This project describes Dr. Stukenborg's plan to develop improved public domain hospital mortality prediction methods for use with administrative data, in collaboration with faculty sponsors who have substantial expertise in health care outcomes research. The research is organized in six stages: In stage one, California hospital data from 1996 to 1997 will be used to define study populations for patients in five disease categories using both narrow and broad criteria. The secondary diagnoses reported for patients will be individually evaluated to determine the proportion coded as 'present on admission.' All secondary diagnoses reported as 'present on admission' for 90% or more patients will be reviewed by an expert physician panel to identify secondary diagnoses that are appropriate for use as indicators of co-morbidities or as indicators of the severity of the primary diagnosis. Diagnoses that are present on admission 90% of the time and considered appropriate for use as indicators by the physician panel will be organized into categories of co- morbid illness or into categories of primary diagnosis severity. In stage two, multivariable models will be developed to determine the relationship between inpatient mortality and all patient characteristics with the potential for use as predictor variables. In stage three, the developed mortality risk adjustment models will be validated by using them to forecast inpatient mortality in each of the five disease groups in two different data sets: first to California hospital data from 1998 and then to the Nationwide Inpatient Sample of the Healthcare Cost and Utilization Project (HCUP) data from 1993 to 1997. In stage four, the performance of the developed mortality risk adjustment models will be compared to that of two existing methods of adjusting for the influence of comorbid illness: the Deyo/Charlson method and the Elixhauser method. In stage five, the level of construct validity achieved by the developed model will be compared to that of existing methods. In stage six, results from risk adjustment models developed in the narrowly defined study populations will be compared with models developed in the broadly defined study populations. This research could substantially improve our ability to evaluate the quality of care, by allowing more accurate estimates of the effects of therapy and of the effects of health care programs at the population level.
