintervention, 4777 postintervention) on 1639 patients. Four hundred eighty-nine (10.2%) vascular anomalies were identified within the intended needle path in the postintervention group (Figure). More patients in the preintervention group were on aspirin (93 [9.3%] vs 230 [4.8%]; P < 0.001) and therapeutic intravenous anticoagulants (33 [3.3%] vs 89 [1.9%]; P = 0.004), while more patients in the postintervention group were on both an antiplatelet and oral anticoagulant (1 [0.1%] vs 38 [0.8%]; P = 0.015) and subcutaneous prophylactic anticoagulants (184 [18.4%] vs 1120 [23.4%]; P = 0.001) at the time of the procedure. There were no other differences between groups with antiplatelet or anticoagulant drugs. We found no difference in minor bleeding between pre- and postintervention groups. Major bleeding was lower after the 2-probe technique was implemented (3 [0.3%] vs 4 [0.08%]; P = 0.07). There were no between-group differences in INR, PTT, or platelet counts among major bleeders. One patient in the postintervention group had hemodynamic instability and dropped his hemoglobin by 3.8 g/dl at 7 hours after the procedure. This was unexplained, as the patient had no abdominal symptoms or findings on examination. The patient received several liters of fluid before ultimately dying, and the primary team considered sepsis as a possible cause, but no postmortem examination was performed. This was the only death attributed to a major bleeding complication. We included this patient in our analysis because the cause of his demise was not completely clear. However, excluding this patient would change the results from a trend to a statistically significant difference between groups (3 [0.3%] vs 3 [0.06%]; P = 0.03).

To our knowledge, we report the largest series of paracentesis prospectively evaluated for bleeding complications, and this is the first study to evaluate whether adding a vascular ultrasound (high-frequency probe) avoids major bleeding. In our series, up to 10% of patients had abnormal vessels seen with a vascular ultrasound that were within the original intended trajectory path of the needle. These vessels were also likely present yet invisible when ultrasound-guided paracentesis using only the standard, low-frequency probe was being performed. It is unknown whether these vessels are routinely traversed with the needle, nicked, or narrowly avoided during paracenteses performed using only a low-frequency probe.

Procedure-related bleeding may not be completely avoidable, despite using the vascular probe. Some authors have suggested that the mechanism of bleeding is more related to the rapid reduction in intraperitoneal pressure, which increases the gradient across vessel walls, resulting in rupture and bleeding.⁶ However, in our series, using vascular ultrasound also reduced major bleeding to numbers lower than those historically reported in the literature (0.2%).^3-4 Our preintervention number needed to harm was 333 procedures to cause 1 major bleed, compared to 1250 (or 1666 using the 3-patient bleeding analysis) in the postintervention group. In 2008, 150,000 Medicare beneficiaries underwent paracentesis.¹³ Using our study analysis, if vascular ultrasound was used on these patients, up to 360 major bleeds may have been prevented, along with a corresponding reduction in unnecessary morbidity and mortality.

Our study has several limitations. First, it was limited to 1 center with 1 very experienced proceduralist. Although it is possible that the reduction in major bleeding may have been due to the increasing experience of the proceduralist over time, we do not think that this is likely because he had already performed thousands of paracenteses over 9 years before the start of our study. Second, major bleeding was rare and therefore precluded a multivariate analysis to control for temporal trends that might have occurred in our pre- to poststudy design. Statistically significant demographic and clinical variable differences between groups were likely not clinically meaningful. Although more patients were on intravenous anticoagulants in the preintervention group, coagulopathy or low platelets do not increase the bleeding risk during paracenteses,^1,8 and there was no clinical difference in INR, PTT, or platelets between groups (Table). Third, it is possible that unmeasured characteristics contributed to more patient complications in the preintervention group. Finally, we were unable to evaluate length of stay and mortality differences between groups that might have been attributable to the procedure because of the low number of major bleeding complications and the inability to perform a multivariate analysis.

Our results suggest that using the 2-probe technique to predetermine the needle path before performing paracentesis might prevent major bleeding. Based on our findings, we believe that the addition of a vascular ultrasound during paracentesis should be considered by all hospitalists.

Acknowledgments

The authors acknowledge Drs. Douglas Vaughan and Kevin O’Leary for their support and encouragement of this work. They would also like to thank the Cedars-Sinai Enterprise Information Systems Department for assistance with their data query.

Disclosure

The authors have no relevant financial disclosures or conflicts of interest to report.

