Author + information
- Received August 28, 1998
- Revision received January 5, 1999
- Accepted January 21, 1999
- Published online May 1, 1999.
- ↵*Reprint requests and correspondence: Dr. Edward F. Philbin, Head, Section of Heart Failure and Cardiac Transplantation, Division of Cardiovascular Medicine, Henry Ford Hospital, 2799 West Grand Blvd, Detroit, Michigan 48202
The purpose of this study was to develop a convenient and inexpensive method for identifying an individual’s risk for hospital readmission for congestive heart failure (CHF) using information derived exclusively from administrative data sources and available at the time of an index hospital discharge.
Rates of readmission are high after hospitalization for CHF. The significant determinants of rehospitalization are debated.
Administrative information on all 1995 hospital discharges in New York State which were assigned International Classification of Diseases–9–Clinical Modification codes indicative of CHF in the principal diagnosis position were obtained. The following were compared among hospital survivors who did and did not experience readmission: demographics, comorbid illness, hospital type and location, processes of care, length of stay and hospital charges.
A total of 42,731 black or white patients were identified. The subgroup of 9,112 patients (21.3%) who were readmitted were distinguished by a greater proportion of blacks, a higher prevalence of Medicare and Medicaid insurance, more comorbid illnesses and the use of telemetry monitoring during their index hospitalization. Patients treated at rural hospitals, those discharged to skilled nursing facilities and those having echocardiograms or cardiac catheterization were less likely to be readmitted. Using multiple regression methods, a simple methodology was devised that segregated patients into low, intermediate and high risk for readmission.
Patient characteristics, hospital features, processes of care and clinical outcomes may be used to estimate the risk of hospital readmission for CHF. However, some of the variation in rehospitalization risk remains unexplained and may be the result of discretionary behavior by physicians and patients.
Congestive heart failure (CHF) is the leading cause of hospital admission among patients over the age of 65 years (1). This syndrome is associated with high rates of mortality and morbidity (2), including hospital readmission (3–6), thereby posing a significant societal burden (7). Substantial effort is being devoted to devise safe and effective programs to reduce preventable hospital admissions and readmissions for CHF (8–13). Logically, these efforts should focus on individuals at greatest risk for rehospitalization for CHF, as CHF is the most common cause for readmission among this group of patients (3,4,6,8), and this end point is the clinical outcome most affected by disease state management programs (9,11). However, the prediction of readmission, like the prediction of mortality (14), may be imprecise in the hospital setting. Previous studies of determinants of hospital admission among patients with CHF have focused on all-cause readmission rates (3,4), have been limited by small sample size (4–6)or have been single-center investigations (4–6). The current study utilized an existing hospital discharge data set and retrospective analyses to examine the prediction of rehospitalization for CHF among a large and diverse group of patients. We sought to determine whether an individual’s risk for readmission could be calculated at the time of hospital discharge in a convenient and inexpensive way using information derived exclusively from administrative data sources.
Information on all 1995 New York State hospital discharges assigned International Classification of Diseases–9–Clinical Modification (ICD-9-CM) codes indicative of CHF in the principal diagnosis position were obtained from the Statewide Planning and Research Cooperative System (SPARCS) database. SPARCS is an agency of the New York State Department of Health which, by law, incorporates information or all patients hospitalized in acute care facilities from various sources including the uniform bill and uniform discharge abstract submitted by hospitals. Clinical investigators may petition the Department of Health for access to some of all of these administrative records. The application process includes review of the scientific merit of the proposed project and approval by an institutional review board or ethics committee. This study was approved by the institutional review board of the Massachusetts General Hospital.
The codes used were 428.0, 402.91, 404.93, 428.1, 402.11, 398.91, 404.91, 404.13, 402.01, 404.03, 404.11, 404.01 and 428.9. This method of case selection defined a group of patients whose primary diagnosis was CHF, irrespective of procedures performed. The chronologically first hospital admission during 1995 for each patient was considered the index admission (or discharge). Patients who died during their index admission were excluded from this analysis. To simplify analyses of potentially complex relationships between ethnicity and other predictors of readmission (5,15), patients whose race was reported “unknown,” “other” or any race other than black or white were also excluded.
