Sumit R. Majumdar, MD, MPH

Study Design and Population

As described in detail elsewhere,[14] this was a prospective cohort study that enrolled adult patients (any age older than 18 years) at the time of discharge back to the community from 7 general internal medicine wards in 2 teaching hospitals in Edmonton, Alberta between October 2013 and November 2014. We excluded patients admitted from, or being discharged back to, long‐term care facilities or other acute care hospitals, or from out of the province; patients who were unable to communicate in English; patients with moderate or severe cognitive impairment (scoring 5 or more on the Short Portable Mental Status Questionnaire); or patients with projected life expectancy of less than 3 months. All patients provided written consent, and the study was approved by the Health Research Ethics board of the University of Alberta (project ID Pro00036880).

We assessed the degree of frailty within 24 hours of discharge in 3 ways. First, we used the CFS[9, 15] with patients being asked to rate their best functional status in the week prior to admission. As per the CFS validations studies, scores 5 were defined as frail.[9, 15] Second, we used the TUGT as a proxy for slow gait speed/dysmobility (with >20 seconds defined as abnormal).[13] The TUGT was recorded as the shortest recorded time of the 2 timed trials to get up from a seated position, walk 10 feet and back, and then sit in the chair again. Third, we also determined their Fried score[6] (using the modifications outlined below) and categorized the patients as frail if they scored 3 or more. Of the 5 Fried categories, we assessed weakness by grip strength in their dominant hand using a Jamar handheld dynamometer and weight loss of 10 lb or more in the past year based on patient self‐report; these are identical to the original Fried scale description. Grip strength in the lowest quintile for sex and body mass index was defined as weak grip strength as per convention in the literature, which corresponded to less than 28.5 kg for men and less than 18.5 kg for women.[16, 17] We assessed the other 3 Fried categories in modified fashion as follows. For slow gait, rather than assessing time to walk 15 feet as in the original study and assigning a point to those testing in the lowest quintile for their age/sex, we used the TUGT, because our research personnel were already trained in this test, and we were doing it already as part of the discharge package for all patients.[13] For the Fried category of low activity, we based this on patient self‐report using the relevant questions in the EuroQoL Questionnaire (EQ‐5D); the Fried score used self‐report with a different questionnaire. Finally, for self‐reported exhaustion we used the questions in the Patient Health Questionnaire 9 (PHQ‐9)[18] analogous to those used from the Center for Epidemiological Studies depression scale in the original Fried description. We did this as we were evaluating the PHQ‐9 in our cohort already, and did not want to increase responder burden by presenting them with 2 depression questionnaires.

We followed all patients until 30 days after discharge, and outcome data (all‐cause mortality or all‐cause readmission) were collected by research personnel blinded to the patient's frailty status at discharge using patient/caregiver self‐report and analysis of the provincial electronic health record. We included deaths in or out of the hospital, and all readmissions were unplanned.

We examined the correlation between the CFS score (5 vs <5) and (1) the modified Fried score (3 vs <3) and (2) TUGT (20 seconds vs >20 seconds) using chance corrected kappa coefficients. In our previous article[14] we reported the association between the CFS and readmissions/hospitalizations within 30 days of discharge. In this article we examine whether either the Fried score or TUGT accurately and independently predict postdischarge readmissions/deaths, and whether they add additional prognostic information to the CFS assessment by comparing models with/without each definition using the C statistic and the Integrated Discrimination Improvement index. All analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC), with P values of <0.05 considered statistically significant. Subgroup analysis was done in patients older than 65 years.

Prevalence of Frailty According to Different Definitions

Although the CFS assessment resulted in 162 (33%) patients being deemed frail, only 82 (51%) of those patients also met the phenotype frailty definition using either the Fried model or the TUGT, and 49 (10%) patients who were not classified as frail on the CFS met either of the phenotypic definitions of frailty (Figure 1). Overall, 211 (43%) patients were frail according to at least 1 assessment, and 46 (9%) met all 3 frailty definitions. In the subgroup of 245 patients older than 65 years, 137 (56%) were frail according to at least 1 assessment, 38 (16%) met all 3 frailty definitions, and 27 (11%) of those patients classified as not frail on the CFS met either phenotypic definition of frailty. Agreement between TUGT and CFS or CFS and Fried was relatively poor with kappas of 0.31 (95% confidence interval [CI]: 0.23‐0.40) and 0.33 (95% CI: 0.25‐0.42), respectively. It is noteworthy that some patients deemed nonfrail on the CFS had slow gait speeds, and most CFS‐frail patients had gait speeds in the nonfrail range (Figure 2).

Figure 1

Figure 2

Characteristics According to Frailty Status

Although frail patients were generally similar across definitions (Table 1) in that they were older, had more comorbidities, more hospitalizations in the prior year, and longer index hospitalization lengths of stay than nonfrail patients, patients meeting phenotypic definitions of frailty but not classified as frail on the CFS were younger, had lower Charlson scores, higher EQ‐5D scores, and were discharged with less medications (Table 1).

Outcomes According to Frailty Status

The overall rate of 30‐day death or hospital readmission was 17.1% (85 patients), primarily as a result of hospital readmissions (81, 16.4%) (Table 2). Although patients classified as frail on the CFS exhibited significantly higher 30‐day readmission/death rates (24.1% vs 13.8% for not frail, P = 0.005) even after adjusting for age and sex (adjusted odds ratio [aOR]: 2.02, 95% CI: 1.19‐3.41) (Table 3), patients meeting either of the phenotypic definitions for frailty but not the CFS definition were not at higher risk for 30‐day readmission/death (aOR: 0.87, 95% CI: 0.34‐2.19) (Table 3). The group at highest risk for 30‐day readmissions/death were those meeting both the CFS and either phenotypic definition of frailty (25.6% vs 13.8% for those not frail, aOR: 2.15, 95% CI: 1.10‐4.19) (Tables 2 and 3). None of the Integrated Discrimination Improvement indices (for modified Fried added to CFS or TUGT added to CFS) were statistically significant, suggesting no net new information was added to predictive models, and there were no appreciable changes in C statistics (Table 3). Neither the modified Fried score nor the TUGT on their own added independent prognostic information to age/sex alone as predictors of postdischarge outcomes. It is noteworthy that the areas under the curve for models using any combination of the frailty definitions plus age and sex were not high (all ranged between 0.55 and 0.60 for the overall cohort and from 0.52 and 0.65 in the elderly). If the frailty definitions were examined as continuous variables rather than dichotomized into frail/not frail, the C statistics were not appreciably better: 0.65 for CFS, 0.58 for TUGT, and 0.60 for modified Fried. Of note, the CFS score with the published cutoff of 5 demonstrated the highest kappa, sensitivity, specificity, and positive predictive value in relation to outcomes.

Outcomes According to Frailty Status in the Elderly Subgroup

Although absolute risks of readmission or death were higher in elderly patients than younger patients, the excess risk was largely seen in those elderly patients classified as frail on the CFS. In fact, all of the associations reported above for the entire cohort were in the same direction in the elderly subgroup (Tables 2 and 3).

Strengths and Limitations

Although this was a prospective cohort study with blinded ascertainment of endpoints (30‐day outcome data were collected by observers who were unaware of the patients' CFS or phenotypic model scores), it is not without limitations. First, the only postdischarge outcomes we assessed were readmission and death, and it would be interesting to evaluate which frailty tools best predict those who are most likely to benefit from home‐care services in the community. Second, as we were interested in 30‐day readmission rates, we excluded long‐term care residents from our study and patients who had foreshortened life expectancy, in essence, the frailest of the frail. Although this reduced the size of any association between frailty and adverse outcomes, we focused this study on the situations where there is clinical equipoise and there is rarely a diagnostic dilemma around the identification of frailty and need for increased services in palliative or long‐term care patients. Third, we did not use exactly the same questionnaires or gait speed assessments as used in the original Fried score description, but as outlined in the Methods section, we used analogous questions on closely related questionnaires to extract the same information. Fourth, some might consider our comparisons biased toward the CFS, as it reflects gestalt clinical impressions (informed by patients and proxies) of frailty status before hospital admission while the Fried score and TUGT were based on patient status just prior to discharge, it may be that the former is a better measure of eventual recovery (and ongoing risk) than the latter measures. If this is the case, for the purposes of targeting interventions to prevent postdischarge complications, it would suggest to us that the CFS is better suited, whereas phenotype tools can be reserved for the postdischarge phase of recovery. By the same token, perhaps serial measures of the CFS and phenotypic tools are more important, as the trajectory of recovery may be most informative for risk prediction.[7] Certainly, if one were interested in changes in functional status during hospitalization,[29] then objective phenotypic measures such as grip strength or TUGT times would seem more appropriate choices. Fifth, some may perceive it as a weakness that we did not restrict our cohort to elderly patients; however, we actually view this as a strength, because frailty is not exclusive to older patients. Sixth, although we restricted this study to patients being discharged from general internal medicine wards, it is worth mentioning that previous studies have shown similar associations between frailty and outcomes in nonmedical hospitalized patients.[19, 22, 23, 24]

In conclusion, we looked at 3 different ways of screening for frailty, 1 being a subjective but well‐validated tool (the CFS) and the other 2 being objective assessments that look at specific phenotypic characteristics. There is a compelling need to find a standardized assessment to determine frailty in both research and clinical settings, and our study provides support for use of the CFS over the Fried or TUGT as screening tools. Standardized frailty assessments should be part of the discharge planning for all medical patients so that extra resources can be properly targeted at those patients at greatest risk for suboptimal transition back to community living.

Study Design and Population

As described in detail elsewhere,[14] this was a prospective cohort study that enrolled adult patients (any age older than 18 years) at the time of discharge back to the community from 7 general internal medicine wards in 2 teaching hospitals in Edmonton, Alberta between October 2013 and November 2014. We excluded patients admitted from, or being discharged back to, long‐term care facilities or other acute care hospitals, or from out of the province; patients who were unable to communicate in English; patients with moderate or severe cognitive impairment (scoring 5 or more on the Short Portable Mental Status Questionnaire); or patients with projected life expectancy of less than 3 months. All patients provided written consent, and the study was approved by the Health Research Ethics board of the University of Alberta (project ID Pro00036880).

