HHS Quality of Care Measurement and Patient Classification This project, called the Nonintrusive Outcomes Study, is exploring several measurement issues regarding the use of Medicare data for measuring quality of care. The primary objective of this study is to determine whether patient outcomes that can be defined through Medicare part A data adequately and accurately reflect the quality of inpatient care. One aspect of the study focuses on the hospital-specific mortality rates for a set of 48 medical conditions. Data from all Medicare-certified acute care hospitals were analyzed to determine whether individual hospitals have mortality rates (for particular conditions) that vary significantly from expected rates. Further analyses are being performed to determine whether patterns among hospitals identified as having significant differences between their observed and expected mortality rates reflect more than random variation. This second aspect of the study involves comparing the outcomes data derived from Medicare part A records to information abstracted from medical records for two conditions (myocardial infarction and congestive heart failure). Approximately 3,000 medical charts from over 800 hospitals in four states will be reviewed to assess whether outcome indicators based upon claims data can properly classify patients. Patient classification systems are of interest for two reasons. First, they may be used as the basis for determining appropriate levels of reim bursement for different clinical conditions. Second, they may be used to adjust for case-mix differences among hospitals in quality assessment studies. Medicare's patient classification system based on diagnosis-related groups (DRGS) is used primarily for reimbursement purposes. The system assumes that patients with similar ages, diagnoses, and use of surgical procedures tend to use similar levels of hospital resources. Though this system is widely used, there is recognition that it does not control for within-DRG variation in treatment strategies and patient severity of illness.2 Because the DRG system does not adjust within DRGS for severity of illness or for variations in the underlying clinical condition, it is difficult to attribute variations in patient outcomes solely to hospital treatment effects. Only a small number of existing patient classification systems attempt to adjust for variations in the severity levels of patients. In the 2See, for example, S. Jencks, et al., "Evaluating and Improving the Measurement of Hospital Case Mix,” Health Care Financing Review Annual Supplement (November 1984), pp. 1-11. HHS Quality of Care Measurement and past, the methodological sophistication of many of these systems has been hampered by data availability and difficulty in specifying the appropriate variables to use in adjusting risk. However, in the last few years, research on patient classification systems has made significant strides toward identifying similar risk groups within the Medicare patient population. HCFA and NCHSR&HCTA have both supported research on patient classification systems. NCHSR&HCTA has been actively involved in the funding of research on case-mix and severity-of-illness measures, including support of the development of the Acute Physiology and Chronic Health Evaluation System. This system is used to evaluate and classify patients treated in intensive care units and is designed to provide predictive information on morbidity and mortality that can be used in assessing both aberrant outcomes and unnecessary use of intensive care services. A series of HCFA-funded projects, two of which are near completion, should provide further insight into the use of patient classification systems for assessing the quality of medical care. Both SysteMetrics and the Commission on Professional and Hospital Activities (CPHA) have developed patient classification systems that adjust for disease severity as it might relate to hospital mortality. These studies are similar in that they use categorical data analysis techniques to determine risk adjustment factors associated with hospital mortality and use the conceptual logic of the standardized mortality ratio to assess mortality outcomes. However, they use different data bases and have different definitions of hospital-attributable mortality. The SysteMetrics study involves the construction of models to estimate the effects of age, sex, admission diagnoses, disease severity, and the presence of high-risk comorbidities (additional diseases or conditions) on a 30-day postadmission mortality rate. Disease severity is measured by a technique called “disease staging," which groups clinically-similar diagnostic codes into diagnostic clusters. These clusters are then assigned a severity-of-illness score. The clusters are constructed so that each patient may be assigned mutually exclusive severity scores for principal diagnoses and unrelated comorbidities. Logistic regression models are used to estimate an equation showing the relationship of mortality to the previously described variables. The equations and the coefficients of the variables were estimated from data contained in the 1984 MEDPAR file. (See appendix V.) Average values for these variables in the 1985 Medicare provider analysis and review file were used in the equations to determine risk-adjusted expected mortality counts. Quality HHS Quality of Care Measurement and of care is then assessed by comparing the expected mortality counts with the observed mortality counts for individual providers. The Severity Adjusted Mortality Index was developed as part of a larger CPHA project to develop indexes of hospital efficiency and quality. The conceptual framework underlying the construction of this index is similar to the one employed by SysteMetrics in the development of their mortality index. Like the SysteMetrics study, this mortality index uses logistic regression to adjust for patient characteristics, such as principal diagnosis, severity and number of comorbid conditions, age, poverty status, presence of cancer, and race. The index uses the DRG system as a starting point by collapsing into single clusters DRG categories containing similar underlying clinical conditions but differing in terms of age and presence of comorbidities. Within DRG clusters, risk estimates are made by using logistic regression to determine the effects of the variables identified above on the risk of patient mortality. These estimates are derived from data on 6 million patient records contained in the 1983 Professional Activities Survey file maintained by CPHA. Specifically, applying to the survey file data the coefficients of the variables estimated by the logistic regression equation allows CPHA to derive expected mortality counts for each DRG cluster. Quality of care can then be assessed by comparing the observed mortality counts with the expected mortality counts. Other HHS work is addressing issues of measurement in posthospital care settings. A study at Duke University will apply a methodology known as grade of membership analysis and life table analysis methods to estimating PPS impacts on the use of skilled nursing facilities and home health services by members of the Medicare population. The integrated use of these two methods represents a methodologically sophisticated attempt to estimate time-dependent, risk-homogeneous probabilities of postacute care service utilization. Grade of membership techniques will be used to identify subpopulations who are differentiated by their probability of having a set of related medical conditions and physical impairments. By using maximum likelihood estimation techniques inherent in the grade of membership approach, the researchers will be able to identify a number of subgroups that exhibit distinct sets of transition probabilities regarding the use of post-acute health services. These group-related probabilities will then be applied to a Medicare population in the framework of a life table to derive time-specific probabilities of service utilization. Estimates of how individuals with particular characteristics might be likely to use post-acute services will then be calculated from these time-specific transition probabilities. HHS Quality of Care Measurement and This project represents an innovative attempt to use multiple data sources; that is, Medicare part A data, Medicaid data, National Long Term Care Surveys, and demonstration project data, to derive probability estimates of post-acute service utilization for an elderly population. It will provide information on PPS impacts on changes in the timing of, selection, and the intensity of service utilization by members of the Medicare population. The Secretary asked that I respond to your request for the We appreciate the opportunity to comment on this draft report Sincerely yours, В Кижин Richard P. Kusserow Enclosure |