Ning Tang, MD

Reducing hospital readmissions is a national healthcare priority. In October 2012, the Centers for Medicare and Medicaid Services (CMS) enacted financial penalties on hospitals with higher than average risk‐adjusted readmissions, offering an incentive pool of $850 million in the first year.[1] As a result, a wide range of activities to understand and reduce readmissions among patients with congestive heart failure (CHF), acute myocardial infarction (AMI), and pneumonia have emerged.[2, 3, 4, 5] Some have proposed that to effectively reduce hospital readmissions, a community of inpatient and outpatient providers and local support organizations must coordinate efforts.[6, 7] In fact, CMS has recognized the important role of community support groups and outpatient providers in safe discharge transitions in 2 ways. First, in 2011 CMS launched the Community‐based Care Transitions Program to fund community‐based organizations to assist Medicare patients with care transitions.[8] Second, CMS introduced 2 new reimbursement codes for primary care providers (PCPs) to perform care coordination immediately after hospital discharge.[9] Both of these new payment programs represent an evolving perspective that reducing hospital readmissions requires active participation among outpatient partners.

As leaders of the outpatient care team, PCPs play a significant role in reducing hospital readmissions. One way in which PCPs can begin to understand the magnitude of the issue within their clinic is to evaluate the clinic's 30‐day readmission rates. Currently, CMS calculates readmission rates at the hospital level. However, understanding these rates at the clinic level is critical for developing strategies for improvement across the care continuum. Our current understanding of effective outpatient interventions to reduce hospital readmission is limited.[3] As clinics introduce and refine strategies to reduce readmissions, tracking the impact of these strategies on readmission rates will be critical for identifying effective outpatient interventions. Clinics with similar patient case‐mix can also benchmark readmission rates, sharing best practices from clinics with lower‐than‐expected rates.

There are no available studies or proposed methodologies to guide primary care clinics in calculating their 30‐day readmission rates. A particularly difficult challenge is obtaining the admission information when patients may be admitted to 1 of several area hospitals. For large integrated delivery networks where primary care patients are relatively loyal to the network of physicians and hospitals, an opportunity exists to explore the data. Furthermore, variations in readmission rates across primary care specialties (such as internal medicine and family practice) are not well understood. In this study, we set out to develop a methodology for calculating all‐cause 30‐day hospital readmission rates at the level of individual primary care practices and to identify factors associated with variations in these rates.

METHODS

RESULTS

During the study period, there were 12,564 discharges from UCSF Medical Center for primary care patients belonging to the 7 clinics. Of these, 8685 were index discharges and 1032 were readmissions within 30 days. Table 1 shows the characteristics of the patients who had at least 1 admission during the study period. In all but 2 clinics (HIV Primary Care and Concierge Internal Medicine), there were more women hospitalized than men. Age and gender differences across clinics are consistent with the patient populations served by these clinics, with Women's Health having more female patients and younger patients hospitalized compared to Geriatrics.

All‐cause 30‐day readmission rates varied across practices, with HIV Primary Care being the highest at 14.9%, followed by Geriatrics at 13.7%, General Internal Medicine at 13.3%, Concierge Internal Medicine at 9.7%, combined IM/FP at 8.9%, Family Practice at 8.6%, and Women's Health at 7.7% (Figure 1). Despite HIV Primary Care having the highest readmission rate, the number of index discharges during the 3‐year period was relatively low (249) compared to General Internal Medicine (5388). For the index admission, the top 5 admitting services were medicine, obstetrics, cardiology, orthopedic surgery, and general surgery (Table 2). Medicine was the primary admitting service for patients in the following clinics: General Internal Medicine, Geriatrics, HIV Primary Care, and Concierge Internal Medicine. Obstetrics was the primary admitting service for patients in the Family Practice, Women's Health, and combined IM/FP clinics. Vaginal delivery was the top discharge diagnosis‐related group (DRG) for patients in the General Internal Medicine, Family Practice, Women's Health, and combined IM/FP clinics. The top discharge DRG was urinary tract infection for Geriatrics, HIV for the HIV clinic (13.6%), and chemotherapy (8.0%) for Concierge Internal Medicine. Average LOS varied from 4.7 days for patients in the HIV Primary Care clinic to 2.8 days for patients in the Concierge Internal Medicine clinic. Average LOS was 3.4, 3.1, and 3.2 days in the Family Practice, Women's Health, and Geriatrics clinics, respectively, and 3.8 days in the General Internal Medicine and combined IM/FP clinics. For all clinics except Geriatrics, the majority of patients were discharged home without home health. A larger proportion of patients in the Geriatrics clinic were discharged home with home health or discharged to a skilled nursing facility.

Figure 1

Factors associated with variation in readmission rates included: male gender (OR: 1.21, 95% CI: 1.051.40), Medicare (OR: 1.31, 95% CI: 1.051.64; Ref=private) and dual‐eligible Medicare‐Medicaid insurance (OR: 1.26, 95% CI: 1.011.56), unknown primary language (OR: 0.06, 95% CI: 0.020.25; Ref=English), and the following comorbidities: peripheral vascular disease, renal failure, lymphoma, fluid and electrolyte disorders, and anemia. Multivariable logistic regression modeling results are listed in Table 3. Patients having a resident PCP showed no increased odds of readmission (OR: 1.13, 95% CI: 0.931.37; Ref=attending PCP). However, patients with a graduated resident PCP or departed faculty PCP awaiting transfer to a new PCP had an OR of 1.59 (95% CI: 1.162.17) compared with having a current faculty PCP. The C‐statistic for this model was 0.67.

DISCUSSION

In this study we introduce a complementary way to view hospital readmissions, from the perspective of primary care practices. We found variation in readmission rates across primary care practices. After controlling for admitting service, clinic, provider, and patient factors, the specific characteristics of male gender, patients with Medicare or Medicare with Medicaid, and patients in our General Internal Medicine clinic with a departed PCP were independent risk factors for hospital readmission. Patients with specific comorbidities were also at increased risk for admissions, including those with peripheral vascular disease, renal failure, lymphoma, fluid and electrolyte disorders, and anemia. Because this is the first study viewing readmissions from the perspective of primary care practices, our findings are unique in the literature. However, hospital‐based studies have found similar relationships between readmission rates and these specific comorbidities.[11, 12, 15] Unlike other studies, our study cohort did not show CHF as an independent risk factor. We hypothesize this is because in 2008, UCSF Medical Center introduced an inpatient‐based intervention to reduce readmissions in patients with heart failure. By 2011, the readmission rates of patients with heart failure (primary or secondary diagnoses) had dropped by 30%. The success of this program, focused only on patients with heart failure, likely affected our analysis of comorbidities as risk factors for readmissions.

Models developed to predict hospital readmissions, overall and for specific disease conditions, have inconsistently identified predictive factors, and there is not a specific set of variables that dominate.[16, 17, 18, 19] A recent review of readmission risk prediction models suggested that models that take into account psychosocial factors such as social support, substance abuse, and functional status improved model performance.[16] We hypothesize that the reason why male gender was significant in our model may be related to lack of social support, especially among those who may be single or widowed. Other studies have also showed male gender as a predictor for hospital readmissions.[14, 20]

People with Medicare as the primary payor or Medicare with Medicaid also tended to have higher risk of readmission. We believe that this may be a proxy for the combined effect of age, multiple comorbidities, and psychosocial factors. In a multicenter study of general medicine inpatients, Medicare, but not age, was also found to be a predictive variable.[21] Unlike other studies, our study did not find Medicaid alone as a significant predictive variable for readmissions.[15, 21] One explanation may be that in hospital‐based studies, people with Medicaid who are at high risk for readmission may be high‐risk because they do not have a PCP or good access to outpatient care. In our study, all patients have a PCP in 1 of the UCSF clinics, and access to care is improved with this established relationship. These other studies did not examine Medicare‐Medicaid dual eligibility status. Our results are consistent with a national study on avoidable hospital admissions that showed the dual eligibility population experiencing 60% higher avoidable admission rates compared to the Medicaid‐only population.[22]

An interesting finding was that there were no statistically significant differences in readmissions among patients who report a language other than English as their primary language. Although language barriers and health literacy can affect a patient's ability to understand discharge instructions, the use of translators at UCSF Medical Center may have decreased the risk of readmissions. For a small number of patients whose primary language was not recorded (unknown language in our model), they appeared to have a lower risk of readmissions. Language preferences are recorded by our admitting staff during the process of admission. This step may have been skipped after hours or if the patient was not able to answer the question. However, we do not have a good hypothesis as to why these patients may have a reduced risk of readmissions.

Having a departed PCP in the General Internal Medicine clinic was an independent predictor of readmissions. These patients have access to primary care; they can schedule acute appointments with covering providers for new medical issues or follow‐up of chronic conditions. However, until they are transferred to a new PCP, they do not have a provider who is proactively managing their preventive and chronic disease care, including follow‐up and coordinating care after hospital discharge. Our study is not the first to suggest adverse outcomes for patients who are in transition from 1 primary care provider to the next.[23] Studies conducted in General Internal Medicine clinics have shown missed opportunities for cancer screening and overlooked test results during the transition period,[24] and as many as one‐fifth of patients whom residents identified as high‐risk were lost to follow‐up.[25] However, our study is the first to show the link between PCP transition in a teaching clinic and hospital readmissions. This finding underscores the importance of continuity of care in the optimal management of patients following hospital discharge.

