Mortality in a Public and a Private Hospital Compared: The Severity of Antecedent Disorders in Medicare

Transcription

1 Mortality in a Public and a Private Hospital Compared: The Severity of Antecedent Disorders in Medicare Patients Robert Bums, MD, Linda Olivia Nichols, PhD, Marshall J. Graney, PhD, and William B. Applegate, MD Introduction Public hospitals have long argued that their patient populations are different from those in other facilities and that these patient characteristics may influence outcomes.' This issue is particularly important when outcomes are used to assess quality of care.'-3 Differences in patient outcomes have often been attributed to facility and physician characteristics.4'- However, hospital mortality rates should be adjusted for patient characteristics before the quality of care is assessed. The Health Care Financing Administration (HCFA) uses patient characteristics (e.g., diagnoses, age, gender) to calculate individual hospital mortality rates. However, the relationship between these rates and the quality of care at a hospital remains elusive. For example, more severely ill patients may skew mortality rates for certain types of hospitals (e.g., public hospitals or tertiary care hospitals), but current assessment techniques may not adequately adjust for patient characteristics. For example, diagnostic coding errors or inaccuracies may limit the value of the administrative data set for outcome assessment.7,8 Because of the limitations of currently used data sources, it may be difficult to identify hospitals with qualityof-care problems vs those in which random events may have increased the mortality rate.9 The HCFA has stated that hospitals identified as mortality outliers do not necessarily have quality-of-care problems.1' To investigate the importance of patient characteristics, particularly severity of illness, on any comparison of interhospital mortality rates, we examined data from both a university and a public hospital. The university hospital exists in a competitive health care market and should be considered comparable to a private community hospital. The two hospitals present a unique opportunity to examine hospital mortality because they share the same medical and surgical staff and many of the same ancillary services. However, the two patient populations, although drawn from the same community, are different. The public hospital serves a poor, medically indigent, predominantly (80% to 90%) Black patient population whereas the university hospital cares for a more heterogeneous group (50% Black, higher socioeconomic status). These hospitals also differ in their HCFA-calculated mortality rates, with the public hospital designated a mortality outlier in HCFA mortality analyses. Using variables from the first HCFA mortality analysis, this study addressed two research questions: (1) Were Medicare patients who died after being admitted to the public hospital more severely ill at admission than those admitted to the university hospital? (2) Does the inclusion of data on severity of illness at admission improve the accuracy of mortality predic- Robert Bums and Linda Olivia Nichols are with the Memphis Veterans Affairs Medical Center and the Departments of Internal Medicine and Preventive Medicine, University of Tennessee in Memphis. William B. Applegate is with the Department of Internal Medicine and, with Marshall J. Graney, the Department of Preventive Medicine and the Department of Biostatistics and Epidemiology, University of Tennessee. Requests for reprints should be sent to Robert Bums, MD, Veterans Affairs Medical Center (lib), 1030 Jefferson Ave, Memphis, TN This paper was accepted February 22, Note. The opinions expressed in this paper are those of the authors and do not necessarily reflect those of the Department of Veterans Affairs. July 1993, Vol. 83, No. 7

2 tion for both hospitals? The primary objective of this research was to determine if there were clinically significant differences between the two hospital populations, including differences in severity of illness, and to determine if those differences were associated with an increased risk of death. An assessment of the quality of care at the two hospitals was beyond the scope of this research. Methods Data Collection Retrospective chart review was undertaken at two hospitals providing secondary and tertiary care services to residents of Shelby County, Tennessee: the Regional Medical Center, a 437-bed public hospital, and the University of Tennessee Medical Center, a 126-bed university hospital. Both are sites for house staff training programs of the University of Tennessee at Memphis and share the same medical and resident staffs, subspecialty consultants, and multiple ancillary services, including radiology, special diagnostics (cardiology and gastroenterology laboratories), dialysis, and pathology. Resident house staff are the primary inpatient caregivers