We assessed the degree of frailty within 24 hours of discharge in 3 ways. First, we used the CFS[9, 15] with patients being asked to rate their best functional status in the week prior to admission. As per the CFS validations studies, scores 5 were defined as frail.[9, 15] Second, we used the TUGT as a proxy for slow gait speed/dysmobility (with >20 seconds defined as abnormal).[13] The TUGT was recorded as the shortest recorded time of the 2 timed trials to get up from a seated position, walk 10 feet and back, and then sit in the chair again. Third, we also determined their Fried score[6] (using the modifications outlined below) and categorized the patients as frail if they scored 3 or more. Of the 5 Fried categories, we assessed weakness by grip strength in their dominant hand using a Jamar handheld dynamometer and weight loss of 10 lb or more in the past year based on patient self‐report; these are identical to the original Fried scale description. Grip strength in the lowest quintile for sex and body mass index was defined as weak grip strength as per convention in the literature, which corresponded to less than 28.5 kg for men and less than 18.5 kg for women.[16, 17] We assessed the other 3 Fried categories in modified fashion as follows. For slow gait, rather than assessing time to walk 15 feet as in the original study and assigning a point to those testing in the lowest quintile for their age/sex, we used the TUGT, because our research personnel were already trained in this test, and we were doing it already as part of the discharge package for all patients.[13] For the Fried category of low activity, we based this on patient self‐report using the relevant questions in the EuroQoL Questionnaire (EQ‐5D); the Fried score used self‐report with a different questionnaire. Finally, for self‐reported exhaustion we used the questions in the Patient Health Questionnaire 9 (PHQ‐9)[18] analogous to those used from the Center for Epidemiological Studies depression scale in the original Fried description. We did this as we were evaluating the PHQ‐9 in our cohort already, and did not want to increase responder burden by presenting them with 2 depression questionnaires.

We followed all patients until 30 days after discharge, and outcome data (all‐cause mortality or all‐cause readmission) were collected by research personnel blinded to the patient's frailty status at discharge using patient/caregiver self‐report and analysis of the provincial electronic health record. We included deaths in or out of the hospital, and all readmissions were unplanned.

We examined the correlation between the CFS score (5 vs <5) and (1) the modified Fried score (3 vs <3) and (2) TUGT (20 seconds vs >20 seconds) using chance corrected kappa coefficients. In our previous article[14] we reported the association between the CFS and readmissions/hospitalizations within 30 days of discharge. In this article we examine whether either the Fried score or TUGT accurately and independently predict postdischarge readmissions/deaths, and whether they add additional prognostic information to the CFS assessment by comparing models with/without each definition using the C statistic and the Integrated Discrimination Improvement index. All analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC), with P values of <0.05 considered statistically significant. Subgroup analysis was done in patients older than 65 years.

Prevalence of Frailty According to Different Definitions

Although the CFS assessment resulted in 162 (33%) patients being deemed frail, only 82 (51%) of those patients also met the phenotype frailty definition using either the Fried model or the TUGT, and 49 (10%) patients who were not classified as frail on the CFS met either of the phenotypic definitions of frailty (Figure 1). Overall, 211 (43%) patients were frail according to at least 1 assessment, and 46 (9%) met all 3 frailty definitions. In the subgroup of 245 patients older than 65 years, 137 (56%) were frail according to at least 1 assessment, 38 (16%) met all 3 frailty definitions, and 27 (11%) of those patients classified as not frail on the CFS met either phenotypic definition of frailty. Agreement between TUGT and CFS or CFS and Fried was relatively poor with kappas of 0.31 (95% confidence interval [CI]: 0.23‐0.40) and 0.33 (95% CI: 0.25‐0.42), respectively. It is noteworthy that some patients deemed nonfrail on the CFS had slow gait speeds, and most CFS‐frail patients had gait speeds in the nonfrail range (Figure 2).

Figure 1

Figure 2

Characteristics According to Frailty Status

Although frail patients were generally similar across definitions (Table 1) in that they were older, had more comorbidities, more hospitalizations in the prior year, and longer index hospitalization lengths of stay than nonfrail patients, patients meeting phenotypic definitions of frailty but not classified as frail on the CFS were younger, had lower Charlson scores, higher EQ‐5D scores, and were discharged with less medications (Table 1).

Outcomes According to Frailty Status

The overall rate of 30‐day death or hospital readmission was 17.1% (85 patients), primarily as a result of hospital readmissions (81, 16.4%) (Table 2). Although patients classified as frail on the CFS exhibited significantly higher 30‐day readmission/death rates (24.1% vs 13.8% for not frail, P = 0.005) even after adjusting for age and sex (adjusted odds ratio [aOR]: 2.02, 95% CI: 1.19‐3.41) (Table 3), patients meeting either of the phenotypic definitions for frailty but not the CFS definition were not at higher risk for 30‐day readmission/death (aOR: 0.87, 95% CI: 0.34‐2.19) (Table 3). The group at highest risk for 30‐day readmissions/death were those meeting both the CFS and either phenotypic definition of frailty (25.6% vs 13.8% for those not frail, aOR: 2.15, 95% CI: 1.10‐4.19) (Tables 2 and 3). None of the Integrated Discrimination Improvement indices (for modified Fried added to CFS or TUGT added to CFS) were statistically significant, suggesting no net new information was added to predictive models, and there were no appreciable changes in C statistics (Table 3). Neither the modified Fried score nor the TUGT on their own added independent prognostic information to age/sex alone as predictors of postdischarge outcomes. It is noteworthy that the areas under the curve for models using any combination of the frailty definitions plus age and sex were not high (all ranged between 0.55 and 0.60 for the overall cohort and from 0.52 and 0.65 in the elderly). If the frailty definitions were examined as continuous variables rather than dichotomized into frail/not frail, the C statistics were not appreciably better: 0.65 for CFS, 0.58 for TUGT, and 0.60 for modified Fried. Of note, the CFS score with the published cutoff of 5 demonstrated the highest kappa, sensitivity, specificity, and positive predictive value in relation to outcomes.

Outcomes According to Frailty Status in the Elderly Subgroup

Although absolute risks of readmission or death were higher in elderly patients than younger patients, the excess risk was largely seen in those elderly patients classified as frail on the CFS. In fact, all of the associations reported above for the entire cohort were in the same direction in the elderly subgroup (Tables 2 and 3).

Strengths and Limitations

Although this was a prospective cohort study with blinded ascertainment of endpoints (30‐day outcome data were collected by observers who were unaware of the patients' CFS or phenotypic model scores), it is not without limitations. First, the only postdischarge outcomes we assessed were readmission and death, and it would be interesting to evaluate which frailty tools best predict those who are most likely to benefit from home‐care services in the community. Second, as we were interested in 30‐day readmission rates, we excluded long‐term care residents from our study and patients who had foreshortened life expectancy, in essence, the frailest of the frail. Although this reduced the size of any association between frailty and adverse outcomes, we focused this study on the situations where there is clinical equipoise and there is rarely a diagnostic dilemma around the identification of frailty and need for increased services in palliative or long‐term care patients. Third, we did not use exactly the same questionnaires or gait speed assessments as used in the original Fried score description, but as outlined in the Methods section, we used analogous questions on closely related questionnaires to extract the same information. Fourth, some might consider our comparisons biased toward the CFS, as it reflects gestalt clinical impressions (informed by patients and proxies) of frailty status before hospital admission while the Fried score and TUGT were based on patient status just prior to discharge, it may be that the former is a better measure of eventual recovery (and ongoing risk) than the latter measures. If this is the case, for the purposes of targeting interventions to prevent postdischarge complications, it would suggest to us that the CFS is better suited, whereas phenotype tools can be reserved for the postdischarge phase of recovery. By the same token, perhaps serial measures of the CFS and phenotypic tools are more important, as the trajectory of recovery may be most informative for risk prediction.[7] Certainly, if one were interested in changes in functional status during hospitalization,[29] then objective phenotypic measures such as grip strength or TUGT times would seem more appropriate choices. Fifth, some may perceive it as a weakness that we did not restrict our cohort to elderly patients; however, we actually view this as a strength, because frailty is not exclusive to older patients. Sixth, although we restricted this study to patients being discharged from general internal medicine wards, it is worth mentioning that previous studies have shown similar associations between frailty and outcomes in nonmedical hospitalized patients.[19, 22, 23, 24]

In conclusion, we looked at 3 different ways of screening for frailty, 1 being a subjective but well‐validated tool (the CFS) and the other 2 being objective assessments that look at specific phenotypic characteristics. There is a compelling need to find a standardized assessment to determine frailty in both research and clinical settings, and our study provides support for use of the CFS over the Fried or TUGT as screening tools. Standardized frailty assessments should be part of the discharge planning for all medical patients so that extra resources can be properly targeted at those patients at greatest risk for suboptimal transition back to community living.

Study Design and Population

As described in detail elsewhere,[14] this was a prospective cohort study that enrolled adult patients (any age older than 18 years) at the time of discharge back to the community from 7 general internal medicine wards in 2 teaching hospitals in Edmonton, Alberta between October 2013 and November 2014. We excluded patients admitted from, or being discharged back to, long‐term care facilities or other acute care hospitals, or from out of the province; patients who were unable to communicate in English; patients with moderate or severe cognitive impairment (scoring 5 or more on the Short Portable Mental Status Questionnaire); or patients with projected life expectancy of less than 3 months. All patients provided written consent, and the study was approved by the Health Research Ethics board of the University of Alberta (project ID Pro00036880).

We assessed the degree of frailty within 24 hours of discharge in 3 ways. First, we used the CFS[9, 15] with patients being asked to rate their best functional status in the week prior to admission. As per the CFS validations studies, scores 5 were defined as frail.[9, 15] Second, we used the TUGT as a proxy for slow gait speed/dysmobility (with >20 seconds defined as abnormal).[13] The TUGT was recorded as the shortest recorded time of the 2 timed trials to get up from a seated position, walk 10 feet and back, and then sit in the chair again. Third, we also determined their Fried score[6] (using the modifications outlined below) and categorized the patients as frail if they scored 3 or more. Of the 5 Fried categories, we assessed weakness by grip strength in their dominant hand using a Jamar handheld dynamometer and weight loss of 10 lb or more in the past year based on patient self‐report; these are identical to the original Fried scale description. Grip strength in the lowest quintile for sex and body mass index was defined as weak grip strength as per convention in the literature, which corresponded to less than 28.5 kg for men and less than 18.5 kg for women.[16, 17] We assessed the other 3 Fried categories in modified fashion as follows. For slow gait, rather than assessing time to walk 15 feet as in the original study and assigning a point to those testing in the lowest quintile for their age/sex, we used the TUGT, because our research personnel were already trained in this test, and we were doing it already as part of the discharge package for all patients.[13] For the Fried category of low activity, we based this on patient self‐report using the relevant questions in the EuroQoL Questionnaire (EQ‐5D); the Fried score used self‐report with a different questionnaire. Finally, for self‐reported exhaustion we used the questions in the Patient Health Questionnaire 9 (PHQ‐9)[18] analogous to those used from the Center for Epidemiological Studies depression scale in the original Fried description. We did this as we were evaluating the PHQ‐9 in our cohort already, and did not want to increase responder burden by presenting them with 2 depression questionnaires.