There are several limitations to this study. First, our study only considers hospitalizations to UCSF Medical Center, potentially undercounting readmissions to area hospitals. Because our study population are patients with PCPs at UCSF, these patients tend to seek specialty and acute care at UCSF Medical Center as well. We obtained payor data from our medical group (Hill Physicians), which covers our largest private payors, to understand whether our results can be applied on a global basis. We found that in 87.4% of index admissions with readmissions from January 1, 2010 through May 30, 2012, the readmission occurred at UCSF Medical Center. We conducted a similar analysis with CMS data from October 1, 2008 to June 30, 2011, the latest data we have from CMS. For patients with UCSF PCPs and a diagnosis of AMI or CHF, 100% were readmitted back to UCSF Medical Center. For patients with UCSF PCPs and a diagnosis of pneumonia, 89% were readmitted back to UCSF Medical Center. Given that only a small percentage of patients with PCPs at UCSF present for readmission at other area hospitals, we believe that limiting our analysis to UCSF Medical Center is reasonable.

In our study, we did not remove vaginal deliveries or Caesarian sections prior to building the model. Primary care physicians and their clinic leadership are accustomed to taking a population‐health perspective. We anticipate they would be interested in designing interventions and addressing readmission for the entire primary care panel. Although readmissions after delivery are not frequent, they can still occur, and interventions should not necessarily exclude this population. We did run a sensitivity analysis by removing vaginal delivery and Caesarian sections from the analysis. As expected, readmission rates for all practices increased except for Geriatrics. However, the independent predictors of readmissions did not change.

Our study is based on a population of patients at 1 urban academic medical center and may not be generalizable across all delivery systems. Our population is racially and ethnically diverse, and many do not speak English as their primary language. The study also spans different types of primary care clinics, capturing a wide range of ages and case mix. As PCP assignments fluctuate over time, there may be errors with PCP attribution in the UCSF Medical Center data systems. We do not believe errors in PCP attribution would differ across primary care practices. Because the primary care practices' performance reports on quality measures are based on PCP assignment, each clinic regularly updates their clinic panels and has specific protocols to address errors with PCP attribution.

Finally, our study includes only the variables that we were able to extract from administrative claims. Other explanatory variables that have been suggested as important for evaluation, such as social support, functional assessment, access to care, hospital discharge process, and posthospitalization follow‐up, were not included. Each of these could be explored in future studies.

This study offers a unique perspective of hospital readmission by introducing a new methodology for primary care clinics to calculate and evaluate their all‐cause 30‐day readmission rates. Although this methodology is not intended to provide real‐time feedback to clinics on readmitted patients, it opens the door for benchmarking based on specific case‐mix indices. Another direction for future research is to design robust evaluations of the impact of interventions, both inpatient and outpatient, on primary care clinic readmission rates. Finally, future research should replicate this analysis across teaching clinics to identify whether provider turnover is a consistent independent predictor of hospital readmissions.

This study also has implications for inpatient interventions. For example, discharging physicians may consider proactively identifying whether the patient has a continuous primary care provider. Patients who are in between PCPs may need closer follow‐up after discharge, until they re‐establish with a new PCP. This can be accomplished through a postdischarge clinic visit, either run by inpatient providers or covering physicians in the primary care clinic. In addition, the discharging physician may work with the case manager to increase the level of care coordination. The case manager could contact the primary care clinic and proactively ask for immediate PCP re‐assignment. Once a new PCP has been identified, the discharging physician could consider a warm hand‐off. With a warm hand‐off, the new PCP may feel more comfortable managing problems that may arise after hospital discharge, and especially before the first outpatient visit with the new PCP. Future research can test whether these interventions could effectively reduce hospital readmissions across a broad primary care population.

CONCLUSION

Primary care providers and their clinics play an important role in managing population health, decreasing healthcare spending, and keeping patients out of the hospital. In this study, we introduce a tool in which primary care clinics can begin to understand their hospital readmission rates. This may be particularly valuable as primary care providers enter global payment arrangements such as accountable care organizations or bundled payments and are responsible for a population of patients across the continuum of care. We found significant variation in readmission rates between different primary care practices, but much of this variation appears to be due to differences between practices in patient demographics, comorbidities, and hospitalization factors. Our study is the first to show the association between provider transitions and higher hospital readmissions. Continuity of care is critical for the optimal management of patients following hospital discharge. More attention will need to be focused on providing good continuity outpatient care for patients at high risk for readmissions.

Reducing hospital readmissions is a national healthcare priority. In October 2012, the Centers for Medicare and Medicaid Services (CMS) enacted financial penalties on hospitals with higher than average risk‐adjusted readmissions, offering an incentive pool of $850 million in the first year.[1] As a result, a wide range of activities to understand and reduce readmissions among patients with congestive heart failure (CHF), acute myocardial infarction (AMI), and pneumonia have emerged.[2, 3, 4, 5] Some have proposed that to effectively reduce hospital readmissions, a community of inpatient and outpatient providers and local support organizations must coordinate efforts.[6, 7] In fact, CMS has recognized the important role of community support groups and outpatient providers in safe discharge transitions in 2 ways. First, in 2011 CMS launched the Community‐based Care Transitions Program to fund community‐based organizations to assist Medicare patients with care transitions.[8] Second, CMS introduced 2 new reimbursement codes for primary care providers (PCPs) to perform care coordination immediately after hospital discharge.[9] Both of these new payment programs represent an evolving perspective that reducing hospital readmissions requires active participation among outpatient partners.

As leaders of the outpatient care team, PCPs play a significant role in reducing hospital readmissions. One way in which PCPs can begin to understand the magnitude of the issue within their clinic is to evaluate the clinic's 30‐day readmission rates. Currently, CMS calculates readmission rates at the hospital level. However, understanding these rates at the clinic level is critical for developing strategies for improvement across the care continuum. Our current understanding of effective outpatient interventions to reduce hospital readmission is limited.[3] As clinics introduce and refine strategies to reduce readmissions, tracking the impact of these strategies on readmission rates will be critical for identifying effective outpatient interventions. Clinics with similar patient case‐mix can also benchmark readmission rates, sharing best practices from clinics with lower‐than‐expected rates.

There are no available studies or proposed methodologies to guide primary care clinics in calculating their 30‐day readmission rates. A particularly difficult challenge is obtaining the admission information when patients may be admitted to 1 of several area hospitals. For large integrated delivery networks where primary care patients are relatively loyal to the network of physicians and hospitals, an opportunity exists to explore the data. Furthermore, variations in readmission rates across primary care specialties (such as internal medicine and family practice) are not well understood. In this study, we set out to develop a methodology for calculating all‐cause 30‐day hospital readmission rates at the level of individual primary care practices and to identify factors associated with variations in these rates.

METHODS

RESULTS

During the study period, there were 12,564 discharges from UCSF Medical Center for primary care patients belonging to the 7 clinics. Of these, 8685 were index discharges and 1032 were readmissions within 30 days. Table 1 shows the characteristics of the patients who had at least 1 admission during the study period. In all but 2 clinics (HIV Primary Care and Concierge Internal Medicine), there were more women hospitalized than men. Age and gender differences across clinics are consistent with the patient populations served by these clinics, with Women's Health having more female patients and younger patients hospitalized compared to Geriatrics.

All‐cause 30‐day readmission rates varied across practices, with HIV Primary Care being the highest at 14.9%, followed by Geriatrics at 13.7%, General Internal Medicine at 13.3%, Concierge Internal Medicine at 9.7%, combined IM/FP at 8.9%, Family Practice at 8.6%, and Women's Health at 7.7% (Figure 1). Despite HIV Primary Care having the highest readmission rate, the number of index discharges during the 3‐year period was relatively low (249) compared to General Internal Medicine (5388). For the index admission, the top 5 admitting services were medicine, obstetrics, cardiology, orthopedic surgery, and general surgery (Table 2). Medicine was the primary admitting service for patients in the following clinics: General Internal Medicine, Geriatrics, HIV Primary Care, and Concierge Internal Medicine. Obstetrics was the primary admitting service for patients in the Family Practice, Women's Health, and combined IM/FP clinics. Vaginal delivery was the top discharge diagnosis‐related group (DRG) for patients in the General Internal Medicine, Family Practice, Women's Health, and combined IM/FP clinics. The top discharge DRG was urinary tract infection for Geriatrics, HIV for the HIV clinic (13.6%), and chemotherapy (8.0%) for Concierge Internal Medicine. Average LOS varied from 4.7 days for patients in the HIV Primary Care clinic to 2.8 days for patients in the Concierge Internal Medicine clinic. Average LOS was 3.4, 3.1, and 3.2 days in the Family Practice, Women's Health, and Geriatrics clinics, respectively, and 3.8 days in the General Internal Medicine and combined IM/FP clinics. For all clinics except Geriatrics, the majority of patients were discharged home without home health. A larger proportion of patients in the Geriatrics clinic were discharged home with home health or discharged to a skilled nursing facility.

