Inpatient medication errors represent an important patient safety issue. The magnitude of the problem is staggering, with 1 review finding almost 1 in every 5 medication doses in error, with 7% having potential for adverse drug events.1 While mistakes made at the ordering stage are frequently intercepted by pharmacist or nursing review, administration errors are particularly difficult to prevent.2 The patient, as the last link in the medication administration chain, represents the final individual capable of preventing an incorrect medication administration. It is perhaps surprising then that patients generally lack a formal role in detecting and preventing adverse medication administration events.3

There have been some ambitious attempts to improve patient education regarding hospital medications and involve selected patients in the medication administration process. Such initiatives may result in increased patient participation and satisfaction.47 There is also potential that increased patient knowledge of their hospital medications could promote the goal of medication safety, as the actively involved patient may be able to catch medication errors in the hospital.

Knowledge of prescribed medications is a prerequisite to patient involvement in prevention of inpatient medication errors and yet there is little research on patient knowledge of their hospital medications. Furthermore, as the experience of hospitalization may be disorienting and disempowering for patients, it remains to be seen if patient attitudes toward participation in inpatient medication safety are favorable. To that end, we conducted a pilot study in which we assessed current patient awareness of their in‐hospital medications and surveyed attitudes toward increased patient knowledge of hospital medications.

PATIENTS AND METHODS

We conducted a cross‐sectional study of 50 cognitively intact adult internal medicine inpatients at the University of Colorado Hospital, a tertiary‐care academic teaching hospital. This study was part of a larger project designed to examine potential for patient involvement in the medication reconciliation process. A professional research assistant approached eligible patients within 24 hours of admission. To be eligible, patients had to self‐identify as knowing their outpatient medications, speak English, and have been admitted from the community. Nursing home residents and patients with a past medical history of dementia were excluded. Enrollment was tracked during the first half of the study to estimate effect of inclusion/exclusion criteria. Thirty‐eight percent of hospital admissions to medicine services were excluded based on the specified criteria. Thirty‐four percent of eligible patients were approached and 50% of approached patients agreed to participate in the study. Patient knowledge of their outpatient medication regimen was compared to admitting physician medication reconciliation to assess accuracy of patient self‐report of outpatient medication knowledge.

After consenting to participate, study patients completed a structured list of their outpatient medications and a survey of attitudes about being shown their in‐hospital medications, hospital medication errors, and patient involvement in hospital safety. They then completed a list of the medications they believed to be prescribed to them in the hospital.

The primary outcomes were the proportions of as needed (PRN), scheduled, and total hospital medications omitted by the patient, compared to the inpatient medication administration record (MAR) (patient errors of omission). Secondary outcomes included the number of in‐hospital medications listed by the patient that did not appear on the inpatient MAR (patient errors of commission), as well as patient attitudes measured on a 5‐point Likert scale (1 indicated strongly disagree and 5 indicated strongly agree.) Descriptive data included age, race, gender, and number of inpatient medications prescribed. Separate analysis of variance (ANOVA) models provided mean estimates of the primary outcomes and tested differences according to each of the patient characteristics: age in years (<65 or 65), self‐reported knowledge of hospital medications, and self‐reported desire to be involved in medication safety. Similar ANOVA models adjusted for number of medications were also examined to determine whether the relationship between the primary outcomes according to patient characteristics were altered by the number of medications. The protocol was approved by the Colorado Multiple Institutional Review Board.

RESULTS

Participants averaged 54 years of age (standard deviation [SD] = 17, range = 21‐89). Forty‐six percent (23/50) were male, and 74% (37/50) were non‐Hispanic white. Using a structured, patient‐completed, outpatient medication list, patients in the study were on an average of 5.3 outpatient prescription medications (range = 0‐17), 2.2 over‐the‐counter medications (range = 0‐8), and 0.2 herbal medications (range = 0‐7). The admitting physician's medication reconciliation list demonstrated similar number of outpatient prescription medications (average = 5.7) to the patient‐generated list. Fifty‐four percent of patient‐completed home medication lists included all of the prescription medications on the physician's medication reconciliation at admission. According to the inpatient MAR, study patients were prescribed an average of 11.3 scheduled and PRN hospital medications (range = 2‐26) at time of study enrollment.

Patient Knowledge of Their Hospital Medication List

Ninety‐six percent (48/50) of study patients omitted 1 or more of their hospital medications. On average, patients omitted 6.8 medications (range = 0‐22) (Table 1). Among scheduled medications, patients most commonly omitted antibiotics (17%), cardiovascular medications (16%), and antithrombotics (15%) (Figure 1). Among PRN medications, patients most commonly omitted analgesics (33%) and gastrointestinal medications (29%) (Figure 2).

Table 1.Patient Knowledge of Their Hospital Medications List

	Total Medications	Scheduled Medications	PRN Medications
NOTE: n = 50 patients. Abbreviations: CI, confidence interval; PRN, as needed.
Percent of patients with at least 1 hospital medication they could not name (95% CI)	96% (90‐100%)	94% (87‐100%)	80% (69‐92%)
Average number of hospital medications omitted by patient (range)	6.8 (0‐22)	5.2 (0‐15)	1.6 (0‐7)
Percentage of hospital medications omitted by patient (95% CI)	60% (52‐67%)	60% (52‐67%)	68% (57‐78%)

Figure 1

Figure 2

Patients less than 65 years omitted 60% of their PRN medications whereas patients greater than 65 years omitted 88% (P = 0.01). This difference remained even after adjustment for number of medications. There were no significant differences, based on age, in ability to name scheduled or total medications. Forty‐four percent of patients (22/50) believed they were receiving a medication in the hospital that was not actually prescribed.

DISCUSSION

Overall, patients in the study were able to name fewer than one‐half of their hospital medications. Our study suggests that adult medicine inpatients believe learning about their hospital medications would increase their satisfaction and has potential to promote medication safety. At the same time, patients did not know many of their hospital medications and this would limit their ability to fully participate in the medication safety process. Study patients frequently committed both errors of omission (ie, they did not know which medications were prescribed), and errors of commission (ie, they believed they were prescribed medications that were not prescribed). Younger patients were aware of more of their PRN medications than older patients, potentially reflecting greater patient care involvement in younger generations. However, study patients, regardless of age, were able to name fewer than one‐half of their PRN hospital medications. The most common scheduled hospital medications that patients were unable to name come from medication classes which can be associated with significant adverse events, including antibiotics, cardiovascular medications, and antithrombotics.

We posit that without systematically educating patients about their hospital medications, significant deficits in patient knowledge are inevitable. Some might argue that patients should not be asked to know their hospital medications or identify medication errors while sick and vulnerable. Certainly with multiple medication changes, formulary substitutions, and frequent modifications based on changes in clinical status, inpatient medication education could be time consuming and potentially introduce patient confusion or anxiety. Incorrect patient feedback could have potential to introduce new errors. An educational program might use graded participation based on patient interest and ability. Models for this exist in the literature, even extending to patient medication self‐administration.57 In our sample of inpatients, the majority desired a more active role in learning about their hospital medications and believed that their involvement might prevent hospital medication errors from occurring.

Medication literacy, education, and active patient involvement in medication monitoring as a means to improve patient outcomes has received significant attention in the outpatient setting, with lessons applicable to the hospital.8, 9 More broadly, the Joint Commission has established a Hospital National Patient Safety Goal to encourage patients' active involvement in their own care as a patient safety strategy.10 Examples set forth by the Joint Commission include involving patients in infection control measures, marking of procedural sites, and reporting of safety concerns relating to treatment.

While this study identifies patient knowledge deficit as a barrier to utilizing patients as part of the hospital medication safety process, it does not test whether reducing this knowledge deficit would actually reduce medication error. Our study population was limited to cognitively intact adult medicine patients at a single institution, limiting the generalizability of our conclusions. Our enrollment process may have resulted in a study population with less serious illness, greater knowledge of their hospital medications, and greater interest in participating in medication safety potentially overestimating patient knowledge of hospital medications. Finally, our small sample size limits the power to find differences in study comparisons.

Our findings are striking in that we found significant deficits in patient understanding of their hospital medications even among patients who believed they knew, or desired to know, what is being prescribed to them in the hospital. Without a system to incorporate the patient into hospital medication management, these patients will be disenfranchised from participating in inpatient medication safety. These results are a call to reexamine how we educate and involve patients regarding hospital medications. Mechanisms to allow patients to provide feedback to the medical team on their hospital medications might identify errors or improve patient satisfaction with their care. However, the systems and cultural changes needed to provide education on inpatient medications are considerable. Future research is needed to determine if increasing patient knowledge regarding their hospital medications would reduce medication errors in the inpatient setting and how this could be effectively implemented.

Inpatient medication errors represent an important patient safety issue. The magnitude of the problem is staggering, with 1 review finding almost 1 in every 5 medication doses in error, with 7% having potential for adverse drug events.1 While mistakes made at the ordering stage are frequently intercepted by pharmacist or nursing review, administration errors are particularly difficult to prevent.2 The patient, as the last link in the medication administration chain, represents the final individual capable of preventing an incorrect medication administration. It is perhaps surprising then that patients generally lack a formal role in detecting and preventing adverse medication administration events.3

There have been some ambitious attempts to improve patient education regarding hospital medications and involve selected patients in the medication administration process. Such initiatives may result in increased patient participation and satisfaction.47 There is also potential that increased patient knowledge of their hospital medications could promote the goal of medication safety, as the actively involved patient may be able to catch medication errors in the hospital.

Knowledge of prescribed medications is a prerequisite to patient involvement in prevention of inpatient medication errors and yet there is little research on patient knowledge of their hospital medications. Furthermore, as the experience of hospitalization may be disorienting and disempowering for patients, it remains to be seen if patient attitudes toward participation in inpatient medication safety are favorable. To that end, we conducted a pilot study in which we assessed current patient awareness of their in‐hospital medications and surveyed attitudes toward increased patient knowledge of hospital medications.

PATIENTS AND METHODS

We conducted a cross‐sectional study of 50 cognitively intact adult internal medicine inpatients at the University of Colorado Hospital, a tertiary‐care academic teaching hospital. This study was part of a larger project designed to examine potential for patient involvement in the medication reconciliation process. A professional research assistant approached eligible patients within 24 hours of admission. To be eligible, patients had to self‐identify as knowing their outpatient medications, speak English, and have been admitted from the community. Nursing home residents and patients with a past medical history of dementia were excluded. Enrollment was tracked during the first half of the study to estimate effect of inclusion/exclusion criteria. Thirty‐eight percent of hospital admissions to medicine services were excluded based on the specified criteria. Thirty‐four percent of eligible patients were approached and 50% of approached patients agreed to participate in the study. Patient knowledge of their outpatient medication regimen was compared to admitting physician medication reconciliation to assess accuracy of patient self‐report of outpatient medication knowledge.

After consenting to participate, study patients completed a structured list of their outpatient medications and a survey of attitudes about being shown their in‐hospital medications, hospital medication errors, and patient involvement in hospital safety. They then completed a list of the medications they believed to be prescribed to them in the hospital.

The primary outcomes were the proportions of as needed (PRN), scheduled, and total hospital medications omitted by the patient, compared to the inpatient medication administration record (MAR) (patient errors of omission). Secondary outcomes included the number of in‐hospital medications listed by the patient that did not appear on the inpatient MAR (patient errors of commission), as well as patient attitudes measured on a 5‐point Likert scale (1 indicated strongly disagree and 5 indicated strongly agree.) Descriptive data included age, race, gender, and number of inpatient medications prescribed. Separate analysis of variance (ANOVA) models provided mean estimates of the primary outcomes and tested differences according to each of the patient characteristics: age in years (<65 or 65), self‐reported knowledge of hospital medications, and self‐reported desire to be involved in medication safety. Similar ANOVA models adjusted for number of medications were also examined to determine whether the relationship between the primary outcomes according to patient characteristics were altered by the number of medications. The protocol was approved by the Colorado Multiple Institutional Review Board.

RESULTS

Participants averaged 54 years of age (standard deviation [SD] = 17, range = 21‐89). Forty‐six percent (23/50) were male, and 74% (37/50) were non‐Hispanic white. Using a structured, patient‐completed, outpatient medication list, patients in the study were on an average of 5.3 outpatient prescription medications (range = 0‐17), 2.2 over‐the‐counter medications (range = 0‐8), and 0.2 herbal medications (range = 0‐7). The admitting physician's medication reconciliation list demonstrated similar number of outpatient prescription medications (average = 5.7) to the patient‐generated list. Fifty‐four percent of patient‐completed home medication lists included all of the prescription medications on the physician's medication reconciliation at admission. According to the inpatient MAR, study patients were prescribed an average of 11.3 scheduled and PRN hospital medications (range = 2‐26) at time of study enrollment.

Patient Knowledge of Their Hospital Medication List

Ninety‐six percent (48/50) of study patients omitted 1 or more of their hospital medications. On average, patients omitted 6.8 medications (range = 0‐22) (Table 1). Among scheduled medications, patients most commonly omitted antibiotics (17%), cardiovascular medications (16%), and antithrombotics (15%) (Figure 1). Among PRN medications, patients most commonly omitted analgesics (33%) and gastrointestinal medications (29%) (Figure 2).