Readmission for CHF was determined by searching the same data set for subsequent hospitalizations occurring before December 31, 1995 for each individual who qualified for inclusion in the study. Readmission was coded as present or absent for each patient. The records of 3% of the patients were manually reviewed by trained chart auditors to confirm the accuracy of the unique patient identifier used to identify hospital readmission, and to confirm the presence of CHF based on the documentation of typical symptoms, physical findings, diagnostic test results and clinical response to appropriate therapy.
Hospital length of stay was defined as the date of discharge or death minus the date of admission. Discharges were classified as “urban” if they occurred at a hospital located in a county which is part of a Federal Metropolitan Statistical Area. All other discharges were classified as “rural.” Discharges were classified as “teaching” if they occurred at a hospital listed as a primary or affiliated institution of an accredited internal medicine or family practice residency program according to the American Medical Association’s directory of postgraduate medical training programs (16). All other discharges were classified as “nonteaching.”
Coexistent illnesses were determined by searching the principal ICD-9-CM diagnosis code and up to 14 secondary diagnosis codes for each patient. Total comorbid disease was quantified according to the method of Charlson. To achieve this, a Charlson Comorbidity Index (17), and its age-modified variant (18), were calculated for each patient. Process of care was determined by searching the principal procedure code and up to 14 secondary procedure codes for each patient. In some cases, similar procedures with closely related codes were combined to yield clinically relevant composite rates of technical services. Medical specialty codes for a maximum of three physicians caring for each patient were examined. A patient was classified as receiving care from a cardiologist if any of his or her providers was listed as a specialist in cardiovascular diseases.
Differences between patients readmitted and those not readmitted were analyzed using chi-square tables (categorical data) and Student unpaired ttest (interval data). To provide the opportunity to both develop risk scores and test their validity within a single data set, all patients were randomly assigned to either a derivation subset or a validation subset. Among patients in the derivation subset only, SAS’s PROC LOGISTIC(19)was used to determine which of the patient characteristics, hospital features, processes of care and clinical outcomes had independent predictive value for hospital readmission. All predictors with a significant or borderline statistical relationship with rehospitalization at the univariate level (p ≤ 0.10) were entered as independent variables in a logistic regression model for readmission. This technique yielded odds ratios and confidence intervals for readmission for every predictor variable tested in the model. After identifying those variables with significant independent predictive value (p ≤ 0.05 in the logistic regression model), two risk scoring systems based on these variables were derived. The simple risk score was computed by adding the number of positive predictors for readmission present for each patient, then subtracting the number of negative predictors present. The modified risk score assigned weights to each positive and negative predictor based on the odds ratio for that variable derived from the logistic regression model. The precision of the single best prediction method developed using data from the derivation subset only was then tested among patients in the validation subset only by entering each patient’s risk score as the independent variable in a logistic regression model for readmission. In interpreting results, a p value ≤ 0.05 was considered statistically significant. In this report, results are displayed as mean ± standard deviation.
Patients and hospitals
A total of 52,010 individual patients were hospitalized at least once in New York State during 1995 with the chosen ICD-9-CM codes. Of these, 6,116 were excluded because race was reported as neither black nor white (Native American, 102 patients; Asian, 518; other race, 3,330; unknown race, 2,196). Of the remaining 45,894 patients, 3,163 died during their index hospitalization and were also excluded from this analysis. Thus, the study sample comprised 42,731 black or white patients. Of these, 9,112 (21.3%) were rehospitalized at least once for CHF after their index discharge; 33,619 (78.7%) were not. In the opinion of the chart auditors, CHF was present and was a primary reason for hospitalization in 96% of the 1,220 cases reviewed manually.
A total of 236 of New York’s 243 acute care hospitals contributed at least one patient to the study sample of 42,731. Among these 236 institutions, the caseload per hospital ranged from 1 to 590 patients, with a median of 274. Of all discharges, 10.2% occurred at rural hospitals; 51.9% of them occurred at teaching institutions.
Among the 42,731 patients included in this study, mean age was 73.7 ± 13.2 years. Women comprised the majority of the cohort (56.4%). Blacks formed 18.2% of the group. The mean Charlson Comorbidity Index was 2.5 ± 1.6, and mean age-modified Charlson Comorbidity Index was 5.5 ± 2.0. Mean length of stay during the index admission was 9.2 ± 13.5 days, and mean hospital charges were $11,026 ± 14,748. The median postdischarge “follow-up” period for all patients was 6.9 months. Because patients in the rehospitalized group tended to experience their index discharge somewhat earlier during calendar year 1995, their median period of follow-up was 8.5 months. The same time period was 6.4 months for the nonrehospitalized group. Among patients who had one or more rehospitalizations, the median time from their index discharge to their first readmission was 91 days. Mortality rate during the index admission among black or white patients was 6.9%; the rate of death during subsequent hospitalizations was 7.9% (p < 0.0001).