Figure 1

Factors associated with variation in readmission rates included: male gender (OR: 1.21, 95% CI: 1.051.40), Medicare (OR: 1.31, 95% CI: 1.051.64; Ref=private) and dual‐eligible Medicare‐Medicaid insurance (OR: 1.26, 95% CI: 1.011.56), unknown primary language (OR: 0.06, 95% CI: 0.020.25; Ref=English), and the following comorbidities: peripheral vascular disease, renal failure, lymphoma, fluid and electrolyte disorders, and anemia. Multivariable logistic regression modeling results are listed in Table 3. Patients having a resident PCP showed no increased odds of readmission (OR: 1.13, 95% CI: 0.931.37; Ref=attending PCP). However, patients with a graduated resident PCP or departed faculty PCP awaiting transfer to a new PCP had an OR of 1.59 (95% CI: 1.162.17) compared with having a current faculty PCP. The C‐statistic for this model was 0.67.

DISCUSSION

In this study we introduce a complementary way to view hospital readmissions, from the perspective of primary care practices. We found variation in readmission rates across primary care practices. After controlling for admitting service, clinic, provider, and patient factors, the specific characteristics of male gender, patients with Medicare or Medicare with Medicaid, and patients in our General Internal Medicine clinic with a departed PCP were independent risk factors for hospital readmission. Patients with specific comorbidities were also at increased risk for admissions, including those with peripheral vascular disease, renal failure, lymphoma, fluid and electrolyte disorders, and anemia. Because this is the first study viewing readmissions from the perspective of primary care practices, our findings are unique in the literature. However, hospital‐based studies have found similar relationships between readmission rates and these specific comorbidities.[11, 12, 15] Unlike other studies, our study cohort did not show CHF as an independent risk factor. We hypothesize this is because in 2008, UCSF Medical Center introduced an inpatient‐based intervention to reduce readmissions in patients with heart failure. By 2011, the readmission rates of patients with heart failure (primary or secondary diagnoses) had dropped by 30%. The success of this program, focused only on patients with heart failure, likely affected our analysis of comorbidities as risk factors for readmissions.

Models developed to predict hospital readmissions, overall and for specific disease conditions, have inconsistently identified predictive factors, and there is not a specific set of variables that dominate.[16, 17, 18, 19] A recent review of readmission risk prediction models suggested that models that take into account psychosocial factors such as social support, substance abuse, and functional status improved model performance.[16] We hypothesize that the reason why male gender was significant in our model may be related to lack of social support, especially among those who may be single or widowed. Other studies have also showed male gender as a predictor for hospital readmissions.[14, 20]

People with Medicare as the primary payor or Medicare with Medicaid also tended to have higher risk of readmission. We believe that this may be a proxy for the combined effect of age, multiple comorbidities, and psychosocial factors. In a multicenter study of general medicine inpatients, Medicare, but not age, was also found to be a predictive variable.[21] Unlike other studies, our study did not find Medicaid alone as a significant predictive variable for readmissions.[15, 21] One explanation may be that in hospital‐based studies, people with Medicaid who are at high risk for readmission may be high‐risk because they do not have a PCP or good access to outpatient care. In our study, all patients have a PCP in 1 of the UCSF clinics, and access to care is improved with this established relationship. These other studies did not examine Medicare‐Medicaid dual eligibility status. Our results are consistent with a national study on avoidable hospital admissions that showed the dual eligibility population experiencing 60% higher avoidable admission rates compared to the Medicaid‐only population.[22]

An interesting finding was that there were no statistically significant differences in readmissions among patients who report a language other than English as their primary language. Although language barriers and health literacy can affect a patient's ability to understand discharge instructions, the use of translators at UCSF Medical Center may have decreased the risk of readmissions. For a small number of patients whose primary language was not recorded (unknown language in our model), they appeared to have a lower risk of readmissions. Language preferences are recorded by our admitting staff during the process of admission. This step may have been skipped after hours or if the patient was not able to answer the question. However, we do not have a good hypothesis as to why these patients may have a reduced risk of readmissions.

Having a departed PCP in the General Internal Medicine clinic was an independent predictor of readmissions. These patients have access to primary care; they can schedule acute appointments with covering providers for new medical issues or follow‐up of chronic conditions. However, until they are transferred to a new PCP, they do not have a provider who is proactively managing their preventive and chronic disease care, including follow‐up and coordinating care after hospital discharge. Our study is not the first to suggest adverse outcomes for patients who are in transition from 1 primary care provider to the next.[23] Studies conducted in General Internal Medicine clinics have shown missed opportunities for cancer screening and overlooked test results during the transition period,[24] and as many as one‐fifth of patients whom residents identified as high‐risk were lost to follow‐up.[25] However, our study is the first to show the link between PCP transition in a teaching clinic and hospital readmissions. This finding underscores the importance of continuity of care in the optimal management of patients following hospital discharge.

There are several limitations to this study. First, our study only considers hospitalizations to UCSF Medical Center, potentially undercounting readmissions to area hospitals. Because our study population are patients with PCPs at UCSF, these patients tend to seek specialty and acute care at UCSF Medical Center as well. We obtained payor data from our medical group (Hill Physicians), which covers our largest private payors, to understand whether our results can be applied on a global basis. We found that in 87.4% of index admissions with readmissions from January 1, 2010 through May 30, 2012, the readmission occurred at UCSF Medical Center. We conducted a similar analysis with CMS data from October 1, 2008 to June 30, 2011, the latest data we have from CMS. For patients with UCSF PCPs and a diagnosis of AMI or CHF, 100% were readmitted back to UCSF Medical Center. For patients with UCSF PCPs and a diagnosis of pneumonia, 89% were readmitted back to UCSF Medical Center. Given that only a small percentage of patients with PCPs at UCSF present for readmission at other area hospitals, we believe that limiting our analysis to UCSF Medical Center is reasonable.

In our study, we did not remove vaginal deliveries or Caesarian sections prior to building the model. Primary care physicians and their clinic leadership are accustomed to taking a population‐health perspective. We anticipate they would be interested in designing interventions and addressing readmission for the entire primary care panel. Although readmissions after delivery are not frequent, they can still occur, and interventions should not necessarily exclude this population. We did run a sensitivity analysis by removing vaginal delivery and Caesarian sections from the analysis. As expected, readmission rates for all practices increased except for Geriatrics. However, the independent predictors of readmissions did not change.

Our study is based on a population of patients at 1 urban academic medical center and may not be generalizable across all delivery systems. Our population is racially and ethnically diverse, and many do not speak English as their primary language. The study also spans different types of primary care clinics, capturing a wide range of ages and case mix. As PCP assignments fluctuate over time, there may be errors with PCP attribution in the UCSF Medical Center data systems. We do not believe errors in PCP attribution would differ across primary care practices. Because the primary care practices' performance reports on quality measures are based on PCP assignment, each clinic regularly updates their clinic panels and has specific protocols to address errors with PCP attribution.

Finally, our study includes only the variables that we were able to extract from administrative claims. Other explanatory variables that have been suggested as important for evaluation, such as social support, functional assessment, access to care, hospital discharge process, and posthospitalization follow‐up, were not included. Each of these could be explored in future studies.

This study offers a unique perspective of hospital readmission by introducing a new methodology for primary care clinics to calculate and evaluate their all‐cause 30‐day readmission rates. Although this methodology is not intended to provide real‐time feedback to clinics on readmitted patients, it opens the door for benchmarking based on specific case‐mix indices. Another direction for future research is to design robust evaluations of the impact of interventions, both inpatient and outpatient, on primary care clinic readmission rates. Finally, future research should replicate this analysis across teaching clinics to identify whether provider turnover is a consistent independent predictor of hospital readmissions.

This study also has implications for inpatient interventions. For example, discharging physicians may consider proactively identifying whether the patient has a continuous primary care provider. Patients who are in between PCPs may need closer follow‐up after discharge, until they re‐establish with a new PCP. This can be accomplished through a postdischarge clinic visit, either run by inpatient providers or covering physicians in the primary care clinic. In addition, the discharging physician may work with the case manager to increase the level of care coordination. The case manager could contact the primary care clinic and proactively ask for immediate PCP re‐assignment. Once a new PCP has been identified, the discharging physician could consider a warm hand‐off. With a warm hand‐off, the new PCP may feel more comfortable managing problems that may arise after hospital discharge, and especially before the first outpatient visit with the new PCP. Future research can test whether these interventions could effectively reduce hospital readmissions across a broad primary care population.

CONCLUSION

Primary care providers and their clinics play an important role in managing population health, decreasing healthcare spending, and keeping patients out of the hospital. In this study, we introduce a tool in which primary care clinics can begin to understand their hospital readmission rates. This may be particularly valuable as primary care providers enter global payment arrangements such as accountable care organizations or bundled payments and are responsible for a population of patients across the continuum of care. We found significant variation in readmission rates between different primary care practices, but much of this variation appears to be due to differences between practices in patient demographics, comorbidities, and hospitalization factors. Our study is the first to show the association between provider transitions and higher hospital readmissions. Continuity of care is critical for the optimal management of patients following hospital discharge. More attention will need to be focused on providing good continuity outpatient care for patients at high risk for readmissions.