Table 1.Patient Knowledge of Their Hospital Medications List

	Total Medications	Scheduled Medications	PRN Medications
NOTE: n = 50 patients. Abbreviations: CI, confidence interval; PRN, as needed.
Percent of patients with at least 1 hospital medication they could not name (95% CI)	96% (90‐100%)	94% (87‐100%)	80% (69‐92%)
Average number of hospital medications omitted by patient (range)	6.8 (0‐22)	5.2 (0‐15)	1.6 (0‐7)
Percentage of hospital medications omitted by patient (95% CI)	60% (52‐67%)	60% (52‐67%)	68% (57‐78%)

Figure 1

Figure 2

Patients less than 65 years omitted 60% of their PRN medications whereas patients greater than 65 years omitted 88% (P = 0.01). This difference remained even after adjustment for number of medications. There were no significant differences, based on age, in ability to name scheduled or total medications. Forty‐four percent of patients (22/50) believed they were receiving a medication in the hospital that was not actually prescribed.

DISCUSSION

Overall, patients in the study were able to name fewer than one‐half of their hospital medications. Our study suggests that adult medicine inpatients believe learning about their hospital medications would increase their satisfaction and has potential to promote medication safety. At the same time, patients did not know many of their hospital medications and this would limit their ability to fully participate in the medication safety process. Study patients frequently committed both errors of omission (ie, they did not know which medications were prescribed), and errors of commission (ie, they believed they were prescribed medications that were not prescribed). Younger patients were aware of more of their PRN medications than older patients, potentially reflecting greater patient care involvement in younger generations. However, study patients, regardless of age, were able to name fewer than one‐half of their PRN hospital medications. The most common scheduled hospital medications that patients were unable to name come from medication classes which can be associated with significant adverse events, including antibiotics, cardiovascular medications, and antithrombotics.

We posit that without systematically educating patients about their hospital medications, significant deficits in patient knowledge are inevitable. Some might argue that patients should not be asked to know their hospital medications or identify medication errors while sick and vulnerable. Certainly with multiple medication changes, formulary substitutions, and frequent modifications based on changes in clinical status, inpatient medication education could be time consuming and potentially introduce patient confusion or anxiety. Incorrect patient feedback could have potential to introduce new errors. An educational program might use graded participation based on patient interest and ability. Models for this exist in the literature, even extending to patient medication self‐administration.57 In our sample of inpatients, the majority desired a more active role in learning about their hospital medications and believed that their involvement might prevent hospital medication errors from occurring.

Medication literacy, education, and active patient involvement in medication monitoring as a means to improve patient outcomes has received significant attention in the outpatient setting, with lessons applicable to the hospital.8, 9 More broadly, the Joint Commission has established a Hospital National Patient Safety Goal to encourage patients' active involvement in their own care as a patient safety strategy.10 Examples set forth by the Joint Commission include involving patients in infection control measures, marking of procedural sites, and reporting of safety concerns relating to treatment.

While this study identifies patient knowledge deficit as a barrier to utilizing patients as part of the hospital medication safety process, it does not test whether reducing this knowledge deficit would actually reduce medication error. Our study population was limited to cognitively intact adult medicine patients at a single institution, limiting the generalizability of our conclusions. Our enrollment process may have resulted in a study population with less serious illness, greater knowledge of their hospital medications, and greater interest in participating in medication safety potentially overestimating patient knowledge of hospital medications. Finally, our small sample size limits the power to find differences in study comparisons.

Our findings are striking in that we found significant deficits in patient understanding of their hospital medications even among patients who believed they knew, or desired to know, what is being prescribed to them in the hospital. Without a system to incorporate the patient into hospital medication management, these patients will be disenfranchised from participating in inpatient medication safety. These results are a call to reexamine how we educate and involve patients regarding hospital medications. Mechanisms to allow patients to provide feedback to the medical team on their hospital medications might identify errors or improve patient satisfaction with their care. However, the systems and cultural changes needed to provide education on inpatient medications are considerable. Future research is needed to determine if increasing patient knowledge regarding their hospital medications would reduce medication errors in the inpatient setting and how this could be effectively implemented.

Inpatient medication errors represent an important patient safety issue. The magnitude of the problem is staggering, with 1 review finding almost 1 in every 5 medication doses in error, with 7% having potential for adverse drug events.1 While mistakes made at the ordering stage are frequently intercepted by pharmacist or nursing review, administration errors are particularly difficult to prevent.2 The patient, as the last link in the medication administration chain, represents the final individual capable of preventing an incorrect medication administration. It is perhaps surprising then that patients generally lack a formal role in detecting and preventing adverse medication administration events.3

There have been some ambitious attempts to improve patient education regarding hospital medications and involve selected patients in the medication administration process. Such initiatives may result in increased patient participation and satisfaction.47 There is also potential that increased patient knowledge of their hospital medications could promote the goal of medication safety, as the actively involved patient may be able to catch medication errors in the hospital.

Knowledge of prescribed medications is a prerequisite to patient involvement in prevention of inpatient medication errors and yet there is little research on patient knowledge of their hospital medications. Furthermore, as the experience of hospitalization may be disorienting and disempowering for patients, it remains to be seen if patient attitudes toward participation in inpatient medication safety are favorable. To that end, we conducted a pilot study in which we assessed current patient awareness of their in‐hospital medications and surveyed attitudes toward increased patient knowledge of hospital medications.

PATIENTS AND METHODS

We conducted a cross‐sectional study of 50 cognitively intact adult internal medicine inpatients at the University of Colorado Hospital, a tertiary‐care academic teaching hospital. This study was part of a larger project designed to examine potential for patient involvement in the medication reconciliation process. A professional research assistant approached eligible patients within 24 hours of admission. To be eligible, patients had to self‐identify as knowing their outpatient medications, speak English, and have been admitted from the community. Nursing home residents and patients with a past medical history of dementia were excluded. Enrollment was tracked during the first half of the study to estimate effect of inclusion/exclusion criteria. Thirty‐eight percent of hospital admissions to medicine services were excluded based on the specified criteria. Thirty‐four percent of eligible patients were approached and 50% of approached patients agreed to participate in the study. Patient knowledge of their outpatient medication regimen was compared to admitting physician medication reconciliation to assess accuracy of patient self‐report of outpatient medication knowledge.

After consenting to participate, study patients completed a structured list of their outpatient medications and a survey of attitudes about being shown their in‐hospital medications, hospital medication errors, and patient involvement in hospital safety. They then completed a list of the medications they believed to be prescribed to them in the hospital.

The primary outcomes were the proportions of as needed (PRN), scheduled, and total hospital medications omitted by the patient, compared to the inpatient medication administration record (MAR) (patient errors of omission). Secondary outcomes included the number of in‐hospital medications listed by the patient that did not appear on the inpatient MAR (patient errors of commission), as well as patient attitudes measured on a 5‐point Likert scale (1 indicated strongly disagree and 5 indicated strongly agree.) Descriptive data included age, race, gender, and number of inpatient medications prescribed. Separate analysis of variance (ANOVA) models provided mean estimates of the primary outcomes and tested differences according to each of the patient characteristics: age in years (<65 or 65), self‐reported knowledge of hospital medications, and self‐reported desire to be involved in medication safety. Similar ANOVA models adjusted for number of medications were also examined to determine whether the relationship between the primary outcomes according to patient characteristics were altered by the number of medications. The protocol was approved by the Colorado Multiple Institutional Review Board.

RESULTS

Participants averaged 54 years of age (standard deviation [SD] = 17, range = 21‐89). Forty‐six percent (23/50) were male, and 74% (37/50) were non‐Hispanic white. Using a structured, patient‐completed, outpatient medication list, patients in the study were on an average of 5.3 outpatient prescription medications (range = 0‐17), 2.2 over‐the‐counter medications (range = 0‐8), and 0.2 herbal medications (range = 0‐7). The admitting physician's medication reconciliation list demonstrated similar number of outpatient prescription medications (average = 5.7) to the patient‐generated list. Fifty‐four percent of patient‐completed home medication lists included all of the prescription medications on the physician's medication reconciliation at admission. According to the inpatient MAR, study patients were prescribed an average of 11.3 scheduled and PRN hospital medications (range = 2‐26) at time of study enrollment.

Patient Knowledge of Their Hospital Medication List

Ninety‐six percent (48/50) of study patients omitted 1 or more of their hospital medications. On average, patients omitted 6.8 medications (range = 0‐22) (Table 1). Among scheduled medications, patients most commonly omitted antibiotics (17%), cardiovascular medications (16%), and antithrombotics (15%) (Figure 1). Among PRN medications, patients most commonly omitted analgesics (33%) and gastrointestinal medications (29%) (Figure 2).

Table 1.Patient Knowledge of Their Hospital Medications List

	Total Medications	Scheduled Medications	PRN Medications
NOTE: n = 50 patients. Abbreviations: CI, confidence interval; PRN, as needed.
Percent of patients with at least 1 hospital medication they could not name (95% CI)	96% (90‐100%)	94% (87‐100%)	80% (69‐92%)
Average number of hospital medications omitted by patient (range)	6.8 (0‐22)	5.2 (0‐15)	1.6 (0‐7)
Percentage of hospital medications omitted by patient (95% CI)	60% (52‐67%)	60% (52‐67%)	68% (57‐78%)

Figure 1

Figure 2

Patients less than 65 years omitted 60% of their PRN medications whereas patients greater than 65 years omitted 88% (P = 0.01). This difference remained even after adjustment for number of medications. There were no significant differences, based on age, in ability to name scheduled or total medications. Forty‐four percent of patients (22/50) believed they were receiving a medication in the hospital that was not actually prescribed.

DISCUSSION

Overall, patients in the study were able to name fewer than one‐half of their hospital medications. Our study suggests that adult medicine inpatients believe learning about their hospital medications would increase their satisfaction and has potential to promote medication safety. At the same time, patients did not know many of their hospital medications and this would limit their ability to fully participate in the medication safety process. Study patients frequently committed both errors of omission (ie, they did not know which medications were prescribed), and errors of commission (ie, they believed they were prescribed medications that were not prescribed). Younger patients were aware of more of their PRN medications than older patients, potentially reflecting greater patient care involvement in younger generations. However, study patients, regardless of age, were able to name fewer than one‐half of their PRN hospital medications. The most common scheduled hospital medications that patients were unable to name come from medication classes which can be associated with significant adverse events, including antibiotics, cardiovascular medications, and antithrombotics.

We posit that without systematically educating patients about their hospital medications, significant deficits in patient knowledge are inevitable. Some might argue that patients should not be asked to know their hospital medications or identify medication errors while sick and vulnerable. Certainly with multiple medication changes, formulary substitutions, and frequent modifications based on changes in clinical status, inpatient medication education could be time consuming and potentially introduce patient confusion or anxiety. Incorrect patient feedback could have potential to introduce new errors. An educational program might use graded participation based on patient interest and ability. Models for this exist in the literature, even extending to patient medication self‐administration.57 In our sample of inpatients, the majority desired a more active role in learning about their hospital medications and believed that their involvement might prevent hospital medication errors from occurring.

Medication literacy, education, and active patient involvement in medication monitoring as a means to improve patient outcomes has received significant attention in the outpatient setting, with lessons applicable to the hospital.8, 9 More broadly, the Joint Commission has established a Hospital National Patient Safety Goal to encourage patients' active involvement in their own care as a patient safety strategy.10 Examples set forth by the Joint Commission include involving patients in infection control measures, marking of procedural sites, and reporting of safety concerns relating to treatment.

While this study identifies patient knowledge deficit as a barrier to utilizing patients as part of the hospital medication safety process, it does not test whether reducing this knowledge deficit would actually reduce medication error. Our study population was limited to cognitively intact adult medicine patients at a single institution, limiting the generalizability of our conclusions. Our enrollment process may have resulted in a study population with less serious illness, greater knowledge of their hospital medications, and greater interest in participating in medication safety potentially overestimating patient knowledge of hospital medications. Finally, our small sample size limits the power to find differences in study comparisons.

Our findings are striking in that we found significant deficits in patient understanding of their hospital medications even among patients who believed they knew, or desired to know, what is being prescribed to them in the hospital. Without a system to incorporate the patient into hospital medication management, these patients will be disenfranchised from participating in inpatient medication safety. These results are a call to reexamine how we educate and involve patients regarding hospital medications. Mechanisms to allow patients to provide feedback to the medical team on their hospital medications might identify errors or improve patient satisfaction with their care. However, the systems and cultural changes needed to provide education on inpatient medications are considerable. Future research is needed to determine if increasing patient knowledge regarding their hospital medications would reduce medication errors in the inpatient setting and how this could be effectively implemented.

In the evaluation of patients presenting with complaints referable to the chest, the chest radiograph (CXR) is an important and almost universal component of the initial assessment.

Chest radiography is normally performed with both posterior‐anterior (PA) and lateral projections.1 The lateral projection is generally accepted as an indispensable component because it allows better visualization of certain structures including the lower lobes, areas of which are partially obscured by the heart or hemidiaphragms on the PA projection. As such, some radiographic findings are only apparent on the lateral projection. As well, when an abnormality is discovered on the PA projection, the orthogonal orientation of the lateral projection often allows lesion localization.