We followed all patients until 30 days after discharge, and outcome data (all‐cause mortality or all‐cause readmission) were collected by research personnel blinded to the patient's frailty status at discharge using patient/caregiver self‐report and analysis of the provincial electronic health record. We included deaths in or out of the hospital, and all readmissions were unplanned.

We examined the correlation between the CFS score (5 vs <5) and (1) the modified Fried score (3 vs <3) and (2) TUGT (20 seconds vs >20 seconds) using chance corrected kappa coefficients. In our previous article[14] we reported the association between the CFS and readmissions/hospitalizations within 30 days of discharge. In this article we examine whether either the Fried score or TUGT accurately and independently predict postdischarge readmissions/deaths, and whether they add additional prognostic information to the CFS assessment by comparing models with/without each definition using the C statistic and the Integrated Discrimination Improvement index. All analyses were performed using SAS version 9.4 (SAS Institute, Cary, NC), with P values of <0.05 considered statistically significant. Subgroup analysis was done in patients older than 65 years.

Prevalence of Frailty According to Different Definitions

Although the CFS assessment resulted in 162 (33%) patients being deemed frail, only 82 (51%) of those patients also met the phenotype frailty definition using either the Fried model or the TUGT, and 49 (10%) patients who were not classified as frail on the CFS met either of the phenotypic definitions of frailty (Figure 1). Overall, 211 (43%) patients were frail according to at least 1 assessment, and 46 (9%) met all 3 frailty definitions. In the subgroup of 245 patients older than 65 years, 137 (56%) were frail according to at least 1 assessment, 38 (16%) met all 3 frailty definitions, and 27 (11%) of those patients classified as not frail on the CFS met either phenotypic definition of frailty. Agreement between TUGT and CFS or CFS and Fried was relatively poor with kappas of 0.31 (95% confidence interval [CI]: 0.23‐0.40) and 0.33 (95% CI: 0.25‐0.42), respectively. It is noteworthy that some patients deemed nonfrail on the CFS had slow gait speeds, and most CFS‐frail patients had gait speeds in the nonfrail range (Figure 2).

Figure 1

Figure 2

Characteristics According to Frailty Status

Although frail patients were generally similar across definitions (Table 1) in that they were older, had more comorbidities, more hospitalizations in the prior year, and longer index hospitalization lengths of stay than nonfrail patients, patients meeting phenotypic definitions of frailty but not classified as frail on the CFS were younger, had lower Charlson scores, higher EQ‐5D scores, and were discharged with less medications (Table 1).

Outcomes According to Frailty Status

The overall rate of 30‐day death or hospital readmission was 17.1% (85 patients), primarily as a result of hospital readmissions (81, 16.4%) (Table 2). Although patients classified as frail on the CFS exhibited significantly higher 30‐day readmission/death rates (24.1% vs 13.8% for not frail, P = 0.005) even after adjusting for age and sex (adjusted odds ratio [aOR]: 2.02, 95% CI: 1.19‐3.41) (Table 3), patients meeting either of the phenotypic definitions for frailty but not the CFS definition were not at higher risk for 30‐day readmission/death (aOR: 0.87, 95% CI: 0.34‐2.19) (Table 3). The group at highest risk for 30‐day readmissions/death were those meeting both the CFS and either phenotypic definition of frailty (25.6% vs 13.8% for those not frail, aOR: 2.15, 95% CI: 1.10‐4.19) (Tables 2 and 3). None of the Integrated Discrimination Improvement indices (for modified Fried added to CFS or TUGT added to CFS) were statistically significant, suggesting no net new information was added to predictive models, and there were no appreciable changes in C statistics (Table 3). Neither the modified Fried score nor the TUGT on their own added independent prognostic information to age/sex alone as predictors of postdischarge outcomes. It is noteworthy that the areas under the curve for models using any combination of the frailty definitions plus age and sex were not high (all ranged between 0.55 and 0.60 for the overall cohort and from 0.52 and 0.65 in the elderly). If the frailty definitions were examined as continuous variables rather than dichotomized into frail/not frail, the C statistics were not appreciably better: 0.65 for CFS, 0.58 for TUGT, and 0.60 for modified Fried. Of note, the CFS score with the published cutoff of 5 demonstrated the highest kappa, sensitivity, specificity, and positive predictive value in relation to outcomes.

Outcomes According to Frailty Status in the Elderly Subgroup

Although absolute risks of readmission or death were higher in elderly patients than younger patients, the excess risk was largely seen in those elderly patients classified as frail on the CFS. In fact, all of the associations reported above for the entire cohort were in the same direction in the elderly subgroup (Tables 2 and 3).

Strengths and Limitations

Although this was a prospective cohort study with blinded ascertainment of endpoints (30‐day outcome data were collected by observers who were unaware of the patients' CFS or phenotypic model scores), it is not without limitations. First, the only postdischarge outcomes we assessed were readmission and death, and it would be interesting to evaluate which frailty tools best predict those who are most likely to benefit from home‐care services in the community. Second, as we were interested in 30‐day readmission rates, we excluded long‐term care residents from our study and patients who had foreshortened life expectancy, in essence, the frailest of the frail. Although this reduced the size of any association between frailty and adverse outcomes, we focused this study on the situations where there is clinical equipoise and there is rarely a diagnostic dilemma around the identification of frailty and need for increased services in palliative or long‐term care patients. Third, we did not use exactly the same questionnaires or gait speed assessments as used in the original Fried score description, but as outlined in the Methods section, we used analogous questions on closely related questionnaires to extract the same information. Fourth, some might consider our comparisons biased toward the CFS, as it reflects gestalt clinical impressions (informed by patients and proxies) of frailty status before hospital admission while the Fried score and TUGT were based on patient status just prior to discharge, it may be that the former is a better measure of eventual recovery (and ongoing risk) than the latter measures. If this is the case, for the purposes of targeting interventions to prevent postdischarge complications, it would suggest to us that the CFS is better suited, whereas phenotype tools can be reserved for the postdischarge phase of recovery. By the same token, perhaps serial measures of the CFS and phenotypic tools are more important, as the trajectory of recovery may be most informative for risk prediction.[7] Certainly, if one were interested in changes in functional status during hospitalization,[29] then objective phenotypic measures such as grip strength or TUGT times would seem more appropriate choices. Fifth, some may perceive it as a weakness that we did not restrict our cohort to elderly patients; however, we actually view this as a strength, because frailty is not exclusive to older patients. Sixth, although we restricted this study to patients being discharged from general internal medicine wards, it is worth mentioning that previous studies have shown similar associations between frailty and outcomes in nonmedical hospitalized patients.[19, 22, 23, 24]

In conclusion, we looked at 3 different ways of screening for frailty, 1 being a subjective but well‐validated tool (the CFS) and the other 2 being objective assessments that look at specific phenotypic characteristics. There is a compelling need to find a standardized assessment to determine frailty in both research and clinical settings, and our study provides support for use of the CFS over the Fried or TUGT as screening tools. Standardized frailty assessments should be part of the discharge planning for all medical patients so that extra resources can be properly targeted at those patients at greatest risk for suboptimal transition back to community living.

Data Sources and Search Methods

We searched the Cumulative Index to Nursing and Allied Health Literature, Ovid MEDLINE, Ovid Embase, and PsycINFO from inception to January 9, 2015, and the last 5 years of PubMed for full publications with any of the following Medical Subject Headings: depressive disorder, depression, patient readmission, interviews, psychological, inpatients, with restrictions for peer‐reviewed publication, humans, adults aged 18 years, and the English language. Search strategies were developed with a librarian (available upon request). We manually searched reference lists of all included studies and relevant review articles and contacted content experts to identify additional publications.

Eligibility Criteria and Selection of Studies

Two authors (J.L.P. and L.M.W.) independently screened full texts of all relevant articles for inclusion. Disagreements were resolved by consensus or a third reviewer (S.R.M.). We considered any original research that compared readmission or mortality after discharge for hospitalized medical patients (ie, general patients or subgroups thereof) with versus without depression identified by any validated depression measure,[16] including any study design that incorporated at least 30‐day follow‐up postdischarge. We excluded studies that examined patients hospitalized in nonacute care settings or on surgical, psychiatric, obstetric, or intensive care services. We calculated Cohen's coefficient to evaluate inter‐rater agreement on study selection.

Data Extraction

Data were abstracted by 2 authors (J.L.P. and L.M.W.). Disagreements were resolved by consensus or a third reviewer (S.R.M.). We contacted authors of all included studies to obtain missing data. If unavailable, crude data were estimated from published survival curves employing validated techniques in R (version 3.1.2; R Foundation for Statistical Computing, Vienna, Austria) and Digitizeit (http://www.digitieit.de; DigitizeIt, Braunschweig, Germany).[17, 18] We sought information on trial characteristics (country, type of hospital, inclusion and exclusion criteria, sample size, follow‐up duration, attrition), participants (age, sex, ethnicity, level of education and social support, comorbidities, marital status), exposure ascertainment (self‐report depression screening tool or diagnostic interview for depression), and outcomes (primary: 30‐day all‐cause readmission or mortality, secondary: 90‐day all‐cause readmission or mortality, ED visits, primary care physician [PCP] visits).

Data Synthesis and Statistical Analysis

Where possible, we calculated the pooled risk ratio (RR) with 95% confidence interval (95% CI) using a random effects models in Review Manager (RevMan) 5.3 (The Nordic Cochrane Centre, Copenhagen, Denmark). The random effects approach that we employed assumes heterogeneity (ie, underlying parameters vary between individual studies) and is distributed around a mean or population average effect, and results in more conservative (wider) confidence intervals, wherein larger cohorts (or studies with smaller standard errors) are given more weight. Heterogeneity was assessed using the I² statistic, with values of <25%, 25% to 50%, and >50% representing low, moderate, and high heterogeneity.[19] As per the guidance of Higgins et al., we did not a priori define any degree of heterogeneity that would preclude pooling of the data; the expectation would be that heterogeneity would be substantially higher pooling observational studies rather than randomized trials.[19] Statistical significance was considered a 2‐sided P value 0.05.