Reducing hospital readmissions is a national healthcare priority. In October 2012, the Centers for Medicare and Medicaid Services (CMS) enacted financial penalties on hospitals with higher than average risk‐adjusted readmissions, offering an incentive pool of $850 million in the first year.[1] As a result, a wide range of activities to understand and reduce readmissions among patients with congestive heart failure (CHF), acute myocardial infarction (AMI), and pneumonia have emerged.[2, 3, 4, 5] Some have proposed that to effectively reduce hospital readmissions, a community of inpatient and outpatient providers and local support organizations must coordinate efforts.[6, 7] In fact, CMS has recognized the important role of community support groups and outpatient providers in safe discharge transitions in 2 ways. First, in 2011 CMS launched the Community‐based Care Transitions Program to fund community‐based organizations to assist Medicare patients with care transitions.[8] Second, CMS introduced 2 new reimbursement codes for primary care providers (PCPs) to perform care coordination immediately after hospital discharge.[9] Both of these new payment programs represent an evolving perspective that reducing hospital readmissions requires active participation among outpatient partners.

As leaders of the outpatient care team, PCPs play a significant role in reducing hospital readmissions. One way in which PCPs can begin to understand the magnitude of the issue within their clinic is to evaluate the clinic's 30‐day readmission rates. Currently, CMS calculates readmission rates at the hospital level. However, understanding these rates at the clinic level is critical for developing strategies for improvement across the care continuum. Our current understanding of effective outpatient interventions to reduce hospital readmission is limited.[3] As clinics introduce and refine strategies to reduce readmissions, tracking the impact of these strategies on readmission rates will be critical for identifying effective outpatient interventions. Clinics with similar patient case‐mix can also benchmark readmission rates, sharing best practices from clinics with lower‐than‐expected rates.

There are no available studies or proposed methodologies to guide primary care clinics in calculating their 30‐day readmission rates. A particularly difficult challenge is obtaining the admission information when patients may be admitted to 1 of several area hospitals. For large integrated delivery networks where primary care patients are relatively loyal to the network of physicians and hospitals, an opportunity exists to explore the data. Furthermore, variations in readmission rates across primary care specialties (such as internal medicine and family practice) are not well understood. In this study, we set out to develop a methodology for calculating all‐cause 30‐day hospital readmission rates at the level of individual primary care practices and to identify factors associated with variations in these rates.

METHODS

RESULTS

During the study period, there were 12,564 discharges from UCSF Medical Center for primary care patients belonging to the 7 clinics. Of these, 8685 were index discharges and 1032 were readmissions within 30 days. Table 1 shows the characteristics of the patients who had at least 1 admission during the study period. In all but 2 clinics (HIV Primary Care and Concierge Internal Medicine), there were more women hospitalized than men. Age and gender differences across clinics are consistent with the patient populations served by these clinics, with Women's Health having more female patients and younger patients hospitalized compared to Geriatrics.

All‐cause 30‐day readmission rates varied across practices, with HIV Primary Care being the highest at 14.9%, followed by Geriatrics at 13.7%, General Internal Medicine at 13.3%, Concierge Internal Medicine at 9.7%, combined IM/FP at 8.9%, Family Practice at 8.6%, and Women's Health at 7.7% (Figure 1). Despite HIV Primary Care having the highest readmission rate, the number of index discharges during the 3‐year period was relatively low (249) compared to General Internal Medicine (5388). For the index admission, the top 5 admitting services were medicine, obstetrics, cardiology, orthopedic surgery, and general surgery (Table 2). Medicine was the primary admitting service for patients in the following clinics: General Internal Medicine, Geriatrics, HIV Primary Care, and Concierge Internal Medicine. Obstetrics was the primary admitting service for patients in the Family Practice, Women's Health, and combined IM/FP clinics. Vaginal delivery was the top discharge diagnosis‐related group (DRG) for patients in the General Internal Medicine, Family Practice, Women's Health, and combined IM/FP clinics. The top discharge DRG was urinary tract infection for Geriatrics, HIV for the HIV clinic (13.6%), and chemotherapy (8.0%) for Concierge Internal Medicine. Average LOS varied from 4.7 days for patients in the HIV Primary Care clinic to 2.8 days for patients in the Concierge Internal Medicine clinic. Average LOS was 3.4, 3.1, and 3.2 days in the Family Practice, Women's Health, and Geriatrics clinics, respectively, and 3.8 days in the General Internal Medicine and combined IM/FP clinics. For all clinics except Geriatrics, the majority of patients were discharged home without home health. A larger proportion of patients in the Geriatrics clinic were discharged home with home health or discharged to a skilled nursing facility.

Figure 1

Factors associated with variation in readmission rates included: male gender (OR: 1.21, 95% CI: 1.051.40), Medicare (OR: 1.31, 95% CI: 1.051.64; Ref=private) and dual‐eligible Medicare‐Medicaid insurance (OR: 1.26, 95% CI: 1.011.56), unknown primary language (OR: 0.06, 95% CI: 0.020.25; Ref=English), and the following comorbidities: peripheral vascular disease, renal failure, lymphoma, fluid and electrolyte disorders, and anemia. Multivariable logistic regression modeling results are listed in Table 3. Patients having a resident PCP showed no increased odds of readmission (OR: 1.13, 95% CI: 0.931.37; Ref=attending PCP). However, patients with a graduated resident PCP or departed faculty PCP awaiting transfer to a new PCP had an OR of 1.59 (95% CI: 1.162.17) compared with having a current faculty PCP. The C‐statistic for this model was 0.67.

DISCUSSION

In this study we introduce a complementary way to view hospital readmissions, from the perspective of primary care practices. We found variation in readmission rates across primary care practices. After controlling for admitting service, clinic, provider, and patient factors, the specific characteristics of male gender, patients with Medicare or Medicare with Medicaid, and patients in our General Internal Medicine clinic with a departed PCP were independent risk factors for hospital readmission. Patients with specific comorbidities were also at increased risk for admissions, including those with peripheral vascular disease, renal failure, lymphoma, fluid and electrolyte disorders, and anemia. Because this is the first study viewing readmissions from the perspective of primary care practices, our findings are unique in the literature. However, hospital‐based studies have found similar relationships between readmission rates and these specific comorbidities.[11, 12, 15] Unlike other studies, our study cohort did not show CHF as an independent risk factor. We hypothesize this is because in 2008, UCSF Medical Center introduced an inpatient‐based intervention to reduce readmissions in patients with heart failure. By 2011, the readmission rates of patients with heart failure (primary or secondary diagnoses) had dropped by 30%. The success of this program, focused only on patients with heart failure, likely affected our analysis of comorbidities as risk factors for readmissions.

Models developed to predict hospital readmissions, overall and for specific disease conditions, have inconsistently identified predictive factors, and there is not a specific set of variables that dominate.[16, 17, 18, 19] A recent review of readmission risk prediction models suggested that models that take into account psychosocial factors such as social support, substance abuse, and functional status improved model performance.[16] We hypothesize that the reason why male gender was significant in our model may be related to lack of social support, especially among those who may be single or widowed. Other studies have also showed male gender as a predictor for hospital readmissions.[14, 20]

People with Medicare as the primary payor or Medicare with Medicaid also tended to have higher risk of readmission. We believe that this may be a proxy for the combined effect of age, multiple comorbidities, and psychosocial factors. In a multicenter study of general medicine inpatients, Medicare, but not age, was also found to be a predictive variable.[21] Unlike other studies, our study did not find Medicaid alone as a significant predictive variable for readmissions.[15, 21] One explanation may be that in hospital‐based studies, people with Medicaid who are at high risk for readmission may be high‐risk because they do not have a PCP or good access to outpatient care. In our study, all patients have a PCP in 1 of the UCSF clinics, and access to care is improved with this established relationship. These other studies did not examine Medicare‐Medicaid dual eligibility status. Our results are consistent with a national study on avoidable hospital admissions that showed the dual eligibility population experiencing 60% higher avoidable admission rates compared to the Medicaid‐only population.[22]

An interesting finding was that there were no statistically significant differences in readmissions among patients who report a language other than English as their primary language. Although language barriers and health literacy can affect a patient's ability to understand discharge instructions, the use of translators at UCSF Medical Center may have decreased the risk of readmissions. For a small number of patients whose primary language was not recorded (unknown language in our model), they appeared to have a lower risk of readmissions. Language preferences are recorded by our admitting staff during the process of admission. This step may have been skipped after hours or if the patient was not able to answer the question. However, we do not have a good hypothesis as to why these patients may have a reduced risk of readmissions.

Having a departed PCP in the General Internal Medicine clinic was an independent predictor of readmissions. These patients have access to primary care; they can schedule acute appointments with covering providers for new medical issues or follow‐up of chronic conditions. However, until they are transferred to a new PCP, they do not have a provider who is proactively managing their preventive and chronic disease care, including follow‐up and coordinating care after hospital discharge. Our study is not the first to suggest adverse outcomes for patients who are in transition from 1 primary care provider to the next.[23] Studies conducted in General Internal Medicine clinics have shown missed opportunities for cancer screening and overlooked test results during the transition period,[24] and as many as one‐fifth of patients whom residents identified as high‐risk were lost to follow‐up.[25] However, our study is the first to show the link between PCP transition in a teaching clinic and hospital readmissions. This finding underscores the importance of continuity of care in the optimal management of patients following hospital discharge.