Together with information gleaned from a thorough history and physical examination, the results of chest radiography often inform initial management when a diagnosis has been established, and the need for additional investigations when the diagnosis remains in question. In the hospital setting, the CXR is often reviewed first by physicians who are not radiologists (eg, internists, emergency physicians, and trainees at various stages of training) when evaluating a patient.

We undertook the current study to investigate the test characteristics (sensitivity, specificity, and likelihood ratio [LR]), and precision of 1 particular finding on lateral chest radiography as interpreted by nonradiologist physicians in the hospital setting. On a normal lateral CXR, one should observe progressive superior‐inferior vertebral radiolucency (Figure 1A). Observed opacity overlying the vertebral column obscuring this progression is usually abnormal and suggestive of pathology in the lower lobes of the lungs or associated structures (Figure 1B). A review of the literature yielded only 1 study of this finding,2 which used a case‐control design and lacked a true gold standard investigation necessary for calculation of meaningful test characteristics. In fact, few studies have compared findings on chest radiography with more definitive investigations,3, 4 and none have examined the predictive value of this finding by nonradiologist observers using a reference standard investigation such as computed tomography (CT) of the chest.

Methods

The radiology Picture Archiving and Communication System (PACS) used at our institution allows us to search for exams by date and study type. We retrospectively identified all patients seen at 1 of 3 university‐affiliated tertiary care adult teaching hospitals (Toronto General, Toronto Western, and Mount Sinai Hospitals) within an 8‐month period (January 1, 2006 to August 31, 2006) who underwent a 2‐view CXR (PA and lateral views). (Note that in this study, the terms radiograph, x‐ray, and plain film are used synonymously.) We then determined which of these patients had a subsequent CT within 24 hours of the x‐ray, resulting in a sample of 370 patients for this study. These patients primarily included patients presenting to the emergency department, and inpatients, with a very small number of outpatients. The majority of the index CXRs were performed for chief complaints of dyspnea, chest pain, cough, or for follow‐up of a previous CXR. However, many were simply performed routinely for admission. Patients with prosthetic devices or appliances obscuring the vertebral column were excluded.

After several training sessions by an experienced internist (A.S.D.), 2 authors (D.R.M., M.E.D.) independently reviewed each lateral CXR using standard 17‐inch displays and documented the presence or absence of abnormal radioopacity obscuring the superior to inferior progression of vertebral radiolucency. These 2 authors were fourth‐year medical students at the time the study began and first‐year trainees in internal medicine when it ended. The presence of abnormal opacity overlying the vertebral column was recorded as a positive test while the absence of this finding was recorded as a negative test.

Observed opacity overlying the vertebral column on lateral CXRs was considered abnormal when it did not represent manifestations of normal anatomical structures. However, the finding of opacity overlying the vertebral column of little diagnostic significance, such as prominent pulmonary vessels, degenerative bony changes, or the finding of a tortuous aorta, were considered normal in this study. Corresponding PA CXRs were also available for viewing. In most cases, the authors viewed both the lateral and PA CXRs, reflecting their use in clinical practice. However, in cases of obvious abnormality on the lateral CXR, only that projection was viewed. No clinical information was made available to the observers of the lateral CXR and they were blinded to the results of CT imaging of the chest. All 370 cases were reviewed by both observers (D.R.M. and M.E.D.). For the purpose of calculating test characteristics and LRs, cases of disagreement between the 3 lateral CXR observers were resolved by independent review by a third author (A.S.D.), a general internist with over 20 years of experience interpreting the lateral CXR.

A fourth author (M.O.B) reviewed the chest CT reports for each patient and recorded the mention of the presence or absence of various pathologies in the lower lobes of the lungs and associated structures in those reports. No clinical information was made available to this author and he was blinded to the results of lateral CXR. All CT investigations were originally interpreted by a university‐affiliated chest radiology faculty member at the time of the investigation. Table 1 lists all relevant chest CT findings in our sample that were recorded as disease‐positive for the purpose of dichotomizing the results of the reference standard, and enabling calculation of test characteristics (Table 2). Notable chest CT findings that were not recorded as disease‐positive for this purpose included mediastinal lymphadenopathy, subpleural density, lytic vertebral lesions, cystic or emphysematous changes, and pneumothorax. Dependent atelectasis was included within the disease‐positive category, though some cases may not have been pathological. It should be pointed out that there may be some variation in terminology used between staff radiologists (eg, reticulation by one radiologist may be called minor densities by another radiologist).

Using the chest CT report as the reference standard for abnormal opacity overlying the vertebral column on lateral chest radiography, we calculated the sensitivity, specificity, and positive and negative LRs (LR+ and LR, respectively) with 95% confidence intervals (CIs) for individual and summary CT‐documented pathologies.5 For this purpose, we constructed a 2 2 table (Table 2) for summary CT‐documented abnormal findings, in which patients with any abnormal CT finding were considered disease‐positive and compared with patients whose CTs were interpreted as normal, considered disease‐negative. We also constructed 2 2 tables for each of the individual CT‐documented pathologies using data from Table 1, in which only the patients with the abnormal CT finding of interest (eg, consolidation) were considered disease‐positive and compared with patients whose CTs were interpreted as normal, considered disease‐negative. In this case, patients with abnormal CT findings (eg, atelectasis, effusion) other than the finding of interest were excluded from the analysis. This secondary analysis is an attempt to estimate the variability of the accuracy of the finding in question across different diagnoses, and not to derive precise estimates of LRs given the small sample sizes for some individual findings.

Of the 370 original patients, we selected a sample of 100 patients by random number assignment whose lateral CXRs were reviewed a second time by the same observers to quantify intraobserver variability. Interobserver variability was quantified by comparing the data of the 2 independent lateral CXR observers on all 370 patients. In both cases, we calculated simple agreement and kappa statistics as measures of precision.6 Our chest CT observer also identified a sample of 10 CT investigations by random number assignment and reviewed the images in a blinded fashion to quantify interobserver variability in CT findings (ie, a comparison of the original CT report with our chest CT observer's interpretation).

We obtained approval from the relevant research ethics boards for the hospitals in which our study population was identified and have endeavored to comply with the Standards for Reporting of Diagnostic Accuracy (STARD) initiative.7 All statistical analyses were performed using R version 2.018 (Free Software Foundation, Boston, MA) and WinBUGS version 1.4. (MRC Biostatistics Unit, Cambridge, UK)9

Results

The identified study sample of 370 patients was 52% male and had an average age of 58 17 years (range, 18 to 96 years). Of the 370 patients, 81 (21.9%) were found to have a normal chest CT, 118 (31.9%) had a single CT finding in the lower lobes designated as disease‐positive, and 171 (46.2%) had 2 or more lower‐lobe CT findings. Overall, 78.1% had 1 or more CT findings considered disease‐positive.

Abnormal opacity overlying the vertebral column on lateral chest radiography had a sensitivity of 86.9% (95% CI, 82.5%‐90.3%) and specificity of 70.4% (95% CI, 59.7%‐79.2%) for CT‐documented lower‐lobe and associated structural pathology (Table 2). The summary LR+ for abnormal opacity overlying the vertebral column on lateral chest radiography was 2.9 (95% CI, 2.1‐4.1) and the summary LR for the absence of this finding was 0.19 (95% CI, 0.13‐0.26). LRs for individual CT‐documented pathologies were very similar to the summary LRs, with a range for LR+s between 2.8 and 3.4, and a range for LRs between 0 and 0.26 (Table 1).

Intraobserver simple agreement and kappa statistics for each of the lateral CXR observers were 79% ( = 0.56) and 81% ( = 0.58), respectively. Interobserver simple agreement between the lateral CXR observers, as well as the associated kappa statistic, were similar at 77% ( = 0.52). Compared with the original chest CT reports generated by university‐affiliated radiology faculty members, the blinded review of 10 randomly‐identified CT investigations by our chest CT observer (M.O.B.) yielded 100% agreement.

Discussion

This study fills a gap in the literature by providing evidence of the accuracy and precision of a particular finding on lateral chest radiography: namely, observed radioopacity obscuring the normal succession of superior‐inferior vertebral radiolucency.

Our investigation of this finding's test characteristics reveal that abnormal opacity overlying the vertebral column on lateral chest radiography is a more sensitive than specific finding, and thus in general more useful for ruling out the presence of disease than ruling it in. But it is our calculated LRs that allow application of this finding's predictive value to clinical scenarios in practice.

LRs are a powerful method of applying new information to the pretest probability of disease, to arrive at the posttest probability. If the summary point estimate LRs of our study are applied to a hypothetical pretest probability of 50% for any CT‐documented pathology, abnormal opacity overlying the vertebral column (LR+ 2.9) gives a posttest probability of 75%, and the absence of this finding (LR 0.19) gives a posttest probability of 16%. In some cases, these posttest probabilities may be high enough to stop investigating and start treating, or low enough to stop investigating.

We also calculated LRs for each subgroup of CT‐documented pathology by comparing only patients with the CT finding of interest and patients with CTs interpreted as normal. While the validity of these calculations is compromised by ignoring the patients in the other subgroups of diagnoses in the calculation, the stability of these LR estimates suggests that the finding and summary LRs can be used for a variety of diagnoses. The individual LRs, however, should not be used in arriving at posttest probabilities of individual pathologies.

Our calculated kappa statistics, a measure of chance‐corrected agreement, quantified the precision of abnormal opacity overlying the vertebral column noted by nonradiologist observers. The kappa statistics associated with intraobserver and interobserver variability for abnormal opacity overlying the vertebral column are indicative of moderate agreement, which is similar to the precision of many other investigational findings in common usage.

This study does have some limitations related to its design. First, CT was used as the gold standard in this study. Ideally, a combination of CT and more invasive measures such as lung biopsy would have been used; however, for ethical and logistical reasons this was obviously not possible. Second, when designing the study we had to decide whether or not to repeat the interpretation of CT images with observers we could ensure were blinded to the corresponding CXRs. We chose not to repeat the interpretation of CT images, and instead used the report of the staff chest radiologist who read the imaging study at the time it was performed. The person reviewing the report of the CT was blinded to the CXR. Our reasons for not rereading each of the CT images with a blinded study radiologist are as follows. First, the chest radiologists who reviewed the CT images at the time they were done were completely unaware of our hypothesis regarding the utility of the lateral CXR (our study took place after the CTs were interpreted). Second, the radiologists tell us that when they interpret CTs they rarely rely on findings in the CXR to help with those interpretations. For these 2 reasons, the original interpretation is very close to complete blinding. In addition, the individuals who interpret and write reports on chest CTs are all expert staff radiologists with considerable experience in this area. A study radiologist (likely a radiology resident) would not have been as proficient. Finally, in performing any study one must weigh the costs with the benefits of any methodological decision, reinterpretation of 370 chest CTs would have required an enormous amount of time. Finally, our small sample of 10 comparing official reports to the reinterpretation of the scans themselves supported the view that we did not need to review all 370 cases again.

Approximately three‐quarters of our study population was found to have CT‐documented disease. However, this is not surprising given our method of patient selection. Because the sample was collected from clinical practice, it is likely that only patients who exhibited a finding on the CXR that required delineation went on to have the reference standard investigation (CT). This study is therefore subject to workup bias. Workup bias in this scenario could work in 1 of 2 directions. In one situation, some patients would have a clear pathology or diagnosis based on the CXR, such that a CT was unnecessary and therefore not performed. In this case, our study would have underestimated the sensitivity of the sign being studied because a group of true positives would have been left out of the sample. In the second situation, patients with true pathology and a normal CXR (false negatives) fail to undergo CT. In this case, our study would have overestimated the sensitivity. We are not sure which effect of workup bias predominates in the study, but in either case an independent, prospective comparison of these imaging modalities in all patients who had CXRs was not feasible for ethical reasons. If we were to apply the reference standard investigation to all those patients, the potential for harm from excess radiation10 would be too great. As such, our cohort of patients is the best possible sample that can be studied.

Another feature of this study is that it intentionally used nonradiologist (budding internist) interpreters of the lateral CXRs, thus defining its generalizability. We did so for 2 reasons. First, the sign studied is likely too basic to be of relevance to radiologists. Second, it is intended to be used by internists, emergency physicians, and nonradiology trainees at all levels, who are required to make initial treatment decisions based on their preliminary interpretation of x‐rays, particularly in the hospital setting. Therefore, we decided our results would be more externally valid and applicable if the interpreters of the x‐rays and use of the x‐ray sign in this study was by trainees.

Abnormal opacity overlying the vertebral column on lateral chest radiography is a clinically useful finding that can help nonradiologist physicians determine initial management or the need for further investigation when diagnostic uncertainty remains. This study provides evidence that this finding is both reliable and useful for ruling the presence of lower‐lobe and associated structural pathology out, and somewhat useful for ruling the presence of such pathology in.

In the evaluation of patients presenting with complaints referable to the chest, the chest radiograph (CXR) is an important and almost universal component of the initial assessment.

Chest radiography is normally performed with both posterior‐anterior (PA) and lateral projections.1 The lateral projection is generally accepted as an indispensable component because it allows better visualization of certain structures including the lower lobes, areas of which are partially obscured by the heart or hemidiaphragms on the PA projection. As such, some radiographic findings are only apparent on the lateral projection. As well, when an abnormality is discovered on the PA projection, the orthogonal orientation of the lateral projection often allows lesion localization.

