Original Research

A continuing focus on readmission prevention as a means to improve quality and reduce waste and costs has led to abundant research on the identification of factors that put older patients at risk for readmission.[1, 2] Recently, research has focused on the development of prediction tools that are based on information from electronic health records (EHR) and that enable early, at‐admission, identification of patients at high risk for readmission.[3, 4, 5] Through such identification, patients are already targeted for inclusion in readmission‐prevention interventions early in their hospital stay.[6, 7] Yet, the ability to rely on these early‐identification tools is contingent on the stability of risk during the hospital stay. Of the in‐hospital factors that can affect readmission, change in functioning has been identified as a potentially major contributor, especially in older patients.[8]

Older adults' deterioration in functioning is a common, troubling phenomenon. Approximately 20% of patients older than 70 years and hospitalized for a medical illness deteriorate in functioning during the hospitalization.[9] Recent evidence points to the contribution of functioning to 30‐day readmission risk, with studies showing that pre‐ or posthospitalization functional impairment is associated with a 1.5 to 3 times greater likelihood of readmission.[10, 11, 12, 13] Whether in‐hospital changes in functioning add to patients' baseline risk, however, is unknown. The only readmission study that examined functional decline during hospitalization was a retrospective study performed in a population of older adults receiving rehabilitation services and therefore may not be generalizable to acutely hospitalized adults.[14] To fill the current knowledge gap, we examined whether in‐hospital functional decline significantly contributes to at‐admission readmission risk factors in its ability to accurately identify high readmission risk in older adults hospitalized in internal medicine units.

METHODS

RESULTS

Table 1 presents the baseline and in‐hospital characteristics of participants with and without readmissions. The sample includes 559 community‐dwelling older adults (49% men) aged 70 to 98 years (mean age 79 years). One‐third (36%) of the participants were fully independent in ADL at admission. Eighty‐five (15.2%) patients were readmitted within 30 days of discharge. Participants who were readmitted had lower at‐admission ADL levels; had 1 more hospitalization in the previous year; and were more likely to have HF, CRF, arrhythmia, and malignancy, and to be at risk of malnutrition, than those who were not readmitted. Participants who were readmitted were more likely to suffer from functional decline during the index hospitalization. No significant differences were found in living arrangements or in at‐admission mental and cognitive status.

Multivariate analysis (Table 2) shows that higher at‐admission mBI score was associated with lower odds of readmission (OR for 1‐unit increase: 0.99, 95% CI: 0.98‐0.99). Other predictors of higher readmission risk were: high or medium at‐admission risk of malnutrition, malignancy, CRF, each additional hospitalization during the previous year, and lower albumin levels. Severity of illness and demographic characteristics were not significantly associated with readmission.

The at‐discharge model that combined the baseline model and in‐hospital risk factors showed that in‐hospital (from admission to discharge) ADL decline was significantly associated with readmission, as a 10‐point decrease in the mBI from admission to discharge was associated with 1.32 (95% CI: 1.02‐1.72) greater odds of readmission. LOS was not significantly associated with readmission, after controlling for baseline health status and in‐hospital ADL decline. All other predictors did not markedly change from the baseline to the at‐discharge model either in significance levels or in magnitude.

The discriminatory power of the baseline model was good (C statistic=0.81). Adding ADL decline and LOS did not change the discriminatory power of the model (C statistic=0.81). The P value of the Hosmer‐Lemeshow test equaled 0.67 for the baseline model and 0.48 for the at‐discharge model, indicating good calibration of both models. The P values for the regression coefficients of bootstrap inference assessing the relationship between the at‐admission and in‐hospital ADL decline odds of readmission remained stable (P 0.05).

Classification of patients into risk categories by the baseline model and the discharge model (Table 3) shows that identifying patients in the top‐tier category (20th highest percentile) according to information available before or at admission does not detect 6/111 (5.4%) of patients who would have been categorized as highest‐risk if information on ADL decline had been incorporated in the predictive algorithm. Additional partitioning of the top fifth group into 2 tiers (8089th and 9099th percentiles) shows that selection of patients in the top 10% of the baseline risk score would not have detected 7/55 (12.7%) patients who would have been identified as high risk at discharge (data not shown).

DISCUSSION

To our knowledge, ours is the first empirical test of the simultaneous role of functioning along the hospitalization course in explaining readmission risk.[8] Our results show that at‐admission lower functional status and in‐hospital functional decline are significant predictors of early unplanned readmission in older adults, beyond other well‐known risk factors.

The major purpose of this study was to examine whether at‐admission data can be used to detect high‐risk patients for potential inclusion in readmission prevention interventions, or whether changes in ADL occurring during the index hospitalization could affect patients' risk, therefore necessitating an additional assessment at discharge. Our results show that some patients would not have been detected at admission, as their in‐hospital ADL decline affects their at‐discharge risk. Nonetheless, this is a small group (only 5% of patients if a targeting threshold of the highest 20% risk is used). Our findings also show that information on ADL decline during the index hospitalization does not contribute to the accuracy of readmission‐risk prediction in a model that utilizes data on prior hospitalizations, baseline nutritional and functional status, and chronic morbidity (CRF and malignancy). Our results are consistent with previous studies showing the association between baseline,[11, 13, 32] or at‐admission[13] functional status and readmission. However, these studies did not analyze the related contribution of in‐hospital functioning to readmission risk, which was recently suggested as a feature that may significantly affect readmission risk, especially in older patients.[8]

Our findings are also congruent with those of a study in which LOS was not significantly associated with readmission in an elderly population.[33] Our null finding can be explained by the broad set of pre‐ and at‐admission variables, such as nutritional and functional status as well as in‐hospital functional decline, included in our model, making LOS a less significant contributor than in more parsimonious models.[28]

Our results also show that malnutrition contributes to readmission risk beyond other well‐known risk factors. Previous studies showed that malnutrition in the elderly is associated with early readmission.[11, 34] These studies, however, did not examine other well‐known risk factors, such as previous hospitalizations, which were tested in our study, precluding identification of the contribution of malnutrition beyond other well‐known risks.

Our findings should be interpreted in light of several limitations. First, the functional, nutritional, and cognitive data were collected from participants' self‐reports, which are prone to recall bias. Nonetheless, self‐report is often used in large‐scale studies, which preclude actual performance measurement.[21] Second, our sample is of adults aged 70 years or older, and may not be representative of the 65 and older population, which is the target population for many readmission reduction interventions.[35] Yet, participants were from a relatively high‐functioning group of patients who were discharged to their homes, thus may resemble the over age 65 years group. Moreover, these inclusion criteria may have affected their readmission rates, which at 15% are lower than the average reported in other older adult populations.[36] Nonetheless, a more heterogenic sample (in terms of baseline functional status) is needed to address the association between in‐hospital functional change and readmissions as well as the discrimination of the model. Third, the attrition rate (16%) might impact the predictive ability of the models, as patients dropped‐out from the study might have had higher in‐hospital deterioration. However, no significant differences between study sample and dropped‐out patients in the wide range of baseline characteristics except for age and baseline albumin levels were found. Fourth, the unique characteristics of the Israeli healthcare system may affect study's generalizability. The high hospital‐bed occupancy rate, stretched to the limit at 99%, which is much higher than in other developed countries,[37] may affect readmission rates and risk. Nonetheless, our findings may be of relevance to other populations and healthcare systems, as variables included in our model have been previously shown to affect readmission risk in other settings,[4, 6] and the percent of in‐hospital ADL decline is similar to that reported by others.[9] Future studies should examine the significance of in‐hospital functioning in other older adult populations, such as greater mix of baseline functioning and myocardial infarction, HF, and COPD patients, that have been emphasized for readmission prevention by the Centers for Medicare and Medicaid Services.

CONCLUSIONS

This study shows that although both functional status and functional decline are significant predictors of readmission, in‐hospital functional decline did not contribute to the discriminative ability of the model, beyond the risk factors known at admission: malnutrition, prior hospitalizations, and being previously diagnosed with CRF or malignancy. These findings call attention to the ability to predict readmission early in the index hospitalization, to enable early intervention in targeted high‐risk patients. Nonetheless, further at‐discharge functional assessment can detect additional patients whose readmission risk changes during the index hospitalization and who should be considered for inclusion in readmission reduction interventions. As suggested in previous prediction models,[3, 38] most of the at‐admission variables examined in this study, including patient‐reported measures such as functioning, are readily available in the EHR or during the at‐admission intake.[39, 40] In settings where these assessments are not routinely performed, their implementation should be considered. These tools could be used to potentially identify patients at high risk of readmission, and accordingly, address physical function as part of routine medical care and during the acute hospitalization, and tailor adequate follow‐up care after discharge.[11]

Disclosures: This work was supported by the Israeli Science Foundation (grant number 565/08); Clalit Health Services(grant number 04‐121/2010); and the Israel National Institute for Health Policy Research (thesis scholarship number 35/2012). The funding agencies had no role in the design and conduct of this study, the analysis or interpretation of the data, or the preparation of the manuscript. The authors report no conflicts of interest.