There are several limitations to this study. First, our study only considers hospitalizations to UCSF Medical Center, potentially undercounting readmissions to area hospitals. Because our study population are patients with PCPs at UCSF, these patients tend to seek specialty and acute care at UCSF Medical Center as well. We obtained payor data from our medical group (Hill Physicians), which covers our largest private payors, to understand whether our results can be applied on a global basis. We found that in 87.4% of index admissions with readmissions from January 1, 2010 through May 30, 2012, the readmission occurred at UCSF Medical Center. We conducted a similar analysis with CMS data from October 1, 2008 to June 30, 2011, the latest data we have from CMS. For patients with UCSF PCPs and a diagnosis of AMI or CHF, 100% were readmitted back to UCSF Medical Center. For patients with UCSF PCPs and a diagnosis of pneumonia, 89% were readmitted back to UCSF Medical Center. Given that only a small percentage of patients with PCPs at UCSF present for readmission at other area hospitals, we believe that limiting our analysis to UCSF Medical Center is reasonable.

In our study, we did not remove vaginal deliveries or Caesarian sections prior to building the model. Primary care physicians and their clinic leadership are accustomed to taking a population‐health perspective. We anticipate they would be interested in designing interventions and addressing readmission for the entire primary care panel. Although readmissions after delivery are not frequent, they can still occur, and interventions should not necessarily exclude this population. We did run a sensitivity analysis by removing vaginal delivery and Caesarian sections from the analysis. As expected, readmission rates for all practices increased except for Geriatrics. However, the independent predictors of readmissions did not change.

Our study is based on a population of patients at 1 urban academic medical center and may not be generalizable across all delivery systems. Our population is racially and ethnically diverse, and many do not speak English as their primary language. The study also spans different types of primary care clinics, capturing a wide range of ages and case mix. As PCP assignments fluctuate over time, there may be errors with PCP attribution in the UCSF Medical Center data systems. We do not believe errors in PCP attribution would differ across primary care practices. Because the primary care practices' performance reports on quality measures are based on PCP assignment, each clinic regularly updates their clinic panels and has specific protocols to address errors with PCP attribution.

Finally, our study includes only the variables that we were able to extract from administrative claims. Other explanatory variables that have been suggested as important for evaluation, such as social support, functional assessment, access to care, hospital discharge process, and posthospitalization follow‐up, were not included. Each of these could be explored in future studies.

This study offers a unique perspective of hospital readmission by introducing a new methodology for primary care clinics to calculate and evaluate their all‐cause 30‐day readmission rates. Although this methodology is not intended to provide real‐time feedback to clinics on readmitted patients, it opens the door for benchmarking based on specific case‐mix indices. Another direction for future research is to design robust evaluations of the impact of interventions, both inpatient and outpatient, on primary care clinic readmission rates. Finally, future research should replicate this analysis across teaching clinics to identify whether provider turnover is a consistent independent predictor of hospital readmissions.

This study also has implications for inpatient interventions. For example, discharging physicians may consider proactively identifying whether the patient has a continuous primary care provider. Patients who are in between PCPs may need closer follow‐up after discharge, until they re‐establish with a new PCP. This can be accomplished through a postdischarge clinic visit, either run by inpatient providers or covering physicians in the primary care clinic. In addition, the discharging physician may work with the case manager to increase the level of care coordination. The case manager could contact the primary care clinic and proactively ask for immediate PCP re‐assignment. Once a new PCP has been identified, the discharging physician could consider a warm hand‐off. With a warm hand‐off, the new PCP may feel more comfortable managing problems that may arise after hospital discharge, and especially before the first outpatient visit with the new PCP. Future research can test whether these interventions could effectively reduce hospital readmissions across a broad primary care population.

CONCLUSION

Primary care providers and their clinics play an important role in managing population health, decreasing healthcare spending, and keeping patients out of the hospital. In this study, we introduce a tool in which primary care clinics can begin to understand their hospital readmission rates. This may be particularly valuable as primary care providers enter global payment arrangements such as accountable care organizations or bundled payments and are responsible for a population of patients across the continuum of care. We found significant variation in readmission rates between different primary care practices, but much of this variation appears to be due to differences between practices in patient demographics, comorbidities, and hospitalization factors. Our study is the first to show the association between provider transitions and higher hospital readmissions. Continuity of care is critical for the optimal management of patients following hospital discharge. More attention will need to be focused on providing good continuity outpatient care for patients at high risk for readmissions.

Despite signs of a slowing trend,[1, 2, 3] US healthcare costs continue to rise, and cost containment remains a major area of concern. Hospital costs are the largest single category of national healthcare expenditures,[4] and the burden of cost containment is increasingly being shifted to hospitals.[5] As such, hospitals are increasingly focusing on implementing interventions to reduce rates of hospitalizations and readmissions as a mechanism to reduce overall healthcare costs.[5, 6, 7, 8, 9]

Multiple factors potentially contribute to patients being high cost, including acute care utilization,[10, 11, 12] pharmaceuticals,[13, 14] procedures,[15] catastrophic illness,[16] and high‐risk chronic conditions.[11, 17, 18] However, many hospitals are implementing interventions focused on a single subset of these high cost patientshigh users of inpatient services. Care management interventions have received particular attention, due to their perceived potential to improve quality of care while reducing costs through the mechanism of reducing hospital admissions.[19, 20, 21] Despite their increasing prevalence, there is limited evidence demonstrating the effectiveness of these programs. Among interventions targeting high cost individuals, the Medicare Care Management for High Cost Beneficiaries showed no effect on hospital admissions.[19] Another high‐profile intervention, the Citywide Care Management System led by the Camden Coalition, showed promising preliminary results, but data from a systematic evaluation are lacking.[22] Conversely, interventions targeting individuals with frequent hospitalizations have similarly shown mixed results in reducing costs.[6, 7, 9, 23]

Taken together, these data suggest that the relationship between high costs and frequent hospital use is complicated, and the population of individuals with frequent hospitalizations may not represent the entire population of high cost individuals. Thus, focusing on reduction of hospitalizations alone may be inadequate to reduce costs. For these reasons, we sought to determine how many high cost individuals would be captured by focusing only on frequently hospitalized (high admit) individuals by examining the overlap between these populations. We also sought to describe the characteristics and distinctions between the resulting subgroups of high users to better inform the design of future interventions.

METHODS

We examined the cross‐sectional relationship between high cost and high admit populations among adult patients 18 years of age hospitalized at the University of California, San Francisco (UCSF) Medical Center, a 660‐bed urban academic medical center from July 1, 2010 to June 30, 2011.

This study was conducted as part of a quality improvement project to identify high user primary care patients for complex care management. Individuals were included in the study if: (1) they had an assigned UCSF primary care provider (PCP), and (2) they had at least 1 hospitalization at UCSF during the study period. PCP assignments were ascertained from panel lists maintained by clinics; lists included individuals with at least 1 visit at any of the 8 primary care clinics at UCSF in the 2 years prior to the end of the study period. Because individuals are dropped from PCP panels at death, we were unable to ascertain or include individuals who died during the study period.

From the initial study population, we defined the high cost group as those who were in the top decile of total hospital costs, and the high admit group as those who were in the top decile of total hospitalizations during the study period. We elected to use the top decile as a cutoff given that it is a common operational definition used to identify high users to target for intervention.[24]

To examine the relationship between high cost and high admits we defined 3 further subgroups: high costlow admits, high costhigh admits, and low costhigh admits (Figure 1). To explore the face validity of these descriptors and classification scheme, we subsequently examined the proportion of total hospital costs and total hospitalizations each subgroup accounted for with respect to the study population.

Figure 1

RESULTS

Of the 2566 unique individuals included in the analysis (Figure 1), 256 individuals were identified as high cost (top decile, $65,000). This group accounted for 45% of all costs and 22% of all hospitalizations (Figure 2). Two hundred fifty individuals were identified as high admits (top decile, 3 hospitalizations). This group accounted for 32% of all costs and 28% of all hospitalizations.

Figure 2

Only 48% of high cost individuals were also high admit ($65,000 and 3 hospitalizations; n=125, Figure 1). Among high users, we subsequently defined 3 subgroups based on the relationship between cost and hospitalizations (high costlow admits, high costhigh admits, and low costhigh admits). Each subgroup accounted for approximately 5% of the overall study population (Figure 2). The high costlow admits subgroup incurred a proportionate share of hospitalizations (6%) but a disproportionate share of costs (20%). The high costhigh admits subgroup had a disproportionate share of both costs (25%) and hospitalizations (16%). The low costhigh admits subgroup had a proportionate share of costs (7%) but a disproportionate share of hospitalizations (12%).

DISCUSSION

In this study, we found that only half of high cost individuals were also high admit. Further categorizing high users into high costlow admit versus high costhigh admit versus low costhigh admit identified distinct patterns between each group. High costhigh admit individuals were more likely to be hospitalized for medical conditions, whereas high costlow admit individuals were more likely to be hospitalized for surgical conditions. CMS core conditions accounted for a low proportion of overall hospitalizations across all groups.