Together with information gleaned from a thorough history and physical examination, the results of chest radiography often inform initial management when a diagnosis has been established, and the need for additional investigations when the diagnosis remains in question. In the hospital setting, the CXR is often reviewed first by physicians who are not radiologists (eg, internists, emergency physicians, and trainees at various stages of training) when evaluating a patient.

We undertook the current study to investigate the test characteristics (sensitivity, specificity, and likelihood ratio [LR]), and precision of 1 particular finding on lateral chest radiography as interpreted by nonradiologist physicians in the hospital setting. On a normal lateral CXR, one should observe progressive superior‐inferior vertebral radiolucency (Figure 1A). Observed opacity overlying the vertebral column obscuring this progression is usually abnormal and suggestive of pathology in the lower lobes of the lungs or associated structures (Figure 1B). A review of the literature yielded only 1 study of this finding,2 which used a case‐control design and lacked a true gold standard investigation necessary for calculation of meaningful test characteristics. In fact, few studies have compared findings on chest radiography with more definitive investigations,3, 4 and none have examined the predictive value of this finding by nonradiologist observers using a reference standard investigation such as computed tomography (CT) of the chest.

Methods

The radiology Picture Archiving and Communication System (PACS) used at our institution allows us to search for exams by date and study type. We retrospectively identified all patients seen at 1 of 3 university‐affiliated tertiary care adult teaching hospitals (Toronto General, Toronto Western, and Mount Sinai Hospitals) within an 8‐month period (January 1, 2006 to August 31, 2006) who underwent a 2‐view CXR (PA and lateral views). (Note that in this study, the terms radiograph, x‐ray, and plain film are used synonymously.) We then determined which of these patients had a subsequent CT within 24 hours of the x‐ray, resulting in a sample of 370 patients for this study. These patients primarily included patients presenting to the emergency department, and inpatients, with a very small number of outpatients. The majority of the index CXRs were performed for chief complaints of dyspnea, chest pain, cough, or for follow‐up of a previous CXR. However, many were simply performed routinely for admission. Patients with prosthetic devices or appliances obscuring the vertebral column were excluded.

After several training sessions by an experienced internist (A.S.D.), 2 authors (D.R.M., M.E.D.) independently reviewed each lateral CXR using standard 17‐inch displays and documented the presence or absence of abnormal radioopacity obscuring the superior to inferior progression of vertebral radiolucency. These 2 authors were fourth‐year medical students at the time the study began and first‐year trainees in internal medicine when it ended. The presence of abnormal opacity overlying the vertebral column was recorded as a positive test while the absence of this finding was recorded as a negative test.

Observed opacity overlying the vertebral column on lateral CXRs was considered abnormal when it did not represent manifestations of normal anatomical structures. However, the finding of opacity overlying the vertebral column of little diagnostic significance, such as prominent pulmonary vessels, degenerative bony changes, or the finding of a tortuous aorta, were considered normal in this study. Corresponding PA CXRs were also available for viewing. In most cases, the authors viewed both the lateral and PA CXRs, reflecting their use in clinical practice. However, in cases of obvious abnormality on the lateral CXR, only that projection was viewed. No clinical information was made available to the observers of the lateral CXR and they were blinded to the results of CT imaging of the chest. All 370 cases were reviewed by both observers (D.R.M. and M.E.D.). For the purpose of calculating test characteristics and LRs, cases of disagreement between the 3 lateral CXR observers were resolved by independent review by a third author (A.S.D.), a general internist with over 20 years of experience interpreting the lateral CXR.

A fourth author (M.O.B) reviewed the chest CT reports for each patient and recorded the mention of the presence or absence of various pathologies in the lower lobes of the lungs and associated structures in those reports. No clinical information was made available to this author and he was blinded to the results of lateral CXR. All CT investigations were originally interpreted by a university‐affiliated chest radiology faculty member at the time of the investigation. Table 1 lists all relevant chest CT findings in our sample that were recorded as disease‐positive for the purpose of dichotomizing the results of the reference standard, and enabling calculation of test characteristics (Table 2). Notable chest CT findings that were not recorded as disease‐positive for this purpose included mediastinal lymphadenopathy, subpleural density, lytic vertebral lesions, cystic or emphysematous changes, and pneumothorax. Dependent atelectasis was included within the disease‐positive category, though some cases may not have been pathological. It should be pointed out that there may be some variation in terminology used between staff radiologists (eg, reticulation by one radiologist may be called minor densities by another radiologist).

Using the chest CT report as the reference standard for abnormal opacity overlying the vertebral column on lateral chest radiography, we calculated the sensitivity, specificity, and positive and negative LRs (LR+ and LR, respectively) with 95% confidence intervals (CIs) for individual and summary CT‐documented pathologies.5 For this purpose, we constructed a 2 2 table (Table 2) for summary CT‐documented abnormal findings, in which patients with any abnormal CT finding were considered disease‐positive and compared with patients whose CTs were interpreted as normal, considered disease‐negative. We also constructed 2 2 tables for each of the individual CT‐documented pathologies using data from Table 1, in which only the patients with the abnormal CT finding of interest (eg, consolidation) were considered disease‐positive and compared with patients whose CTs were interpreted as normal, considered disease‐negative. In this case, patients with abnormal CT findings (eg, atelectasis, effusion) other than the finding of interest were excluded from the analysis. This secondary analysis is an attempt to estimate the variability of the accuracy of the finding in question across different diagnoses, and not to derive precise estimates of LRs given the small sample sizes for some individual findings.

Of the 370 original patients, we selected a sample of 100 patients by random number assignment whose lateral CXRs were reviewed a second time by the same observers to quantify intraobserver variability. Interobserver variability was quantified by comparing the data of the 2 independent lateral CXR observers on all 370 patients. In both cases, we calculated simple agreement and kappa statistics as measures of precision.6 Our chest CT observer also identified a sample of 10 CT investigations by random number assignment and reviewed the images in a blinded fashion to quantify interobserver variability in CT findings (ie, a comparison of the original CT report with our chest CT observer's interpretation).

We obtained approval from the relevant research ethics boards for the hospitals in which our study population was identified and have endeavored to comply with the Standards for Reporting of Diagnostic Accuracy (STARD) initiative.7 All statistical analyses were performed using R version 2.018 (Free Software Foundation, Boston, MA) and WinBUGS version 1.4. (MRC Biostatistics Unit, Cambridge, UK)9

Results

The identified study sample of 370 patients was 52% male and had an average age of 58 17 years (range, 18 to 96 years). Of the 370 patients, 81 (21.9%) were found to have a normal chest CT, 118 (31.9%) had a single CT finding in the lower lobes designated as disease‐positive, and 171 (46.2%) had 2 or more lower‐lobe CT findings. Overall, 78.1% had 1 or more CT findings considered disease‐positive.

Abnormal opacity overlying the vertebral column on lateral chest radiography had a sensitivity of 86.9% (95% CI, 82.5%‐90.3%) and specificity of 70.4% (95% CI, 59.7%‐79.2%) for CT‐documented lower‐lobe and associated structural pathology (Table 2). The summary LR+ for abnormal opacity overlying the vertebral column on lateral chest radiography was 2.9 (95% CI, 2.1‐4.1) and the summary LR for the absence of this finding was 0.19 (95% CI, 0.13‐0.26). LRs for individual CT‐documented pathologies were very similar to the summary LRs, with a range for LR+s between 2.8 and 3.4, and a range for LRs between 0 and 0.26 (Table 1).

Intraobserver simple agreement and kappa statistics for each of the lateral CXR observers were 79% ( = 0.56) and 81% ( = 0.58), respectively. Interobserver simple agreement between the lateral CXR observers, as well as the associated kappa statistic, were similar at 77% ( = 0.52). Compared with the original chest CT reports generated by university‐affiliated radiology faculty members, the blinded review of 10 randomly‐identified CT investigations by our chest CT observer (M.O.B.) yielded 100% agreement.

Discussion

This study fills a gap in the literature by providing evidence of the accuracy and precision of a particular finding on lateral chest radiography: namely, observed radioopacity obscuring the normal succession of superior‐inferior vertebral radiolucency.

Our investigation of this finding's test characteristics reveal that abnormal opacity overlying the vertebral column on lateral chest radiography is a more sensitive than specific finding, and thus in general more useful for ruling out the presence of disease than ruling it in. But it is our calculated LRs that allow application of this finding's predictive value to clinical scenarios in practice.

LRs are a powerful method of applying new information to the pretest probability of disease, to arrive at the posttest probability. If the summary point estimate LRs of our study are applied to a hypothetical pretest probability of 50% for any CT‐documented pathology, abnormal opacity overlying the vertebral column (LR+ 2.9) gives a posttest probability of 75%, and the absence of this finding (LR 0.19) gives a posttest probability of 16%. In some cases, these posttest probabilities may be high enough to stop investigating and start treating, or low enough to stop investigating.

We also calculated LRs for each subgroup of CT‐documented pathology by comparing only patients with the CT finding of interest and patients with CTs interpreted as normal. While the validity of these calculations is compromised by ignoring the patients in the other subgroups of diagnoses in the calculation, the stability of these LR estimates suggests that the finding and summary LRs can be used for a variety of diagnoses. The individual LRs, however, should not be used in arriving at posttest probabilities of individual pathologies.

Our calculated kappa statistics, a measure of chance‐corrected agreement, quantified the precision of abnormal opacity overlying the vertebral column noted by nonradiologist observers. The kappa statistics associated with intraobserver and interobserver variability for abnormal opacity overlying the vertebral column are indicative of moderate agreement, which is similar to the precision of many other investigational findings in common usage.

This study does have some limitations related to its design. First, CT was used as the gold standard in this study. Ideally, a combination of CT and more invasive measures such as lung biopsy would have been used; however, for ethical and logistical reasons this was obviously not possible. Second, when designing the study we had to decide whether or not to repeat the interpretation of CT images with observers we could ensure were blinded to the corresponding CXRs. We chose not to repeat the interpretation of CT images, and instead used the report of the staff chest radiologist who read the imaging study at the time it was performed. The person reviewing the report of the CT was blinded to the CXR. Our reasons for not rereading each of the CT images with a blinded study radiologist are as follows. First, the chest radiologists who reviewed the CT images at the time they were done were completely unaware of our hypothesis regarding the utility of the lateral CXR (our study took place after the CTs were interpreted). Second, the radiologists tell us that when they interpret CTs they rarely rely on findings in the CXR to help with those interpretations. For these 2 reasons, the original interpretation is very close to complete blinding. In addition, the individuals who interpret and write reports on chest CTs are all expert staff radiologists with considerable experience in this area. A study radiologist (likely a radiology resident) would not have been as proficient. Finally, in performing any study one must weigh the costs with the benefits of any methodological decision, reinterpretation of 370 chest CTs would have required an enormous amount of time. Finally, our small sample of 10 comparing official reports to the reinterpretation of the scans themselves supported the view that we did not need to review all 370 cases again.

Approximately three‐quarters of our study population was found to have CT‐documented disease. However, this is not surprising given our method of patient selection. Because the sample was collected from clinical practice, it is likely that only patients who exhibited a finding on the CXR that required delineation went on to have the reference standard investigation (CT). This study is therefore subject to workup bias. Workup bias in this scenario could work in 1 of 2 directions. In one situation, some patients would have a clear pathology or diagnosis based on the CXR, such that a CT was unnecessary and therefore not performed. In this case, our study would have underestimated the sensitivity of the sign being studied because a group of true positives would have been left out of the sample. In the second situation, patients with true pathology and a normal CXR (false negatives) fail to undergo CT. In this case, our study would have overestimated the sensitivity. We are not sure which effect of workup bias predominates in the study, but in either case an independent, prospective comparison of these imaging modalities in all patients who had CXRs was not feasible for ethical reasons. If we were to apply the reference standard investigation to all those patients, the potential for harm from excess radiation10 would be too great. As such, our cohort of patients is the best possible sample that can be studied.

Another feature of this study is that it intentionally used nonradiologist (budding internist) interpreters of the lateral CXRs, thus defining its generalizability. We did so for 2 reasons. First, the sign studied is likely too basic to be of relevance to radiologists. Second, it is intended to be used by internists, emergency physicians, and nonradiology trainees at all levels, who are required to make initial treatment decisions based on their preliminary interpretation of x‐rays, particularly in the hospital setting. Therefore, we decided our results would be more externally valid and applicable if the interpreters of the x‐rays and use of the x‐ray sign in this study was by trainees.

Abnormal opacity overlying the vertebral column on lateral chest radiography is a clinically useful finding that can help nonradiologist physicians determine initial management or the need for further investigation when diagnostic uncertainty remains. This study provides evidence that this finding is both reliable and useful for ruling the presence of lower‐lobe and associated structural pathology out, and somewhat useful for ruling the presence of such pathology in.

In the evaluation of patients presenting with complaints referable to the chest, the chest radiograph (CXR) is an important and almost universal component of the initial assessment.

Chest radiography is normally performed with both posterior‐anterior (PA) and lateral projections.1 The lateral projection is generally accepted as an indispensable component because it allows better visualization of certain structures including the lower lobes, areas of which are partially obscured by the heart or hemidiaphragms on the PA projection. As such, some radiographic findings are only apparent on the lateral projection. As well, when an abnormality is discovered on the PA projection, the orthogonal orientation of the lateral projection often allows lesion localization.