Quality Assessment and Risk of Bias

We assessed study quality using the 9‐item Newcastle‐Ottawa scale with 0 to 3, 4 to 6, and 7 to 9 stars considered low, moderate, and high quality, respectively, and criteria for external and internal validity, including group selection and comparability, outcome assessment, and adequacy of follow‐up.[20] Adjusted estimates published in individual reports (or obtained directly from authors) were compared wherever possible with unadjusted estimates to assess the degree of confounding. We generated funnel plots in RevMan 5.3 and conducted Egger tests using Stata 13 (StataCorp LP, College Station, TX) to assess for publication bias.[21]

Study Selection

After removing duplicate publications, we identified 4066 reports and reviewed 133 reports in full text (see Supporting Figure 1 in the online version of this article). Despite our broad study inclusion criteria, we found only 35 longitudinal studies addressing this question. All 35 authors were contacted for additional outcomes data and other missing information (response rate of 34%). We had to exclude 17 studies as they did not provide 30 or 90‐day post‐discharge outcomes. Only 4 studies had published crude data for outcomes within 90 days,[22, 23, 24, 25] but after contact with authors, we received unpublished data for a further 7 studies[26, 27, 28, 29, 30, 31, 32] (including individual level data for 2 cohorts).[31, 32] We were able to estimate crude data from Kaplan‐Meier curves for another 3 studies.[33, 34, 35] Another 4 studies did not collect the outcomes we were interested in individually. These studies were included in this systematic review but are not poolable in our models: 3 authors could only provide composite endpoint data,[36, 37, 38] and 1 author provided unadjusted hazard ratios.[39] Inter‐reviewer agreement for inclusion was 80% (Cohen's = 0.60).

Characteristics of Included Studies

The 18 studies ranged in size from 58 to 1418 patients; 13 were cohort studies and 5 included secondary data from randomized control trials.[22, 27, 30, 34, 36] All studies ascertained depressive status by screening during index medical admission with either diagnostic interview or self‐report questionnaires, although a variety of scales and definitions for depression were used (Beck Depression Inventory [BDI] in 6 studies, Geriatric Depression Scale in 5 studies, Patient Health Questionnaire in another 4 studies, Medical Outcomes Study‐Depression Questionnaire in 1 study, and Center for Epidemiologic Studies Depression Scale in another study) (Table 1). Screening interviews were conducted mostly by research assistants or nurses (68%) or self‐administered (21%). Most studies examined specific medical patient subgroups (10 cardiac, 3 pulmonary, and 2 elderly). Major exclusion criteria reported were terminal illness (4 studies), unstable condition (6 studies), severe cognitive impairment (5 studies), and suicidal ideation or known depression (4 studies); 1 study enrolled patients with suspected depression (Table 1). Patient cohorts were on average older (range, 5082 years) (Table 1). Attrition rates for readmission and mortality data were low (average <1% among entire sample of studies). All studies scored at least 5 on the Newcastle‐Ottawa scale and were thus considered of at least moderate quality (see Supporting Table 1 in the online version of this article).

Prevalence and Recognition of Depressive Symptoms

The range of depression prevalence in hospitalized medical patients was 14% to 79%, with a median of 32% (interquartile range, 27%40%) (Table 1). In those studies that used a diagnostic interview, the prevalence tended to be lower for major depression, with a median of 17% (interquartile range, 16%22%) (Table 1). None of the included studies reported frequency of clinically recognized depression (ie, prior to screening for the study). Only 2 studies assessed the persistence of depression after discharge: 1 reported that depression persisted in 53% (by screening questionnaire) and 34% (by diagnostic interview) of patients at 30 days,[38] whereas the other reported 48% persistence at 90 days after discharge according to a combined screening method.[37]

Hospital Readmission

Overall, 8 studies provided readmission data. Among patients discharged from acute care medical wards (4 studies reporting on 5 cohorts), 395 of 2433 (16.2%) patients were readmitted within 30 days (Figure 1). Hospitalized patients with depressive symptoms were more likely to be readmitted within 30 days after discharge (20.4% vs 13.7%, RR: 1.73, 95% CI: 1.16‐2.58, P = 0.007, I² = 55%) (Figure 1), compared to those without depression. Results were consistent for 90‐day readmissions (39.8% vs 31.0%, RR: 1.68, 95% CI: 1.13‐2.50, P = 0.01, I² = 76%, n = 1543 patients) (see Supporting Figure 2 in the online version of this article) in 6 studies. One individual study examined readmission within 6 months after discharge, but was not poolable in this model, as it presented only hazard ratios and not raw data; however, it did report a 50% increased risk of readmission in medical inpatients aged 75 years (adjusted hazard ratio: 1.50, 95% CI: 1.03‐2.17, n = 401).[39]

Figure 1

Mortality After Discharge

Overall, 11 studies provided all‐cause mortality data. Among medical patients discharged from acute care in 9 studies, 69 of 3397 (2.0%) patients died within 30 days (Figure 2). Medical patients discharged with depressive symptoms were more likely to die within 30 days (2.8% vs 1.5%, RR: 2.13, 95% CI: 1.31‐3.44, P = 0.002, I² = 0%) (Figure 2) compared to those without depression. Similar results were found for 90‐day mortality (7.7% vs 4.1%, RR: 2.01, 95% CI: 1.47‐2.76, P < 0.001, I² = 4%, n = 3784 patients) (see Supporting Figure 3 in the online version of this article) in 11 studies.

Figure 2

ED and PCP Visits

Four studies examined the use of ED or PCP services within 90 days of discharge, but 3 did not have extractable data for meta‐analysis. All showed increased utilization of health services for depressed compared to nondepressed patients after discharge.[22, 36, 37, 38] Depressed patients were more likely to visit the ED (adjusted incidence rate ratio: 1.73, 95% CI: 1.27‐2.36),[36] had significantly more medical encounters (eg, PCP, ED visits, hospital admissions, laboratory tests, and home care [mean 2.9 vs 2.6, P = 0.05])[38] and had a greater number of ED visits alone (27 vs 15 per 100 patients, P = 0.007)[22] within 30 days of hospital discharge compared to nondepressed patients. Similar results were found at 90 days.[36]

Sensitivity Analyses

All told, most studies reported a positive association between depression and adverse events, and this was true regardless of how much adjustment for potential confounding had been undertaken by the authors. Although all studies were qualitatively in the same direction, the magnitude of the association varied due to methodological and/or clinical heterogeneity. Sensitivity analysis revealed no overall difference in pooled risk ratios or heterogeneity between Mantel‐Haenszel fixed effects versus random effects models or with the addition of 0.5 to cells to permit inclusion of zero‐event data. There was no evidence of publication bias; funnel plots and Egger test results are available upon request. There were no statistically significant differences in the risk associated with depressive symptoms whether studies used Diagnostic and Statistical Manual of Mental Disorders (DSM)‐III or DSM‐IV criteria, whether the study samples were disease specific or unselected general medical cohorts, whether studies were of moderate or high quality, or regardless of the severity of depressive symptoms.

Summary of Evidence

We found that depression was common in medical inpatients (about one‐third of all patients) and persisted for at least 30 days in up to half of those patients after discharge. We found strong evidence of an association between depressive symptoms and poor short‐term prognosis after discharge from the hospital: a 73% increased risk of readmission and a 2‐fold risk of death within 30 days compared to patients without depressive symptoms with similar results at 90 days.

Our meta‐analysis complements a recent systematic review that found concomitant depression to be a risk factor for poor prognosis among inpatients and outpatients with acute coronary syndrome,[40] and a meta‐analysis that demonstrated an increased risk of 2‐year mortality among patients with depression after myocardial infarction.[41] To our knowledge, our study is the first to quantify the short‐term postdischarge risks across a diverse group of medical inpatients.

The potential mechanisms underlying the observed relationship between depression and adverse patient outcomes after discharge are likely multiple. We believe there are 2 main possibilities. First, the increased risk associated with depression might be due to residual confounding, even though many of these studies did adjust for extensive lists of comorbidities,[22, 24, 26, 27, 29, 30, 33, 35, 36, 39] including functional status[39] and prior health services utilization.[22, 34, 36] This could occur if other risk factors were not sufficiently adjusted for, such as unrecognized comorbidities or concomitant disability, which are often present among chronically ill patients,[42] or if depression were a marker of psychosocial risk factors such as anxiety,[43] stress or poor resiliency,[44] or low social support,[45] though a few adjusted for psychosocial factors such as social support[26] or anxiety.[35] Confounding could also occur if symptoms of acute illness inflate reports of somatic symptoms of depression on self‐report questionnaires. Recent studies on the BDI, found that scores were higher in postmyocardial infarction patients when compared to outpatient controls,[46] but with no differences between those groups in scores for the BDI‐II,[47] a version with fewer somatic symptom questions.

Second, depression may cause adverse outcomes through indirect or direct pathways. Indirect causation could occur if depression hindered self‐care behaviors such as medication adherence.[42] Depression could also act directly through pathophysiological changes. Some studies have suggested that depression is associated with metabolic abnormalities, including alterations in glucose transport[42, 48] and increased vulnerability to obesity, type 2 diabetes mellitus, and/or diabetic complications, common conditions among hospitalized patients that also adversely affect postdischarge outcomes.[40, 48]

Strengths and Limitations

This review has multiple strengths. We cast a broad search and included studies that examined a wide range of medical patient subgroups, thus increasing the generalizability of our findings. We identified a general scarcity of studies on this topic and obtained additional unpublished data for 10 of the 18 relevant studies, and our response rate of 34% is compatible with the 37% response rate reported for Cochrane reviews when seeking additional data from authors.[49] Whether examined qualitatively (vote counting of the number of studies that showed an association) or quantitatively (via formal meta‐analysis), it seems apparent that there is a clinically important association between depression and postdischarge adverse events, but given the number, quality, and heterogeneity of the studies we examined, there may be some ongoing dispute about exactly how strong this association is and the degree of bias contributed by a couple of large studies of the topic.

There are limitations to our review. First, as we did not have individual‐level patient data, we could not use metaregression to explore sources of heterogeneity (clinical or methodological) or adjust for confounding, and this likely contributes to observed differences between individual estimates. For instance, the included studies had heterogeneous screening measures and cutoffs; thus, all cases of depression in these studies might not be equivalent. Some of the included studies assessed depression early during admission where psychological distress may be greatest; others assessed symptoms closer to discharge. Most studies included patients with specific conditions like heart failure or chronic obstructive pulmonary disease rather than a wide spectrum of medical inpatients. Moreover, few studies adjusted for psychosocial risk factors such as social support, anxiety, and functional status, and only 2 studies assessed the persistence of depressive symptoms after discharge. Second, we did not explore quantitative measures of between‐study variation (eg, I²), because experts question its utility given the expected heterogeneity in meta‐analyses of observational studies.[50] Third, although the included studies were deemed to be of at least moderate quality, they could be at risk for sources of bias that may not be sufficiently appraised by the current version of the Newscastle‐Ottawa scale for observational studies. Finally, we excluded grey literature (eg, conference proceedings or technical reports) that could potentially exclude null findings, although we did contact authors in this field to identify additional unpublished data relevant to this topic.