Annual healthcare costs in the United States are over $3 trillion and are garnering significant national attention.¹ The United States spends approximately 2.5 times more per capita on healthcare when compared to other developed nations.² One source of unnecessary cost in healthcare is defensive medicine. Defensive medicine has been defined by Congress as occurring “when doctors order tests, procedures, or visits, or avoid certain high-risk patients or procedures, primarily (but not necessarily) because of concern about malpractice liability.”³

Though difficult to assess, in 1 study, defensive medicine was estimated to cost $45 billion annually.⁴ While general agreement exists that physicians practice defensive medicine, the extent of defensive practices and the subsequent impact on healthcare costs remain unclear. This is especially true for a group of clinicians that is rapidly increasing in number: hospitalists. Currently, there are more than 50,000 hospitalists in the United States,⁵ yet the prevalence of defensive medicine in this relatively new specialty is unknown. Inpatient care is complex and time constraints can impede establishing an optimal therapeutic relationship with the patient, potentially raising liability fears. We therefore sought to quantify hospitalist physician estimates of the cost of defensive medicine and assess correlates of their estimates. As being sued might spur defensive behaviors, we also assessed how many hospitalists reported being sued and whether this was associated with their estimates of defensive medicine.

Survey Questionnaire

In a previously published survey-based analysis, we reported on physician practice and overuse for 2 common scenarios in hospital medicine: preoperative evaluation and management of uncomplicated syncope.⁶ After responding to the vignettes, each physician was asked to provide demographic and employment information and malpractice history. In addition, they were asked the following: In your best estimation, what percentage of healthcare-related resources (eg, hospital admissions, diagnostic testing, treatment) are spent purely because of defensive medicine concerns? __________% resources

Survey Sample & Administration

The survey was sent to a sample of 1753 hospitalists, randomly identified through the Society of Hospital Medicine’s (SHM) database of members and annual meeting attendees. It is estimated that almost 30% of practicing hospitalists in the United States are members of the SHM.⁵ A full description of the sampling methodology was previously published.⁶ Selected hospitalists were mailed surveys, a $20 financial incentive, and subsequent reminders between June and October 2011.

The study was exempted from institutional review board review by the University of Michigan and the VA Ann Arbor Healthcare System.

Variables

The primary outcome of interest was the response to the “% resources” estimated to be spent on defensive medicine. This was analyzed as a continuous variable. Independent variables included the following: VA employment, malpractice insurance payer, employer, history of malpractice lawsuit, sex, race, and years practicing as a physician.

Statistical Analysis

Analyses were conducted using SAS, version 9.4 (SAS Institute). Descriptive statistics were first calculated for all variables. Next, bivariable comparisons between the outcome variables and other variables of interest were performed. Multivariable comparisons were made using linear regression for the outcome of estimated resources spent on defensive medicine. A P value of < 0.05 was considered statistically significant.

Of the 1753 surveys mailed, 253 were excluded due to incorrect addresses or because the recipients were not practicing hospitalists. A total of 1020 were completed and returned, yielding a 68% response rate (1020 out of 1500 eligible). The hospitalist respondents were in practice for an average of 11 years (range 1-40 years). Respondents represented all 50 states and had a diverse background of experience and demographic characteristics, which has been previously described.⁶

Resources Estimated Spent on Defensive Medicine

Hospitalists reported, on average, that they believed defensive medicine accounted for 37.5% (standard deviation, 20.2%) of all healthcare spending. Results from the multivariable regression are presented in the Table. Hospitalists affiliated with a VA hospital reported 5.5% less in resources spent on defensive medicine than those not affiliated with a VA hospital (32.2% VA vs 37.7% non-VA, P = 0.025). For every 10 years in practice, the estimate of resources spent on defensive medicine decreased by 3% (P = 0.003). Those who were male (36.4% male vs 39.4% female, P = 0.023) and non-Hispanic white (32.5% non-Hispanic white vs 44.7% other, P ≤ 0.001) also estimated less resources spent on defensive medicine. We did not find an association between a hospitalist reporting being sued and their perception of resources spent on defensive medicine.  

Risk of Being Sued

Over a quarter of our sample (25.6%) reported having been sued at least once for medical malpractice. The proportion of hospitalists that reported a history of being sued generally increased with more years of practice (Figure). For those who had been in practice for at least 20 years, more than half (55%) had been sued at least once during their career.