Together with information gleaned from a thorough history and physical examination, the results of chest radiography often inform initial management when a diagnosis has been established, and the need for additional investigations when the diagnosis remains in question. In the hospital setting, the CXR is often reviewed first by physicians who are not radiologists (eg, internists, emergency physicians, and trainees at various stages of training) when evaluating a patient.

We undertook the current study to investigate the test characteristics (sensitivity, specificity, and likelihood ratio [LR]), and precision of 1 particular finding on lateral chest radiography as interpreted by nonradiologist physicians in the hospital setting. On a normal lateral CXR, one should observe progressive superior‐inferior vertebral radiolucency (Figure 1A). Observed opacity overlying the vertebral column obscuring this progression is usually abnormal and suggestive of pathology in the lower lobes of the lungs or associated structures (Figure 1B). A review of the literature yielded only 1 study of this finding,2 which used a case‐control design and lacked a true gold standard investigation necessary for calculation of meaningful test characteristics. In fact, few studies have compared findings on chest radiography with more definitive investigations,3, 4 and none have examined the predictive value of this finding by nonradiologist observers using a reference standard investigation such as computed tomography (CT) of the chest.

Methods

The radiology Picture Archiving and Communication System (PACS) used at our institution allows us to search for exams by date and study type. We retrospectively identified all patients seen at 1 of 3 university‐affiliated tertiary care adult teaching hospitals (Toronto General, Toronto Western, and Mount Sinai Hospitals) within an 8‐month period (January 1, 2006 to August 31, 2006) who underwent a 2‐view CXR (PA and lateral views). (Note that in this study, the terms radiograph, x‐ray, and plain film are used synonymously.) We then determined which of these patients had a subsequent CT within 24 hours of the x‐ray, resulting in a sample of 370 patients for this study. These patients primarily included patients presenting to the emergency department, and inpatients, with a very small number of outpatients. The majority of the index CXRs were performed for chief complaints of dyspnea, chest pain, cough, or for follow‐up of a previous CXR. However, many were simply performed routinely for admission. Patients with prosthetic devices or appliances obscuring the vertebral column were excluded.

After several training sessions by an experienced internist (A.S.D.), 2 authors (D.R.M., M.E.D.) independently reviewed each lateral CXR using standard 17‐inch displays and documented the presence or absence of abnormal radioopacity obscuring the superior to inferior progression of vertebral radiolucency. These 2 authors were fourth‐year medical students at the time the study began and first‐year trainees in internal medicine when it ended. The presence of abnormal opacity overlying the vertebral column was recorded as a positive test while the absence of this finding was recorded as a negative test.

Observed opacity overlying the vertebral column on lateral CXRs was considered abnormal when it did not represent manifestations of normal anatomical structures. However, the finding of opacity overlying the vertebral column of little diagnostic significance, such as prominent pulmonary vessels, degenerative bony changes, or the finding of a tortuous aorta, were considered normal in this study. Corresponding PA CXRs were also available for viewing. In most cases, the authors viewed both the lateral and PA CXRs, reflecting their use in clinical practice. However, in cases of obvious abnormality on the lateral CXR, only that projection was viewed. No clinical information was made available to the observers of the lateral CXR and they were blinded to the results of CT imaging of the chest. All 370 cases were reviewed by both observers (D.R.M. and M.E.D.). For the purpose of calculating test characteristics and LRs, cases of disagreement between the 3 lateral CXR observers were resolved by independent review by a third author (A.S.D.), a general internist with over 20 years of experience interpreting the lateral CXR.

A fourth author (M.O.B) reviewed the chest CT reports for each patient and recorded the mention of the presence or absence of various pathologies in the lower lobes of the lungs and associated structures in those reports. No clinical information was made available to this author and he was blinded to the results of lateral CXR. All CT investigations were originally interpreted by a university‐affiliated chest radiology faculty member at the time of the investigation. Table 1 lists all relevant chest CT findings in our sample that were recorded as disease‐positive for the purpose of dichotomizing the results of the reference standard, and enabling calculation of test characteristics (Table 2). Notable chest CT findings that were not recorded as disease‐positive for this purpose included mediastinal lymphadenopathy, subpleural density, lytic vertebral lesions, cystic or emphysematous changes, and pneumothorax. Dependent atelectasis was included within the disease‐positive category, though some cases may not have been pathological. It should be pointed out that there may be some variation in terminology used between staff radiologists (eg, reticulation by one radiologist may be called minor densities by another radiologist).

Using the chest CT report as the reference standard for abnormal opacity overlying the vertebral column on lateral chest radiography, we calculated the sensitivity, specificity, and positive and negative LRs (LR+ and LR, respectively) with 95% confidence intervals (CIs) for individual and summary CT‐documented pathologies.5 For this purpose, we constructed a 2 2 table (Table 2) for summary CT‐documented abnormal findings, in which patients with any abnormal CT finding were considered disease‐positive and compared with patients whose CTs were interpreted as normal, considered disease‐negative. We also constructed 2 2 tables for each of the individual CT‐documented pathologies using data from Table 1, in which only the patients with the abnormal CT finding of interest (eg, consolidation) were considered disease‐positive and compared with patients whose CTs were interpreted as normal, considered disease‐negative. In this case, patients with abnormal CT findings (eg, atelectasis, effusion) other than the finding of interest were excluded from the analysis. This secondary analysis is an attempt to estimate the variability of the accuracy of the finding in question across different diagnoses, and not to derive precise estimates of LRs given the small sample sizes for some individual findings.

Of the 370 original patients, we selected a sample of 100 patients by random number assignment whose lateral CXRs were reviewed a second time by the same observers to quantify intraobserver variability. Interobserver variability was quantified by comparing the data of the 2 independent lateral CXR observers on all 370 patients. In both cases, we calculated simple agreement and kappa statistics as measures of precision.6 Our chest CT observer also identified a sample of 10 CT investigations by random number assignment and reviewed the images in a blinded fashion to quantify interobserver variability in CT findings (ie, a comparison of the original CT report with our chest CT observer's interpretation).

We obtained approval from the relevant research ethics boards for the hospitals in which our study population was identified and have endeavored to comply with the Standards for Reporting of Diagnostic Accuracy (STARD) initiative.7 All statistical analyses were performed using R version 2.018 (Free Software Foundation, Boston, MA) and WinBUGS version 1.4. (MRC Biostatistics Unit, Cambridge, UK)9

Results

The identified study sample of 370 patients was 52% male and had an average age of 58 17 years (range, 18 to 96 years). Of the 370 patients, 81 (21.9%) were found to have a normal chest CT, 118 (31.9%) had a single CT finding in the lower lobes designated as disease‐positive, and 171 (46.2%) had 2 or more lower‐lobe CT findings. Overall, 78.1% had 1 or more CT findings considered disease‐positive.

Abnormal opacity overlying the vertebral column on lateral chest radiography had a sensitivity of 86.9% (95% CI, 82.5%‐90.3%) and specificity of 70.4% (95% CI, 59.7%‐79.2%) for CT‐documented lower‐lobe and associated structural pathology (Table 2). The summary LR+ for abnormal opacity overlying the vertebral column on lateral chest radiography was 2.9 (95% CI, 2.1‐4.1) and the summary LR for the absence of this finding was 0.19 (95% CI, 0.13‐0.26). LRs for individual CT‐documented pathologies were very similar to the summary LRs, with a range for LR+s between 2.8 and 3.4, and a range for LRs between 0 and 0.26 (Table 1).

Intraobserver simple agreement and kappa statistics for each of the lateral CXR observers were 79% ( = 0.56) and 81% ( = 0.58), respectively. Interobserver simple agreement between the lateral CXR observers, as well as the associated kappa statistic, were similar at 77% ( = 0.52). Compared with the original chest CT reports generated by university‐affiliated radiology faculty members, the blinded review of 10 randomly‐identified CT investigations by our chest CT observer (M.O.B.) yielded 100% agreement.

Discussion

This study fills a gap in the literature by providing evidence of the accuracy and precision of a particular finding on lateral chest radiography: namely, observed radioopacity obscuring the normal succession of superior‐inferior vertebral radiolucency.

Our investigation of this finding's test characteristics reveal that abnormal opacity overlying the vertebral column on lateral chest radiography is a more sensitive than specific finding, and thus in general more useful for ruling out the presence of disease than ruling it in. But it is our calculated LRs that allow application of this finding's predictive value to clinical scenarios in practice.

LRs are a powerful method of applying new information to the pretest probability of disease, to arrive at the posttest probability. If the summary point estimate LRs of our study are applied to a hypothetical pretest probability of 50% for any CT‐documented pathology, abnormal opacity overlying the vertebral column (LR+ 2.9) gives a posttest probability of 75%, and the absence of this finding (LR 0.19) gives a posttest probability of 16%. In some cases, these posttest probabilities may be high enough to stop investigating and start treating, or low enough to stop investigating.

We also calculated LRs for each subgroup of CT‐documented pathology by comparing only patients with the CT finding of interest and patients with CTs interpreted as normal. While the validity of these calculations is compromised by ignoring the patients in the other subgroups of diagnoses in the calculation, the stability of these LR estimates suggests that the finding and summary LRs can be used for a variety of diagnoses. The individual LRs, however, should not be used in arriving at posttest probabilities of individual pathologies.

Our calculated kappa statistics, a measure of chance‐corrected agreement, quantified the precision of abnormal opacity overlying the vertebral column noted by nonradiologist observers. The kappa statistics associated with intraobserver and interobserver variability for abnormal opacity overlying the vertebral column are indicative of moderate agreement, which is similar to the precision of many other investigational findings in common usage.

This study does have some limitations related to its design. First, CT was used as the gold standard in this study. Ideally, a combination of CT and more invasive measures such as lung biopsy would have been used; however, for ethical and logistical reasons this was obviously not possible. Second, when designing the study we had to decide whether or not to repeat the interpretation of CT images with observers we could ensure were blinded to the corresponding CXRs. We chose not to repeat the interpretation of CT images, and instead used the report of the staff chest radiologist who read the imaging study at the time it was performed. The person reviewing the report of the CT was blinded to the CXR. Our reasons for not rereading each of the CT images with a blinded study radiologist are as follows. First, the chest radiologists who reviewed the CT images at the time they were done were completely unaware of our hypothesis regarding the utility of the lateral CXR (our study took place after the CTs were interpreted). Second, the radiologists tell us that when they interpret CTs they rarely rely on findings in the CXR to help with those interpretations. For these 2 reasons, the original interpretation is very close to complete blinding. In addition, the individuals who interpret and write reports on chest CTs are all expert staff radiologists with considerable experience in this area. A study radiologist (likely a radiology resident) would not have been as proficient. Finally, in performing any study one must weigh the costs with the benefits of any methodological decision, reinterpretation of 370 chest CTs would have required an enormous amount of time. Finally, our small sample of 10 comparing official reports to the reinterpretation of the scans themselves supported the view that we did not need to review all 370 cases again.

Approximately three‐quarters of our study population was found to have CT‐documented disease. However, this is not surprising given our method of patient selection. Because the sample was collected from clinical practice, it is likely that only patients who exhibited a finding on the CXR that required delineation went on to have the reference standard investigation (CT). This study is therefore subject to workup bias. Workup bias in this scenario could work in 1 of 2 directions. In one situation, some patients would have a clear pathology or diagnosis based on the CXR, such that a CT was unnecessary and therefore not performed. In this case, our study would have underestimated the sensitivity of the sign being studied because a group of true positives would have been left out of the sample. In the second situation, patients with true pathology and a normal CXR (false negatives) fail to undergo CT. In this case, our study would have overestimated the sensitivity. We are not sure which effect of workup bias predominates in the study, but in either case an independent, prospective comparison of these imaging modalities in all patients who had CXRs was not feasible for ethical reasons. If we were to apply the reference standard investigation to all those patients, the potential for harm from excess radiation10 would be too great. As such, our cohort of patients is the best possible sample that can be studied.

Another feature of this study is that it intentionally used nonradiologist (budding internist) interpreters of the lateral CXRs, thus defining its generalizability. We did so for 2 reasons. First, the sign studied is likely too basic to be of relevance to radiologists. Second, it is intended to be used by internists, emergency physicians, and nonradiology trainees at all levels, who are required to make initial treatment decisions based on their preliminary interpretation of x‐rays, particularly in the hospital setting. Therefore, we decided our results would be more externally valid and applicable if the interpreters of the x‐rays and use of the x‐ray sign in this study was by trainees.

Abnormal opacity overlying the vertebral column on lateral chest radiography is a clinically useful finding that can help nonradiologist physicians determine initial management or the need for further investigation when diagnostic uncertainty remains. This study provides evidence that this finding is both reliable and useful for ruling the presence of lower‐lobe and associated structural pathology out, and somewhat useful for ruling the presence of such pathology in.

The past decade has seen an increase in the number of hospital discharges associated with a diabetes diagnosis.1, 2 Diabetes is the fourth leading comorbid condition associated with any hospital discharge in the United States.3 Nearly one‐third of diabetes patients require 2 or more hospitalizations in any given year,4 and inpatient stays account for the largest proportion of direct medical expenses incurred by persons with the disease.5

The hospital component of diabetes care has been receiving considerable attention. The advantage of effective inpatient diabetes managementwith particular attention to improving glycemic controlis evident for a number of clinical situations (eg, acute myocardial infarction, critically ill patients).68 National and regional organizations,912 and professional societies68, 12 have developed guidelines about management of inpatient hyperglycemia.

