We used readily available demographic, health system, and clinical comorbidity data to develop and validate a model to predict 30-day hospital readmission for elders admitted to medical or surgical services of acute care hospitals. We found that older age, male sex, African American race, Medicare-only insurance without supplemental health insurance, medical service admission, and discharge to long-term care were independently associated with increased risk of 30-day hospital readmission. Major comorbid conditions similarly predicted 30-day readmission using either the Elixhauser or the HRDES classification. The highest PAR for 30-day hospital admission was due to discharge to long-term care. High-priority conditions for interventions to reduce 30-day readmission can be identified using either the Elixhauser or HRDES classifications. The specific magnitude of risk and rank order of the PAR for individual categories differ between the two classifications, resulting in part from the specific ICD-9-CM codes included in each category and in part from the inclusion of systemic condition categories, such as fluid and electrolyte disorders, which are associated with other disease categories.
The Elixhauser comorbidity index was designed to assess the impact of comorbid conditions on outcomes independent of the diagnoses explaining the hospital admission (DRG) (15
). In a separate analysis of predictors of 30-day readmission, models with Elixhauser categories based on primary and comorbid diagnoses (implemented by removing the DRG exclusion from the Elixhauser algorithm) had similar discrimination as models with the comorbidity and DRG exclusion (receiver operating characteristic curve area, 0.65 vs 0.63).
Predictive models for 30-day readmission based on demographic variables (age, sex, and race), health system variables (health insurance, hospital service, and discharge location), and either 10 Elixhauser major comorbidity categories or 13 HRDES comorbidity categories were valid predictors of early (30-day) hospital readmission. Alternatively, models based on demographic variables, health system variables, and the total count of either the number of Elixhauser comorbidity categories or HRDES comorbidity categories were also valid predictors of 30-day hospital readmission.
For each model the probability of readmission can be easily obtained by summing the baseline 30-day readmission probability and the covariate-specific incremental probabilities (risk differences). For example, using the risk difference estimates for the Elixhauser comorbidity in Table , the probability of 30-day hospital readmission for a 78-year-old African American man with Medicare insurance admitted to a medical service for heart failure and diabetes mellitus without chronic complications would be 0.154, or 15.4%. This is calculated as the sum of the baseline risk (0.058) and the age category risk (0.027), male sex (0.008), African American race (0.010), Medicare insurance referent category risk (0), medical service referent category risk (0), Elixhauser comorbidity category risk for congestive heart failure (0.034), and diabetes mellitus without chronic complications (0.017).
The predictive models for 30-day hospital readmission may be most useful in two areas. First, the probability of readmission can be used early during the patient's hospital admission to estimate the risk of hospital readmission and identify elders who may benefit from more coordinated care management, intensive assessment, and additional services after hospital discharge. An elder's demographic characteristics (age, sex, race, health insurance, and anticipated discharge location) and major diagnoses are likely to be known in the first day or two after admission, and elders at risk of readmission can be identified at the time of discharge planning. Second, to reduce elders' risk of hospital readmission, hospital administrators and others responsible for discharge planning and care coordination programs can use the PAR in setting priorities for allocating personnel and resources to discharge planning and postdischarge care programs. Our results suggest that interventions in the long-term care setting may be effective in reducing hospital readmissions. Patients with cardiovascular disease (heart failure and peripheral vascular disease), chronic lung disease, renal failure, cancer, and diabetes mellitus were identified as having a high PAR of readmission. Previous studies have shown that patient interventions for heart failure reduce hospital readmissions (28
Previous studies of hospital readmission in the Medicare population found that male sex, Medicaid insurance, prior admission, and admission to hospitals with fewer beds were significantly associated with a higher risk of 60-day readmission, while younger age, nonwhite race, self-limited disease, surgery performed, and urban hospitals were associated with a lower risk of 60-day readmission (4
). Our study contained more detailed comorbidity and discharge location information and found a similar relationship for age and a different relationship for African American race. A 1991 meta-analysis of 44 studies reported that diagnoses, age, initial length of hospital stay, and prior use of hospital resources were related to readmission, but the strength of the relationship was trivial (31
