Algorithms for Health Care Quality Management and Outcomes Assessment

American health care is in the midst of momentous change. Quality improvement principles are being used increasingly to enhance productivity and efficiency of health care delivery and to help contain costs. In producing patient outcomes and resource utilization measures, it is vital to control appropriately for case mix (i.e., differences in patients due, e.g., to illness severity) in the predictive model. If the severity of illness among patients can be accounted for accurately, differences in patient outcomes--mortality, resource utilization, morbidity, and patient function (both at discharge and over the long term)--will reflect, in large measure, differences in the quality-of-care received. With an accurate patient outcomes model, trained and validated using high-integrity hospital patient records, control for illness severity is explicit: the predictive model captures the relationship between observables and outcomes. Quality outcomes differing significantly from the expected values, as predicted by the model, reflect differences in quality of care. Static and dynamic polynomial neural networks (PNNs) will be used to synthesize accurate prediction of patient outcomes based on the most cost-effective patient observables. For categorical variables, PNNs are trained using a minimum logistic-loss fitting criterion; for continuous variables, a minimum squared- error criterion is used.