In a national survey, hospitalists estimated that almost 40% of all healthcare-related resources are spent purely because of defensive medicine concerns. This estimate was affected by personal demographic and employment factors. Our second major finding is that over one-quarter of a large random sample of hospitalist physicians reported being sued for malpractice.

Hospitalist perceptions of defensive medicine varied significantly based on employment at a VA hospital, with VA-affiliated hospitalists reporting less estimated spending on defensive medicine. This effect may reflect a less litigious environment within the VA, even though physicians practicing within the VA can be reported to the National Practitioner Data Bank.⁷ The different environment may be due to the VA’s patient mix (VA patients tend to be poorer, older, sicker, and have more mental illness)⁸; however, it could also be due to its de facto practice of a form of enterprise liability, in which, by law, the VA assumes responsibility for negligence, sheltering its physicians from direct liability.

We also found that the higher the number of years a hospitalist reported practicing, the lower the perception of resources being spent on defensive medicine. The reason for this finding is unclear. There has been a recent focus on high-value care and overspending, and perhaps younger hospitalists are more aware of these initiatives and thus have higher estimates. Additionally, non-Hispanic white male respondents estimated a lower amount spent on defensive medicine compared with other respondents. This is consistent with previous studies of risk perception which have noted a “white male effect” in which white males generally perceive a wide range of risks to be lower than female and non-white individuals, likely due to sociopolitical factors.⁹ Here, the white male effect is particularly interesting, considering that male physicians are almost 2.5 times as likely as female physicians to report being sued.¹⁰

Similar to prior studies,¹¹ there was no association with personal liability claim experience and perceived resources spent on defensive medicine. It is unclear why personal experience of being sued does not appear to be associated with perceptions of defensive medicine practice. It is possible that the fear of being sued is worse than the actual experience or that physicians believe that lawsuits are either random events or inevitable and, as a result, do not change their practice patterns.

The lifetime risk of being named in a malpractice suit is substantial for hospitalists: in our study, over half of hospitalists in practice for 20 years or more reported they had been sued. This corresponds with the projection made by Jena and colleagues,¹² which estimated that 55% of internal medicine physicians will be sued by the age of 45, a number just slightly higher than the average for all physicians.

Our study has important limitations. Our sample was of hospitalists and therefore may not be reflective of other medical specialties. Second, due to the nature of the study design, the responses to spending on defensive medicine may not represent actual practice. Third, we did not confirm details such as place of employment or history of lawsuit, and this may be subject to recall bias. However, physicians are unlikely to forget having been sued. Finally, this survey is observational and cross-sectional. Our data imply association rather than causation. Without longitudinal data, it is impossible to know if years of practice correlate with perceived defensive medicine spending due to a generational effect or a longitudinal effect (such as more confidence in diagnostic skills with more years of practice).

Despite these limitations, our survey has important policy implications. First, we found that defensive medicine is perceived by hospitalists to be costly. Although physicians likely overestimated the cost (37.5%, or an estimated $1 trillion is far higher than previous estimates of approximately 2% of all healthcare spending),⁴ it also demonstrates the extent to which physicians feel as though the medical care that is provided may be unnecessary. Second, at least a quarter of hospitalist physicians have been sued, and the risk of being named as a defendant in a lawsuit increases the longer they have been in clinical practice.

Given these findings, policies aimed to reduce the practice of defensive medicine may help the rising costs of healthcare. Reducing defensive medicine requires decreasing physician fears of liability and related reporting. Traditional tort reforms (with the exception of damage caps) have not been proven to do this. And damage caps can be inequitable, hard to pass, and even found to be unconstitutional in some states.¹³ However, other reform options hold promise in reducing liability fears, including enterprise liability, safe harbor legislation, and health courts.¹³ Finally, shared decision-making models may also provide a method to reduce defensive fears as well.⁶

Acknowledgments

The authors thank the Society of Hospital Medicine, Dr. Scott Flanders, Andrew Hickner, and David Ratz for their assistance with this project.

Disclosure

The authors received financial support from the Blue Cross Blue Shield of Michigan Foundation, the Department of Veterans Affairs Health Services Research and Development Center for Clinical Management Research, the University of Michigan Specialist-Hospitalist Allied Research Program, and the Ann Arbor University of Michigan VA Patient Safety Enhancement Program.

Disclaimer

The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of Blue Cross Blue Shield of Michigan Foundation, the Department of Veterans Affairs, or the Society of Hospital Medicine.