Data Sources and Search Methods

We searched the Cumulative Index to Nursing and Allied Health Literature, Ovid MEDLINE, Ovid Embase, and PsycINFO from inception to January 9, 2015, and the last 5 years of PubMed for full publications with any of the following Medical Subject Headings: depressive disorder, depression, patient readmission, interviews, psychological, inpatients, with restrictions for peer‐reviewed publication, humans, adults aged 18 years, and the English language. Search strategies were developed with a librarian (available upon request). We manually searched reference lists of all included studies and relevant review articles and contacted content experts to identify additional publications.

Eligibility Criteria and Selection of Studies

Two authors (J.L.P. and L.M.W.) independently screened full texts of all relevant articles for inclusion. Disagreements were resolved by consensus or a third reviewer (S.R.M.). We considered any original research that compared readmission or mortality after discharge for hospitalized medical patients (ie, general patients or subgroups thereof) with versus without depression identified by any validated depression measure,[16] including any study design that incorporated at least 30‐day follow‐up postdischarge. We excluded studies that examined patients hospitalized in nonacute care settings or on surgical, psychiatric, obstetric, or intensive care services. We calculated Cohen's coefficient to evaluate inter‐rater agreement on study selection.

Data Extraction

Data were abstracted by 2 authors (J.L.P. and L.M.W.). Disagreements were resolved by consensus or a third reviewer (S.R.M.). We contacted authors of all included studies to obtain missing data. If unavailable, crude data were estimated from published survival curves employing validated techniques in R (version 3.1.2; R Foundation for Statistical Computing, Vienna, Austria) and Digitizeit (http://www.digitieit.de; DigitizeIt, Braunschweig, Germany).[17, 18] We sought information on trial characteristics (country, type of hospital, inclusion and exclusion criteria, sample size, follow‐up duration, attrition), participants (age, sex, ethnicity, level of education and social support, comorbidities, marital status), exposure ascertainment (self‐report depression screening tool or diagnostic interview for depression), and outcomes (primary: 30‐day all‐cause readmission or mortality, secondary: 90‐day all‐cause readmission or mortality, ED visits, primary care physician [PCP] visits).

Data Synthesis and Statistical Analysis

Where possible, we calculated the pooled risk ratio (RR) with 95% confidence interval (95% CI) using a random effects models in Review Manager (RevMan) 5.3 (The Nordic Cochrane Centre, Copenhagen, Denmark). The random effects approach that we employed assumes heterogeneity (ie, underlying parameters vary between individual studies) and is distributed around a mean or population average effect, and results in more conservative (wider) confidence intervals, wherein larger cohorts (or studies with smaller standard errors) are given more weight. Heterogeneity was assessed using the I² statistic, with values of <25%, 25% to 50%, and >50% representing low, moderate, and high heterogeneity.[19] As per the guidance of Higgins et al., we did not a priori define any degree of heterogeneity that would preclude pooling of the data; the expectation would be that heterogeneity would be substantially higher pooling observational studies rather than randomized trials.[19] Statistical significance was considered a 2‐sided P value 0.05.

Quality Assessment and Risk of Bias

We assessed study quality using the 9‐item Newcastle‐Ottawa scale with 0 to 3, 4 to 6, and 7 to 9 stars considered low, moderate, and high quality, respectively, and criteria for external and internal validity, including group selection and comparability, outcome assessment, and adequacy of follow‐up.[20] Adjusted estimates published in individual reports (or obtained directly from authors) were compared wherever possible with unadjusted estimates to assess the degree of confounding. We generated funnel plots in RevMan 5.3 and conducted Egger tests using Stata 13 (StataCorp LP, College Station, TX) to assess for publication bias.[21]

Study Selection

After removing duplicate publications, we identified 4066 reports and reviewed 133 reports in full text (see Supporting Figure 1 in the online version of this article). Despite our broad study inclusion criteria, we found only 35 longitudinal studies addressing this question. All 35 authors were contacted for additional outcomes data and other missing information (response rate of 34%). We had to exclude 17 studies as they did not provide 30 or 90‐day post‐discharge outcomes. Only 4 studies had published crude data for outcomes within 90 days,[22, 23, 24, 25] but after contact with authors, we received unpublished data for a further 7 studies[26, 27, 28, 29, 30, 31, 32] (including individual level data for 2 cohorts).[31, 32] We were able to estimate crude data from Kaplan‐Meier curves for another 3 studies.[33, 34, 35] Another 4 studies did not collect the outcomes we were interested in individually. These studies were included in this systematic review but are not poolable in our models: 3 authors could only provide composite endpoint data,[36, 37, 38] and 1 author provided unadjusted hazard ratios.[39] Inter‐reviewer agreement for inclusion was 80% (Cohen's = 0.60).

Characteristics of Included Studies

The 18 studies ranged in size from 58 to 1418 patients; 13 were cohort studies and 5 included secondary data from randomized control trials.[22, 27, 30, 34, 36] All studies ascertained depressive status by screening during index medical admission with either diagnostic interview or self‐report questionnaires, although a variety of scales and definitions for depression were used (Beck Depression Inventory [BDI] in 6 studies, Geriatric Depression Scale in 5 studies, Patient Health Questionnaire in another 4 studies, Medical Outcomes Study‐Depression Questionnaire in 1 study, and Center for Epidemiologic Studies Depression Scale in another study) (Table 1). Screening interviews were conducted mostly by research assistants or nurses (68%) or self‐administered (21%). Most studies examined specific medical patient subgroups (10 cardiac, 3 pulmonary, and 2 elderly). Major exclusion criteria reported were terminal illness (4 studies), unstable condition (6 studies), severe cognitive impairment (5 studies), and suicidal ideation or known depression (4 studies); 1 study enrolled patients with suspected depression (Table 1). Patient cohorts were on average older (range, 5082 years) (Table 1). Attrition rates for readmission and mortality data were low (average <1% among entire sample of studies). All studies scored at least 5 on the Newcastle‐Ottawa scale and were thus considered of at least moderate quality (see Supporting Table 1 in the online version of this article).

Prevalence and Recognition of Depressive Symptoms

The range of depression prevalence in hospitalized medical patients was 14% to 79%, with a median of 32% (interquartile range, 27%40%) (Table 1). In those studies that used a diagnostic interview, the prevalence tended to be lower for major depression, with a median of 17% (interquartile range, 16%22%) (Table 1). None of the included studies reported frequency of clinically recognized depression (ie, prior to screening for the study). Only 2 studies assessed the persistence of depression after discharge: 1 reported that depression persisted in 53% (by screening questionnaire) and 34% (by diagnostic interview) of patients at 30 days,[38] whereas the other reported 48% persistence at 90 days after discharge according to a combined screening method.[37]

Hospital Readmission

Overall, 8 studies provided readmission data. Among patients discharged from acute care medical wards (4 studies reporting on 5 cohorts), 395 of 2433 (16.2%) patients were readmitted within 30 days (Figure 1). Hospitalized patients with depressive symptoms were more likely to be readmitted within 30 days after discharge (20.4% vs 13.7%, RR: 1.73, 95% CI: 1.16‐2.58, P = 0.007, I² = 55%) (Figure 1), compared to those without depression. Results were consistent for 90‐day readmissions (39.8% vs 31.0%, RR: 1.68, 95% CI: 1.13‐2.50, P = 0.01, I² = 76%, n = 1543 patients) (see Supporting Figure 2 in the online version of this article) in 6 studies. One individual study examined readmission within 6 months after discharge, but was not poolable in this model, as it presented only hazard ratios and not raw data; however, it did report a 50% increased risk of readmission in medical inpatients aged 75 years (adjusted hazard ratio: 1.50, 95% CI: 1.03‐2.17, n = 401).[39]

Figure 1

Mortality After Discharge

Overall, 11 studies provided all‐cause mortality data. Among medical patients discharged from acute care in 9 studies, 69 of 3397 (2.0%) patients died within 30 days (Figure 2). Medical patients discharged with depressive symptoms were more likely to die within 30 days (2.8% vs 1.5%, RR: 2.13, 95% CI: 1.31‐3.44, P = 0.002, I² = 0%) (Figure 2) compared to those without depression. Similar results were found for 90‐day mortality (7.7% vs 4.1%, RR: 2.01, 95% CI: 1.47‐2.76, P < 0.001, I² = 4%, n = 3784 patients) (see Supporting Figure 3 in the online version of this article) in 11 studies.

Figure 2

ED and PCP Visits

Four studies examined the use of ED or PCP services within 90 days of discharge, but 3 did not have extractable data for meta‐analysis. All showed increased utilization of health services for depressed compared to nondepressed patients after discharge.[22, 36, 37, 38] Depressed patients were more likely to visit the ED (adjusted incidence rate ratio: 1.73, 95% CI: 1.27‐2.36),[36] had significantly more medical encounters (eg, PCP, ED visits, hospital admissions, laboratory tests, and home care [mean 2.9 vs 2.6, P = 0.05])[38] and had a greater number of ED visits alone (27 vs 15 per 100 patients, P = 0.007)[22] within 30 days of hospital discharge compared to nondepressed patients. Similar results were found at 90 days.[36]

Sensitivity Analyses

All told, most studies reported a positive association between depression and adverse events, and this was true regardless of how much adjustment for potential confounding had been undertaken by the authors. Although all studies were qualitatively in the same direction, the magnitude of the association varied due to methodological and/or clinical heterogeneity. Sensitivity analysis revealed no overall difference in pooled risk ratios or heterogeneity between Mantel‐Haenszel fixed effects versus random effects models or with the addition of 0.5 to cells to permit inclusion of zero‐event data. There was no evidence of publication bias; funnel plots and Egger test results are available upon request. There were no statistically significant differences in the risk associated with depressive symptoms whether studies used Diagnostic and Statistical Manual of Mental Disorders (DSM)‐III or DSM‐IV criteria, whether the study samples were disease specific or unselected general medical cohorts, whether studies were of moderate or high quality, or regardless of the severity of depressive symptoms.

Summary of Evidence

We found that depression was common in medical inpatients (about one‐third of all patients) and persisted for at least 30 days in up to half of those patients after discharge. We found strong evidence of an association between depressive symptoms and poor short‐term prognosis after discharge from the hospital: a 73% increased risk of readmission and a 2‐fold risk of death within 30 days compared to patients without depressive symptoms with similar results at 90 days.

Our meta‐analysis complements a recent systematic review that found concomitant depression to be a risk factor for poor prognosis among inpatients and outpatients with acute coronary syndrome,[40] and a meta‐analysis that demonstrated an increased risk of 2‐year mortality among patients with depression after myocardial infarction.[41] To our knowledge, our study is the first to quantify the short‐term postdischarge risks across a diverse group of medical inpatients.

