The primary objective of this study is to evaluate the usefulness of neurocomputational techniques for extracting clinically meaningful information from massive health care databases. The secondary objective is to exploit the ability of neural networks to discern multivariate patterns (i.e., cluster meaningful local interactions among predictor variables) in order to select important predictors and detect heretofore unanticipated patterns of comorbidity, service utilization, and practice style. The third objective is to obtain practical health services results useful for concurrently evaluating and improving the practice of coronary artery bypass (CABG) surgery. Databases derived from patients undergoing CABG surgery will be the initial focus of this project. Artificial neural network processing will be compared to conventional statistical methods for defining risk-adjusted service utilization and outcomes of CABG surgery in the settings of chronic ischemic heart disease and acute myocardial infarction. Meaningful variation will be sought with respect to geography, institutional features, and practice styles. To accomplish these objectives, we pose the following specific aims: *DATA-To obtain and utilize electronic data, and cross-validate our findings using approximately 30,000 clinically-rich case records and an additional 200,000 administrative records abstracted from patients who underwent CABG in the United States. The sources of clinical records include: (1) the Medicare UCDS field collection phases I and II, and thereafter the ongoing UCDS national repository. (2) the PORT evaluating acute myocardial infarction, (3) the Dept. of veterans Affairs Health Services studies Operative Risk, Mortality. and Quality Assurance in Cardiac Surgery Study and Processes, Structures, and Outcomes of Care in Cardiac Surgery', (4) the Boston University/HCFA CABG database, (5) the Cleveland Clinic Foundation CABG database, and (6) the SMS national hospital archive. *PREDICTION-To compare neural network models (including advances in adaptive resonance theory mapping and backpropagation) to parametric and other established statistical techniques in the analysis of massive health care databases for determining severity-adjusted mortality, adverse events, length of stay, disposition, readmission, and functional outcome of CABG surgery. *PATTERNS-To compare neural networks to other methods, and published findings and recommendations for (a) defining geographic, inter- institutional, and sociodemographic variation in patterns of risk-adjusted resource utilization and outcome; and discovering emerging patterns of comorbidity, utilization, and practice style. *PUBLIC DOMAIN-To support neural network software dissemination and to draw upon external expertise by establishing a National Advisory Panel for Healthcare Database Neurocomputing"""""""".