A continuing focus on readmission prevention as a means to improve quality and reduce waste and costs has led to abundant research on the identification of factors that put older patients at risk for readmission.[1, 2] Recently, research has focused on the development of prediction tools that are based on information from electronic health records (EHR) and that enable early, at‐admission, identification of patients at high risk for readmission.[3, 4, 5] Through such identification, patients are already targeted for inclusion in readmission‐prevention interventions early in their hospital stay.[6, 7] Yet, the ability to rely on these early‐identification tools is contingent on the stability of risk during the hospital stay. Of the in‐hospital factors that can affect readmission, change in functioning has been identified as a potentially major contributor, especially in older patients.[8]

Older adults' deterioration in functioning is a common, troubling phenomenon. Approximately 20% of patients older than 70 years and hospitalized for a medical illness deteriorate in functioning during the hospitalization.[9] Recent evidence points to the contribution of functioning to 30‐day readmission risk, with studies showing that pre‐ or posthospitalization functional impairment is associated with a 1.5 to 3 times greater likelihood of readmission.[10, 11, 12, 13] Whether in‐hospital changes in functioning add to patients' baseline risk, however, is unknown. The only readmission study that examined functional decline during hospitalization was a retrospective study performed in a population of older adults receiving rehabilitation services and therefore may not be generalizable to acutely hospitalized adults.[14] To fill the current knowledge gap, we examined whether in‐hospital functional decline significantly contributes to at‐admission readmission risk factors in its ability to accurately identify high readmission risk in older adults hospitalized in internal medicine units.

METHODS

RESULTS

Table 1 presents the baseline and in‐hospital characteristics of participants with and without readmissions. The sample includes 559 community‐dwelling older adults (49% men) aged 70 to 98 years (mean age 79 years). One‐third (36%) of the participants were fully independent in ADL at admission. Eighty‐five (15.2%) patients were readmitted within 30 days of discharge. Participants who were readmitted had lower at‐admission ADL levels; had 1 more hospitalization in the previous year; and were more likely to have HF, CRF, arrhythmia, and malignancy, and to be at risk of malnutrition, than those who were not readmitted. Participants who were readmitted were more likely to suffer from functional decline during the index hospitalization. No significant differences were found in living arrangements or in at‐admission mental and cognitive status.

Multivariate analysis (Table 2) shows that higher at‐admission mBI score was associated with lower odds of readmission (OR for 1‐unit increase: 0.99, 95% CI: 0.98‐0.99). Other predictors of higher readmission risk were: high or medium at‐admission risk of malnutrition, malignancy, CRF, each additional hospitalization during the previous year, and lower albumin levels. Severity of illness and demographic characteristics were not significantly associated with readmission.

The at‐discharge model that combined the baseline model and in‐hospital risk factors showed that in‐hospital (from admission to discharge) ADL decline was significantly associated with readmission, as a 10‐point decrease in the mBI from admission to discharge was associated with 1.32 (95% CI: 1.02‐1.72) greater odds of readmission. LOS was not significantly associated with readmission, after controlling for baseline health status and in‐hospital ADL decline. All other predictors did not markedly change from the baseline to the at‐discharge model either in significance levels or in magnitude.

The discriminatory power of the baseline model was good (C statistic=0.81). Adding ADL decline and LOS did not change the discriminatory power of the model (C statistic=0.81). The P value of the Hosmer‐Lemeshow test equaled 0.67 for the baseline model and 0.48 for the at‐discharge model, indicating good calibration of both models. The P values for the regression coefficients of bootstrap inference assessing the relationship between the at‐admission and in‐hospital ADL decline odds of readmission remained stable (P 0.05).

Classification of patients into risk categories by the baseline model and the discharge model (Table 3) shows that identifying patients in the top‐tier category (20th highest percentile) according to information available before or at admission does not detect 6/111 (5.4%) of patients who would have been categorized as highest‐risk if information on ADL decline had been incorporated in the predictive algorithm. Additional partitioning of the top fifth group into 2 tiers (8089th and 9099th percentiles) shows that selection of patients in the top 10% of the baseline risk score would not have detected 7/55 (12.7%) patients who would have been identified as high risk at discharge (data not shown).

DISCUSSION

To our knowledge, ours is the first empirical test of the simultaneous role of functioning along the hospitalization course in explaining readmission risk.[8] Our results show that at‐admission lower functional status and in‐hospital functional decline are significant predictors of early unplanned readmission in older adults, beyond other well‐known risk factors.

The major purpose of this study was to examine whether at‐admission data can be used to detect high‐risk patients for potential inclusion in readmission prevention interventions, or whether changes in ADL occurring during the index hospitalization could affect patients' risk, therefore necessitating an additional assessment at discharge. Our results show that some patients would not have been detected at admission, as their in‐hospital ADL decline affects their at‐discharge risk. Nonetheless, this is a small group (only 5% of patients if a targeting threshold of the highest 20% risk is used). Our findings also show that information on ADL decline during the index hospitalization does not contribute to the accuracy of readmission‐risk prediction in a model that utilizes data on prior hospitalizations, baseline nutritional and functional status, and chronic morbidity (CRF and malignancy). Our results are consistent with previous studies showing the association between baseline,[11, 13, 32] or at‐admission[13] functional status and readmission. However, these studies did not analyze the related contribution of in‐hospital functioning to readmission risk, which was recently suggested as a feature that may significantly affect readmission risk, especially in older patients.[8]

Our findings are also congruent with those of a study in which LOS was not significantly associated with readmission in an elderly population.[33] Our null finding can be explained by the broad set of pre‐ and at‐admission variables, such as nutritional and functional status as well as in‐hospital functional decline, included in our model, making LOS a less significant contributor than in more parsimonious models.[28]

Our results also show that malnutrition contributes to readmission risk beyond other well‐known risk factors. Previous studies showed that malnutrition in the elderly is associated with early readmission.[11, 34] These studies, however, did not examine other well‐known risk factors, such as previous hospitalizations, which were tested in our study, precluding identification of the contribution of malnutrition beyond other well‐known risks.

Our findings should be interpreted in light of several limitations. First, the functional, nutritional, and cognitive data were collected from participants' self‐reports, which are prone to recall bias. Nonetheless, self‐report is often used in large‐scale studies, which preclude actual performance measurement.[21] Second, our sample is of adults aged 70 years or older, and may not be representative of the 65 and older population, which is the target population for many readmission reduction interventions.[35] Yet, participants were from a relatively high‐functioning group of patients who were discharged to their homes, thus may resemble the over age 65 years group. Moreover, these inclusion criteria may have affected their readmission rates, which at 15% are lower than the average reported in other older adult populations.[36] Nonetheless, a more heterogenic sample (in terms of baseline functional status) is needed to address the association between in‐hospital functional change and readmissions as well as the discrimination of the model. Third, the attrition rate (16%) might impact the predictive ability of the models, as patients dropped‐out from the study might have had higher in‐hospital deterioration. However, no significant differences between study sample and dropped‐out patients in the wide range of baseline characteristics except for age and baseline albumin levels were found. Fourth, the unique characteristics of the Israeli healthcare system may affect study's generalizability. The high hospital‐bed occupancy rate, stretched to the limit at 99%, which is much higher than in other developed countries,[37] may affect readmission rates and risk. Nonetheless, our findings may be of relevance to other populations and healthcare systems, as variables included in our model have been previously shown to affect readmission risk in other settings,[4, 6] and the percent of in‐hospital ADL decline is similar to that reported by others.[9] Future studies should examine the significance of in‐hospital functioning in other older adult populations, such as greater mix of baseline functioning and myocardial infarction, HF, and COPD patients, that have been emphasized for readmission prevention by the Centers for Medicare and Medicaid Services.

CONCLUSIONS

This study shows that although both functional status and functional decline are significant predictors of readmission, in‐hospital functional decline did not contribute to the discriminative ability of the model, beyond the risk factors known at admission: malnutrition, prior hospitalizations, and being previously diagnosed with CRF or malignancy. These findings call attention to the ability to predict readmission early in the index hospitalization, to enable early intervention in targeted high‐risk patients. Nonetheless, further at‐discharge functional assessment can detect additional patients whose readmission risk changes during the index hospitalization and who should be considered for inclusion in readmission reduction interventions. As suggested in previous prediction models,[3, 38] most of the at‐admission variables examined in this study, including patient‐reported measures such as functioning, are readily available in the EHR or during the at‐admission intake.[39, 40] In settings where these assessments are not routinely performed, their implementation should be considered. These tools could be used to potentially identify patients at high risk of readmission, and accordingly, address physical function as part of routine medical care and during the acute hospitalization, and tailor adequate follow‐up care after discharge.[11]

Disclosures: This work was supported by the Israeli Science Foundation (grant number 565/08); Clalit Health Services(grant number 04‐121/2010); and the Israel National Institute for Health Policy Research (thesis scholarship number 35/2012). The funding agencies had no role in the design and conduct of this study, the analysis or interpretation of the data, or the preparation of the manuscript. The authors report no conflicts of interest.

