Transferability of medical decision support systems based on Bayesian classification.

This study tested the hypothesis that probabilities derived from a large, geographically distant data base of stroke patients could form the basis of an accurate Bayesian decision support system for locally predicting the etiology of strokes. Performance of this "extrainstitutional" system on 100 cases was assessed retrospectively, both by error rate and using a new linear accuracy coefficient. This approach to patient classification was found to be surprisingly accurate when compared to classification by physicians and to Bayesian classification based on "low cost" local and subjective probabilities. We conclude that for some medical problems Bayesian classification systems may be significantly more transferable to new sites than is generally believed. Furthermore, this study provides strong support for the utility of clinical databases in building, transferring, and testing Bayesian classification systems in general.