Annual healthcare costs in the United States are over $3 trillion and are garnering significant national attention.¹ The United States spends approximately 2.5 times more per capita on healthcare when compared to other developed nations.² One source of unnecessary cost in healthcare is defensive medicine. Defensive medicine has been defined by Congress as occurring “when doctors order tests, procedures, or visits, or avoid certain high-risk patients or procedures, primarily (but not necessarily) because of concern about malpractice liability.”³

Though difficult to assess, in 1 study, defensive medicine was estimated to cost $45 billion annually.⁴ While general agreement exists that physicians practice defensive medicine, the extent of defensive practices and the subsequent impact on healthcare costs remain unclear. This is especially true for a group of clinicians that is rapidly increasing in number: hospitalists. Currently, there are more than 50,000 hospitalists in the United States,⁵ yet the prevalence of defensive medicine in this relatively new specialty is unknown. Inpatient care is complex and time constraints can impede establishing an optimal therapeutic relationship with the patient, potentially raising liability fears. We therefore sought to quantify hospitalist physician estimates of the cost of defensive medicine and assess correlates of their estimates. As being sued might spur defensive behaviors, we also assessed how many hospitalists reported being sued and whether this was associated with their estimates of defensive medicine.

Survey Questionnaire

In a previously published survey-based analysis, we reported on physician practice and overuse for 2 common scenarios in hospital medicine: preoperative evaluation and management of uncomplicated syncope.⁶ After responding to the vignettes, each physician was asked to provide demographic and employment information and malpractice history. In addition, they were asked the following: In your best estimation, what percentage of healthcare-related resources (eg, hospital admissions, diagnostic testing, treatment) are spent purely because of defensive medicine concerns? __________% resources

Survey Sample & Administration

The survey was sent to a sample of 1753 hospitalists, randomly identified through the Society of Hospital Medicine’s (SHM) database of members and annual meeting attendees. It is estimated that almost 30% of practicing hospitalists in the United States are members of the SHM.⁵ A full description of the sampling methodology was previously published.⁶ Selected hospitalists were mailed surveys, a $20 financial incentive, and subsequent reminders between June and October 2011.

The study was exempted from institutional review board review by the University of Michigan and the VA Ann Arbor Healthcare System.

Variables

The primary outcome of interest was the response to the “% resources” estimated to be spent on defensive medicine. This was analyzed as a continuous variable. Independent variables included the following: VA employment, malpractice insurance payer, employer, history of malpractice lawsuit, sex, race, and years practicing as a physician.

Statistical Analysis

Analyses were conducted using SAS, version 9.4 (SAS Institute). Descriptive statistics were first calculated for all variables. Next, bivariable comparisons between the outcome variables and other variables of interest were performed. Multivariable comparisons were made using linear regression for the outcome of estimated resources spent on defensive medicine. A P value of < 0.05 was considered statistically significant.

Of the 1753 surveys mailed, 253 were excluded due to incorrect addresses or because the recipients were not practicing hospitalists. A total of 1020 were completed and returned, yielding a 68% response rate (1020 out of 1500 eligible). The hospitalist respondents were in practice for an average of 11 years (range 1-40 years). Respondents represented all 50 states and had a diverse background of experience and demographic characteristics, which has been previously described.⁶

Resources Estimated Spent on Defensive Medicine

Hospitalists reported, on average, that they believed defensive medicine accounted for 37.5% (standard deviation, 20.2%) of all healthcare spending. Results from the multivariable regression are presented in the Table. Hospitalists affiliated with a VA hospital reported 5.5% less in resources spent on defensive medicine than those not affiliated with a VA hospital (32.2% VA vs 37.7% non-VA, P = 0.025). For every 10 years in practice, the estimate of resources spent on defensive medicine decreased by 3% (P = 0.003). Those who were male (36.4% male vs 39.4% female, P = 0.023) and non-Hispanic white (32.5% non-Hispanic white vs 44.7% other, P ≤ 0.001) also estimated less resources spent on defensive medicine. We did not find an association between a hospitalist reporting being sued and their perception of resources spent on defensive medicine.  

Risk of Being Sued

Over a quarter of our sample (25.6%) reported having been sued at least once for medical malpractice. The proportion of hospitalists that reported a history of being sued generally increased with more years of practice (Figure). For those who had been in practice for at least 20 years, more than half (55%) had been sued at least once during their career.