A continuing focus on readmission prevention as a means to improve quality and reduce waste and costs has led to abundant research on the identification of factors that put older patients at risk for readmission.[1, 2] Recently, research has focused on the development of prediction tools that are based on information from electronic health records (EHR) and that enable early, at‐admission, identification of patients at high risk for readmission.[3, 4, 5] Through such identification, patients are already targeted for inclusion in readmission‐prevention interventions early in their hospital stay.[6, 7] Yet, the ability to rely on these early‐identification tools is contingent on the stability of risk during the hospital stay. Of the in‐hospital factors that can affect readmission, change in functioning has been identified as a potentially major contributor, especially in older patients.[8]

Older adults' deterioration in functioning is a common, troubling phenomenon. Approximately 20% of patients older than 70 years and hospitalized for a medical illness deteriorate in functioning during the hospitalization.[9] Recent evidence points to the contribution of functioning to 30‐day readmission risk, with studies showing that pre‐ or posthospitalization functional impairment is associated with a 1.5 to 3 times greater likelihood of readmission.[10, 11, 12, 13] Whether in‐hospital changes in functioning add to patients' baseline risk, however, is unknown. The only readmission study that examined functional decline during hospitalization was a retrospective study performed in a population of older adults receiving rehabilitation services and therefore may not be generalizable to acutely hospitalized adults.[14] To fill the current knowledge gap, we examined whether in‐hospital functional decline significantly contributes to at‐admission readmission risk factors in its ability to accurately identify high readmission risk in older adults hospitalized in internal medicine units.

METHODS

RESULTS

Table 1 presents the baseline and in‐hospital characteristics of participants with and without readmissions. The sample includes 559 community‐dwelling older adults (49% men) aged 70 to 98 years (mean age 79 years). One‐third (36%) of the participants were fully independent in ADL at admission. Eighty‐five (15.2%) patients were readmitted within 30 days of discharge. Participants who were readmitted had lower at‐admission ADL levels; had 1 more hospitalization in the previous year; and were more likely to have HF, CRF, arrhythmia, and malignancy, and to be at risk of malnutrition, than those who were not readmitted. Participants who were readmitted were more likely to suffer from functional decline during the index hospitalization. No significant differences were found in living arrangements or in at‐admission mental and cognitive status.

Multivariate analysis (Table 2) shows that higher at‐admission mBI score was associated with lower odds of readmission (OR for 1‐unit increase: 0.99, 95% CI: 0.98‐0.99). Other predictors of higher readmission risk were: high or medium at‐admission risk of malnutrition, malignancy, CRF, each additional hospitalization during the previous year, and lower albumin levels. Severity of illness and demographic characteristics were not significantly associated with readmission.

The at‐discharge model that combined the baseline model and in‐hospital risk factors showed that in‐hospital (from admission to discharge) ADL decline was significantly associated with readmission, as a 10‐point decrease in the mBI from admission to discharge was associated with 1.32 (95% CI: 1.02‐1.72) greater odds of readmission. LOS was not significantly associated with readmission, after controlling for baseline health status and in‐hospital ADL decline. All other predictors did not markedly change from the baseline to the at‐discharge model either in significance levels or in magnitude.

The discriminatory power of the baseline model was good (C statistic=0.81). Adding ADL decline and LOS did not change the discriminatory power of the model (C statistic=0.81). The P value of the Hosmer‐Lemeshow test equaled 0.67 for the baseline model and 0.48 for the at‐discharge model, indicating good calibration of both models. The P values for the regression coefficients of bootstrap inference assessing the relationship between the at‐admission and in‐hospital ADL decline odds of readmission remained stable (P 0.05).

Classification of patients into risk categories by the baseline model and the discharge model (Table 3) shows that identifying patients in the top‐tier category (20th highest percentile) according to information available before or at admission does not detect 6/111 (5.4%) of patients who would have been categorized as highest‐risk if information on ADL decline had been incorporated in the predictive algorithm. Additional partitioning of the top fifth group into 2 tiers (8089th and 9099th percentiles) shows that selection of patients in the top 10% of the baseline risk score would not have detected 7/55 (12.7%) patients who would have been identified as high risk at discharge (data not shown).

DISCUSSION

To our knowledge, ours is the first empirical test of the simultaneous role of functioning along the hospitalization course in explaining readmission risk.[8] Our results show that at‐admission lower functional status and in‐hospital functional decline are significant predictors of early unplanned readmission in older adults, beyond other well‐known risk factors.

The major purpose of this study was to examine whether at‐admission data can be used to detect high‐risk patients for potential inclusion in readmission prevention interventions, or whether changes in ADL occurring during the index hospitalization could affect patients' risk, therefore necessitating an additional assessment at discharge. Our results show that some patients would not have been detected at admission, as their in‐hospital ADL decline affects their at‐discharge risk. Nonetheless, this is a small group (only 5% of patients if a targeting threshold of the highest 20% risk is used). Our findings also show that information on ADL decline during the index hospitalization does not contribute to the accuracy of readmission‐risk prediction in a model that utilizes data on prior hospitalizations, baseline nutritional and functional status, and chronic morbidity (CRF and malignancy). Our results are consistent with previous studies showing the association between baseline,[11, 13, 32] or at‐admission[13] functional status and readmission. However, these studies did not analyze the related contribution of in‐hospital functioning to readmission risk, which was recently suggested as a feature that may significantly affect readmission risk, especially in older patients.[8]

Our findings are also congruent with those of a study in which LOS was not significantly associated with readmission in an elderly population.[33] Our null finding can be explained by the broad set of pre‐ and at‐admission variables, such as nutritional and functional status as well as in‐hospital functional decline, included in our model, making LOS a less significant contributor than in more parsimonious models.[28]

Our results also show that malnutrition contributes to readmission risk beyond other well‐known risk factors. Previous studies showed that malnutrition in the elderly is associated with early readmission.[11, 34] These studies, however, did not examine other well‐known risk factors, such as previous hospitalizations, which were tested in our study, precluding identification of the contribution of malnutrition beyond other well‐known risks.

Our findings should be interpreted in light of several limitations. First, the functional, nutritional, and cognitive data were collected from participants' self‐reports, which are prone to recall bias. Nonetheless, self‐report is often used in large‐scale studies, which preclude actual performance measurement.[21] Second, our sample is of adults aged 70 years or older, and may not be representative of the 65 and older population, which is the target population for many readmission reduction interventions.[35] Yet, participants were from a relatively high‐functioning group of patients who were discharged to their homes, thus may resemble the over age 65 years group. Moreover, these inclusion criteria may have affected their readmission rates, which at 15% are lower than the average reported in other older adult populations.[36] Nonetheless, a more heterogenic sample (in terms of baseline functional status) is needed to address the association between in‐hospital functional change and readmissions as well as the discrimination of the model. Third, the attrition rate (16%) might impact the predictive ability of the models, as patients dropped‐out from the study might have had higher in‐hospital deterioration. However, no significant differences between study sample and dropped‐out patients in the wide range of baseline characteristics except for age and baseline albumin levels were found. Fourth, the unique characteristics of the Israeli healthcare system may affect study's generalizability. The high hospital‐bed occupancy rate, stretched to the limit at 99%, which is much higher than in other developed countries,[37] may affect readmission rates and risk. Nonetheless, our findings may be of relevance to other populations and healthcare systems, as variables included in our model have been previously shown to affect readmission risk in other settings,[4, 6] and the percent of in‐hospital ADL decline is similar to that reported by others.[9] Future studies should examine the significance of in‐hospital functioning in other older adult populations, such as greater mix of baseline functioning and myocardial infarction, HF, and COPD patients, that have been emphasized for readmission prevention by the Centers for Medicare and Medicaid Services.