The potential mechanisms underlying the observed relationship between depression and adverse patient outcomes after discharge are likely multiple. We believe there are 2 main possibilities. First, the increased risk associated with depression might be due to residual confounding, even though many of these studies did adjust for extensive lists of comorbidities,[22, 24, 26, 27, 29, 30, 33, 35, 36, 39] including functional status[39] and prior health services utilization.[22, 34, 36] This could occur if other risk factors were not sufficiently adjusted for, such as unrecognized comorbidities or concomitant disability, which are often present among chronically ill patients,[42] or if depression were a marker of psychosocial risk factors such as anxiety,[43] stress or poor resiliency,[44] or low social support,[45] though a few adjusted for psychosocial factors such as social support[26] or anxiety.[35] Confounding could also occur if symptoms of acute illness inflate reports of somatic symptoms of depression on self‐report questionnaires. Recent studies on the BDI, found that scores were higher in postmyocardial infarction patients when compared to outpatient controls,[46] but with no differences between those groups in scores for the BDI‐II,[47] a version with fewer somatic symptom questions.

Second, depression may cause adverse outcomes through indirect or direct pathways. Indirect causation could occur if depression hindered self‐care behaviors such as medication adherence.[42] Depression could also act directly through pathophysiological changes. Some studies have suggested that depression is associated with metabolic abnormalities, including alterations in glucose transport[42, 48] and increased vulnerability to obesity, type 2 diabetes mellitus, and/or diabetic complications, common conditions among hospitalized patients that also adversely affect postdischarge outcomes.[40, 48]

Strengths and Limitations

This review has multiple strengths. We cast a broad search and included studies that examined a wide range of medical patient subgroups, thus increasing the generalizability of our findings. We identified a general scarcity of studies on this topic and obtained additional unpublished data for 10 of the 18 relevant studies, and our response rate of 34% is compatible with the 37% response rate reported for Cochrane reviews when seeking additional data from authors.[49] Whether examined qualitatively (vote counting of the number of studies that showed an association) or quantitatively (via formal meta‐analysis), it seems apparent that there is a clinically important association between depression and postdischarge adverse events, but given the number, quality, and heterogeneity of the studies we examined, there may be some ongoing dispute about exactly how strong this association is and the degree of bias contributed by a couple of large studies of the topic.

There are limitations to our review. First, as we did not have individual‐level patient data, we could not use metaregression to explore sources of heterogeneity (clinical or methodological) or adjust for confounding, and this likely contributes to observed differences between individual estimates. For instance, the included studies had heterogeneous screening measures and cutoffs; thus, all cases of depression in these studies might not be equivalent. Some of the included studies assessed depression early during admission where psychological distress may be greatest; others assessed symptoms closer to discharge. Most studies included patients with specific conditions like heart failure or chronic obstructive pulmonary disease rather than a wide spectrum of medical inpatients. Moreover, few studies adjusted for psychosocial risk factors such as social support, anxiety, and functional status, and only 2 studies assessed the persistence of depressive symptoms after discharge. Second, we did not explore quantitative measures of between‐study variation (eg, I²), because experts question its utility given the expected heterogeneity in meta‐analyses of observational studies.[50] Third, although the included studies were deemed to be of at least moderate quality, they could be at risk for sources of bias that may not be sufficiently appraised by the current version of the Newscastle‐Ottawa scale for observational studies. Finally, we excluded grey literature (eg, conference proceedings or technical reports) that could potentially exclude null findings, although we did contact authors in this field to identify additional unpublished data relevant to this topic.

Data Sources and Search Methods

We searched the Cumulative Index to Nursing and Allied Health Literature, Ovid MEDLINE, Ovid Embase, and PsycINFO from inception to January 9, 2015, and the last 5 years of PubMed for full publications with any of the following Medical Subject Headings: depressive disorder, depression, patient readmission, interviews, psychological, inpatients, with restrictions for peer‐reviewed publication, humans, adults aged 18 years, and the English language. Search strategies were developed with a librarian (available upon request). We manually searched reference lists of all included studies and relevant review articles and contacted content experts to identify additional publications.

Eligibility Criteria and Selection of Studies

Two authors (J.L.P. and L.M.W.) independently screened full texts of all relevant articles for inclusion. Disagreements were resolved by consensus or a third reviewer (S.R.M.). We considered any original research that compared readmission or mortality after discharge for hospitalized medical patients (ie, general patients or subgroups thereof) with versus without depression identified by any validated depression measure,[16] including any study design that incorporated at least 30‐day follow‐up postdischarge. We excluded studies that examined patients hospitalized in nonacute care settings or on surgical, psychiatric, obstetric, or intensive care services. We calculated Cohen's coefficient to evaluate inter‐rater agreement on study selection.

Data Extraction

Data were abstracted by 2 authors (J.L.P. and L.M.W.). Disagreements were resolved by consensus or a third reviewer (S.R.M.). We contacted authors of all included studies to obtain missing data. If unavailable, crude data were estimated from published survival curves employing validated techniques in R (version 3.1.2; R Foundation for Statistical Computing, Vienna, Austria) and Digitizeit (http://www.digitieit.de; DigitizeIt, Braunschweig, Germany).[17, 18] We sought information on trial characteristics (country, type of hospital, inclusion and exclusion criteria, sample size, follow‐up duration, attrition), participants (age, sex, ethnicity, level of education and social support, comorbidities, marital status), exposure ascertainment (self‐report depression screening tool or diagnostic interview for depression), and outcomes (primary: 30‐day all‐cause readmission or mortality, secondary: 90‐day all‐cause readmission or mortality, ED visits, primary care physician [PCP] visits).

Data Synthesis and Statistical Analysis

Where possible, we calculated the pooled risk ratio (RR) with 95% confidence interval (95% CI) using a random effects models in Review Manager (RevMan) 5.3 (The Nordic Cochrane Centre, Copenhagen, Denmark). The random effects approach that we employed assumes heterogeneity (ie, underlying parameters vary between individual studies) and is distributed around a mean or population average effect, and results in more conservative (wider) confidence intervals, wherein larger cohorts (or studies with smaller standard errors) are given more weight. Heterogeneity was assessed using the I² statistic, with values of <25%, 25% to 50%, and >50% representing low, moderate, and high heterogeneity.[19] As per the guidance of Higgins et al., we did not a priori define any degree of heterogeneity that would preclude pooling of the data; the expectation would be that heterogeneity would be substantially higher pooling observational studies rather than randomized trials.[19] Statistical significance was considered a 2‐sided P value 0.05.

Quality Assessment and Risk of Bias

We assessed study quality using the 9‐item Newcastle‐Ottawa scale with 0 to 3, 4 to 6, and 7 to 9 stars considered low, moderate, and high quality, respectively, and criteria for external and internal validity, including group selection and comparability, outcome assessment, and adequacy of follow‐up.[20] Adjusted estimates published in individual reports (or obtained directly from authors) were compared wherever possible with unadjusted estimates to assess the degree of confounding. We generated funnel plots in RevMan 5.3 and conducted Egger tests using Stata 13 (StataCorp LP, College Station, TX) to assess for publication bias.[21]

Study Selection

After removing duplicate publications, we identified 4066 reports and reviewed 133 reports in full text (see Supporting Figure 1 in the online version of this article). Despite our broad study inclusion criteria, we found only 35 longitudinal studies addressing this question. All 35 authors were contacted for additional outcomes data and other missing information (response rate of 34%). We had to exclude 17 studies as they did not provide 30 or 90‐day post‐discharge outcomes. Only 4 studies had published crude data for outcomes within 90 days,[22, 23, 24, 25] but after contact with authors, we received unpublished data for a further 7 studies[26, 27, 28, 29, 30, 31, 32] (including individual level data for 2 cohorts).[31, 32] We were able to estimate crude data from Kaplan‐Meier curves for another 3 studies.[33, 34, 35] Another 4 studies did not collect the outcomes we were interested in individually. These studies were included in this systematic review but are not poolable in our models: 3 authors could only provide composite endpoint data,[36, 37, 38] and 1 author provided unadjusted hazard ratios.[39] Inter‐reviewer agreement for inclusion was 80% (Cohen's = 0.60).

Characteristics of Included Studies

The 18 studies ranged in size from 58 to 1418 patients; 13 were cohort studies and 5 included secondary data from randomized control trials.[22, 27, 30, 34, 36] All studies ascertained depressive status by screening during index medical admission with either diagnostic interview or self‐report questionnaires, although a variety of scales and definitions for depression were used (Beck Depression Inventory [BDI] in 6 studies, Geriatric Depression Scale in 5 studies, Patient Health Questionnaire in another 4 studies, Medical Outcomes Study‐Depression Questionnaire in 1 study, and Center for Epidemiologic Studies Depression Scale in another study) (Table 1). Screening interviews were conducted mostly by research assistants or nurses (68%) or self‐administered (21%). Most studies examined specific medical patient subgroups (10 cardiac, 3 pulmonary, and 2 elderly). Major exclusion criteria reported were terminal illness (4 studies), unstable condition (6 studies), severe cognitive impairment (5 studies), and suicidal ideation or known depression (4 studies); 1 study enrolled patients with suspected depression (Table 1). Patient cohorts were on average older (range, 5082 years) (Table 1). Attrition rates for readmission and mortality data were low (average <1% among entire sample of studies). All studies scored at least 5 on the Newcastle‐Ottawa scale and were thus considered of at least moderate quality (see Supporting Table 1 in the online version of this article).