In a national survey, hospitalists estimated that almost 40% of all healthcare-related resources are spent purely because of defensive medicine concerns. This estimate was affected by personal demographic and employment factors. Our second major finding is that over one-quarter of a large random sample of hospitalist physicians reported being sued for malpractice.

Hospitalist perceptions of defensive medicine varied significantly based on employment at a VA hospital, with VA-affiliated hospitalists reporting less estimated spending on defensive medicine. This effect may reflect a less litigious environment within the VA, even though physicians practicing within the VA can be reported to the National Practitioner Data Bank.⁷ The different environment may be due to the VA’s patient mix (VA patients tend to be poorer, older, sicker, and have more mental illness)⁸; however, it could also be due to its de facto practice of a form of enterprise liability, in which, by law, the VA assumes responsibility for negligence, sheltering its physicians from direct liability.

We also found that the higher the number of years a hospitalist reported practicing, the lower the perception of resources being spent on defensive medicine. The reason for this finding is unclear. There has been a recent focus on high-value care and overspending, and perhaps younger hospitalists are more aware of these initiatives and thus have higher estimates. Additionally, non-Hispanic white male respondents estimated a lower amount spent on defensive medicine compared with other respondents. This is consistent with previous studies of risk perception which have noted a “white male effect” in which white males generally perceive a wide range of risks to be lower than female and non-white individuals, likely due to sociopolitical factors.⁹ Here, the white male effect is particularly interesting, considering that male physicians are almost 2.5 times as likely as female physicians to report being sued.¹⁰

Similar to prior studies,¹¹ there was no association with personal liability claim experience and perceived resources spent on defensive medicine. It is unclear why personal experience of being sued does not appear to be associated with perceptions of defensive medicine practice. It is possible that the fear of being sued is worse than the actual experience or that physicians believe that lawsuits are either random events or inevitable and, as a result, do not change their practice patterns.

The lifetime risk of being named in a malpractice suit is substantial for hospitalists: in our study, over half of hospitalists in practice for 20 years or more reported they had been sued. This corresponds with the projection made by Jena and colleagues,¹² which estimated that 55% of internal medicine physicians will be sued by the age of 45, a number just slightly higher than the average for all physicians.

Our study has important limitations. Our sample was of hospitalists and therefore may not be reflective of other medical specialties. Second, due to the nature of the study design, the responses to spending on defensive medicine may not represent actual practice. Third, we did not confirm details such as place of employment or history of lawsuit, and this may be subject to recall bias. However, physicians are unlikely to forget having been sued. Finally, this survey is observational and cross-sectional. Our data imply association rather than causation. Without longitudinal data, it is impossible to know if years of practice correlate with perceived defensive medicine spending due to a generational effect or a longitudinal effect (such as more confidence in diagnostic skills with more years of practice).

Despite these limitations, our survey has important policy implications. First, we found that defensive medicine is perceived by hospitalists to be costly. Although physicians likely overestimated the cost (37.5%, or an estimated $1 trillion is far higher than previous estimates of approximately 2% of all healthcare spending),⁴ it also demonstrates the extent to which physicians feel as though the medical care that is provided may be unnecessary. Second, at least a quarter of hospitalist physicians have been sued, and the risk of being named as a defendant in a lawsuit increases the longer they have been in clinical practice.

Given these findings, policies aimed to reduce the practice of defensive medicine may help the rising costs of healthcare. Reducing defensive medicine requires decreasing physician fears of liability and related reporting. Traditional tort reforms (with the exception of damage caps) have not been proven to do this. And damage caps can be inequitable, hard to pass, and even found to be unconstitutional in some states.¹³ However, other reform options hold promise in reducing liability fears, including enterprise liability, safe harbor legislation, and health courts.¹³ Finally, shared decision-making models may also provide a method to reduce defensive fears as well.⁶

Acknowledgments

The authors thank the Society of Hospital Medicine, Dr. Scott Flanders, Andrew Hickner, and David Ratz for their assistance with this project.

Disclosure

The authors received financial support from the Blue Cross Blue Shield of Michigan Foundation, the Department of Veterans Affairs Health Services Research and Development Center for Clinical Management Research, the University of Michigan Specialist-Hospitalist Allied Research Program, and the Ann Arbor University of Michigan VA Patient Safety Enhancement Program.

Disclaimer

The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of Blue Cross Blue Shield of Michigan Foundation, the Department of Veterans Affairs, or the Society of Hospital Medicine.