CONCLUSIONS

This study shows that although both functional status and functional decline are significant predictors of readmission, in‐hospital functional decline did not contribute to the discriminative ability of the model, beyond the risk factors known at admission: malnutrition, prior hospitalizations, and being previously diagnosed with CRF or malignancy. These findings call attention to the ability to predict readmission early in the index hospitalization, to enable early intervention in targeted high‐risk patients. Nonetheless, further at‐discharge functional assessment can detect additional patients whose readmission risk changes during the index hospitalization and who should be considered for inclusion in readmission reduction interventions. As suggested in previous prediction models,[3, 38] most of the at‐admission variables examined in this study, including patient‐reported measures such as functioning, are readily available in the EHR or during the at‐admission intake.[39, 40] In settings where these assessments are not routinely performed, their implementation should be considered. These tools could be used to potentially identify patients at high risk of readmission, and accordingly, address physical function as part of routine medical care and during the acute hospitalization, and tailor adequate follow‐up care after discharge.[11]

Disclosures: This work was supported by the Israeli Science Foundation (grant number 565/08); Clalit Health Services(grant number 04‐121/2010); and the Israel National Institute for Health Policy Research (thesis scholarship number 35/2012). The funding agencies had no role in the design and conduct of this study, the analysis or interpretation of the data, or the preparation of the manuscript. The authors report no conflicts of interest.

Antipsychotic medications are frequently used off label for management of behavioral symptoms associated with delirium and/or dementia. Despite regulations designed to curb inappropriate prescribing of these medications in nursing homes, substantial levels of use and variation in use have been observed in this setting.[1] Although antipsychotic medications are also frequently used in the hospital, the scope and variation in use have not been adequately investigated. Given the lack of oversight for medication prescribing in the hospital setting and the frequency of delirium, occurring in 15% to 26% of hospitalized older adults,[2, 3, 4] off‐label use of antipsychotic medications and variation in use could be substantial.

Because variation in practice is known to increase in the setting of controversy or lack of clarity regarding appropriate management,[5] large degrees of variation can draw attention to priority areas for clinical effectiveness studies, and the need for guidelines, clinical decision support, or regulatory oversight. In the absence of clear guidelines for the use of antipsychotic medication in nonpsychiatric hospitalized patients, we hypothesized that significant variation in use would persist after controlling for patient characteristics. Using a large, nationally representative cohort of admissions to 300 hospitals from July 2009 to June 2010, we sought to investigate prescribing patterns and hospital variation in use of antipsychotic medications in nonpsychiatric admissions to US hospitals.

METHODS

RESULTS

Hospital Variation in Antipsychotic Use

Figure 1 demonstrates the antipsychotic prescribing rate at each hospital in our cohort, and the corresponding quintiles. Patients admitted to hospitals in the highest prescribing quintile were more than twice as likely to be exposed to antipsychotics compared to patients admitted to hospitals in the lowest prescribing quintile, even after adjustment for patient characteristics and comorbidities (Table 4). This relationship was similar across subgroups of admissions with delirium and dementia (see Supporting Information, Appendix, in the online version of this article for the distribution of hospital antipsychotic prescribing rates before and after adjustment for patient characteristics).

Table 4.Relative Risk of Antipsychotic Exposure for Admissions Within Each Hospital Prescribing Quintile, Overall, and Stratified by Delirium and Dementia

	Admissions, No. (% of Total)	Unadjusted RR of Exposure [95% CI]	Adjusted RR of exposure [95% CI]*
NOTE: Abbreviations: CI, confidence interval; Q, quintile; RR, relative risk. *Adjusted for all patient characteristics listed in Table 1 (except discharge disposition), as well as the Elixhauser comorbidities, using a generalized estimating equation with a Poisson error term, a log link, and an exchangeable correlation structure to account for repeated admissions of the same patient during the study. Antipsychotic prescribing rates for hospitals within each quintile were: Q1: 4.6%, Q2: 4.6%5.6%, Q3: 5.7%6.3%, Q4: 6.4%7.4%, Q5: >7.4%.
Overall
Q1	431,017 (16%)	Reference	Reference
Q2	630,486 (23%)	1.67 [1.63, 1.71]	1.59 [1.55, 1.62]
Q3	548,337 (20%)	1.93 [1.88, 1.97]	1.84 [1.80, 1.88]
Q4	639,027 (24%)	2.16 [2.12, 2.21]	2.07 [2.03, 2.12]
Q5	446,214 (17%)	2.83 [2.77, 2.89]	2.56 [2.50, 2.61]
Delirium
Q1	12,878 (15%)	Reference	Reference
Q2	20,588 (24%)	1.58 [1.51, 1.65]	1.58 [1.51, 1.65]
Q3	17,402 (20%)	1.71 [1.64, 1.80]	1.73 [1.65, 1.82]
Q4	20,943 (24%)	2.01 [1.92, 2.10]	1.99 [1.91, 2.08]
Q5	14,883 (17%)	2.15 [2.05, 2.25]	2.16 [2.07, 2.26]
Dementia
Q1	28,290 (15%)	Reference	Reference
Q2	42,018 (22%)	1.43 [1.36, 1.50]	1.40 [1.34, 1.47]
Q3	38,593 (21%)	1.61 [1.53, 1.69]	1.59 [1.51, 1.66]
Q4	44,638 (24%)	1.69 [1.62, 1.77]	1.69 [1.61, 1.77]
Q5	34,442 (18%)	1.92 [1.83, 2.01]	1.90 [1.81, 1.99]

Figure 1

DISCUSSION

In this cohort of nonpsychiatric admissions to 300 US hospitals, antipsychotic medications were used in 6% of admissions, with atypical antipsychotics representing the majority of use. Potentially excessive daily doses based on CMS recommendations for long‐term care facilities occurred in almost half of admissions with any antipsychotic exposure, and in 87% of admissions with haloperidol exposure specifically. We found variation in hospital use of antipsychotics that was not fully accounted for by measured patient characteristics, and which persisted among subgroups of admissions with delirium and/or dementia. Although unmeasured patient characteristics or different billing practices between hospitals are potential explanations, our findings also raise the possibility of different hospital antipsychotic prescribing cultures. These findings provide new information regarding the scope of prescribing in US hospitals, and draw attention to the need for additional studies to better define what constitutes appropriate use of antipsychotics in the hospital setting.

A recent single‐center study at a large academic medical center found an overall antipsychotic exposure rate of 9% of nonpsychiatric admissions.[12] Our finding that 6% of admissions in this multicenter cohort were exposed to antipsychotics is slightly lower, but similar to the previous estimate. Assuming 37 million discharges from US hospitals each year,[13] our study suggests that more than 2 million hospitalized patients receive antipsychotics annually. With around 1.4 million residents in nursing homes on any given day,[14] and an exposure rate of 25% to 30% in that setting,[15, 16, 17] our study suggests that the number of patients exposed in the hospital setting is greater than the number exposed in the nursing home setting, the site of care for which prescribing regulations have been focused thus far.

Because our dataset does not contain preadmission medications, we were unable to specifically investigate new initiation. In the prior single‐center study, approximately 55% of overall use in the hospital setting was new initiation,[12] which would suggest that antipsychotics are newly initiated in around 1 million admissions each year in the hospital. Although we are unable to determine reason for use in our analysis, delirium was a strong predictor of antipsychotic use in our multivariable model, and prior studies have demonstrated delirium to be the most common reason for antipsychotic initiation in hospitalized patients,[12, 18] an indication for which efficacy/effectiveness data are lacking. A recent systematic review of antipsychotics for the treatment of delirium in older adults concluded that because of severe methodological limitations, the small number of existing studies on this topic do not support the use of antipsychotics in the treatment of delirium in older hospitalized adults.[19] Our results further highlight the need for randomized placebo‐controlled trials of antipsychotics in treatment of delirium.