Table 1displays the basic demographic characteristics, comorbid illnesses and comorbidity scores of the entire cohort, stratified by readmission status. Clinically relevant features distinguishing the rehospitalized group included a higher proportion of blacks, a higher prevalence of Medicare and Medicaid insurance, a lower prevalence of health maintenance organization and indemnity insurance and a higher prevalence of ischemic heart disease, idiopathic cardiomyopathy, prior cardiac surgery, peripheral vascular disease, renal disease, diabetes mellitus and anemia. The rehospitalized group also had higher mean comorbidity scores.
Process of care
Processes of care during the index hospital admission are shown in Table 2, stratified by readmission status. Patients who were readmitted were less likely to undergo the following procedures during their index hospitalization: echocardiography, exercise stress testing, cardiac catheterization, coronary revascularization or any cardiac surgical procedure. Readmitted patients were somewhat more likely to have telemetry electrocardiographic monitoring during their index hospitalization.
Resource utilization and crude clinical outcomes
Table 3displays the resource utilization and crude clinical outcomes during the index hospital admission, stratified by readmission status. Patients rehospitalized were less likely to be discharged to nursing homes, and more likely to receive home health care services, after their index discharge. Length of stay and hospital charges during the index admission were not different between groups.
The logistic regression model for rehospitalization, with a cstatistic of 0.62, retained 12 predictors of higher risk of readmission and four predictors of lower risk of readmission. The predictors of higher risk were black race (odds ratio = 1.28, 95% confidence interval = 1.16 to 1.41), Medicare insurance (1.66, 1.38 to 2.00), Medicaid insurance (1.92, 1.57 to 2.36), home health care services after discharge (1.10, 1.01 to 1.21), ischemic heart disease (1.25, 1.16 to 1.34), valvular heart disease (1.19, 1.09 to 1.29), diabetes mellitus (1.45, 1.33 to 1.58), renal disease (1.35, 1.23 to 1.49), chronic lung disease (1.10, 1.02 to 1.20), idiopathic cardiomyopathy (1.46, 1.32 to 1.61), prior cardiac surgery (1.16, 1.04 to 1.29) and use of telemetry monitoring during index hospitalization (1.13, 1.01 to 1.27). The predictors of lower risk were treatment in a rural hospital (0.87, 0.78 to 0.98), discharge to a skilled nursing facility (0.68, 0.59 to 0.79), performance of echocardiography during the index admission (0.78, 0.73 to 0.85) and performance of cardiac catheterization during the index admission (0.60, 0.49 to 0.73).
Composite risk scores
In the simple scoring system (Table 4), one integer was added for each predictor of higher risk present, and one subtracted for each predictor of lower risk present. To avoid negative values, a correction factor of +4 was added to this sum. Thus, the range of potential values for the simple risk score was 0 to 15. (Because a patient’s primary insurance could not have been both Medicare and Medicaid, a maximum of 11 positive predictors could be present for any individual. A single variable “Medicare or Medicaid” was not created because of the different odds ratios associated with these two predictors.) Among patients in the derivation subset only, a logistic regression model for readmission with each patient’s simple risk score as the independent variable had a cstatistic of 0.60, indicating that the risk score was moderately but significantly predictive of readmission (p < 0.001). The predictive value of this simple risk score was not reduced by grouping all patients with a score of 0 to 3 into a single unit and all patients with a score ≥11 into another single unit (cstatistic of 0.60, p < 0.001). The modified scoring system based on weighting the 16 predictor variables proportional to their odds ratios performed no better than the simple system (cstatistic of 0.61, p < 0.001).
Validation of risk scores
Among patients in the validation subset only, a logistic regression model for readmission with each patient’s simple risk score as the independent variable had a cstatistic of 0.60, indicating that the risk score was moderately but significantly predictive of readmission (p < 0.001). Thus, this model performed equally as well in the validation subset as it did in the derivation subset. The actual readmission rates observed among patients in the validation subset as a function of incrementally higher simple risk scores is shown in Figure 1. As shown, the observed readmission rates ranged from 9.8% among patients with simple risk scores from 0 to 3 to 45.4% among patients with scores ≥11.