Our findings suggest several distinct types of high users with different clinical characteristics, utilization, and cost patterns. From a hospital perspective, one implication is that a multifaceted approach to cost containment, rather than the one‐size‐fits‐all strategy of reducing hospitalizations, may be more effective in reducing costs. For example, our findings show that high costlow admit individuals have a disproportionate number of hospitalizations for surgical conditions, longer LOS, and more ICU days. Costs incurred by this subgroup may be more responsive to in‐hospital interventions aimed at reducing procedural costs, LOS, unnecessary use of the ICU, and minimizing postoperative infections and complications rather than to a care management approach.

In contrast, care management strategies such as improving postdischarge care and chronic disease management, which aim to achieve cost savings through reducing hospitalizations, may be more effective in reducing costs among high costhigh admit individuals, who have a high proportion of hospitalizations for medical conditions and the highest proportion of 30‐day readmissions. Such strategies may also be important in optimizing the quality of care for low costhigh admit individuals, who have the highest proportion of medical hospitalizations among all high users, though the potential for cost savings may be more limited in this subgroup.

Our results suggest that current hospital‐based approachesdriven by readmissions penalties for CMS core conditionsmay have less than the expected impact on costs. For example, although high costhigh admit individuals had the highest proportion of 30‐day readmissions, readmissions specifically for CMS core conditions accounted for <1% of costs in this subgroup. Thus, the potential return on an expensive investment in a care management intervention is unclear, given the small number of readmissions for these select conditions. From a broader perspective, the focus on readmissions for CMS core conditions, which overall contribute relatively little to high hospital costs, may not be a comprehensive enough strategy for cost containment. To date, there have been limited policies targeting factors contributing to high hospital costs outside of frequent medical hospitalizations. Medicare's nonpayment policy for treatment of preventable hospital conditions, including surgical site infections, translates prevention of these conditions into cost savings for hospitals.[30] However, this rule has been criticized for not going far enough to drive substantial savings.[31] A new CMS rule authorizes states to identify other provider‐preventable conditions for which Medicaid payment will be prohibited.[32] Future policy efforts should further emphasize a comprehensive, multipronged approach beyond readmissions penalties for select conditions if healthcare cost containment remains a policy priority.

Our results should be interpreted in light of several limitations. First, this was a single‐site study at an academic medical center; the generalizability of our results to other settings is unclear. Our cost data likely reflect local market factors, including the highest wage rates for skilled healthcare labor in the United States.[33] Although the explicit distribution of high user subgroups may be institution‐specific due to variations in our cost structure, we anticipate that the general classification paradigm will be similar in other health systems. Second, we captured utilization and costs only at a single hospital. However, our study population includes only individuals with PCPs at UCSF, and internal data from both Medicare and UCSF's largest private payer show that over 85% of hospitalizations among this population are to UCSF Medical Center. Third, we were able to capture only hospital costs rather than overall healthcare costs. Given that hospital costs account for the single largest category of total national health costs,[4] we expect that future studies examining total health costs will show similar findings. Fourth, our data did not include measures of health status, socioeconomic status, housing, or mental health comorbidities to permit an analysis of these factors, which have been previously related to frequent hospitalizations and high costs.[34, 35, 36, 37, 38, 39] Fifth, due to resource constraints, we were unable to conduct a longitudinal analysis to examine the extent to which individuals are consistently high users over time. Previous studies have described that 20% to 30% of individuals are consistently high users; the remainder have discrete periods of high utilization.[34, 40] This may be an important consideration in the design of future interventions.

Finally, our analysis was limited to individuals with a PCP to allow identification of an accessible cohort for care management. Thus, we did not capture individuals without a PCP and individuals who died during the study period, because these individuals no longer had an assigned PCP following death. Although this approach is consistent with that of many care management programs,[19] these populations are likely to incur higher than average utilization and healthcare costs, and represent important areas for future investigation.

In summary, our study identifies three types of high‐user populations that differ in the proportion of costs attributable to frequent hospitalizations, clinical conditions associated with hospital use, and frequency of 30‐day readmissions. Stratifying high users by both costs and hospitalizations may help identify tailored strategies to more effectively reduce costs and utilization.

Despite signs of a slowing trend,[1, 2, 3] US healthcare costs continue to rise, and cost containment remains a major area of concern. Hospital costs are the largest single category of national healthcare expenditures,[4] and the burden of cost containment is increasingly being shifted to hospitals.[5] As such, hospitals are increasingly focusing on implementing interventions to reduce rates of hospitalizations and readmissions as a mechanism to reduce overall healthcare costs.[5, 6, 7, 8, 9]

Multiple factors potentially contribute to patients being high cost, including acute care utilization,[10, 11, 12] pharmaceuticals,[13, 14] procedures,[15] catastrophic illness,[16] and high‐risk chronic conditions.[11, 17, 18] However, many hospitals are implementing interventions focused on a single subset of these high cost patientshigh users of inpatient services. Care management interventions have received particular attention, due to their perceived potential to improve quality of care while reducing costs through the mechanism of reducing hospital admissions.[19, 20, 21] Despite their increasing prevalence, there is limited evidence demonstrating the effectiveness of these programs. Among interventions targeting high cost individuals, the Medicare Care Management for High Cost Beneficiaries showed no effect on hospital admissions.[19] Another high‐profile intervention, the Citywide Care Management System led by the Camden Coalition, showed promising preliminary results, but data from a systematic evaluation are lacking.[22] Conversely, interventions targeting individuals with frequent hospitalizations have similarly shown mixed results in reducing costs.[6, 7, 9, 23]

Taken together, these data suggest that the relationship between high costs and frequent hospital use is complicated, and the population of individuals with frequent hospitalizations may not represent the entire population of high cost individuals. Thus, focusing on reduction of hospitalizations alone may be inadequate to reduce costs. For these reasons, we sought to determine how many high cost individuals would be captured by focusing only on frequently hospitalized (high admit) individuals by examining the overlap between these populations. We also sought to describe the characteristics and distinctions between the resulting subgroups of high users to better inform the design of future interventions.

METHODS

We examined the cross‐sectional relationship between high cost and high admit populations among adult patients 18 years of age hospitalized at the University of California, San Francisco (UCSF) Medical Center, a 660‐bed urban academic medical center from July 1, 2010 to June 30, 2011.

This study was conducted as part of a quality improvement project to identify high user primary care patients for complex care management. Individuals were included in the study if: (1) they had an assigned UCSF primary care provider (PCP), and (2) they had at least 1 hospitalization at UCSF during the study period. PCP assignments were ascertained from panel lists maintained by clinics; lists included individuals with at least 1 visit at any of the 8 primary care clinics at UCSF in the 2 years prior to the end of the study period. Because individuals are dropped from PCP panels at death, we were unable to ascertain or include individuals who died during the study period.

From the initial study population, we defined the high cost group as those who were in the top decile of total hospital costs, and the high admit group as those who were in the top decile of total hospitalizations during the study period. We elected to use the top decile as a cutoff given that it is a common operational definition used to identify high users to target for intervention.[24]

To examine the relationship between high cost and high admits we defined 3 further subgroups: high costlow admits, high costhigh admits, and low costhigh admits (Figure 1). To explore the face validity of these descriptors and classification scheme, we subsequently examined the proportion of total hospital costs and total hospitalizations each subgroup accounted for with respect to the study population.

Figure 1

RESULTS

Of the 2566 unique individuals included in the analysis (Figure 1), 256 individuals were identified as high cost (top decile, $65,000). This group accounted for 45% of all costs and 22% of all hospitalizations (Figure 2). Two hundred fifty individuals were identified as high admits (top decile, 3 hospitalizations). This group accounted for 32% of all costs and 28% of all hospitalizations.

Figure 2

Only 48% of high cost individuals were also high admit ($65,000 and 3 hospitalizations; n=125, Figure 1). Among high users, we subsequently defined 3 subgroups based on the relationship between cost and hospitalizations (high costlow admits, high costhigh admits, and low costhigh admits). Each subgroup accounted for approximately 5% of the overall study population (Figure 2). The high costlow admits subgroup incurred a proportionate share of hospitalizations (6%) but a disproportionate share of costs (20%). The high costhigh admits subgroup had a disproportionate share of both costs (25%) and hospitalizations (16%). The low costhigh admits subgroup had a proportionate share of costs (7%) but a disproportionate share of hospitalizations (12%).

DISCUSSION

In this study, we found that only half of high cost individuals were also high admit. Further categorizing high users into high costlow admit versus high costhigh admit versus low costhigh admit identified distinct patterns between each group. High costhigh admit individuals were more likely to be hospitalized for medical conditions, whereas high costlow admit individuals were more likely to be hospitalized for surgical conditions. CMS core conditions accounted for a low proportion of overall hospitalizations across all groups.

Our findings suggest several distinct types of high users with different clinical characteristics, utilization, and cost patterns. From a hospital perspective, one implication is that a multifaceted approach to cost containment, rather than the one‐size‐fits‐all strategy of reducing hospitalizations, may be more effective in reducing costs. For example, our findings show that high costlow admit individuals have a disproportionate number of hospitalizations for surgical conditions, longer LOS, and more ICU days. Costs incurred by this subgroup may be more responsive to in‐hospital interventions aimed at reducing procedural costs, LOS, unnecessary use of the ICU, and minimizing postoperative infections and complications rather than to a care management approach.