Despite increased awareness of the value of treating inpatient hyperglycemia, little is known about glucose control in U.S. hospitals. As hospitals begin to develop programs to improve inpatient glucose management, some method of standardized benchmarking should be put in place. Using information systems solutions to obtain point‐of‐care bedside glucose (POC‐BG) data, we previously reported on inpatient glucose control from a smaller number of U.S. hospitals.13, 14 We now provide data on a larger, more representative number of U.S. hospitals that gives a broader national view of the current status of inpatient glycemic control.

Patients and Methods

Results

Overall Glycemic Control

A total of 12,559,305 POC‐BG measurements (2,935,167 from the ICU and 9,624,138 from the non‐ICU) from 1,010,705 patients with 3,973,460 patient days were analyzed from 126 hospitals. The mean number of measurements was 20 per ICU patient and 9.5 for non‐ICU patients. The average number of measurements taken per patient‐day was 5 for the ICU patient and 3 for the non‐ICU patient.

Hospital hyperglycemia (>180 mg/dL) was 46.0% for ICU and 31.7% for non‐ICU. The patient‐day‐weighted mean POC‐BG for ICU measurements was 165 mg/dL (median = 164 mg/dL, SD 14.5) and 166 mg/dL (median = 167 mg/dL, SD 8) for non‐ICU. The distributions of patient‐day‐weighted mean POC‐BG values for ICU and non‐ICU settings are shown in Figure 1. The range of patient‐day‐weighted mean values was much wider for the ICU (126‐203 mg/dL) than in the non‐ICU (139‐186 mg/dL).

Figure 1

Hyperglycemia Prevalence

Of ICU patients, 60.6% had at least 1 POC‐BG value >180 mg/dL, as did 46.4% of non‐ICU patients. The proportion of patient‐days with a patient‐day‐weighted mean POC‐BG >180 mg/dL was 26.3% in the ICU setting (Figure 2A) and 31.3% in the non‐ICU (Figure 2B); the other cut points are also shown in Figure 2. The prevalence of patient‐days where hyperglycemia was more severe (>300 mg/dL) was low but nonetheless still detected in both the ICU and non‐ICU settings, although these differences appear to be less pronounced than in the ICU.

Figure 2

Hypoglycemia Rates

There were 21.3% of patients who had at least 1 POC‐BG value <70 mg/dL. Hospital hypoglycemia was low in both the ICU and non‐ICU measurement data, although the proportion of patient days with POC‐BG <70 mg/dL was higher in the ICU vs. the non‐ICU setting (Figure 3A,B). Hypoglycemia (<70 mg/dL) was detected in 10.1% of patient‐days (3.2% of all measures) in the ICU setting (Figure 3A) and 3.5% of patient‐days (4.2% of all measures) in the non‐ICU (Figure 3B). Moderate (<60 mg/dL) and more severe (<50 mg/dL and <40 mg/dL) hypoglycemia were very uncommon in both the ICU and non‐ICU.

Figure 3

Relationship of Glucose Control with Hospital Characteristics

There was a significant relationship between the total number of hospital beds and patient‐day‐weighted mean POC‐BG values in the ICU (Figure 4A). In the ICU, hospitals with <200 beds had significantly higher patient‐day‐weighted mean POC‐BG levels than those with 200 to 299 beds (P < 0.05), 300 to 399 beds (P < 0.01), and 400 beds (P < 0.001). Rural hospitals (Figure 4B) also had higher patient‐day‐weighted mean POC‐BG values compared to urban community and academic hospitals (both P < 0.001). Finally, ICUs in hospitals in the West (Figure 4C, bottom panel) had significantly lower values than those in the Midwest and South (both P < 0.01).

Figure 4

Differences in patient‐day‐weighted mean POC‐BG levels based on hospital characteristics were also observed for the non‐ICU (Figure 5), although these differences appear to be less pronounced than in the ICU. Hospitals with <200 beds (Figure 5A) had significantly higher patient‐day‐weighted mean POC‐BG values compared to hospitals with 300 to 399 beds (P < 0.05) and 400 beds (P < 0.001). Rural hospitals (Figure 5B) had significantly higher values than academic (P < 0.05) and urban community (P < 0.001) hospitals, and hospitals in the West (Figure 5C) had significantly lower values than those in the South and Northeast (both P < 0.05).

Figure 5

Discussion

Hospitalizations associated with diabetes pose a substantial burden on the U.S. health system.15 Recent consensus advocates good glucose control in the hospital to optimize outcomes for a number of clinical scenarios.68 Aside from a few institution‐specific studies,2527 the quality of diabetes treatment in U.S. hospitals is mostly unknown, but assessing the level of glycemic control will be a key metric that hospitals will need to track as they implement improvement programs targeting hospital hyperglycemia. Hospitals will need a way not just to track overall glucose levels, but also to monitor whether hypoglycemic events rise as they implement tight glycemic control initiatives. To our knowledge this is the first report on glycemic control from a large number of U.S. hospitals with diverse characteristics and from different geographic regions.

Debate continues as to what glucose targets for inpatients should be attained.28, 29 The overall patient‐day‐weighted mean POC‐BG was 170 mg/dL for the non‐ICU, and only a moderately lower 162 mg/dL in the ICU, despite much lower thresholds for ICU measurements in current suggested guidelines.8, 30 For the average hospital, over one‐third of non‐ICUs had patient‐day‐weighted mean POC‐BG levels that were >180 mg/dL and nearly one‐quarter had values >200 mg/dL. Similarly, nearly 40% of ICUs had patient‐day‐weighted mean POC‐BGs >180 mg/dL and over 30% were >200 mg/dL, indicating room for improvement in hospital ICU glucose control, at least in the hospitals sampled here. The range of patient‐day‐weighted mean POC‐BG levels for the ICU was broader than what was seen in the non‐ICU data, with the ICU data containing lower weighted mean POC‐BG values, and may indicate that hospitals are concentrating their efforts on adopting stricter glucose control measures in their ICUs.

Whether examining data from a single institution,17 from a larger group of hospitals,14 or now from 126 hospitals, one consistent finding has been the low prevalence of hypoglycemiaparticularly severe hypoglycemia (glucose <50 mg/dL or <40 mg/dL). Based on this larger sampling, however, hypoglycemia in the ICU, while still uncommon with respect to hyperglycemia, is more than double that of the non‐ICU. Fear of hypoglycemia is frequently mentioned as a barrier to attaining lower inpatient glucose levels.31 Although hypoglycemia frequency in the hospital is low, and even though recent data indicates that hypoglycemia is not perceived by practitioners as the number 1 barrier to successful inpatient diabetes management,3234 the possible association of severe low glucose levels to inpatient mortality18, 19, 21, 22, 24, 35 makes hypoglycemia a key counterbalance metric that hospitals will need to track as they implement glycemic control programs. In the ICU, higher glycemic targets may be needed to allay practitioner fears, and insulin administration protocols that have the best track record for minimizing hypoglycemia should be identified and promulgated.

Recent data showing increased risk of hospital hypoglycemia with attempts to better control hyperglycemia may unjustifiably deter practitioners and hospitals from implementing programs to better control inpatient glucose levels.24, 36 Unlike the outpatient setting, where patients can take measures to prevent hypoglycemia, hospitalized patients surrender control of their diabetes management to staff. Inpatient tight glycemic control initiatives cannot be instituted unless they are coupled with efforts to understand and correct system‐based problems that increase the risk of hypoglycemia. Recently published reports demonstrate that hypoglycemic events can be kept very low during treatment with an intensive insulin infusion protocol if expert rules are built into the algorithm that address hypoglycemia.37, 38 Thus, rather than abandon efforts at improving inpatient hyperglycemia over concerns about hypoglycemia, hospitals will need to develop methods to change their hypoglycemia policies from ones that typically just guide treatment to ones that incorporate preventive strategies.

Our data suggest a relationship between POC‐BG levels and hospital characteristics. Rural hospitals and hospitals with the least number of beds had higher POC‐BG levels compared to urban, academic, or larger hospitals, especially in the ICU setting. The reasons underlying these findings cannot be determined from this analysis, but it is possible that smaller hospitals and those located in rural areas do not have access to the diabetes experts (eg, endocrinologists or diabetes educators) to assist them in developing tight glycemic control programs. We also detected differences in patient‐day‐weighted mean POC‐BG data based on geographic region. Whether considering ICU or non‐ICU data, hospitals located in the West had lower glucose values compared to other regions. As with the other hospital characteristics, the explanation underlying these observations cannot be determined. It is possible that hospitals in the West are earlier‐adopters of tight glycemic control programs compared to other U.S. geographic regions. Further study is needed in a larger number of hospitals to confirm these findings.

These findings should be considered in light of the following limitations: unavailability to us of specific patient‐level information that would allow adjustment of data for such as variables as comorbidity; the fact that recommendations about glycemic targets in the hospital vary by organization,810, 30 which may result in hospitals aiming for different targets in different populations; and the controversy that continues on the benefits of glycemic control in the ICU, which may be dissuading facilities from implementing glucose control programs.39, 40 All that can be concluded from our analysis is that there is variation in the POC‐BG data based on hospital characteristics. We cannot state that one type of hospital is performing glycemic control better than another, particularly as some hospital types are underrepresented in our sample, and we cannot control for patient‐level data. Moreover, this statistical variation seen between different hospital types may not be of clinical importance in terms of being associated with different outcomes, or may simply be a result of different patterns of glucose monitoring in individual hospitals. However, the observed variation should prompt further investigation into the basis of differences (eg, some hospital types or regions may be further ahead in inpatient diabetes quality improvement initiatives than others).

There is no consensus about how best to summarize and report glycemic control in the hospital (so called glucometrics),41 and a variety of reporting measures have been suggested.20, 4245 We show data using one method: with the mean BG normalized to patient‐day as the unit of analysis; however, we found similar results when we used the patient or the glucose reading as the unit of analysis. As organizations move to develop standards for summarizing inpatient glucose data, consideration must be given to which measure is best correlated with hospital outcomes. In addition, when developing standards, it will be important to determine what type of data hospitals will find most clinically useful to track the impact of glucose control interventions. For instance, hospitals may wish to see data on the frequencies of glucose measurements that are above and below certain desired thresholds, which is one of the approaches that we have used in previous publications,14, 17, 26, 46 and which is currently provided as feedback to hospitals participating in RALS reporting.

The other issue to address in development of standards in inpatient glycemic control reporting is what method of glucose measurement should be used. Correlation between whole‐blood vs. POC‐BG values can be imprecise in the intensive care setting.41, 47 We have previously utilized bedside glucose measurements as our means of evaluating the status of inpatient glucose control,14, 17, 26 and bedside glucose measurements remain the mainstay of how practitioners judge the status of inpatient hyperglycemia and make therapeutic decisions about management. The hospitals participating in the process reviewed here all use the same system of bedside glucose monitoring and glucometer‐laboratory electronic interface. Until alternative clinical methods are developed to frequently sample glucose levels in a convenient and minimally invasive way at the bedside, current POC‐BG technology will continue to be the most utilized means of assessing hospital glucose management in the inpatient setting.

Electronic data warehouses such as RALS‐Plus are convenient sources of information in which to store data on the quality of inpatient diabetes care. Unlike chart abstraction which requires extensive man‐hours to extract data on a few patients, use of electronic data allows examination of large numbers of hospital cases. Queries of information systems could be automated, report cards potentially generated, and feedback given to providers and hospitals on the status of inpatient glycemic control.

Nonetheless, there are limitations to using electronic records as the sole method to assess inpatient diabetes care. Analysis of electronic records does not allow assessment of reasons underlying decision‐making behavior of clinicians (eg, why they did or did not change hyperglycemic therapy). Moreover, our electronic data does not permit an assessment of who had preexisting diabetes, who was admitted with new onset diabetes, or who developed hyperglycemia as a result of the hospital stay.

In addition to the above, while our sample was representative of other RALs participating hospitals, it was not entirely representative of all U.S. hospitals. Hospitals contributing data to this report were chosen by self‐selection rather than by random methods. Expanding hospital participation in this inpatient glucose assessment benchmarking process will be needed to determine if findings in this work can be generalized. Finally, our study was conducted using the hospital, rather than the patient, as the unit of analysis, as patient‐level characteristics (age, sex, race/ethnicity) were not provided by participating hospitals.

Despite these limitations and issues noted above, to our knowledge this report is the most extensive review of the state of blood glucose control in hospitals across the United States. While other commercial laboratory data management systems may exist in hospitals, their data has not been reported to date. Additionally, our analysis provides a first glimpse of inpatient glycemic control of a large number of U.S. hospitals of varying characteristics and different national regions. Increased hospital participation in data collection may allow the creation of a national benchmarking process for the development of best practices and improved inpatient hyperglycemia management.