Prevalence and Recognition of Depressive Symptoms

The range of depression prevalence in hospitalized medical patients was 14% to 79%, with a median of 32% (interquartile range, 27%40%) (Table 1). In those studies that used a diagnostic interview, the prevalence tended to be lower for major depression, with a median of 17% (interquartile range, 16%22%) (Table 1). None of the included studies reported frequency of clinically recognized depression (ie, prior to screening for the study). Only 2 studies assessed the persistence of depression after discharge: 1 reported that depression persisted in 53% (by screening questionnaire) and 34% (by diagnostic interview) of patients at 30 days,[38] whereas the other reported 48% persistence at 90 days after discharge according to a combined screening method.[37]

Hospital Readmission

Overall, 8 studies provided readmission data. Among patients discharged from acute care medical wards (4 studies reporting on 5 cohorts), 395 of 2433 (16.2%) patients were readmitted within 30 days (Figure 1). Hospitalized patients with depressive symptoms were more likely to be readmitted within 30 days after discharge (20.4% vs 13.7%, RR: 1.73, 95% CI: 1.16‐2.58, P = 0.007, I² = 55%) (Figure 1), compared to those without depression. Results were consistent for 90‐day readmissions (39.8% vs 31.0%, RR: 1.68, 95% CI: 1.13‐2.50, P = 0.01, I² = 76%, n = 1543 patients) (see Supporting Figure 2 in the online version of this article) in 6 studies. One individual study examined readmission within 6 months after discharge, but was not poolable in this model, as it presented only hazard ratios and not raw data; however, it did report a 50% increased risk of readmission in medical inpatients aged 75 years (adjusted hazard ratio: 1.50, 95% CI: 1.03‐2.17, n = 401).[39]

Figure 1

Mortality After Discharge

Overall, 11 studies provided all‐cause mortality data. Among medical patients discharged from acute care in 9 studies, 69 of 3397 (2.0%) patients died within 30 days (Figure 2). Medical patients discharged with depressive symptoms were more likely to die within 30 days (2.8% vs 1.5%, RR: 2.13, 95% CI: 1.31‐3.44, P = 0.002, I² = 0%) (Figure 2) compared to those without depression. Similar results were found for 90‐day mortality (7.7% vs 4.1%, RR: 2.01, 95% CI: 1.47‐2.76, P < 0.001, I² = 4%, n = 3784 patients) (see Supporting Figure 3 in the online version of this article) in 11 studies.

Figure 2

ED and PCP Visits

Four studies examined the use of ED or PCP services within 90 days of discharge, but 3 did not have extractable data for meta‐analysis. All showed increased utilization of health services for depressed compared to nondepressed patients after discharge.[22, 36, 37, 38] Depressed patients were more likely to visit the ED (adjusted incidence rate ratio: 1.73, 95% CI: 1.27‐2.36),[36] had significantly more medical encounters (eg, PCP, ED visits, hospital admissions, laboratory tests, and home care [mean 2.9 vs 2.6, P = 0.05])[38] and had a greater number of ED visits alone (27 vs 15 per 100 patients, P = 0.007)[22] within 30 days of hospital discharge compared to nondepressed patients. Similar results were found at 90 days.[36]

Sensitivity Analyses

All told, most studies reported a positive association between depression and adverse events, and this was true regardless of how much adjustment for potential confounding had been undertaken by the authors. Although all studies were qualitatively in the same direction, the magnitude of the association varied due to methodological and/or clinical heterogeneity. Sensitivity analysis revealed no overall difference in pooled risk ratios or heterogeneity between Mantel‐Haenszel fixed effects versus random effects models or with the addition of 0.5 to cells to permit inclusion of zero‐event data. There was no evidence of publication bias; funnel plots and Egger test results are available upon request. There were no statistically significant differences in the risk associated with depressive symptoms whether studies used Diagnostic and Statistical Manual of Mental Disorders (DSM)‐III or DSM‐IV criteria, whether the study samples were disease specific or unselected general medical cohorts, whether studies were of moderate or high quality, or regardless of the severity of depressive symptoms.

Summary of Evidence

We found that depression was common in medical inpatients (about one‐third of all patients) and persisted for at least 30 days in up to half of those patients after discharge. We found strong evidence of an association between depressive symptoms and poor short‐term prognosis after discharge from the hospital: a 73% increased risk of readmission and a 2‐fold risk of death within 30 days compared to patients without depressive symptoms with similar results at 90 days.

Our meta‐analysis complements a recent systematic review that found concomitant depression to be a risk factor for poor prognosis among inpatients and outpatients with acute coronary syndrome,[40] and a meta‐analysis that demonstrated an increased risk of 2‐year mortality among patients with depression after myocardial infarction.[41] To our knowledge, our study is the first to quantify the short‐term postdischarge risks across a diverse group of medical inpatients.

The potential mechanisms underlying the observed relationship between depression and adverse patient outcomes after discharge are likely multiple. We believe there are 2 main possibilities. First, the increased risk associated with depression might be due to residual confounding, even though many of these studies did adjust for extensive lists of comorbidities,[22, 24, 26, 27, 29, 30, 33, 35, 36, 39] including functional status[39] and prior health services utilization.[22, 34, 36] This could occur if other risk factors were not sufficiently adjusted for, such as unrecognized comorbidities or concomitant disability, which are often present among chronically ill patients,[42] or if depression were a marker of psychosocial risk factors such as anxiety,[43] stress or poor resiliency,[44] or low social support,[45] though a few adjusted for psychosocial factors such as social support[26] or anxiety.[35] Confounding could also occur if symptoms of acute illness inflate reports of somatic symptoms of depression on self‐report questionnaires. Recent studies on the BDI, found that scores were higher in postmyocardial infarction patients when compared to outpatient controls,[46] but with no differences between those groups in scores for the BDI‐II,[47] a version with fewer somatic symptom questions.

Second, depression may cause adverse outcomes through indirect or direct pathways. Indirect causation could occur if depression hindered self‐care behaviors such as medication adherence.[42] Depression could also act directly through pathophysiological changes. Some studies have suggested that depression is associated with metabolic abnormalities, including alterations in glucose transport[42, 48] and increased vulnerability to obesity, type 2 diabetes mellitus, and/or diabetic complications, common conditions among hospitalized patients that also adversely affect postdischarge outcomes.[40, 48]

Strengths and Limitations

This review has multiple strengths. We cast a broad search and included studies that examined a wide range of medical patient subgroups, thus increasing the generalizability of our findings. We identified a general scarcity of studies on this topic and obtained additional unpublished data for 10 of the 18 relevant studies, and our response rate of 34% is compatible with the 37% response rate reported for Cochrane reviews when seeking additional data from authors.[49] Whether examined qualitatively (vote counting of the number of studies that showed an association) or quantitatively (via formal meta‐analysis), it seems apparent that there is a clinically important association between depression and postdischarge adverse events, but given the number, quality, and heterogeneity of the studies we examined, there may be some ongoing dispute about exactly how strong this association is and the degree of bias contributed by a couple of large studies of the topic.

There are limitations to our review. First, as we did not have individual‐level patient data, we could not use metaregression to explore sources of heterogeneity (clinical or methodological) or adjust for confounding, and this likely contributes to observed differences between individual estimates. For instance, the included studies had heterogeneous screening measures and cutoffs; thus, all cases of depression in these studies might not be equivalent. Some of the included studies assessed depression early during admission where psychological distress may be greatest; others assessed symptoms closer to discharge. Most studies included patients with specific conditions like heart failure or chronic obstructive pulmonary disease rather than a wide spectrum of medical inpatients. Moreover, few studies adjusted for psychosocial risk factors such as social support, anxiety, and functional status, and only 2 studies assessed the persistence of depressive symptoms after discharge. Second, we did not explore quantitative measures of between‐study variation (eg, I²), because experts question its utility given the expected heterogeneity in meta‐analyses of observational studies.[50] Third, although the included studies were deemed to be of at least moderate quality, they could be at risk for sources of bias that may not be sufficiently appraised by the current version of the Newscastle‐Ottawa scale for observational studies. Finally, we excluded grey literature (eg, conference proceedings or technical reports) that could potentially exclude null findings, although we did contact authors in this field to identify additional unpublished data relevant to this topic.

Study Setting

The Canadian province of Alberta has a single vertically integrated healthcare system that is government‐funded and provides universal access to hospitals, emergency departments (EDs), and outpatient physician services for all 4.1 million Albertans as well as all prescription medications for the poor, socially disadvantaged, disabled, or those age 65 years and older. This study received approval from the University of Alberta Health Research Ethics Board with waiver of informed consent.

Data Sources

This study used deidentified linked data from 3 Alberta Health administrative databases that capture vital status and all hospital or ED visits and have previously been shown to have high accuracy for medical diagnoses.[13] The Alberta Health Care Insurance Plan Registry tracks date of death or emigration from the province. The Discharge Abstract Database includes the most responsible diagnosis identified by the hospital attending physician, up to 25 other diagnoses coded by nosologists in each hospital, the admission and discharge dates, and the admission category (elective or urgent/emergent) for all acute care hospitalizations. Of note, unlike US studies, the hospital databases are able to distinguish in‐hospital (eg, adverse events) versus premorbid diagnoses (eg, preexisting comorbidities). The Ambulatory Care Database captures all patient visits to EDs with coding for up to 10 conditions per encounter.

Study Cohort

We identified all adults with an acute care hospitalization on the GIM services at all 7 Alberta teaching hospitals (ie, defined as those with Royal College of Physicians and Surgeons of Canadaapproved residency training programs in internal medicine, the equivalent of the Association of American Medical Colleges certification in the United States) between October 1, 2009 and September 30, 2010 and between April 1, 2011 and December 1, 2011 (these 20 months covered most of the pre/post intervals for a recently reported quality improvement initiative at 1 of the teaching hospitals that had no significant impact on postdischarge outcomes).[14] Patients from out of the province or transferred from/to another inpatient service (eg, the intensive care unit, a different service in the same hospital [such as surgery], another acute care hospital, or rehabilitation hospital) or with lengths of stay greater than 30 days were excluded. We only included the first hospitalization for any patient in our study timeframe and thus excluded repeat discharges of the same patient.

Explanatory Variable of Interest

The independent variable of interest was calendar day of discharge, stratified according to weekday (Monday thru Friday) versus weekend (Saturday and Sunday). Only 1.4% of weekday discharges occurred on a statutory holiday, and for the purposes of this study, these discharges were also considered weekend discharges. At the 7 teaching hospitals in Alberta, nursing staffing ratios do not differ between weekend and weekday, but availability of all other members of the healthcare team does. Physician census decreases from 4 to 5 per ward to 1 to 2, and ward‐based social workers, occupational therapists, physiotherapists, and pharmacist educators are generally not available on weekends.

Outcomes

Our primary outcome of interest was the composite outcome of death or all‐cause nonelective readmission within 30 days of discharge (ie, not including in‐hospital events prior to discharge or elective readmissions after discharge for planned procedures such as chemotherapy); hereafter we refer to this as death or readmission. This is a patient‐relevant outcome that is highlighted in the Affordable Care Act and for which there are several validated risk adjustment models.[15] We chose a composite outcome to deal with the issue of competing risks; if weekend discharges were more likely to die then we could observe a spurious association between weekend discharge and reduced readmissions if we focused on only that outcome.