In contrast, care management strategies such as improving postdischarge care and chronic disease management, which aim to achieve cost savings through reducing hospitalizations, may be more effective in reducing costs among high costhigh admit individuals, who have a high proportion of hospitalizations for medical conditions and the highest proportion of 30‐day readmissions. Such strategies may also be important in optimizing the quality of care for low costhigh admit individuals, who have the highest proportion of medical hospitalizations among all high users, though the potential for cost savings may be more limited in this subgroup.

Our results suggest that current hospital‐based approachesdriven by readmissions penalties for CMS core conditionsmay have less than the expected impact on costs. For example, although high costhigh admit individuals had the highest proportion of 30‐day readmissions, readmissions specifically for CMS core conditions accounted for <1% of costs in this subgroup. Thus, the potential return on an expensive investment in a care management intervention is unclear, given the small number of readmissions for these select conditions. From a broader perspective, the focus on readmissions for CMS core conditions, which overall contribute relatively little to high hospital costs, may not be a comprehensive enough strategy for cost containment. To date, there have been limited policies targeting factors contributing to high hospital costs outside of frequent medical hospitalizations. Medicare's nonpayment policy for treatment of preventable hospital conditions, including surgical site infections, translates prevention of these conditions into cost savings for hospitals.[30] However, this rule has been criticized for not going far enough to drive substantial savings.[31] A new CMS rule authorizes states to identify other provider‐preventable conditions for which Medicaid payment will be prohibited.[32] Future policy efforts should further emphasize a comprehensive, multipronged approach beyond readmissions penalties for select conditions if healthcare cost containment remains a policy priority.

Our results should be interpreted in light of several limitations. First, this was a single‐site study at an academic medical center; the generalizability of our results to other settings is unclear. Our cost data likely reflect local market factors, including the highest wage rates for skilled healthcare labor in the United States.[33] Although the explicit distribution of high user subgroups may be institution‐specific due to variations in our cost structure, we anticipate that the general classification paradigm will be similar in other health systems. Second, we captured utilization and costs only at a single hospital. However, our study population includes only individuals with PCPs at UCSF, and internal data from both Medicare and UCSF's largest private payer show that over 85% of hospitalizations among this population are to UCSF Medical Center. Third, we were able to capture only hospital costs rather than overall healthcare costs. Given that hospital costs account for the single largest category of total national health costs,[4] we expect that future studies examining total health costs will show similar findings. Fourth, our data did not include measures of health status, socioeconomic status, housing, or mental health comorbidities to permit an analysis of these factors, which have been previously related to frequent hospitalizations and high costs.[34, 35, 36, 37, 38, 39] Fifth, due to resource constraints, we were unable to conduct a longitudinal analysis to examine the extent to which individuals are consistently high users over time. Previous studies have described that 20% to 30% of individuals are consistently high users; the remainder have discrete periods of high utilization.[34, 40] This may be an important consideration in the design of future interventions.

Finally, our analysis was limited to individuals with a PCP to allow identification of an accessible cohort for care management. Thus, we did not capture individuals without a PCP and individuals who died during the study period, because these individuals no longer had an assigned PCP following death. Although this approach is consistent with that of many care management programs,[19] these populations are likely to incur higher than average utilization and healthcare costs, and represent important areas for future investigation.

In summary, our study identifies three types of high‐user populations that differ in the proportion of costs attributable to frequent hospitalizations, clinical conditions associated with hospital use, and frequency of 30‐day readmissions. Stratifying high users by both costs and hospitalizations may help identify tailored strategies to more effectively reduce costs and utilization.

Despite signs of a slowing trend,[1, 2, 3] US healthcare costs continue to rise, and cost containment remains a major area of concern. Hospital costs are the largest single category of national healthcare expenditures,[4] and the burden of cost containment is increasingly being shifted to hospitals.[5] As such, hospitals are increasingly focusing on implementing interventions to reduce rates of hospitalizations and readmissions as a mechanism to reduce overall healthcare costs.[5, 6, 7, 8, 9]

Multiple factors potentially contribute to patients being high cost, including acute care utilization,[10, 11, 12] pharmaceuticals,[13, 14] procedures,[15] catastrophic illness,[16] and high‐risk chronic conditions.[11, 17, 18] However, many hospitals are implementing interventions focused on a single subset of these high cost patientshigh users of inpatient services. Care management interventions have received particular attention, due to their perceived potential to improve quality of care while reducing costs through the mechanism of reducing hospital admissions.[19, 20, 21] Despite their increasing prevalence, there is limited evidence demonstrating the effectiveness of these programs. Among interventions targeting high cost individuals, the Medicare Care Management for High Cost Beneficiaries showed no effect on hospital admissions.[19] Another high‐profile intervention, the Citywide Care Management System led by the Camden Coalition, showed promising preliminary results, but data from a systematic evaluation are lacking.[22] Conversely, interventions targeting individuals with frequent hospitalizations have similarly shown mixed results in reducing costs.[6, 7, 9, 23]

Taken together, these data suggest that the relationship between high costs and frequent hospital use is complicated, and the population of individuals with frequent hospitalizations may not represent the entire population of high cost individuals. Thus, focusing on reduction of hospitalizations alone may be inadequate to reduce costs. For these reasons, we sought to determine how many high cost individuals would be captured by focusing only on frequently hospitalized (high admit) individuals by examining the overlap between these populations. We also sought to describe the characteristics and distinctions between the resulting subgroups of high users to better inform the design of future interventions.

METHODS

We examined the cross‐sectional relationship between high cost and high admit populations among adult patients 18 years of age hospitalized at the University of California, San Francisco (UCSF) Medical Center, a 660‐bed urban academic medical center from July 1, 2010 to June 30, 2011.

This study was conducted as part of a quality improvement project to identify high user primary care patients for complex care management. Individuals were included in the study if: (1) they had an assigned UCSF primary care provider (PCP), and (2) they had at least 1 hospitalization at UCSF during the study period. PCP assignments were ascertained from panel lists maintained by clinics; lists included individuals with at least 1 visit at any of the 8 primary care clinics at UCSF in the 2 years prior to the end of the study period. Because individuals are dropped from PCP panels at death, we were unable to ascertain or include individuals who died during the study period.

From the initial study population, we defined the high cost group as those who were in the top decile of total hospital costs, and the high admit group as those who were in the top decile of total hospitalizations during the study period. We elected to use the top decile as a cutoff given that it is a common operational definition used to identify high users to target for intervention.[24]

To examine the relationship between high cost and high admits we defined 3 further subgroups: high costlow admits, high costhigh admits, and low costhigh admits (Figure 1). To explore the face validity of these descriptors and classification scheme, we subsequently examined the proportion of total hospital costs and total hospitalizations each subgroup accounted for with respect to the study population.

Figure 1

RESULTS

Of the 2566 unique individuals included in the analysis (Figure 1), 256 individuals were identified as high cost (top decile, $65,000). This group accounted for 45% of all costs and 22% of all hospitalizations (Figure 2). Two hundred fifty individuals were identified as high admits (top decile, 3 hospitalizations). This group accounted for 32% of all costs and 28% of all hospitalizations.

Figure 2

Only 48% of high cost individuals were also high admit ($65,000 and 3 hospitalizations; n=125, Figure 1). Among high users, we subsequently defined 3 subgroups based on the relationship between cost and hospitalizations (high costlow admits, high costhigh admits, and low costhigh admits). Each subgroup accounted for approximately 5% of the overall study population (Figure 2). The high costlow admits subgroup incurred a proportionate share of hospitalizations (6%) but a disproportionate share of costs (20%). The high costhigh admits subgroup had a disproportionate share of both costs (25%) and hospitalizations (16%). The low costhigh admits subgroup had a proportionate share of costs (7%) but a disproportionate share of hospitalizations (12%).

DISCUSSION

In this study, we found that only half of high cost individuals were also high admit. Further categorizing high users into high costlow admit versus high costhigh admit versus low costhigh admit identified distinct patterns between each group. High costhigh admit individuals were more likely to be hospitalized for medical conditions, whereas high costlow admit individuals were more likely to be hospitalized for surgical conditions. CMS core conditions accounted for a low proportion of overall hospitalizations across all groups.

Our findings suggest several distinct types of high users with different clinical characteristics, utilization, and cost patterns. From a hospital perspective, one implication is that a multifaceted approach to cost containment, rather than the one‐size‐fits‐all strategy of reducing hospitalizations, may be more effective in reducing costs. For example, our findings show that high costlow admit individuals have a disproportionate number of hospitalizations for surgical conditions, longer LOS, and more ICU days. Costs incurred by this subgroup may be more responsive to in‐hospital interventions aimed at reducing procedural costs, LOS, unnecessary use of the ICU, and minimizing postoperative infections and complications rather than to a care management approach.

In contrast, care management strategies such as improving postdischarge care and chronic disease management, which aim to achieve cost savings through reducing hospitalizations, may be more effective in reducing costs among high costhigh admit individuals, who have a high proportion of hospitalizations for medical conditions and the highest proportion of 30‐day readmissions. Such strategies may also be important in optimizing the quality of care for low costhigh admit individuals, who have the highest proportion of medical hospitalizations among all high users, though the potential for cost savings may be more limited in this subgroup.