The past decade has seen an increase in the number of hospital discharges associated with a diabetes diagnosis.1, 2 Diabetes is the fourth leading comorbid condition associated with any hospital discharge in the United States.3 Nearly one‐third of diabetes patients require 2 or more hospitalizations in any given year,4 and inpatient stays account for the largest proportion of direct medical expenses incurred by persons with the disease.5

The hospital component of diabetes care has been receiving considerable attention. The advantage of effective inpatient diabetes managementwith particular attention to improving glycemic controlis evident for a number of clinical situations (eg, acute myocardial infarction, critically ill patients).68 National and regional organizations,912 and professional societies68, 12 have developed guidelines about management of inpatient hyperglycemia.

Despite increased awareness of the value of treating inpatient hyperglycemia, little is known about glucose control in U.S. hospitals. As hospitals begin to develop programs to improve inpatient glucose management, some method of standardized benchmarking should be put in place. Using information systems solutions to obtain point‐of‐care bedside glucose (POC‐BG) data, we previously reported on inpatient glucose control from a smaller number of U.S. hospitals.13, 14 We now provide data on a larger, more representative number of U.S. hospitals that gives a broader national view of the current status of inpatient glycemic control.

Patients and Methods

Results

Overall Glycemic Control

A total of 12,559,305 POC‐BG measurements (2,935,167 from the ICU and 9,624,138 from the non‐ICU) from 1,010,705 patients with 3,973,460 patient days were analyzed from 126 hospitals. The mean number of measurements was 20 per ICU patient and 9.5 for non‐ICU patients. The average number of measurements taken per patient‐day was 5 for the ICU patient and 3 for the non‐ICU patient.

Hospital hyperglycemia (>180 mg/dL) was 46.0% for ICU and 31.7% for non‐ICU. The patient‐day‐weighted mean POC‐BG for ICU measurements was 165 mg/dL (median = 164 mg/dL, SD 14.5) and 166 mg/dL (median = 167 mg/dL, SD 8) for non‐ICU. The distributions of patient‐day‐weighted mean POC‐BG values for ICU and non‐ICU settings are shown in Figure 1. The range of patient‐day‐weighted mean values was much wider for the ICU (126‐203 mg/dL) than in the non‐ICU (139‐186 mg/dL).

Figure 1

Hyperglycemia Prevalence

Of ICU patients, 60.6% had at least 1 POC‐BG value >180 mg/dL, as did 46.4% of non‐ICU patients. The proportion of patient‐days with a patient‐day‐weighted mean POC‐BG >180 mg/dL was 26.3% in the ICU setting (Figure 2A) and 31.3% in the non‐ICU (Figure 2B); the other cut points are also shown in Figure 2. The prevalence of patient‐days where hyperglycemia was more severe (>300 mg/dL) was low but nonetheless still detected in both the ICU and non‐ICU settings, although these differences appear to be less pronounced than in the ICU.

Figure 2

Hypoglycemia Rates

There were 21.3% of patients who had at least 1 POC‐BG value <70 mg/dL. Hospital hypoglycemia was low in both the ICU and non‐ICU measurement data, although the proportion of patient days with POC‐BG <70 mg/dL was higher in the ICU vs. the non‐ICU setting (Figure 3A,B). Hypoglycemia (<70 mg/dL) was detected in 10.1% of patient‐days (3.2% of all measures) in the ICU setting (Figure 3A) and 3.5% of patient‐days (4.2% of all measures) in the non‐ICU (Figure 3B). Moderate (<60 mg/dL) and more severe (<50 mg/dL and <40 mg/dL) hypoglycemia were very uncommon in both the ICU and non‐ICU.

Figure 3

Relationship of Glucose Control with Hospital Characteristics

There was a significant relationship between the total number of hospital beds and patient‐day‐weighted mean POC‐BG values in the ICU (Figure 4A). In the ICU, hospitals with <200 beds had significantly higher patient‐day‐weighted mean POC‐BG levels than those with 200 to 299 beds (P < 0.05), 300 to 399 beds (P < 0.01), and 400 beds (P < 0.001). Rural hospitals (Figure 4B) also had higher patient‐day‐weighted mean POC‐BG values compared to urban community and academic hospitals (both P < 0.001). Finally, ICUs in hospitals in the West (Figure 4C, bottom panel) had significantly lower values than those in the Midwest and South (both P < 0.01).

Figure 4

Differences in patient‐day‐weighted mean POC‐BG levels based on hospital characteristics were also observed for the non‐ICU (Figure 5), although these differences appear to be less pronounced than in the ICU. Hospitals with <200 beds (Figure 5A) had significantly higher patient‐day‐weighted mean POC‐BG values compared to hospitals with 300 to 399 beds (P < 0.05) and 400 beds (P < 0.001). Rural hospitals (Figure 5B) had significantly higher values than academic (P < 0.05) and urban community (P < 0.001) hospitals, and hospitals in the West (Figure 5C) had significantly lower values than those in the South and Northeast (both P < 0.05).

Figure 5

Discussion

Hospitalizations associated with diabetes pose a substantial burden on the U.S. health system.15 Recent consensus advocates good glucose control in the hospital to optimize outcomes for a number of clinical scenarios.68 Aside from a few institution‐specific studies,2527 the quality of diabetes treatment in U.S. hospitals is mostly unknown, but assessing the level of glycemic control will be a key metric that hospitals will need to track as they implement improvement programs targeting hospital hyperglycemia. Hospitals will need a way not just to track overall glucose levels, but also to monitor whether hypoglycemic events rise as they implement tight glycemic control initiatives. To our knowledge this is the first report on glycemic control from a large number of U.S. hospitals with diverse characteristics and from different geographic regions.

Debate continues as to what glucose targets for inpatients should be attained.28, 29 The overall patient‐day‐weighted mean POC‐BG was 170 mg/dL for the non‐ICU, and only a moderately lower 162 mg/dL in the ICU, despite much lower thresholds for ICU measurements in current suggested guidelines.8, 30 For the average hospital, over one‐third of non‐ICUs had patient‐day‐weighted mean POC‐BG levels that were >180 mg/dL and nearly one‐quarter had values >200 mg/dL. Similarly, nearly 40% of ICUs had patient‐day‐weighted mean POC‐BGs >180 mg/dL and over 30% were >200 mg/dL, indicating room for improvement in hospital ICU glucose control, at least in the hospitals sampled here. The range of patient‐day‐weighted mean POC‐BG levels for the ICU was broader than what was seen in the non‐ICU data, with the ICU data containing lower weighted mean POC‐BG values, and may indicate that hospitals are concentrating their efforts on adopting stricter glucose control measures in their ICUs.

Whether examining data from a single institution,17 from a larger group of hospitals,14 or now from 126 hospitals, one consistent finding has been the low prevalence of hypoglycemiaparticularly severe hypoglycemia (glucose <50 mg/dL or <40 mg/dL). Based on this larger sampling, however, hypoglycemia in the ICU, while still uncommon with respect to hyperglycemia, is more than double that of the non‐ICU. Fear of hypoglycemia is frequently mentioned as a barrier to attaining lower inpatient glucose levels.31 Although hypoglycemia frequency in the hospital is low, and even though recent data indicates that hypoglycemia is not perceived by practitioners as the number 1 barrier to successful inpatient diabetes management,3234 the possible association of severe low glucose levels to inpatient mortality18, 19, 21, 22, 24, 35 makes hypoglycemia a key counterbalance metric that hospitals will need to track as they implement glycemic control programs. In the ICU, higher glycemic targets may be needed to allay practitioner fears, and insulin administration protocols that have the best track record for minimizing hypoglycemia should be identified and promulgated.

Recent data showing increased risk of hospital hypoglycemia with attempts to better control hyperglycemia may unjustifiably deter practitioners and hospitals from implementing programs to better control inpatient glucose levels.24, 36 Unlike the outpatient setting, where patients can take measures to prevent hypoglycemia, hospitalized patients surrender control of their diabetes management to staff. Inpatient tight glycemic control initiatives cannot be instituted unless they are coupled with efforts to understand and correct system‐based problems that increase the risk of hypoglycemia. Recently published reports demonstrate that hypoglycemic events can be kept very low during treatment with an intensive insulin infusion protocol if expert rules are built into the algorithm that address hypoglycemia.37, 38 Thus, rather than abandon efforts at improving inpatient hyperglycemia over concerns about hypoglycemia, hospitals will need to develop methods to change their hypoglycemia policies from ones that typically just guide treatment to ones that incorporate preventive strategies.

Our data suggest a relationship between POC‐BG levels and hospital characteristics. Rural hospitals and hospitals with the least number of beds had higher POC‐BG levels compared to urban, academic, or larger hospitals, especially in the ICU setting. The reasons underlying these findings cannot be determined from this analysis, but it is possible that smaller hospitals and those located in rural areas do not have access to the diabetes experts (eg, endocrinologists or diabetes educators) to assist them in developing tight glycemic control programs. We also detected differences in patient‐day‐weighted mean POC‐BG data based on geographic region. Whether considering ICU or non‐ICU data, hospitals located in the West had lower glucose values compared to other regions. As with the other hospital characteristics, the explanation underlying these observations cannot be determined. It is possible that hospitals in the West are earlier‐adopters of tight glycemic control programs compared to other U.S. geographic regions. Further study is needed in a larger number of hospitals to confirm these findings.

These findings should be considered in light of the following limitations: unavailability to us of specific patient‐level information that would allow adjustment of data for such as variables as comorbidity; the fact that recommendations about glycemic targets in the hospital vary by organization,810, 30 which may result in hospitals aiming for different targets in different populations; and the controversy that continues on the benefits of glycemic control in the ICU, which may be dissuading facilities from implementing glucose control programs.39, 40 All that can be concluded from our analysis is that there is variation in the POC‐BG data based on hospital characteristics. We cannot state that one type of hospital is performing glycemic control better than another, particularly as some hospital types are underrepresented in our sample, and we cannot control for patient‐level data. Moreover, this statistical variation seen between different hospital types may not be of clinical importance in terms of being associated with different outcomes, or may simply be a result of different patterns of glucose monitoring in individual hospitals. However, the observed variation should prompt further investigation into the basis of differences (eg, some hospital types or regions may be further ahead in inpatient diabetes quality improvement initiatives than others).

There is no consensus about how best to summarize and report glycemic control in the hospital (so called glucometrics),41 and a variety of reporting measures have been suggested.20, 4245 We show data using one method: with the mean BG normalized to patient‐day as the unit of analysis; however, we found similar results when we used the patient or the glucose reading as the unit of analysis. As organizations move to develop standards for summarizing inpatient glucose data, consideration must be given to which measure is best correlated with hospital outcomes. In addition, when developing standards, it will be important to determine what type of data hospitals will find most clinically useful to track the impact of glucose control interventions. For instance, hospitals may wish to see data on the frequencies of glucose measurements that are above and below certain desired thresholds, which is one of the approaches that we have used in previous publications,14, 17, 26, 46 and which is currently provided as feedback to hospitals participating in RALS reporting.

The other issue to address in development of standards in inpatient glycemic control reporting is what method of glucose measurement should be used. Correlation between whole‐blood vs. POC‐BG values can be imprecise in the intensive care setting.41, 47 We have previously utilized bedside glucose measurements as our means of evaluating the status of inpatient glucose control,14, 17, 26 and bedside glucose measurements remain the mainstay of how practitioners judge the status of inpatient hyperglycemia and make therapeutic decisions about management. The hospitals participating in the process reviewed here all use the same system of bedside glucose monitoring and glucometer‐laboratory electronic interface. Until alternative clinical methods are developed to frequently sample glucose levels in a convenient and minimally invasive way at the bedside, current POC‐BG technology will continue to be the most utilized means of assessing hospital glucose management in the inpatient setting.

Electronic data warehouses such as RALS‐Plus are convenient sources of information in which to store data on the quality of inpatient diabetes care. Unlike chart abstraction which requires extensive man‐hours to extract data on a few patients, use of electronic data allows examination of large numbers of hospital cases. Queries of information systems could be automated, report cards potentially generated, and feedback given to providers and hospitals on the status of inpatient glycemic control.

Nonetheless, there are limitations to using electronic records as the sole method to assess inpatient diabetes care. Analysis of electronic records does not allow assessment of reasons underlying decision‐making behavior of clinicians (eg, why they did or did not change hyperglycemic therapy). Moreover, our electronic data does not permit an assessment of who had preexisting diabetes, who was admitted with new onset diabetes, or who developed hyperglycemia as a result of the hospital stay.

In addition to the above, while our sample was representative of other RALs participating hospitals, it was not entirely representative of all U.S. hospitals. Hospitals contributing data to this report were chosen by self‐selection rather than by random methods. Expanding hospital participation in this inpatient glucose assessment benchmarking process will be needed to determine if findings in this work can be generalized. Finally, our study was conducted using the hospital, rather than the patient, as the unit of analysis, as patient‐level characteristics (age, sex, race/ethnicity) were not provided by participating hospitals.

Despite these limitations and issues noted above, to our knowledge this report is the most extensive review of the state of blood glucose control in hospitals across the United States. While other commercial laboratory data management systems may exist in hospitals, their data has not been reported to date. Additionally, our analysis provides a first glimpse of inpatient glycemic control of a large number of U.S. hospitals of varying characteristics and different national regions. Increased hospital participation in data collection may allow the creation of a national benchmarking process for the development of best practices and improved inpatient hyperglycemia management.