Other Measures

Comorbidities for each patient were identified using International Classification of Diseases, Ninth Revision and Tenth Revision codes from the Discharge Abstract Database for the index hospitalization and any hospitalizations in the 12 months prior to their index admission, a method previously validated in Alberta databases.[13] We also recorded health resource use during their index hospitalization and calculated each patient's LACE score at the time of discharge, which is an index for predicting unplanned readmission or early death postdischarge previously validated in Canadian administrative databases.[15] The LACE index includes length of hospital stay (L), acuity of admission (A, based on the admission category variable described earlier), comorbidity burden quantified using the Charlson Comorbidity Index (C), and emergency department visits in the 6 months prior to admission (E); patients with discharge LACE scores >10 (total possible score is 19) are defined as being at high risk of death/readmission within 30 days.[16] As detailed below, to deal with potential concerns that LOS may be a mediator in the causal pathway, we ran 2 sensitivity analyses, 1 in which we excluded LOS from the analyses and 1 in which we included expected LOS rather than the actual LOS. Expected LOS is a data‐driven estimate based on the most current 2 years of patient LOS information available in the Canadian Institute for Health Information discharge abstract database (www.cihi.ca) for all acute care hospitals in Canada, and was generated for each patient independently of our study taking into account case mix group, age, and inpatient resource intensity weights.

Statistical Analysis

Baseline patient characteristics between weekend and weekday discharges were compared with t tests for continuous variables and [2] tests for binary or categorical variables. Logistic regression was used for comparison of death or readmission for weekend versus weekday discharges. Multivariable models were adjusted for age, sex, hospital, and LACE scores (as a continuous variable) at time of discharge; in sensitivity analyses we adjusted for (1) LACE score without including LOS and (2) LACE score using expected LOS rather than actual LOS. In further sensitivity analyses we (1) restricted the analysis to only those patients deemed to be at high risk for events due to LACE scores of 10 or greater and (2) included ED visits as part of the composite endpoint (ie, death, unplanned readmission, or unplanned ED visit within 30 days of discharge). Day of admission (weekend vs weekday) was also considered for the multivariable models, but was not found to be significant and thus was omitted from final models. We do not have any physician identifying variables in our dataset and thus could not investigate the potential correlation among patients discharged by the same physician. We did explore the hospital intraclass correlation coefficient, and as it was very small (0.001), we did not utilize models to account for the hierarchical nature of the data, but did include hospital as a fixed effect in the logistic models. The results were virtually identical whether we did or did not include hospital in the models. Adjusted odds ratios (aORs) are displayed with 95% confidence intervals (CI) and P values. Average LOS was calculated for weekend and weekday discharges with 95% CIs. P values for adjusted length of stay were calculated using multivariable linear regression adjusting for age, sex, day of admission, and Charlson score. All statistical analyses were done using SAS for Windows version 9.4 (SAS Institute, Inc., Cary, NC).

Patient Characteristics

Of the 7991 patients discharged during our study interval, 1146 (14.3%) were discharged on weekend or holiday days (Table 1). In contrast, 2180 of our cohort were admitted on a weekend (27.3%). The mean age of our study population was 62.1 years, 51.9% were men, mean Charlson score was 2.56, and 4591 (57.5%) had LACE scores of at least 10 at discharge.

Weekday Versus Weekend Discharge

Although patients admitted on weekdays and weekends were very similar (data available upon request), patients discharged on weekends (compared to those discharged on weekdays) were younger, more likely to be discharged home without additional support, and had fewer comorbidities (Table 1, Figure 1). Patients discharged on weekends had shorter lengths of stay than those discharged on weekdays (5.6 days vs 7.9 days, P<0.0001). In adjusted linear regression analyses, this 2.3‐day difference remained statistically significant (adjusted P value <0.0001).

Figure 1

Patients discharged on a weekend exhibited lower unadjusted 30‐day rates of death or readmission than those discharged on a weekday (10.6% vs 13.2%), but these differences disappeared after multivariable adjustment that accounted for differences in risk profile (aOR: 0.94, 95% CI: 0.771.16 (Table 2). Results were similar in sensitivity analyses adjusting for LACE scores without LOS included (aOR: 0.88, 95% CI: 0.711.08) or adjusting for LACE scores using expected LOS rather than actual LOS (aOR: 0.90, 95% CI: 0.731.10). Restricting the analysis to only those patients deemed to be at high risk for events due to LACE scores of 10 or greater confirmed that weekend and weekday discharges had similar outcomes in the first 30 days after discharge (aOR: 1.09, 95% CI: 0.851.41, Table 2). Similar patterns were seen when we included ED visits as part of the composite endpoint (ie, death, unplanned readmission, or unplanned ED visit within 30 days of discharge) (Table 2).

Study Setting

The Canadian province of Alberta has a single vertically integrated healthcare system that is government‐funded and provides universal access to hospitals, emergency departments (EDs), and outpatient physician services for all 4.1 million Albertans as well as all prescription medications for the poor, socially disadvantaged, disabled, or those age 65 years and older. This study received approval from the University of Alberta Health Research Ethics Board with waiver of informed consent.

Data Sources

This study used deidentified linked data from 3 Alberta Health administrative databases that capture vital status and all hospital or ED visits and have previously been shown to have high accuracy for medical diagnoses.[13] The Alberta Health Care Insurance Plan Registry tracks date of death or emigration from the province. The Discharge Abstract Database includes the most responsible diagnosis identified by the hospital attending physician, up to 25 other diagnoses coded by nosologists in each hospital, the admission and discharge dates, and the admission category (elective or urgent/emergent) for all acute care hospitalizations. Of note, unlike US studies, the hospital databases are able to distinguish in‐hospital (eg, adverse events) versus premorbid diagnoses (eg, preexisting comorbidities). The Ambulatory Care Database captures all patient visits to EDs with coding for up to 10 conditions per encounter.

Study Cohort

We identified all adults with an acute care hospitalization on the GIM services at all 7 Alberta teaching hospitals (ie, defined as those with Royal College of Physicians and Surgeons of Canadaapproved residency training programs in internal medicine, the equivalent of the Association of American Medical Colleges certification in the United States) between October 1, 2009 and September 30, 2010 and between April 1, 2011 and December 1, 2011 (these 20 months covered most of the pre/post intervals for a recently reported quality improvement initiative at 1 of the teaching hospitals that had no significant impact on postdischarge outcomes).[14] Patients from out of the province or transferred from/to another inpatient service (eg, the intensive care unit, a different service in the same hospital [such as surgery], another acute care hospital, or rehabilitation hospital) or with lengths of stay greater than 30 days were excluded. We only included the first hospitalization for any patient in our study timeframe and thus excluded repeat discharges of the same patient.

Explanatory Variable of Interest

The independent variable of interest was calendar day of discharge, stratified according to weekday (Monday thru Friday) versus weekend (Saturday and Sunday). Only 1.4% of weekday discharges occurred on a statutory holiday, and for the purposes of this study, these discharges were also considered weekend discharges. At the 7 teaching hospitals in Alberta, nursing staffing ratios do not differ between weekend and weekday, but availability of all other members of the healthcare team does. Physician census decreases from 4 to 5 per ward to 1 to 2, and ward‐based social workers, occupational therapists, physiotherapists, and pharmacist educators are generally not available on weekends.

Outcomes

Our primary outcome of interest was the composite outcome of death or all‐cause nonelective readmission within 30 days of discharge (ie, not including in‐hospital events prior to discharge or elective readmissions after discharge for planned procedures such as chemotherapy); hereafter we refer to this as death or readmission. This is a patient‐relevant outcome that is highlighted in the Affordable Care Act and for which there are several validated risk adjustment models.[15] We chose a composite outcome to deal with the issue of competing risks; if weekend discharges were more likely to die then we could observe a spurious association between weekend discharge and reduced readmissions if we focused on only that outcome.

Other Measures

Comorbidities for each patient were identified using International Classification of Diseases, Ninth Revision and Tenth Revision codes from the Discharge Abstract Database for the index hospitalization and any hospitalizations in the 12 months prior to their index admission, a method previously validated in Alberta databases.[13] We also recorded health resource use during their index hospitalization and calculated each patient's LACE score at the time of discharge, which is an index for predicting unplanned readmission or early death postdischarge previously validated in Canadian administrative databases.[15] The LACE index includes length of hospital stay (L), acuity of admission (A, based on the admission category variable described earlier), comorbidity burden quantified using the Charlson Comorbidity Index (C), and emergency department visits in the 6 months prior to admission (E); patients with discharge LACE scores >10 (total possible score is 19) are defined as being at high risk of death/readmission within 30 days.[16] As detailed below, to deal with potential concerns that LOS may be a mediator in the causal pathway, we ran 2 sensitivity analyses, 1 in which we excluded LOS from the analyses and 1 in which we included expected LOS rather than the actual LOS. Expected LOS is a data‐driven estimate based on the most current 2 years of patient LOS information available in the Canadian Institute for Health Information discharge abstract database (www.cihi.ca) for all acute care hospitals in Canada, and was generated for each patient independently of our study taking into account case mix group, age, and inpatient resource intensity weights.

Statistical Analysis

Baseline patient characteristics between weekend and weekday discharges were compared with t tests for continuous variables and [2] tests for binary or categorical variables. Logistic regression was used for comparison of death or readmission for weekend versus weekday discharges. Multivariable models were adjusted for age, sex, hospital, and LACE scores (as a continuous variable) at time of discharge; in sensitivity analyses we adjusted for (1) LACE score without including LOS and (2) LACE score using expected LOS rather than actual LOS. In further sensitivity analyses we (1) restricted the analysis to only those patients deemed to be at high risk for events due to LACE scores of 10 or greater and (2) included ED visits as part of the composite endpoint (ie, death, unplanned readmission, or unplanned ED visit within 30 days of discharge). Day of admission (weekend vs weekday) was also considered for the multivariable models, but was not found to be significant and thus was omitted from final models. We do not have any physician identifying variables in our dataset and thus could not investigate the potential correlation among patients discharged by the same physician. We did explore the hospital intraclass correlation coefficient, and as it was very small (0.001), we did not utilize models to account for the hierarchical nature of the data, but did include hospital as a fixed effect in the logistic models. The results were virtually identical whether we did or did not include hospital in the models. Adjusted odds ratios (aORs) are displayed with 95% confidence intervals (CI) and P values. Average LOS was calculated for weekend and weekday discharges with 95% CIs. P values for adjusted length of stay were calculated using multivariable linear regression adjusting for age, sex, day of admission, and Charlson score. All statistical analyses were done using SAS for Windows version 9.4 (SAS Institute, Inc., Cary, NC).

Patient Characteristics

Of the 7991 patients discharged during our study interval, 1146 (14.3%) were discharged on weekend or holiday days (Table 1). In contrast, 2180 of our cohort were admitted on a weekend (27.3%). The mean age of our study population was 62.1 years, 51.9% were men, mean Charlson score was 2.56, and 4591 (57.5%) had LACE scores of at least 10 at discharge.