at both hospitals, supervised by faculty attending physicians. The university hospital patients are the private patients of the university faculty, whereas the public hospital patients are cared for in the facility's outpatient clinics. Data were collected for the term of the HCFA's 1986 mortality reporting period (January 1, 1986, to January 31, 1987). Although the mortalitymodel has changed each year, this study is based on the first model. Diagnostic categories studiedpulmonary disease, sepsis, cancer, and acute heart disease-were those in which the Regional Medical Center was near or had exceeded the upper limits of the predicted mortality rates. The mortality group included all Medicare patients in these diagnostic categories who died within 30 days following admission to either hospital (public = 104 cases; university = 26 cases). Comparison patients for both hospitals were randomly selected Medicare patients in the same diagnostic categories, during the same time frame, who did not die within 30 days following admission. Comparison charts were oversampled to ensure approximately the same number in the comparison group as in the mortality group. A high chart-retrieval rate resulted in more comparison patients (n = 145) than mortality cases (n = 130). A total of 275 charts were reviewed (public hospital = 184, university hospital = 91). In model development, the cases were weighted to provide equal numbers from each facility. Vanables Chart data were collected by trained utilization review personnel who were not blinded to patient outcome. Data included age, gender, race, principal and comorbid diagnoses, admission type (emergent or elective) and source, preadmission function (ambulatory, wheelchair, or bedbound), and number of hospitalizations in the past year. Admission severity was measured using the Acute Physiology Scale of APACHE II at admission (first recorded values)." Patients with missing data were categorized for analysis as having the mortality risk factor (i.e., normal lab value, ambulatory, home admission) absent. Most subjects (95%) were established patients at each hospital, and prior hospitalization data were available in their charts. Three patients were missing one variable each-type and source of admission and lab values-and 16% of subjects were missing functional status. Course severity was also included, determined by the highest Acute Physiology Scale score in any 24-hour period. The HCFA groups comorbid diagnoses into nine categories, reflecting broad diagnostic risk groups (e.g., renal disease).'2 However, not all patient comorbid diagnoses are used by the HCFA in mortality analysis. This is because not all comorbid diagnoses fit into one of the nine risk categories and because Medicare claims forms only list a maximum of four diagnoses, which possibly excludes some diagnoses that are listed on the discharge summary. To examine the effect of more extensive coding of comorbid disease, three variables were developed: HCFA comorbid diagnoses (up to four, as done by the HCFA), HCFA-eligible comorbid diagnoses (up to six), and total comorbid diagnoses (including those not part of the HCFA's nine mortality risk categories). For this sample, 228 (83%) patients had all their allowable comorbid diagnoses listed in the first four diagnoses on the hospital discharge summary (and thus available to the HCFA for analysis), and 260 (95%) patients had all HCFA-eligible comorbid diagnoses listed in the first six diagnoses. Statistical Analysis To test for differences between patients who lived and those who died and Patit Severity Differences between hospitals for patients who lived and those who died, variables were compared by means of chi-square test, Student's t test, or Fisher's Exact Test, as appropriate. Death within 30 days of admission was the dependent variable. Two logistic regression models were developed, using cases from both hospitals. The HCFA variables model was the model used by the HCFA and included age, gender, past year hospitalizations, transfer from another hospital, and HCFA comorbid diseases. The severity-adjusted HCFA model added severity. Admission severity (Acute Physiology Scale score) was used for two reasons. First, although there is now research demonstrating the increased utility of repeated measurements, APACHE II was developed for use in the first 24 hours of admission to an intensive care unit.1' Second, using admission severity minimizes the influence of quality of care during hospital stay. Although it is not used in the predictive models, course severity is also reported. The two models were tested on each hospital's data. Assessments of logistic regression models were performed in the same manner as assessments of diagnostic medical tests. That is, analysis of the validity of the prediction models was based on tables that cross-tabulated model-predicted outcomes (probability of death was <.5 or 2.5) and actual