In this study, we examined how patient characteristics, hospital features, process of care and clinical outcomes, as derived from an administrative data set, relate to the risk for readmission among a large and diverse group of patients treated for CHF. The principal findings of this study are: 1) 21.3% of the 42,731 black or white hospital survivors admitted for CHF in New York State during 1995 were readmitted with the same diagnosis during a median follow-up period of 6.9 months after their first discharge that year; 2) using univariate statistical techniques, we found black race, Medicare and Medicaid insurance, ischemic heart disease, idiopathic cardiomyopathy, prior cardiac surgery, peripheral vascular disease, renal disease, diabetes mellitus and anemia to be related to a greater risk of readmission; 3) using univariate statistical techniques, we found that patients undergoing echocardiography, exercise stress testing, cardiac catheterization, coronary revascularization or any cardiac surgical procedure were less likely to be readmitted; 4) a simple and convenient scoring system based on administrative data, which can be calculated at the time of a patient’s index hospital discharge, may be used to estimate his or her risk of readmission, and 5) there is residual variation in hospital readmission not explained by the regression models developed in this study.
In analyzing a large administrative data set, Krumholz et al. found male gender, prior hospitalization, higher comorbidity score, treatment in a tertiary care hospital and prolonged length of stay during the index hospitalization to be important predictors of all-cause readmission among Medicare patients hospitalized in Connecticut (3). These investigators could not study the effect of invasive cardiac procedures because of their choice of Diagnosis-Related Group inclusion criteria. In examining outcomes among a small number of patients treated at a single urban university teaching center, Vinson et al. found that past hospitalizations, prior CHF, acute myocardial ischemia and uncontrolled hypertension were significant predictors of all-cause readmission (4). In both studies (3,4), age was not an important predictor. In examining CHF admissions among a small number of black patients treated at a single urban university teaching hospital, Ghali et al. found that hospitalization for decompensated CHF could be explained in most cases by poor patient compliance, cardiac arrhythmias, emotional and environmental factors, poor medical care, pulmonary infection or thyrotoxicosis (5). Unfortunately, Ghali et al. did not include a comparison group of nonhospitalized patients to allow for determination of the significant predictors of admission or readmission. Chin and Goldman, examining a small number of patients treated at a single urban university teaching hospital, found that single marital status, hypotension, absent electrocardiographic repolarization abnormalities and higher Charlson Comorbidity Index were significant multivariable correlates of the combined end point of postdischarge death or hospital readmission (6). These authors did not report the determinants of readmission as an isolated end point.
From a methodologic viewpoint, our study differed from the others by including a much larger number of patients, perhaps with more diffuse demographic characteristics. Processes of care during the index hospitalization were examined as potential predictors based on their putative relationship with clinical events (20). Conversely, our data set lacked some of the richness of clinical information available to Vinson and colleagues (4), Ghali et al. (5)and Chin and Goldman (6). The results of our study were concordant with the prior investigations in describing comorbid illness as a major determinant of hospitalization risk. We, too, found that age was not a significant predictor of hospitalization after adjustment for other factors. Like Krumholz et al. (3), we found that patients treated at urban hospitals were more likely to be readmitted. In contrast, we found no relationship between readmission risk and gender or prior length of stay. The limitations of our data set precluded the opportunity to study prior hospitalization, social support and compliance as predictors of rehospitalization. Our study provided the additional insight that, after adjustment for other factors, race, medical insurance type and certain processes of care are significant positive or negative predictors of the risk for readmission for CHF. Moreover, our study yielded a framework for combining these multiple risk states to formulate a composite CHF readmission risk score—a utility not provided by previous authors.
Unexplained variation in readmission
Our results indicate that a prediction rule derived from administrative data collected during a patient’s index hospitalization only partially explains the risk for readmission. Our simple risk scoring system is convenient and inexpensive, as it is based on data that are readily available and not dependent upon case-by-case chart review (21). More precise mathematical prediction of readmission risk that reduces the unexplained residual variation in this end point would require additional data input from two broad domains: 1) patient- and disease-specific characteristics identifiable during the index hospitalization; and 2) factors related to the patient’s CHF and other medical conditions and their management assessed after discharge.