Annual healthcare costs in the United States are over $3 trillion and are garnering significant national attention.¹ The United States spends approximately 2.5 times more per capita on healthcare when compared to other developed nations.² One source of unnecessary cost in healthcare is defensive medicine. Defensive medicine has been defined by Congress as occurring “when doctors order tests, procedures, or visits, or avoid certain high-risk patients or procedures, primarily (but not necessarily) because of concern about malpractice liability.”³

Though difficult to assess, in 1 study, defensive medicine was estimated to cost $45 billion annually.⁴ While general agreement exists that physicians practice defensive medicine, the extent of defensive practices and the subsequent impact on healthcare costs remain unclear. This is especially true for a group of clinicians that is rapidly increasing in number: hospitalists. Currently, there are more than 50,000 hospitalists in the United States,⁵ yet the prevalence of defensive medicine in this relatively new specialty is unknown. Inpatient care is complex and time constraints can impede establishing an optimal therapeutic relationship with the patient, potentially raising liability fears. We therefore sought to quantify hospitalist physician estimates of the cost of defensive medicine and assess correlates of their estimates. As being sued might spur defensive behaviors, we also assessed how many hospitalists reported being sued and whether this was associated with their estimates of defensive medicine.

Survey Questionnaire

In a previously published survey-based analysis, we reported on physician practice and overuse for 2 common scenarios in hospital medicine: preoperative evaluation and management of uncomplicated syncope.⁶ After responding to the vignettes, each physician was asked to provide demographic and employment information and malpractice history. In addition, they were asked the following: In your best estimation, what percentage of healthcare-related resources (eg, hospital admissions, diagnostic testing, treatment) are spent purely because of defensive medicine concerns? __________% resources

Survey Sample & Administration

The survey was sent to a sample of 1753 hospitalists, randomly identified through the Society of Hospital Medicine’s (SHM) database of members and annual meeting attendees. It is estimated that almost 30% of practicing hospitalists in the United States are members of the SHM.⁵ A full description of the sampling methodology was previously published.⁶ Selected hospitalists were mailed surveys, a $20 financial incentive, and subsequent reminders between June and October 2011.

The study was exempted from institutional review board review by the University of Michigan and the VA Ann Arbor Healthcare System.

Variables

The primary outcome of interest was the response to the “% resources” estimated to be spent on defensive medicine. This was analyzed as a continuous variable. Independent variables included the following: VA employment, malpractice insurance payer, employer, history of malpractice lawsuit, sex, race, and years practicing as a physician.

Statistical Analysis

Analyses were conducted using SAS, version 9.4 (SAS Institute). Descriptive statistics were first calculated for all variables. Next, bivariable comparisons between the outcome variables and other variables of interest were performed. Multivariable comparisons were made using linear regression for the outcome of estimated resources spent on defensive medicine. A P value of < 0.05 was considered statistically significant.

Of the 1753 surveys mailed, 253 were excluded due to incorrect addresses or because the recipients were not practicing hospitalists. A total of 1020 were completed and returned, yielding a 68% response rate (1020 out of 1500 eligible). The hospitalist respondents were in practice for an average of 11 years (range 1-40 years). Respondents represented all 50 states and had a diverse background of experience and demographic characteristics, which has been previously described.⁶

Resources Estimated Spent on Defensive Medicine

Hospitalists reported, on average, that they believed defensive medicine accounted for 37.5% (standard deviation, 20.2%) of all healthcare spending. Results from the multivariable regression are presented in the Table. Hospitalists affiliated with a VA hospital reported 5.5% less in resources spent on defensive medicine than those not affiliated with a VA hospital (32.2% VA vs 37.7% non-VA, P = 0.025). For every 10 years in practice, the estimate of resources spent on defensive medicine decreased by 3% (P = 0.003). Those who were male (36.4% male vs 39.4% female, P = 0.023) and non-Hispanic white (32.5% non-Hispanic white vs 44.7% other, P ≤ 0.001) also estimated less resources spent on defensive medicine. We did not find an association between a hospitalist reporting being sued and their perception of resources spent on defensive medicine.  

Risk of Being Sued

Over a quarter of our sample (25.6%) reported having been sued at least once for medical malpractice. The proportion of hospitalists that reported a history of being sued generally increased with more years of practice (Figure). For those who had been in practice for at least 20 years, more than half (55%) had been sued at least once during their career.