Our results suggest that current hospital‐based approachesdriven by readmissions penalties for CMS core conditionsmay have less than the expected impact on costs. For example, although high costhigh admit individuals had the highest proportion of 30‐day readmissions, readmissions specifically for CMS core conditions accounted for <1% of costs in this subgroup. Thus, the potential return on an expensive investment in a care management intervention is unclear, given the small number of readmissions for these select conditions. From a broader perspective, the focus on readmissions for CMS core conditions, which overall contribute relatively little to high hospital costs, may not be a comprehensive enough strategy for cost containment. To date, there have been limited policies targeting factors contributing to high hospital costs outside of frequent medical hospitalizations. Medicare's nonpayment policy for treatment of preventable hospital conditions, including surgical site infections, translates prevention of these conditions into cost savings for hospitals.[30] However, this rule has been criticized for not going far enough to drive substantial savings.[31] A new CMS rule authorizes states to identify other provider‐preventable conditions for which Medicaid payment will be prohibited.[32] Future policy efforts should further emphasize a comprehensive, multipronged approach beyond readmissions penalties for select conditions if healthcare cost containment remains a policy priority.

Our results should be interpreted in light of several limitations. First, this was a single‐site study at an academic medical center; the generalizability of our results to other settings is unclear. Our cost data likely reflect local market factors, including the highest wage rates for skilled healthcare labor in the United States.[33] Although the explicit distribution of high user subgroups may be institution‐specific due to variations in our cost structure, we anticipate that the general classification paradigm will be similar in other health systems. Second, we captured utilization and costs only at a single hospital. However, our study population includes only individuals with PCPs at UCSF, and internal data from both Medicare and UCSF's largest private payer show that over 85% of hospitalizations among this population are to UCSF Medical Center. Third, we were able to capture only hospital costs rather than overall healthcare costs. Given that hospital costs account for the single largest category of total national health costs,[4] we expect that future studies examining total health costs will show similar findings. Fourth, our data did not include measures of health status, socioeconomic status, housing, or mental health comorbidities to permit an analysis of these factors, which have been previously related to frequent hospitalizations and high costs.[34, 35, 36, 37, 38, 39] Fifth, due to resource constraints, we were unable to conduct a longitudinal analysis to examine the extent to which individuals are consistently high users over time. Previous studies have described that 20% to 30% of individuals are consistently high users; the remainder have discrete periods of high utilization.[34, 40] This may be an important consideration in the design of future interventions.

Finally, our analysis was limited to individuals with a PCP to allow identification of an accessible cohort for care management. Thus, we did not capture individuals without a PCP and individuals who died during the study period, because these individuals no longer had an assigned PCP following death. Although this approach is consistent with that of many care management programs,[19] these populations are likely to incur higher than average utilization and healthcare costs, and represent important areas for future investigation.

In summary, our study identifies three types of high‐user populations that differ in the proportion of costs attributable to frequent hospitalizations, clinical conditions associated with hospital use, and frequency of 30‐day readmissions. Stratifying high users by both costs and hospitalizations may help identify tailored strategies to more effectively reduce costs and utilization.

Across the country, hospitals are rolling out programs to reduce readmissions. These range from patient education around their disease process and medications, improvements in discharge planning, medication reconciliation, outpatient appointments scheduled prior to discharge, and follow‐up phone calls, among others.[1, 2] Several collaboratives such as Project Better Outcomes by Optimizing Safe Transitions[3] and Hospital Medicine Reengineering Network[4] have formed to test, study, and share lessons learned from these inpatient‐based interventions. Because financial penalties thus far have focused on decreases in inpatient reimbursement by the Centers for Medicare and Medicaid Services (CMS), most of the interventions to reduce hospital readmissions have been concentrated in the inpatient domain. It is unclear whether new payment arrangements with CMS or commercial insurers, such as bundled payments and accountable care organizations (ACOs), will pressure primary care physicians (PCPs) to further develop outpatient‐based interventions.

In this article, I provide a PCP's perspective of how inpatient and outpatient providers can partner to create the ideal care transition from hospital to home. Although others have conducted systematic reviews or surveys of interventions to reduce hospital readmissions,[2, 5, 6, 7] I will start from a vision of an ideal transition, and then evaluate evidence supporting each step. I will also highlight areas where new reimbursement codes can help support an idealized transitions‐of‐care program.

ON HOSPITAL ADMISSION

Many PCPs consider the beginning of the care continuum to rest in the primary care practice and relationship. Over time, PCPs develop relationships with their patients, understanding the patients' values toward health and healthcare, learning their social support system and home environment, and documenting a clinical course of chronic diseases, including which therapies have and have not worked well. From a PCP perspective, it seems natural to be involved in care at the point of hospital admission. Some emergency departments (EDs) and hospital admissions offices have automated systems to email or fax PCPs admission notifications. Others rely on providers to make this connection.

Ideally, PCP communication would occur early in the hospitalization, especially for medically or socially complex patients, for 3 main reasons: (1) The PCP can offer insights on goals of care or therapies attempted in the past that may reduce unnecessary procedures, decrease length of stay, and improve patient satisfaction; (2) Reconciling medications that the patient should be taking and the list the patient reports may highlight noncompliance and trigger education around medication compliance prior to discharge; (3) Early PCP involvement may improve discharge planning efficiency, whereby the inpatient and outpatient teams agree on medical and social issues to be addressed for a safe discharge. Although there are no published studies that show communication early in the hospitalization impacts clinical outcomes, a national survey of hospitalists highlighted concerns about poor information exchange, particularly around medical history and outpatient medications, at the time of admission.[8]

As new financial models push healthcare providers to manage a population of patients under a global budget, inpatient and outpatient providers will need to communicate and collaborate at a level that is new for most institutions and providers.[9] Because early markers of success are based on financial savings, this means further reducing length of stay and transferring more care to the outpatient arena, tapping into community and home care resources. Involving PCPs early in the admission may help inpatient providers meet these goals.

DURING HOSPITALIZATION

Several publications from hospitalists have evaluated inpatient‐based interventions to reduce readmissions and improve care transitions.[10, 11, 12] Table 1 summarizes 6 steps that can improve communication and collaboration.

AFTER HOSPITAL DISCHARGE

Immediately after hospital discharge, there are 7 steps that PCPs and their clinic staff can follow to support a safe transition from hospital to home. The literature supports several individual steps, but not the full package. I am proposing that primary care clinics adopt all 7 steps in an ideal transitions‐of‐care program.

IMPLICATIONS

In this vision of the ideal care transition, I am suggesting a shift in culture from a predominantly hospital‐based program to a program that spans the care continuum and requires active participation and ownership from the PCP's team. It will require inpatient and outpatient providers to communicate early and frequently during the hospitalization, sharing patient information efficiently and working collaboratively as part of a larger team to meet the medical and psychosocial needs of the patient. This concept is not new, but has not been supported financially from payors.[1, 9, 40] Most PCPs operate on margins that cannot support additional PCP time to coordinate care for patients or staff to assist (although many PCPs believe this is the role of the primary care medical home).[33] Some payors agree that stipends to support infrastructure change are needed to improve patient outcomes.[17, 38, 39]

Even though every envisioned step does not require additional funding, new payment arrangements under ACOs and bundled payments may offer opportunities for PCPs to assume a larger role in care transitions and secure funding to pay for interventions. However, primary care practices must be positioned to negotiate favorable global payment agreements, be willing to assume risks associated with global payments, and prioritize management of medically and socially complex patients who are at risk for preventable ED visits and hospitalizations. PCPs who are not participating in ACOs or bundled payments, or those who are risk adverse, may be able to finance pieces of this vision with the new care coordination CPT codes supported by Medicare (and possibly commercial payors in the future). They may also partner with community groups participating in CCTP for additional support. Others focus on the long‐term benefits of ACO‐like structures rather than the short‐term investments needed.[41]

Are all 14 steps proposed above essential? Without doubt, this vision will be difficult to fully operationalize and requires coordination and support from many distinct groups. Should all patients be offered a basic package of interventions, reserving the full package for those who are identified as highest risk for poor outcomes after hospital discharge? There is already some support around specialized interventions for patients at high risk for readmissions,[32, 41] and risk prediction models have been introduced to identify these individuals.[42] Or should we approach this as a menu of interventions from which to choose, tailoring interventions to individual patient needs? These questions should be tested, as our experience in coordinating care across the continuum matures. With over 100 ACOs formed in Medicare alone[43] and many more with commercial insurers, our understanding in this area will grow in the next 5 years.

CONCLUSIONS

As cost containment measures in healthcare target preventable readmissions, hospitals and primary care physicians are increasingly encouraged to improve transitions along the care continuum. In this article, I offer 1 PCP's vision of the ideal transitions‐of‐care program from hospital to home. This article focuses on steps that can be taken by PCPs and their clinic staff; it does not address the role of outpatient specialists, home care agencies, or community support groups in care transitions. Operationalizing this vision requires commitment from the hospital and clinic leadership, as well as buy‐in from front‐line providers. More research is required to understand the marginal impact of each component of this vision, as well as the comprehensive package of interventions proposed, on patient outcomes. New financial models with payors and hospitals may make it easier for primary care clinics to test this vision. Current financial incentives are likely still inadequate to fully align care along the continuum, but they offer some support for more PCPs to take an active role. The time has come to shift our traditional view of transitions of care from a hospital‐centric set of interventions toward one that spans the entire care continuum and includes primary care physicians and their clinic staff as key partners.