We found variation in antipsychotic use between hospitals that was not fully explained by patient characteristics. Insufficient data to inform clinical decisions surrounding management of agitated delirium/dementia and lack of clear criteria by which to judge appropriateness of antipsychotic use may contribute to this variation. Some variation may relate to resource allocation at different hospitals, and the feasibility of implementing nonpharmacologic management options across settings. Our results collectively highlight the need for studies evaluating the efficacy/effectiveness of antipsychotics in the treatment of delirium and drivers of physician decision‐making in this realm, as well as the need for greater hospital investment in nonpharmacologic delirium‐prevention programs, which have been shown to be effective in prevention of delirium in hospitalized patients.[20]

We observed high levels of potentially excessive daily dosing using cutoffs applied in the long‐term care setting. The majority of the potentially excessive doses were in the setting of typical antipsychotic use, and haloperidol specifically, where doses exceeded 4 mg on at least 1 day in 87% of exposed admissions. Of note, the threshold for haloperidol dosage above which justification is required was decreased from 4 to 2 mg per day in the 2015 update to the CMS guidelines.[21] For the present analysis, we used the guidelines that were contemporaneous to our cohort; we are unable to determine current rates of potentially inappropriate dosages in the present analysis, but given the high prevalence in 2009 to 2010, and the lowering of the dosage threshold since then, it is unlikely that any decrease in use would be enough to substantially reduce the estimate. Whether these high dosages are actually inappropriate in the hospital setting is not established, and we were not able to review medical records to determine whether justification for use of such doses was documented.[22, 23] It is possible that hospitalized patients with altered pharmacodynamics and greater severity of illness could require larger doses of these medications; however, this is an area in need of further investigation, and current critical care guidelines note the lack of sufficient data upon which to justify use of haloperidol in the prevention or treatment of delirium in ICU patients.[24, 25]

The dosages in use are concerning given that the risk of extrapyramidal side effects increases with increasing dose, and prior studies have demonstrated an association between increased dose of antipsychotics and increased risk of other adverse events, including hip fracture and sudden cardiac death.[22, 23] Further, despite these known risks, studies have demonstrated failure to follow recommendations to mitigate risk,[26] such as electrocardiogram monitoring in individuals receiving intravenous haloperidol.[27] Our results suggest that physicians are similarly not following recommendations to use lower doses of haloperidol when treating older patients, given the almost identical incidence of potentially excessive dosing among admissions of patients aged 65 years and older in our cohort.[25] Clinical decision support prompts have been effective at increasing appropriate use of antipsychotic medications in several single‐center analyses,[28, 29, 30] and widespread implementation of such support with a focus on haloperidol dosing should be considered on the basis of our results.

The patient characteristics associated with antipsychotic use in this large, nationally representative analysis are consistent with those identified in prior single‐center analyses.[12, 18] Both prior analyses identified delirium as the most common reason for antipsychotic use, and dementia, intensive care unit stay, and mechanical ventilation were also previously identified as strong predictors of use that we believe hold face validity for the practicing hospitalist. On the other hand, some of the factors associated with antipsychotic use in our model cannot be readily explained, such as insurance status and race, and may be serving as proxies for other variables not included in our analysis. That nonwhite patients are less likely than white patients to receive antipsychotic medications in the hospital has been previously demonstrated,[12] and further investigation to understand this disparity is warranted.

Our study has several additional limitations. First, because our study is observational, the possibility of residual confounding exists, and we cannot rule out that there are other patient factors driving the hospital variation in antipsychotic use that we observed. Second, because guidelines do not exist for antipsychotic dosing in hospitalized patients, we could only comment on potentially excessive dosing, extrapolating from guidelines in the long‐term care setting. Whether such doses are actually excessive in hospitalized patients is not defined. Third, although Premier performs quality checks on charge and ICD‐9‐CM coding data submitted by participating hospitals, the validity of administrative data is uncertain. For example, the use of administrative data to identify delirium diagnoses is likely to have resulted in underestimation of delirium incidence among our different exposure groups. Delirium is likely to be coded more often in the setting of more severe or hyperactive cases, when antipsychotics are more likely to be utilized. This could result in an overestimation of the association between delirium and antipsychotic use. Additionally, differences in coding practices between hospitals for any of the variables in our models could explain some of the variation in antipsychotic prescribing that we observed. Finally, because we were unable to differentiate between new initiation and continuation of a preadmission antipsychotic, some of the variation that we observed is likely to reflect differences in outpatient antipsychotic prescribing practices.

In conclusion, in this large cohort of nonpsychiatric admissions to 300 US hospitals, we found that antipsychotic medication exposure was common, often at high daily doses. Delirium and dementia were the strongest predictors of use among the patient and hospital characteristics examined. The variation in antipsychotic prescribing that we observed was not fully accounted for by measured patient characteristics, and raises the possibility of differing hospital prescribing cultures. Our results draw attention to the need for additional research to better define what constitutes appropriate use of these potentially harmful medications in the hospital setting.

Disclosures: Dr. Herzig had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Drs. Herzig, Rothberg, Gurwitz, and Marcantonio. Acquisition of data: Dr. Herzig. Analysis of data: Mr. Guess. Interpretation of data: Drs. Herzig, Rothberg, Gurwitz, Marcantonio, and Mr. Guess. Drafting of the manuscript: Dr. Herzig. Critical revision of the manuscript for important intellectual content: Drs. Rothberg, Gurwitz, Marcantonio, and Mr. Guess. Dr. Herzig was funded by grant number K23AG042459 from the National Institute on Aging. Dr. Marcantonio was funded by grant number K24AG035075 from the National Institute on Aging. The funding organizations had no involvement in any aspect of the study, including design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript. The authors report no conflicts of interest.

Antipsychotic medications are frequently used off label for management of behavioral symptoms associated with delirium and/or dementia. Despite regulations designed to curb inappropriate prescribing of these medications in nursing homes, substantial levels of use and variation in use have been observed in this setting.[1] Although antipsychotic medications are also frequently used in the hospital, the scope and variation in use have not been adequately investigated. Given the lack of oversight for medication prescribing in the hospital setting and the frequency of delirium, occurring in 15% to 26% of hospitalized older adults,[2, 3, 4] off‐label use of antipsychotic medications and variation in use could be substantial.

Because variation in practice is known to increase in the setting of controversy or lack of clarity regarding appropriate management,[5] large degrees of variation can draw attention to priority areas for clinical effectiveness studies, and the need for guidelines, clinical decision support, or regulatory oversight. In the absence of clear guidelines for the use of antipsychotic medication in nonpsychiatric hospitalized patients, we hypothesized that significant variation in use would persist after controlling for patient characteristics. Using a large, nationally representative cohort of admissions to 300 hospitals from July 2009 to June 2010, we sought to investigate prescribing patterns and hospital variation in use of antipsychotic medications in nonpsychiatric admissions to US hospitals.

METHODS

RESULTS

Hospital Variation in Antipsychotic Use

Figure 1 demonstrates the antipsychotic prescribing rate at each hospital in our cohort, and the corresponding quintiles. Patients admitted to hospitals in the highest prescribing quintile were more than twice as likely to be exposed to antipsychotics compared to patients admitted to hospitals in the lowest prescribing quintile, even after adjustment for patient characteristics and comorbidities (Table 4). This relationship was similar across subgroups of admissions with delirium and dementia (see Supporting Information, Appendix, in the online version of this article for the distribution of hospital antipsychotic prescribing rates before and after adjustment for patient characteristics).

Table 4.Relative Risk of Antipsychotic Exposure for Admissions Within Each Hospital Prescribing Quintile, Overall, and Stratified by Delirium and Dementia

	Admissions, No. (% of Total)	Unadjusted RR of Exposure [95% CI]	Adjusted RR of exposure [95% CI]*
NOTE: Abbreviations: CI, confidence interval; Q, quintile; RR, relative risk. *Adjusted for all patient characteristics listed in Table 1 (except discharge disposition), as well as the Elixhauser comorbidities, using a generalized estimating equation with a Poisson error term, a log link, and an exchangeable correlation structure to account for repeated admissions of the same patient during the study. Antipsychotic prescribing rates for hospitals within each quintile were: Q1: 4.6%, Q2: 4.6%5.6%, Q3: 5.7%6.3%, Q4: 6.4%7.4%, Q5: >7.4%.
Overall
Q1	431,017 (16%)	Reference	Reference
Q2	630,486 (23%)	1.67 [1.63, 1.71]	1.59 [1.55, 1.62]
Q3	548,337 (20%)	1.93 [1.88, 1.97]	1.84 [1.80, 1.88]
Q4	639,027 (24%)	2.16 [2.12, 2.21]	2.07 [2.03, 2.12]
Q5	446,214 (17%)	2.83 [2.77, 2.89]	2.56 [2.50, 2.61]
Delirium
Q1	12,878 (15%)	Reference	Reference
Q2	20,588 (24%)	1.58 [1.51, 1.65]	1.58 [1.51, 1.65]
Q3	17,402 (20%)	1.71 [1.64, 1.80]	1.73 [1.65, 1.82]
Q4	20,943 (24%)	2.01 [1.92, 2.10]	1.99 [1.91, 2.08]
Q5	14,883 (17%)	2.15 [2.05, 2.25]	2.16 [2.07, 2.26]
Dementia
Q1	28,290 (15%)	Reference	Reference
Q2	42,018 (22%)	1.43 [1.36, 1.50]	1.40 [1.34, 1.47]
Q3	38,593 (21%)	1.61 [1.53, 1.69]	1.59 [1.51, 1.66]
Q4	44,638 (24%)	1.69 [1.62, 1.77]	1.69 [1.61, 1.77]
Q5	34,442 (18%)	1.92 [1.83, 2.01]	1.90 [1.81, 1.99]

Figure 1

DISCUSSION

In this cohort of nonpsychiatric admissions to 300 US hospitals, antipsychotic medications were used in 6% of admissions, with atypical antipsychotics representing the majority of use. Potentially excessive daily doses based on CMS recommendations for long‐term care facilities occurred in almost half of admissions with any antipsychotic exposure, and in 87% of admissions with haloperidol exposure specifically. We found variation in hospital use of antipsychotics that was not fully accounted for by measured patient characteristics, and which persisted among subgroups of admissions with delirium and/or dementia. Although unmeasured patient characteristics or different billing practices between hospitals are potential explanations, our findings also raise the possibility of different hospital antipsychotic prescribing cultures. These findings provide new information regarding the scope of prescribing in US hospitals, and draw attention to the need for additional studies to better define what constitutes appropriate use of antipsychotics in the hospital setting.