The past decade has seen an increase in the number of hospital discharges associated with a diabetes diagnosis.1, 2 Diabetes is the fourth leading comorbid condition associated with any hospital discharge in the United States.3 Nearly one‐third of diabetes patients require 2 or more hospitalizations in any given year,4 and inpatient stays account for the largest proportion of direct medical expenses incurred by persons with the disease.5

The hospital component of diabetes care has been receiving considerable attention. The advantage of effective inpatient diabetes managementwith particular attention to improving glycemic controlis evident for a number of clinical situations (eg, acute myocardial infarction, critically ill patients).68 National and regional organizations,912 and professional societies68, 12 have developed guidelines about management of inpatient hyperglycemia.

Despite increased awareness of the value of treating inpatient hyperglycemia, little is known about glucose control in U.S. hospitals. As hospitals begin to develop programs to improve inpatient glucose management, some method of standardized benchmarking should be put in place. Using information systems solutions to obtain point‐of‐care bedside glucose (POC‐BG) data, we previously reported on inpatient glucose control from a smaller number of U.S. hospitals.13, 14 We now provide data on a larger, more representative number of U.S. hospitals that gives a broader national view of the current status of inpatient glycemic control.

Patients and Methods

Results

Overall Glycemic Control

A total of 12,559,305 POC‐BG measurements (2,935,167 from the ICU and 9,624,138 from the non‐ICU) from 1,010,705 patients with 3,973,460 patient days were analyzed from 126 hospitals. The mean number of measurements was 20 per ICU patient and 9.5 for non‐ICU patients. The average number of measurements taken per patient‐day was 5 for the ICU patient and 3 for the non‐ICU patient.

Hospital hyperglycemia (>180 mg/dL) was 46.0% for ICU and 31.7% for non‐ICU. The patient‐day‐weighted mean POC‐BG for ICU measurements was 165 mg/dL (median = 164 mg/dL, SD 14.5) and 166 mg/dL (median = 167 mg/dL, SD 8) for non‐ICU. The distributions of patient‐day‐weighted mean POC‐BG values for ICU and non‐ICU settings are shown in Figure 1. The range of patient‐day‐weighted mean values was much wider for the ICU (126‐203 mg/dL) than in the non‐ICU (139‐186 mg/dL).

Figure 1

Hyperglycemia Prevalence

Of ICU patients, 60.6% had at least 1 POC‐BG value >180 mg/dL, as did 46.4% of non‐ICU patients. The proportion of patient‐days with a patient‐day‐weighted mean POC‐BG >180 mg/dL was 26.3% in the ICU setting (Figure 2A) and 31.3% in the non‐ICU (Figure 2B); the other cut points are also shown in Figure 2. The prevalence of patient‐days where hyperglycemia was more severe (>300 mg/dL) was low but nonetheless still detected in both the ICU and non‐ICU settings, although these differences appear to be less pronounced than in the ICU.

Figure 2

Hypoglycemia Rates

There were 21.3% of patients who had at least 1 POC‐BG value <70 mg/dL. Hospital hypoglycemia was low in both the ICU and non‐ICU measurement data, although the proportion of patient days with POC‐BG <70 mg/dL was higher in the ICU vs. the non‐ICU setting (Figure 3A,B). Hypoglycemia (<70 mg/dL) was detected in 10.1% of patient‐days (3.2% of all measures) in the ICU setting (Figure 3A) and 3.5% of patient‐days (4.2% of all measures) in the non‐ICU (Figure 3B). Moderate (<60 mg/dL) and more severe (<50 mg/dL and <40 mg/dL) hypoglycemia were very uncommon in both the ICU and non‐ICU.

Figure 3

Relationship of Glucose Control with Hospital Characteristics

There was a significant relationship between the total number of hospital beds and patient‐day‐weighted mean POC‐BG values in the ICU (Figure 4A). In the ICU, hospitals with <200 beds had significantly higher patient‐day‐weighted mean POC‐BG levels than those with 200 to 299 beds (P < 0.05), 300 to 399 beds (P < 0.01), and 400 beds (P < 0.001). Rural hospitals (Figure 4B) also had higher patient‐day‐weighted mean POC‐BG values compared to urban community and academic hospitals (both P < 0.001). Finally, ICUs in hospitals in the West (Figure 4C, bottom panel) had significantly lower values than those in the Midwest and South (both P < 0.01).

Figure 4

Differences in patient‐day‐weighted mean POC‐BG levels based on hospital characteristics were also observed for the non‐ICU (Figure 5), although these differences appear to be less pronounced than in the ICU. Hospitals with <200 beds (Figure 5A) had significantly higher patient‐day‐weighted mean POC‐BG values compared to hospitals with 300 to 399 beds (P < 0.05) and 400 beds (P < 0.001). Rural hospitals (Figure 5B) had significantly higher values than academic (P < 0.05) and urban community (P < 0.001) hospitals, and hospitals in the West (Figure 5C) had significantly lower values than those in the South and Northeast (both P < 0.05).

Figure 5

Discussion

Hospitalizations associated with diabetes pose a substantial burden on the U.S. health system.15 Recent consensus advocates good glucose control in the hospital to optimize outcomes for a number of clinical scenarios.68 Aside from a few institution‐specific studies,2527 the quality of diabetes treatment in U.S. hospitals is mostly unknown, but assessing the level of glycemic control will be a key metric that hospitals will need to track as they implement improvement programs targeting hospital hyperglycemia. Hospitals will need a way not just to track overall glucose levels, but also to monitor whether hypoglycemic events rise as they implement tight glycemic control initiatives. To our knowledge this is the first report on glycemic control from a large number of U.S. hospitals with diverse characteristics and from different geographic regions.

Debate continues as to what glucose targets for inpatients should be attained.28, 29 The overall patient‐day‐weighted mean POC‐BG was 170 mg/dL for the non‐ICU, and only a moderately lower 162 mg/dL in the ICU, despite much lower thresholds for ICU measurements in current suggested guidelines.8, 30 For the average hospital, over one‐third of non‐ICUs had patient‐day‐weighted mean POC‐BG levels that were >180 mg/dL and nearly one‐quarter had values >200 mg/dL. Similarly, nearly 40% of ICUs had patient‐day‐weighted mean POC‐BGs >180 mg/dL and over 30% were >200 mg/dL, indicating room for improvement in hospital ICU glucose control, at least in the hospitals sampled here. The range of patient‐day‐weighted mean POC‐BG levels for the ICU was broader than what was seen in the non‐ICU data, with the ICU data containing lower weighted mean POC‐BG values, and may indicate that hospitals are concentrating their efforts on adopting stricter glucose control measures in their ICUs.

Whether examining data from a single institution,17 from a larger group of hospitals,14 or now from 126 hospitals, one consistent finding has been the low prevalence of hypoglycemiaparticularly severe hypoglycemia (glucose <50 mg/dL or <40 mg/dL). Based on this larger sampling, however, hypoglycemia in the ICU, while still uncommon with respect to hyperglycemia, is more than double that of the non‐ICU. Fear of hypoglycemia is frequently mentioned as a barrier to attaining lower inpatient glucose levels.31 Although hypoglycemia frequency in the hospital is low, and even though recent data indicates that hypoglycemia is not perceived by practitioners as the number 1 barrier to successful inpatient diabetes management,3234 the possible association of severe low glucose levels to inpatient mortality18, 19, 21, 22, 24, 35 makes hypoglycemia a key counterbalance metric that hospitals will need to track as they implement glycemic control programs. In the ICU, higher glycemic targets may be needed to allay practitioner fears, and insulin administration protocols that have the best track record for minimizing hypoglycemia should be identified and promulgated.

Recent data showing increased risk of hospital hypoglycemia with attempts to better control hyperglycemia may unjustifiably deter practitioners and hospitals from implementing programs to better control inpatient glucose levels.24, 36 Unlike the outpatient setting, where patients can take measures to prevent hypoglycemia, hospitalized patients surrender control of their diabetes management to staff. Inpatient tight glycemic control initiatives cannot be instituted unless they are coupled with efforts to understand and correct system‐based problems that increase the risk of hypoglycemia. Recently published reports demonstrate that hypoglycemic events can be kept very low during treatment with an intensive insulin infusion protocol if expert rules are built into the algorithm that address hypoglycemia.37, 38 Thus, rather than abandon efforts at improving inpatient hyperglycemia over concerns about hypoglycemia, hospitals will need to develop methods to change their hypoglycemia policies from ones that typically just guide treatment to ones that incorporate preventive strategies.

Our data suggest a relationship between POC‐BG levels and hospital characteristics. Rural hospitals and hospitals with the least number of beds had higher POC‐BG levels compared to urban, academic, or larger hospitals, especially in the ICU setting. The reasons underlying these findings cannot be determined from this analysis, but it is possible that smaller hospitals and those located in rural areas do not have access to the diabetes experts (eg, endocrinologists or diabetes educators) to assist them in developing tight glycemic control programs. We also detected differences in patient‐day‐weighted mean POC‐BG data based on geographic region. Whether considering ICU or non‐ICU data, hospitals located in the West had lower glucose values compared to other regions. As with the other hospital characteristics, the explanation underlying these observations cannot be determined. It is possible that hospitals in the West are earlier‐adopters of tight glycemic control programs compared to other U.S. geographic regions. Further study is needed in a larger number of hospitals to confirm these findings.

These findings should be considered in light of the following limitations: unavailability to us of specific patient‐level information that would allow adjustment of data for such as variables as comorbidity; the fact that recommendations about glycemic targets in the hospital vary by organization,810, 30 which may result in hospitals aiming for different targets in different populations; and the controversy that continues on the benefits of glycemic control in the ICU, which may be dissuading facilities from implementing glucose control programs.39, 40 All that can be concluded from our analysis is that there is variation in the POC‐BG data based on hospital characteristics. We cannot state that one type of hospital is performing glycemic control better than another, particularly as some hospital types are underrepresented in our sample, and we cannot control for patient‐level data. Moreover, this statistical variation seen between different hospital types may not be of clinical importance in terms of being associated with different outcomes, or may simply be a result of different patterns of glucose monitoring in individual hospitals. However, the observed variation should prompt further investigation into the basis of differences (eg, some hospital types or regions may be further ahead in inpatient diabetes quality improvement initiatives than others).

There is no consensus about how best to summarize and report glycemic control in the hospital (so called glucometrics),41 and a variety of reporting measures have been suggested.20, 4245 We show data using one method: with the mean BG normalized to patient‐day as the unit of analysis; however, we found similar results when we used the patient or the glucose reading as the unit of analysis. As organizations move to develop standards for summarizing inpatient glucose data, consideration must be given to which measure is best correlated with hospital outcomes. In addition, when developing standards, it will be important to determine what type of data hospitals will find most clinically useful to track the impact of glucose control interventions. For instance, hospitals may wish to see data on the frequencies of glucose measurements that are above and below certain desired thresholds, which is one of the approaches that we have used in previous publications,14, 17, 26, 46 and which is currently provided as feedback to hospitals participating in RALS reporting.

The other issue to address in development of standards in inpatient glycemic control reporting is what method of glucose measurement should be used. Correlation between whole‐blood vs. POC‐BG values can be imprecise in the intensive care setting.41, 47 We have previously utilized bedside glucose measurements as our means of evaluating the status of inpatient glucose control,14, 17, 26 and bedside glucose measurements remain the mainstay of how practitioners judge the status of inpatient hyperglycemia and make therapeutic decisions about management. The hospitals participating in the process reviewed here all use the same system of bedside glucose monitoring and glucometer‐laboratory electronic interface. Until alternative clinical methods are developed to frequently sample glucose levels in a convenient and minimally invasive way at the bedside, current POC‐BG technology will continue to be the most utilized means of assessing hospital glucose management in the inpatient setting.

Electronic data warehouses such as RALS‐Plus are convenient sources of information in which to store data on the quality of inpatient diabetes care. Unlike chart abstraction which requires extensive man‐hours to extract data on a few patients, use of electronic data allows examination of large numbers of hospital cases. Queries of information systems could be automated, report cards potentially generated, and feedback given to providers and hospitals on the status of inpatient glycemic control.

Nonetheless, there are limitations to using electronic records as the sole method to assess inpatient diabetes care. Analysis of electronic records does not allow assessment of reasons underlying decision‐making behavior of clinicians (eg, why they did or did not change hyperglycemic therapy). Moreover, our electronic data does not permit an assessment of who had preexisting diabetes, who was admitted with new onset diabetes, or who developed hyperglycemia as a result of the hospital stay.

In addition to the above, while our sample was representative of other RALs participating hospitals, it was not entirely representative of all U.S. hospitals. Hospitals contributing data to this report were chosen by self‐selection rather than by random methods. Expanding hospital participation in this inpatient glucose assessment benchmarking process will be needed to determine if findings in this work can be generalized. Finally, our study was conducted using the hospital, rather than the patient, as the unit of analysis, as patient‐level characteristics (age, sex, race/ethnicity) were not provided by participating hospitals.

Despite these limitations and issues noted above, to our knowledge this report is the most extensive review of the state of blood glucose control in hospitals across the United States. While other commercial laboratory data management systems may exist in hospitals, their data has not been reported to date. Additionally, our analysis provides a first glimpse of inpatient glycemic control of a large number of U.S. hospitals of varying characteristics and different national regions. Increased hospital participation in data collection may allow the creation of a national benchmarking process for the development of best practices and improved inpatient hyperglycemia management.