Regarding the first of these two domains, the limitations of administrative data are well known (21–23). Illness severity is not fully captured in such sources of data, due to inadequate clinical information such as disease-specific functional status, physical examination findings, laboratory results, medication use, social support structure or valid global measures of risk. Thus, incorporation of more detailed measures of CHF severity and treatment, and global health status, might have improved the accuracy of predicting readmission. For example, information regarding prior hospitalizations (3,4), angiotensin-converting enzyme inhibitor use (24), digitalis use (25), emotional support (26)and quality of care (27)would likely be of value. Unfortunately, with the limitations of existing information systems, costly chart review would most often be required to collect and quantify these putative predictors. Such a process would not be feasible for studying large groups of patients or hospitals (3,21).
Regarding the postdischarge domain, many factors may impact on the risk for readmission. For example, the timing and frequency of medical follow-up (28), medication use (24,25), compliance with medical recommendations (5), participation in a care management program (9–13)and the specialty of the treating physician (29,30)affect the likelihood of hospitalization for CHF. Physician discretionary behavior also affects hospital stays for CHF (31). Other factors that are difficult to measure and quantify, such as a patient’s sense of security or well-being, likely contribute to the variation in rehospitalization. Thus, were it feasible and affordable to measure and quantify these factors, the mathematical prediction of readmission risk would probably become more precise. Of course, those measurements that could only be made after discharge would not be of utility in quantifying risk before a patient leaves the hospital.
The potential implications of our study deserve comment. From the clinical viewpoint, quantification of readmission risk at the time of hospital discharge is of value in that it provides the opportunity to enroll high risk patients into proactive care management programs. Programs of this type (9–13)have been demonstrated to be effective in reducing costs from hospitalization for CHF while improving quality of care and patient functional status. The use of the scoring system developed in the current study brings some refinement to the prediction of readmission risk, as it differentiates those with very low risk from those at intermediate and high risk—a much more precise process than assuming that all previously hospitalized patients have equally high risk (9). From the health services research viewpoint, our findings may be of use in comparing risk-stratified readmission rates among groups of patients, such as those treated within differing health care delivery systems (28,32).
As the present study was based on administrative discharge data, we cannot be sure how our patient groups truly related to one another upon entry into the hospital in terms of CHF severity or global health status. Furthermore, because the statistical methods employed in this study are designed only to examine associations between variables, we cannot draw conclusions about how or why predictor variables (such as hospital type or process of care) caused or prevented readmission. We cannot discount the possibility that biases in hospitals’ coding practices affected the results of this study. Patients living in areas near the geographic borders of New York, or moving from the state after their index discharge, may have been rehospitalized in hospitals outside of the state, thus rendering our study potentially vulnerable to slight underestimation of the rate of readmission, but not necessarily weakening the prediction of this event. As hospital readmissions for CHF were the only data available during the postdischarge period, cumulative mortality rates would by necessity be underestimates and are therefore not reported. As our study sample was limited to black or white patients treated in a single state during a single year, our results may not be generalizable to all patients hospitalized for CHF. Because our study, by design, derived prediction rules exclusively from administrative data, we were forced to exclude potentially important clinical predictors such as a history of prior hospitalizations (3)and psychosocial and behavioral factors (5,26). In point of fact, it is likely that the most powerful prediction methods are derived from a combination of administrative data and a few select clinical variables that are obtained by chart review (33).
- congestive heart failure
- International Classification of Diseases–9–Clinical Modification
- Statewide Planning and Research Cooperative System
- Received August 28, 1998.
- Revision received January 5, 1999.
- Accepted January 21, 1999.
- American College of Cardiology
- Parmley W
- Domanski M.J,
- Garg R,
- Yusuf S
- Hanumanthu S,
- Butler J,
- Chomsky D,
- et al.
- Fonarow G.C,
- Stevenson L.W,
- Walden J.A,
- et al.
- ↵Graduate Medical Education Directory 1995–1996. Chicago, IL: American Medical Association.
- ↵SAS/STAT User’s Guide, Version 6. 4th ed., vol 2. Cary, NC: SAS Institute, 1989:1071–126.
- Daley J,
- Shwartz M
- Krumholz H.M,
- Butler J,
- Miller J,
- et al.
- Reis S.E,
- Holubkov R,
- Edmundowicz D,
- et al.
- Philbin E.F,
- DiSalvo T.G