A recent single‐center study at a large academic medical center found an overall antipsychotic exposure rate of 9% of nonpsychiatric admissions.[12] Our finding that 6% of admissions in this multicenter cohort were exposed to antipsychotics is slightly lower, but similar to the previous estimate. Assuming 37 million discharges from US hospitals each year,[13] our study suggests that more than 2 million hospitalized patients receive antipsychotics annually. With around 1.4 million residents in nursing homes on any given day,[14] and an exposure rate of 25% to 30% in that setting,[15, 16, 17] our study suggests that the number of patients exposed in the hospital setting is greater than the number exposed in the nursing home setting, the site of care for which prescribing regulations have been focused thus far.

Because our dataset does not contain preadmission medications, we were unable to specifically investigate new initiation. In the prior single‐center study, approximately 55% of overall use in the hospital setting was new initiation,[12] which would suggest that antipsychotics are newly initiated in around 1 million admissions each year in the hospital. Although we are unable to determine reason for use in our analysis, delirium was a strong predictor of antipsychotic use in our multivariable model, and prior studies have demonstrated delirium to be the most common reason for antipsychotic initiation in hospitalized patients,[12, 18] an indication for which efficacy/effectiveness data are lacking. A recent systematic review of antipsychotics for the treatment of delirium in older adults concluded that because of severe methodological limitations, the small number of existing studies on this topic do not support the use of antipsychotics in the treatment of delirium in older hospitalized adults.[19] Our results further highlight the need for randomized placebo‐controlled trials of antipsychotics in treatment of delirium.

We found variation in antipsychotic use between hospitals that was not fully explained by patient characteristics. Insufficient data to inform clinical decisions surrounding management of agitated delirium/dementia and lack of clear criteria by which to judge appropriateness of antipsychotic use may contribute to this variation. Some variation may relate to resource allocation at different hospitals, and the feasibility of implementing nonpharmacologic management options across settings. Our results collectively highlight the need for studies evaluating the efficacy/effectiveness of antipsychotics in the treatment of delirium and drivers of physician decision‐making in this realm, as well as the need for greater hospital investment in nonpharmacologic delirium‐prevention programs, which have been shown to be effective in prevention of delirium in hospitalized patients.[20]

We observed high levels of potentially excessive daily dosing using cutoffs applied in the long‐term care setting. The majority of the potentially excessive doses were in the setting of typical antipsychotic use, and haloperidol specifically, where doses exceeded 4 mg on at least 1 day in 87% of exposed admissions. Of note, the threshold for haloperidol dosage above which justification is required was decreased from 4 to 2 mg per day in the 2015 update to the CMS guidelines.[21] For the present analysis, we used the guidelines that were contemporaneous to our cohort; we are unable to determine current rates of potentially inappropriate dosages in the present analysis, but given the high prevalence in 2009 to 2010, and the lowering of the dosage threshold since then, it is unlikely that any decrease in use would be enough to substantially reduce the estimate. Whether these high dosages are actually inappropriate in the hospital setting is not established, and we were not able to review medical records to determine whether justification for use of such doses was documented.[22, 23] It is possible that hospitalized patients with altered pharmacodynamics and greater severity of illness could require larger doses of these medications; however, this is an area in need of further investigation, and current critical care guidelines note the lack of sufficient data upon which to justify use of haloperidol in the prevention or treatment of delirium in ICU patients.[24, 25]

The dosages in use are concerning given that the risk of extrapyramidal side effects increases with increasing dose, and prior studies have demonstrated an association between increased dose of antipsychotics and increased risk of other adverse events, including hip fracture and sudden cardiac death.[22, 23] Further, despite these known risks, studies have demonstrated failure to follow recommendations to mitigate risk,[26] such as electrocardiogram monitoring in individuals receiving intravenous haloperidol.[27] Our results suggest that physicians are similarly not following recommendations to use lower doses of haloperidol when treating older patients, given the almost identical incidence of potentially excessive dosing among admissions of patients aged 65 years and older in our cohort.[25] Clinical decision support prompts have been effective at increasing appropriate use of antipsychotic medications in several single‐center analyses,[28, 29, 30] and widespread implementation of such support with a focus on haloperidol dosing should be considered on the basis of our results.

The patient characteristics associated with antipsychotic use in this large, nationally representative analysis are consistent with those identified in prior single‐center analyses.[12, 18] Both prior analyses identified delirium as the most common reason for antipsychotic use, and dementia, intensive care unit stay, and mechanical ventilation were also previously identified as strong predictors of use that we believe hold face validity for the practicing hospitalist. On the other hand, some of the factors associated with antipsychotic use in our model cannot be readily explained, such as insurance status and race, and may be serving as proxies for other variables not included in our analysis. That nonwhite patients are less likely than white patients to receive antipsychotic medications in the hospital has been previously demonstrated,[12] and further investigation to understand this disparity is warranted.

Our study has several additional limitations. First, because our study is observational, the possibility of residual confounding exists, and we cannot rule out that there are other patient factors driving the hospital variation in antipsychotic use that we observed. Second, because guidelines do not exist for antipsychotic dosing in hospitalized patients, we could only comment on potentially excessive dosing, extrapolating from guidelines in the long‐term care setting. Whether such doses are actually excessive in hospitalized patients is not defined. Third, although Premier performs quality checks on charge and ICD‐9‐CM coding data submitted by participating hospitals, the validity of administrative data is uncertain. For example, the use of administrative data to identify delirium diagnoses is likely to have resulted in underestimation of delirium incidence among our different exposure groups. Delirium is likely to be coded more often in the setting of more severe or hyperactive cases, when antipsychotics are more likely to be utilized. This could result in an overestimation of the association between delirium and antipsychotic use. Additionally, differences in coding practices between hospitals for any of the variables in our models could explain some of the variation in antipsychotic prescribing that we observed. Finally, because we were unable to differentiate between new initiation and continuation of a preadmission antipsychotic, some of the variation that we observed is likely to reflect differences in outpatient antipsychotic prescribing practices.

In conclusion, in this large cohort of nonpsychiatric admissions to 300 US hospitals, we found that antipsychotic medication exposure was common, often at high daily doses. Delirium and dementia were the strongest predictors of use among the patient and hospital characteristics examined. The variation in antipsychotic prescribing that we observed was not fully accounted for by measured patient characteristics, and raises the possibility of differing hospital prescribing cultures. Our results draw attention to the need for additional research to better define what constitutes appropriate use of these potentially harmful medications in the hospital setting.

Disclosures: Dr. Herzig had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Drs. Herzig, Rothberg, Gurwitz, and Marcantonio. Acquisition of data: Dr. Herzig. Analysis of data: Mr. Guess. Interpretation of data: Drs. Herzig, Rothberg, Gurwitz, Marcantonio, and Mr. Guess. Drafting of the manuscript: Dr. Herzig. Critical revision of the manuscript for important intellectual content: Drs. Rothberg, Gurwitz, Marcantonio, and Mr. Guess. Dr. Herzig was funded by grant number K23AG042459 from the National Institute on Aging. Dr. Marcantonio was funded by grant number K24AG035075 from the National Institute on Aging. The funding organizations had no involvement in any aspect of the study, including design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript. The authors report no conflicts of interest.

Antipsychotic medications are frequently used off label for management of behavioral symptoms associated with delirium and/or dementia. Despite regulations designed to curb inappropriate prescribing of these medications in nursing homes, substantial levels of use and variation in use have been observed in this setting.[1] Although antipsychotic medications are also frequently used in the hospital, the scope and variation in use have not been adequately investigated. Given the lack of oversight for medication prescribing in the hospital setting and the frequency of delirium, occurring in 15% to 26% of hospitalized older adults,[2, 3, 4] off‐label use of antipsychotic medications and variation in use could be substantial.