outcomes (no death vs death).'3 The outcome predicted by the logistic regression model was the likelihood of death. Receiver operating characteristics were also computed, and their resultant curves for the prediction models were graphed as (100-specificity) by sensitivity.14 Comparative predictive performance was analyzed using the difference between models' areas under the curve.'5 Results The diagnoses included cancer (115), pulmonary disease (83), acute heart disease (39), and sepsis (36). Mean age was 72.3 years (SD = 11.5, range = 23 to 99). There were 132 (48%) men and 201 (73%) Blacks. There were 130 mortality patients and 145 comparison patients. In our sample, 84% of transfers came from long-term care. However, in the HCFA model, the source of admission was hospital vs all other sources, and this form of the variable was used in model development. July 1993, Vol. 83, No. 7 American Journal of Public Health 967

3 Buns et al. Interhospital Contrasts: Comparison Groups There were no significant differences in age (P =.21), gender (P =.27), bedbound status (P =.40), Acute Physiology Scale course score (P =.23), admission from home (P =.33) or long-term care facility (P =.41), or number of HCFA (P =.65) or HCFA-eligible (P =.55) comorbid diagnoses between comparison groups at the two hospitals. There was, however, a trend toward comparison patients at the public hospital having a higher Acute Physiology Scale admission score than those at the university hospital (P =.08). Regional Medical Center patients did have more previous hospitalizations (P =.04) and total (P <.001) comorbid diagnoses, and theywere more likely to be emergently admitted (P =.003). University oftennessee Medical Center comparison patients were more likely to be transferred from a hospital (P =.03). Most (93%) of the comparison patients in the public hospital were Black compared with 46% of those at the university hospital (P <.001). A similar percentage difference (82% vs 50%) also held in the mortality group (P =.002). Interhospital Contrasts: Mortality Groups Table 1 documents comparisons between the two hospitals' mortality groups. There were no significant differences between the two groups in age (P =.42), gender (P =.93), hospitalizations (P =.33), transfer from another hospital (P =.20), or number ofhcfa comorbidities (P =.10); however, therewas a trend toward public hospital patients having higher course severity (P =.08). Compared with the university hospital mortality group, the public hospital mortality group had significantly more HCFA-eligible (P =.01) and total (P <.001) comorbid diseases; a significantly higher admission severity (P <.001); a significantly higher percentage of patients admitted emergently (P =.02), from long-term care (P =.001), and bedbound (P =.002); and fewer patients admitted from home (P =.05). As mentioned above, Regional Medical Center patients were also more likely to be Black (P =.002). Mortality Group/Comparison Group Contrasts Table 2 documents comparisons between the two mortality groups combined vs the two comparison groups combined. In general, mortality patients at both hospitals presented with more severe illness than comparison patients. In addition to having significantly higher admission and course severity (P <.001), mortality patients were significantly more likely to be older (P =.02) and to have been admitted emergently and from long-term care facilities (both Ps <.001), and they had significantly more HCFA-eligible (P =.01) and total (P <.001) comorbidities; they were also significantly less likely to be admitted from home (P <.001). Bedbound status 968 American Journal of Public Health July 1993, Vol. 83, No. 7

4 approached significance (P =.06), but race (P =.58), gender (P =.48), previous hospitalizations (P =.48), transfer from another hospital (P =.57), and number of HCFA comorbidities (P =.27) were not significantly different between mortality and comparison groups. Mortality Prediction As shown in Table 3, of the HCFA variables (HCFA comorbidities, age, gender, hospitalizations, and transfer from another hospital), only age was significantly different between the mortality and comparison groups (P <.05). However, all HCFA variables were retained in both models. Admission severity was the only other modelvariable significantly different between the two groups (P <.001). Application of the two models to each hospital yielded operating characteristics shown in Table 4 and receiver operating characteristics curves shown in Figure 1. Analysis of areas under the curves found that, for the university hospital, addition of severity did not significantly change model performance (z = 0.69, P =.49). At the public hospital, however, performance was significantly improved with severity (z = 5.50, P '.01). Diwussion Health care outcomes are a function of patient characteristics, quality of care, and chance. Failure to make adequate