The Evaluation of Clinical Predictions

Abstract

Clinical predictions are never certain but are inherently probablisitc. The accuracy coefficient, a measure of probabilistic accuracy based on probability assigned to outcomes that occur, was used to assess the skill of clinical rheumatologists in predicting patient outcomes. Physicians' scores correlated well with degree of clinical experience. An approach to evaluation based on the measure provides a sensitive assessment of marginal benefit of technologies such as laboratory tests, diagnostic procedures or computer consultations. Most currently used methods of computer prediction were not as accurate as the best physicians tested. By allowing measurement of ability to individualize predictions to each patient's unique characteristics, the accuracy-coefficient approach has potential use in physician assessment.