Because variation in practice is known to increase in the setting of controversy or lack of clarity regarding appropriate management,[5] large degrees of variation can draw attention to priority areas for clinical effectiveness studies, and the need for guidelines, clinical decision support, or regulatory oversight. In the absence of clear guidelines for the use of antipsychotic medication in nonpsychiatric hospitalized patients, we hypothesized that significant variation in use would persist after controlling for patient characteristics. Using a large, nationally representative cohort of admissions to 300 hospitals from July 2009 to June 2010, we sought to investigate prescribing patterns and hospital variation in use of antipsychotic medications in nonpsychiatric admissions to US hospitals.

METHODS

RESULTS

Hospital Variation in Antipsychotic Use

Figure 1 demonstrates the antipsychotic prescribing rate at each hospital in our cohort, and the corresponding quintiles. Patients admitted to hospitals in the highest prescribing quintile were more than twice as likely to be exposed to antipsychotics compared to patients admitted to hospitals in the lowest prescribing quintile, even after adjustment for patient characteristics and comorbidities (Table 4). This relationship was similar across subgroups of admissions with delirium and dementia (see Supporting Information, Appendix, in the online version of this article for the distribution of hospital antipsychotic prescribing rates before and after adjustment for patient characteristics).

Table 4.Relative Risk of Antipsychotic Exposure for Admissions Within Each Hospital Prescribing Quintile, Overall, and Stratified by Delirium and Dementia

	Admissions, No. (% of Total)	Unadjusted RR of Exposure [95% CI]	Adjusted RR of exposure [95% CI]*
NOTE: Abbreviations: CI, confidence interval; Q, quintile; RR, relative risk. *Adjusted for all patient characteristics listed in Table 1 (except discharge disposition), as well as the Elixhauser comorbidities, using a generalized estimating equation with a Poisson error term, a log link, and an exchangeable correlation structure to account for repeated admissions of the same patient during the study. Antipsychotic prescribing rates for hospitals within each quintile were: Q1: 4.6%, Q2: 4.6%5.6%, Q3: 5.7%6.3%, Q4: 6.4%7.4%, Q5: >7.4%.
Overall
Q1	431,017 (16%)	Reference	Reference
Q2	630,486 (23%)	1.67 [1.63, 1.71]	1.59 [1.55, 1.62]
Q3	548,337 (20%)	1.93 [1.88, 1.97]	1.84 [1.80, 1.88]
Q4	639,027 (24%)	2.16 [2.12, 2.21]	2.07 [2.03, 2.12]
Q5	446,214 (17%)	2.83 [2.77, 2.89]	2.56 [2.50, 2.61]
Delirium
Q1	12,878 (15%)	Reference	Reference
Q2	20,588 (24%)	1.58 [1.51, 1.65]	1.58 [1.51, 1.65]
Q3	17,402 (20%)	1.71 [1.64, 1.80]	1.73 [1.65, 1.82]
Q4	20,943 (24%)	2.01 [1.92, 2.10]	1.99 [1.91, 2.08]
Q5	14,883 (17%)	2.15 [2.05, 2.25]	2.16 [2.07, 2.26]
Dementia
Q1	28,290 (15%)	Reference	Reference
Q2	42,018 (22%)	1.43 [1.36, 1.50]	1.40 [1.34, 1.47]
Q3	38,593 (21%)	1.61 [1.53, 1.69]	1.59 [1.51, 1.66]
Q4	44,638 (24%)	1.69 [1.62, 1.77]	1.69 [1.61, 1.77]
Q5	34,442 (18%)	1.92 [1.83, 2.01]	1.90 [1.81, 1.99]

Figure 1

DISCUSSION

In this cohort of nonpsychiatric admissions to 300 US hospitals, antipsychotic medications were used in 6% of admissions, with atypical antipsychotics representing the majority of use. Potentially excessive daily doses based on CMS recommendations for long‐term care facilities occurred in almost half of admissions with any antipsychotic exposure, and in 87% of admissions with haloperidol exposure specifically. We found variation in hospital use of antipsychotics that was not fully accounted for by measured patient characteristics, and which persisted among subgroups of admissions with delirium and/or dementia. Although unmeasured patient characteristics or different billing practices between hospitals are potential explanations, our findings also raise the possibility of different hospital antipsychotic prescribing cultures. These findings provide new information regarding the scope of prescribing in US hospitals, and draw attention to the need for additional studies to better define what constitutes appropriate use of antipsychotics in the hospital setting.

A recent single‐center study at a large academic medical center found an overall antipsychotic exposure rate of 9% of nonpsychiatric admissions.[12] Our finding that 6% of admissions in this multicenter cohort were exposed to antipsychotics is slightly lower, but similar to the previous estimate. Assuming 37 million discharges from US hospitals each year,[13] our study suggests that more than 2 million hospitalized patients receive antipsychotics annually. With around 1.4 million residents in nursing homes on any given day,[14] and an exposure rate of 25% to 30% in that setting,[15, 16, 17] our study suggests that the number of patients exposed in the hospital setting is greater than the number exposed in the nursing home setting, the site of care for which prescribing regulations have been focused thus far.

Because our dataset does not contain preadmission medications, we were unable to specifically investigate new initiation. In the prior single‐center study, approximately 55% of overall use in the hospital setting was new initiation,[12] which would suggest that antipsychotics are newly initiated in around 1 million admissions each year in the hospital. Although we are unable to determine reason for use in our analysis, delirium was a strong predictor of antipsychotic use in our multivariable model, and prior studies have demonstrated delirium to be the most common reason for antipsychotic initiation in hospitalized patients,[12, 18] an indication for which efficacy/effectiveness data are lacking. A recent systematic review of antipsychotics for the treatment of delirium in older adults concluded that because of severe methodological limitations, the small number of existing studies on this topic do not support the use of antipsychotics in the treatment of delirium in older hospitalized adults.[19] Our results further highlight the need for randomized placebo‐controlled trials of antipsychotics in treatment of delirium.

We found variation in antipsychotic use between hospitals that was not fully explained by patient characteristics. Insufficient data to inform clinical decisions surrounding management of agitated delirium/dementia and lack of clear criteria by which to judge appropriateness of antipsychotic use may contribute to this variation. Some variation may relate to resource allocation at different hospitals, and the feasibility of implementing nonpharmacologic management options across settings. Our results collectively highlight the need for studies evaluating the efficacy/effectiveness of antipsychotics in the treatment of delirium and drivers of physician decision‐making in this realm, as well as the need for greater hospital investment in nonpharmacologic delirium‐prevention programs, which have been shown to be effective in prevention of delirium in hospitalized patients.[20]

We observed high levels of potentially excessive daily dosing using cutoffs applied in the long‐term care setting. The majority of the potentially excessive doses were in the setting of typical antipsychotic use, and haloperidol specifically, where doses exceeded 4 mg on at least 1 day in 87% of exposed admissions. Of note, the threshold for haloperidol dosage above which justification is required was decreased from 4 to 2 mg per day in the 2015 update to the CMS guidelines.[21] For the present analysis, we used the guidelines that were contemporaneous to our cohort; we are unable to determine current rates of potentially inappropriate dosages in the present analysis, but given the high prevalence in 2009 to 2010, and the lowering of the dosage threshold since then, it is unlikely that any decrease in use would be enough to substantially reduce the estimate. Whether these high dosages are actually inappropriate in the hospital setting is not established, and we were not able to review medical records to determine whether justification for use of such doses was documented.[22, 23] It is possible that hospitalized patients with altered pharmacodynamics and greater severity of illness could require larger doses of these medications; however, this is an area in need of further investigation, and current critical care guidelines note the lack of sufficient data upon which to justify use of haloperidol in the prevention or treatment of delirium in ICU patients.[24, 25]

The dosages in use are concerning given that the risk of extrapyramidal side effects increases with increasing dose, and prior studies have demonstrated an association between increased dose of antipsychotics and increased risk of other adverse events, including hip fracture and sudden cardiac death.[22, 23] Further, despite these known risks, studies have demonstrated failure to follow recommendations to mitigate risk,[26] such as electrocardiogram monitoring in individuals receiving intravenous haloperidol.[27] Our results suggest that physicians are similarly not following recommendations to use lower doses of haloperidol when treating older patients, given the almost identical incidence of potentially excessive dosing among admissions of patients aged 65 years and older in our cohort.[25] Clinical decision support prompts have been effective at increasing appropriate use of antipsychotic medications in several single‐center analyses,[28, 29, 30] and widespread implementation of such support with a focus on haloperidol dosing should be considered on the basis of our results.

The patient characteristics associated with antipsychotic use in this large, nationally representative analysis are consistent with those identified in prior single‐center analyses.[12, 18] Both prior analyses identified delirium as the most common reason for antipsychotic use, and dementia, intensive care unit stay, and mechanical ventilation were also previously identified as strong predictors of use that we believe hold face validity for the practicing hospitalist. On the other hand, some of the factors associated with antipsychotic use in our model cannot be readily explained, such as insurance status and race, and may be serving as proxies for other variables not included in our analysis. That nonwhite patients are less likely than white patients to receive antipsychotic medications in the hospital has been previously demonstrated,[12] and further investigation to understand this disparity is warranted.