In a national survey, hospitalists estimated that almost 40% of all healthcare-related resources are spent purely because of defensive medicine concerns. This estimate was affected by personal demographic and employment factors. Our second major finding is that over one-quarter of a large random sample of hospitalist physicians reported being sued for malpractice.

Hospitalist perceptions of defensive medicine varied significantly based on employment at a VA hospital, with VA-affiliated hospitalists reporting less estimated spending on defensive medicine. This effect may reflect a less litigious environment within the VA, even though physicians practicing within the VA can be reported to the National Practitioner Data Bank.⁷ The different environment may be due to the VA’s patient mix (VA patients tend to be poorer, older, sicker, and have more mental illness)⁸; however, it could also be due to its de facto practice of a form of enterprise liability, in which, by law, the VA assumes responsibility for negligence, sheltering its physicians from direct liability.

We also found that the higher the number of years a hospitalist reported practicing, the lower the perception of resources being spent on defensive medicine. The reason for this finding is unclear. There has been a recent focus on high-value care and overspending, and perhaps younger hospitalists are more aware of these initiatives and thus have higher estimates. Additionally, non-Hispanic white male respondents estimated a lower amount spent on defensive medicine compared with other respondents. This is consistent with previous studies of risk perception which have noted a “white male effect” in which white males generally perceive a wide range of risks to be lower than female and non-white individuals, likely due to sociopolitical factors.⁹ Here, the white male effect is particularly interesting, considering that male physicians are almost 2.5 times as likely as female physicians to report being sued.¹⁰

Similar to prior studies,¹¹ there was no association with personal liability claim experience and perceived resources spent on defensive medicine. It is unclear why personal experience of being sued does not appear to be associated with perceptions of defensive medicine practice. It is possible that the fear of being sued is worse than the actual experience or that physicians believe that lawsuits are either random events or inevitable and, as a result, do not change their practice patterns.

The lifetime risk of being named in a malpractice suit is substantial for hospitalists: in our study, over half of hospitalists in practice for 20 years or more reported they had been sued. This corresponds with the projection made by Jena and colleagues,¹² which estimated that 55% of internal medicine physicians will be sued by the age of 45, a number just slightly higher than the average for all physicians.

Our study has important limitations. Our sample was of hospitalists and therefore may not be reflective of other medical specialties. Second, due to the nature of the study design, the responses to spending on defensive medicine may not represent actual practice. Third, we did not confirm details such as place of employment or history of lawsuit, and this may be subject to recall bias. However, physicians are unlikely to forget having been sued. Finally, this survey is observational and cross-sectional. Our data imply association rather than causation. Without longitudinal data, it is impossible to know if years of practice correlate with perceived defensive medicine spending due to a generational effect or a longitudinal effect (such as more confidence in diagnostic skills with more years of practice).

Despite these limitations, our survey has important policy implications. First, we found that defensive medicine is perceived by hospitalists to be costly. Although physicians likely overestimated the cost (37.5%, or an estimated $1 trillion is far higher than previous estimates of approximately 2% of all healthcare spending),⁴ it also demonstrates the extent to which physicians feel as though the medical care that is provided may be unnecessary. Second, at least a quarter of hospitalist physicians have been sued, and the risk of being named as a defendant in a lawsuit increases the longer they have been in clinical practice.

Given these findings, policies aimed to reduce the practice of defensive medicine may help the rising costs of healthcare. Reducing defensive medicine requires decreasing physician fears of liability and related reporting. Traditional tort reforms (with the exception of damage caps) have not been proven to do this. And damage caps can be inequitable, hard to pass, and even found to be unconstitutional in some states.¹³ However, other reform options hold promise in reducing liability fears, including enterprise liability, safe harbor legislation, and health courts.¹³ Finally, shared decision-making models may also provide a method to reduce defensive fears as well.⁶

Acknowledgments

The authors thank the Society of Hospital Medicine, Dr. Scott Flanders, Andrew Hickner, and David Ratz for their assistance with this project.

Disclosure

The authors received financial support from the Blue Cross Blue Shield of Michigan Foundation, the Department of Veterans Affairs Health Services Research and Development Center for Clinical Management Research, the University of Michigan Specialist-Hospitalist Allied Research Program, and the Ann Arbor University of Michigan VA Patient Safety Enhancement Program.

Disclaimer

The views expressed in this article are those of the authors and do not necessarily reflect the position or policy of Blue Cross Blue Shield of Michigan Foundation, the Department of Veterans Affairs, or the Society of Hospital Medicine.