adjustment for patient characteristics may result in the erroneous conclusion that adverse outcomes such as mortality are due to poor medical care rather than being anticipated outcomes based on the patient population. Our study addressed this issue in a comparison ofmedicare patients admitted to two teaching hospitals served by the same medical and house staff. There were significant differences between the patientswho died at the two hospitals. Public hospital patients had a greater burden of chronic illness with resultant functional impairment, as measured by bedbound status and nursing home residence. In addition, these patients had higher severity and more comorbid diseases at admission, and they were more likely to be admitted emergently. These factors are not influenced by the quality of hospital care. For the university hospital, the addition of severity to the HCFA variables model did not improve outcome prediction. However, for the public hospital, the addition of severity resulted in a significant improvement of outcome prediction. These results suggest that mortality rates at the public hospital may have been higher due to higher admission severity in the mortality group patients. The improvement in mortality prediction with the addition of severity was expected, based on other research In a study of Medicare beneficiaries, Green and colleagues used a methodology similar to ours and computed a multiple correlation coefficient for their dichotomous outcome.19 The explained variance increased 19.5% with the addition of severity to their simulated HCFA model.19 Following their method, our se- Patient Severity Differec verity-adjusted model, using all cases, had a comparable increase in an explained variance of 22.3%. Several limitations of our study must be addressed. We included only two hospitals with a small number of patients; consequently, some of our findings may be specific for our setting. We did not assess the quality of care at either hospital; however, this does not diminish our findings that the two hospitals provide care to different patient populations. Those differences were present at admission and were associated with an increased risk of death Ȯur study should not be generalized July 1993, Vol. 83, No. 7 American Journal of Public Health 969

5 Burns et al. Sensitivity (%) / 50, 40' ' 10' o Specificity (%) Public Hospii Popuifton Sensitvity (%) ( 20/ o SSpmellty(%) Popuiatlo HCFA MODEL SEV. ADJ. MODEL Private Hosphlta Note. Model formulas are as follows. HCFA model: PR(mortality 0.50) = 1/{1 + exp[-1(ijxj)]}, where PI = -2.02, 12 = 0.21, P3 =0.23, P4 = -0.46, P5 = 0.02, 16 = 0.09, X,= unity, X2 = age per decade, X3= gender (female = 1, male = 0), X4 = hospitalizations past year, X5= source (hospital = 1, all others = 0), X6 = HCFA comorbidities. Severity-adjusted HCFA model: PR- (mortality 0.50) = 1/{1 + exp[-1(px3]}, where 13 = -2.77, 12 = 0.14, 13 =0.02, 04 = -0.04, 15 = -0.09, 6 =0.14, 17 = 0.15, X= unity, X2 = age per decade, X3= gender (female = 1, male = 0), X4= hospitalizations past year, X5= source (hospital = 1, all others = 0), X6 = HCFA comorbidities, X7 = APS admission severity, per point. FiGURE 1-Recever opering charactist curves of multiple logisdic regression models predictng mortality within 30 days of hospia admission: public and private hospital to all Medicare patients and other hospitals until further data are available. However, our model's improvement after severity data were incorporated was similar to Green's,19 suggesting comparable populations. Although race has been shown to be associated with a worse outcome for certain diseases,20,21 it was not associated with mortality in our study even though the racial composition of the two hospitals was different. Multiple sites need to be analyzed to determine the influence of race on mortality. Previous research has found that comorbid diseases from claims data have limited predictive value for mortality.77 In our analysis, HCFA's comorbid disease variable, which is limited to four diseases, was not a significant predictor of mortality. Although accurate coding of all comorbid diseases may be difficult,7 limiting the number of diseases used for mortality rate adjustment may further decrease the clinical contribution of comorbid disease data. Nondisease specific markers for severity (e.g., emergent admission or longterm care residency) available on administrative data sets, such as the Uniform Hospital Discharge Data Set, could be used to improve outcome assessment. Previous research has shown that emergent admission may identify patients for adverse outcomes (e.g., mortality, length of stay, and higher charges) 24 and could be particularly important for public hospitals with a large percentage of emergent admissions. In our study, the majority oftransfers in the mortality group (91%) were from long-term care facilities. Too few people were transferred from another