). In our study, patients discharged to a skilled nursing facility had the highest risk of 30-day readmission. In contrast, a study of elders with chronic obstructive pulmonary disease, stroke, or dementia who were discharged to a nursing home were less likely to be readmitted within 30 days than patients discharged to home (32
). Another study of medical patients from a single hospital used recursive partitioning and identified three high-risk diagnoses (AIDS, renal disease, and cancer), albumin level, and prior admission within 60 days as factors associated with 90-day readmission (12
). A study of complicated care transitions in elders found that age, sex, insurance, prior hospitalizations, and three specific comorbidities (heart disease, diabetes mellitus, and cancer) predicted complicated transitions and that the predictions improved when information on health status and activities of daily living were included (23
). In 2000, a narrative review of hospital readmissions as a measure of quality of care concluded that most readmissions seemed to be caused by modifiable factors and that global readmission rates were not a useful indicator of quality of care. The authors noted, however, that a focus on specific needs of patients with defined problems may identify quality of care problems and lead to the creation of a more responsive health care system for the chronically ill (19
). We have developed and validated a method to identify elders at risk of 30-day readmission who may benefit from interventions to reduce readmission.
Our study has several strengths. First, the prediction rule is nonproprietary, and we encourage other investigators to use our prediction models in their settings and replicate our analytic approach with their locally available data. Second, the prediction rules were developed using administrative data readily available in hospital discharge databases. Third, the study included patients in both the medical and surgical services admitted to seven community hospitals of different sizes, including a tertiary care referral hospital. The study findings should therefore generalize to settings outside BHCS. Fourth, our study covers a range of variables in the conceptual domains of predisposing factors, need factors, and enabling factors (33
). The covariates available in the hospital administrative data included demographic factors of age, sex, and race which may predispose
to readmission; clinical conditions and type of hospital service related to the need
for subsequent inpatient care; health insurance that may enable
access to care and readmission; discharge with home care or discharge to a skilled nursing facility that may either substitute for inpatient care or facilitate access to subsequent inpatient care; distance, which may be a barrier to readmission or associated with care in a non-BHCS hospital; and income.
Our study was limited by its reliance on readily available hospital administrative data used to classify the DRG of the admission for billing and reimbursement purposes, and we were unable to fully replicate the covariates used in some of the published prediction models for hospital readmission in elders (9
). Administrative data based on codes from medical record review are known to be less accurate than prospectively collected clinical data and tend to underreport chronic conditions (35
). We included all readmissions because we were unable to identify planned readmissions, and our estimates may overestimate the risks of unplanned 30-day hospital readmission. We found that patients who resided 50 miles or more from the initial hospital (the top decile of distance) were less likely to be readmitted, which could be due to underascertainment of readmission of these patients to non-BHCS hospitals. However, analyses restricted to patients who resided within 50 miles of the index hospital yielded the same predictors of 30-day readmission, so we believe that these predictors are valid. We also used an ecological variable (median income quartile of ZIP code of residence) as a surrogate for income and other resources (social capital) that may be associated with access to care and hospital readmission. These covariates were not included in the final models because of concern about ascertainment of readmissions, missing ZIP code data, and lack of independent statistical significance for median income of ZIP code of residence.
An important limitation of our study was that it did not directly include information on patients' abilities to perform activities of daily living or other measures of physical function. It should be noted, however, that limitations in activities of daily living are required for eligibility for home care and are associated with admission to skilled nursing facilities, which were included in our analyses. Our index was designed to select patients based on information that would be available soon after an elder's admission. Thus, our predictors did not include measures of patients' stability, such as absence of fever for 48 hours prior to discharge or stable medication regimen in the 48 hours prior to discharge.
We believe that our predictive models will be useful in identifying elders who may benefit from interventions early during their hospital course. This could improve elders' transition from the hospital to home or to a skilled nursing care facility and could assist hospital administrators in setting priorities for allocating resources for care management and discharge planning.