Our study has several additional limitations. First, because our study is observational, the possibility of residual confounding exists, and we cannot rule out that there are other patient factors driving the hospital variation in antipsychotic use that we observed. Second, because guidelines do not exist for antipsychotic dosing in hospitalized patients, we could only comment on potentially excessive dosing, extrapolating from guidelines in the long‐term care setting. Whether such doses are actually excessive in hospitalized patients is not defined. Third, although Premier performs quality checks on charge and ICD‐9‐CM coding data submitted by participating hospitals, the validity of administrative data is uncertain. For example, the use of administrative data to identify delirium diagnoses is likely to have resulted in underestimation of delirium incidence among our different exposure groups. Delirium is likely to be coded more often in the setting of more severe or hyperactive cases, when antipsychotics are more likely to be utilized. This could result in an overestimation of the association between delirium and antipsychotic use. Additionally, differences in coding practices between hospitals for any of the variables in our models could explain some of the variation in antipsychotic prescribing that we observed. Finally, because we were unable to differentiate between new initiation and continuation of a preadmission antipsychotic, some of the variation that we observed is likely to reflect differences in outpatient antipsychotic prescribing practices.

In conclusion, in this large cohort of nonpsychiatric admissions to 300 US hospitals, we found that antipsychotic medication exposure was common, often at high daily doses. Delirium and dementia were the strongest predictors of use among the patient and hospital characteristics examined. The variation in antipsychotic prescribing that we observed was not fully accounted for by measured patient characteristics, and raises the possibility of differing hospital prescribing cultures. Our results draw attention to the need for additional research to better define what constitutes appropriate use of these potentially harmful medications in the hospital setting.

Disclosures: Dr. Herzig had full access to all of the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Drs. Herzig, Rothberg, Gurwitz, and Marcantonio. Acquisition of data: Dr. Herzig. Analysis of data: Mr. Guess. Interpretation of data: Drs. Herzig, Rothberg, Gurwitz, Marcantonio, and Mr. Guess. Drafting of the manuscript: Dr. Herzig. Critical revision of the manuscript for important intellectual content: Drs. Rothberg, Gurwitz, Marcantonio, and Mr. Guess. Dr. Herzig was funded by grant number K23AG042459 from the National Institute on Aging. Dr. Marcantonio was funded by grant number K24AG035075 from the National Institute on Aging. The funding organizations had no involvement in any aspect of the study, including design and conduct of the study; collection, management, analysis, and interpretation of the data; and preparation, review, or approval of the manuscript. The authors report no conflicts of interest.

Breathlessness, or dyspnea, is defined as a subjective experience of breathing discomfort that is comprised of qualitatively distinct sensations that vary in intensity.[1] Dyspnea is a leading reason for patients presenting for emergency care,[2] and it is an important predictor for hospitalization and mortality in patients with cardiopulmonary disease.[3, 4, 5]

Several professional societies' guidelines recommend that patients should be asked to quantify the intensity of their breathlessness using a standardized scale, and that these ratings should be documented in medical records to guide dyspnea awareness and management.[1, 6, 7] During the evaluation and treatment of patients with acute cardiopulmonary conditions, the clinician estimates the severity of the illness and response to therapy based on multiple objective measures as well as the patient's perception of dyspnea. A patient‐centered care approach depends upon the physicians having a shared understanding of what the patient is experiencing. Without this appreciation, the healthcare provider cannot make appropriate treatment decisions to ensure alleviation of presenting symptoms. Understanding the severity of patients' dyspnea is critical to avoid under‐ or overtreatment of patients with acute cardiopulmonary conditions, but only a few studies have compared patient and provider perceptions of dyspnea intensity.[8, 9] Discordance between physician's impression of severity of dyspnea and patient's perception may result in suboptimal management and patient dissatisfaction with care. Furthermore, several studies have shown that, when physicians and patients agree with the assessment of well‐being, treatment adherence and outcomes improve.[10, 11]

Therefore, we evaluated the extent and directionality of agreement between patients' perception and healthcare providers' impression of dyspnea and explored which factors contribute to discordance. Additionally, we examined how healthcare providers document dyspnea severity.

METHODS

RESULTS

Patient Characteristics

Among the 219 patients who met the screening criteria, 81 were not enrolled (Figure 1). Data from 138 patients, with both patient information and provider data on dyspnea assessment, were included. The median age of the patients was 72 years (IQR, 5880 years), 56.5% were women, 75.4% were white, and 28.3% were current smokers. Approximately 30% had a diagnosis of HF, 30% of COPD, and 13.0% of pneumonia. The median comorbidity score was 4 (IQR, 26), and 37.0% of the patients had a BMI 30. At admission, the median oxygen saturation index was 346 (IQR, 287.5460) indicating mild to moderate levels of hypoxia. (Table 1).

Table 1.Patient Characteristics (N = 138)

	Value
NOTE: Abbreviations: BMI, body mass index; COPD, chronic obstructive pulmonary disease; IQR, interquartile range; NRS, numerical rating scale.
Age, median (IQR), y	72 (5880)
Gender
Female	78 (56.5)
Male	60 (43.5)
Race
White	104 (75.4)
Black	16 (11.6)
Hispanic	17 (12.3)
Other	1 (0.7)
Body mass index, median (IQR)	28 (23.334.6)
Obese (BMI 30)	51 (37.0)
Smoker, current	39 (28.3)
Admitting diagnosis
Heart failure	46 (33.3)
COPD/asthma	41 (29.7)
Pneumonia	18 (13.0)
Other	33 (23.9)
Depression	32 (23.2)
Comorbidity score, median (IQR)	4 (26)
Respiratory rate at admission, median (IQR)	20 (1924)
Oxygen saturation index at admission, median (IQR)	346.4 (287.5460)
Patient NRS, median (IQR)
At admission	9 (710)
At discharge	2 (14)
Discharged on home oxygen	45 (32.6)
Respiratory rate at discharge, median (IQR)	20 (1820)
Oxygen saturation index at discharge, median (IQR)	475 (350485)

Figure 1

Agreement Between Patients' Self‐Assessment and Providers' Assessment of Dyspnea Severity

Not all patients had complete data points, and more nurses were interviewed than physicians. Overall, 96 patient‐physician and 138 patient‐nurse pairs participated in the study. A total of 336 patient‐nurse rating dyad assessments of dyspnea and 124 patient‐physician rating dyads assessments were collected (Figure 1). The mean difference between patient and physicians and patient and nurses assessments of dyspnea was 1.23 (IQR, 3 to 0) and 0.21 (IQR, 2 to 2) respectively (a negative score means underestimation by the provider, a positive score means overestimation).

The unadjusted agreement on the severity of dyspnea was 36.3% for the patient‐physician dyads and 44.1% for the patient‐nurse dyads. Physicians underestimated their patients' dyspnea 37.9% of the time and overestimated it 25.8% of the time; nurses underestimated it 43.5% of the time and overestimated it in 12.4% of the study patients (Table 2). In 28.2% of the time, physicians were discordant more than 4 points of the patient assessment. Bland‐Altman plots show that there is greater variation in differences of dyspnea assessments with increase in shortness of breath scores (Figure 2). Nurses underestimated more when the dyspnea score was on the lower end. Physicians also tended to estimate either lower or higher when compared to patients when the dyspnea scores were 2 (Figure 2A,B).

Table 2.Underestimation and Overestimation and Concordance of Dyspnea

	Underestimation			Concordance			Overestimation
	3	2	%*	0	1	%	2	3	%
NOTE: NRS scores by nurses and physicians as compared with patients *Percent underestimation out of all dyads. Percent concordance out of all dyads. Percent overestimation out of all dyads.
Patient‐nurse dyads	110	48	43.5	82	78	44.1	17	28	12.4
Patient‐physician dyads	33	14	37.9	21	24	36.3	12	20	25.8

Figure 2

The weighted kappa coefficient for agreement was 0.11 (95% confidence interval [CI]: 0.01 to 0.21) for patient‐physician assessment, 0.18 (95% CI: 0.12 to 0.24) for patient‐nurse, and 0.09 (0.02 to 0.20) for physician‐nurse indicating poor agreement. In a sensitivity analysis in which we used a higher threshold for defining discordance (difference of more than 2 points), the kappa coefficient increased to 0.21 (95% CI: 0.06 to 0.36) for patient‐physician assessments, to 0.24 (95% CI: 0.15 to 0.33) for patient‐nurse, and to 0.24 (95% CI: 0.09 to 0.39) for nurse‐physician assessments.