hospital to draw any conclusions about thisvariable's relationship to mortality. However, nursing home transfer is a severity proxy variable that is available from claims data, and the HCFA has modified its original prediction model to include skilled facility transfer. In our study, however, approximately halfofthe transfers from long-term care facilities were from intermediate care. There may be potential biases in not including all transfers from long-term care facilities because designation of care level (e.g., skilled or intermediate) is a state decision with no federal oversight. Any attempt to assess the quality of care for large groups of patients, such as hospitalized Medicare beneficiaries, must start with an accurate screening process. To date, selected components of the Uniform Hospital Discharge Data Set have been used to calculate adjusted mortality rates for hospitals. Our study, in addition to the work of others,7'8.'9 demonstrates that administrative data sets, as currently used, may not be adequate to explain significant differences in outcomes. Variations between predicted and actual mortality rates may occur because of differences in patient populations, inadequate or inaccurate reimbursement information, or poor quality of care. Future research should be directed at examining improvement in the predictive ability of administrative data sets. Future research should also examine differences between patient populations, especially patients who use public vs private hospitals. O Acknowledgments This work was conducted at the Veterans Affairs Medical Center in Memphis, Tenn. This paper was presented in part at the annual meeting of the American Public Health Association, Boston, Mass, November Wewould like to thank Camille Guyler for data collection, Thomas Cloar for data analysis, and Johnnie Smith for manuscript preparation. References 1. Dubois RW, Rogers WH, MoxleyJH, et al. Hospital inpatient mortality. Is it a predictor of quality? N Engl J Med. 1987;317: Joint Commission on the Accreditation of Healthcare Organizations. The Agenda for Change.AgendaforChange Update. 1987; 1: Health Care Financing Administration. Medicare and Medicaid programs; health care financing research and demonstration; availability of funds for cooperative agreements and grants. Federal Regter. 1987; 52: Luft HS. The relation between surgicalvolume and mortality: an exploration ofcausal factors and alternative models. Med Care. 1980;18: Flood AB, Scott WR, Ewy W. Does practice make perfect? part 2. the relation between volume and outcomes and other hospital characteristics. Med Care. 1984;22: LuftHS, BunkerJP, EnthovenAC. Should operations be regionalized? NEnglJMed. 1979;301: Fisher ES, Whaley FS, Krushat WM, et al. The accuracy ofmedicare's hospital claims data: progress has been made, but prob lemsremain.amjpublichealth. 1992;82: Hsia DC, Krushat M, Fagan AB, et al. Accuracy of diagnostic coding for Medicare patients under prospective pricing system. NEngl JMed. 1988;318: Park RE, Brook RM, KosecoffJ, et al. Explaining variations in hospital death rates. Randomness, severity of illness, quality of care. JAMA. 1990;264: McCormick B. Hospitals find inclusion on HCFA "hit list" damaging. Am Med News, May 27, Knaus WA, Draper EA, Wagner DP, et al. APACHE II: a severity of disease classification system for severey ill patients. Cit Care Med. 1985;13: Federal Register. 1987;52:158: American Joumal of Public Health July 1993, Vol. 83, No. 7

6 Norusis MJ. SPSS/PC+, Base System User's Guide, Version 5.0. Chicago, Ill: SPSS Inc; Sackett DL, Haynes RB, Tugwell P. Clinical Epideniology. Boston, Mass: Little, Brown Co; Rosner B. Fundamentals ofbiostatistics. 2nd ed. Boston, Mass: Duxbuiy Press; Dubois RW, Brook RH, Rogers WH. Adjusted hospital death rates: a potential screen for quality of medical care. Am J Public Health. 1987;77: Daley J, Jencks S, Draper D, et al. Predicting hospital-associated mortality for Medicare patients.jama ;260: Jencks SF, Daley J, Draper D, et al. Interpreting hospital mortality data. JAMA. 1988;260: Green J, Wintfeld N, Sharkey P, et al. The importance of severity of illness in assessing hospital mortality. JAMA ;263: Castaner A, Simmons B, Mar M, et al. Myocardial infarction among Black patients: poor prognosis after hospital discharge. Ann IntemMedl 1988;109: Cooper RS, Simmons B, Castaner A, et al. Patiet Severity Diffens Survival rates and prehospital delay during myocardial infarction among Black persons. Am J CardioL 1986;57: Jencks SF, WilliamsDK, KayTL. Assessing hospital associated deaths from discharge data. JAAL ;260: Munoz E, Soldano R, Laughlin A, et al. Source of admission and cost: public hospitals face financial risk. Am J Public Healtkh 1986;76: Munoz E, Cohen JR, Dietzek A, et al. Route of admission and hospital costs for urologic patients. Urology. 1990;35: X M.M July 1993, Vol. 83, No. 7 American Journal of Public Health 971