Impact of different training strategies on the accuracy of a Bayesian network for predicting hospital admission.

Introduction: The healthcare environment is constantly changing. Probabilistic clinical decision support systems need to recognize and incorporate the changing patterns and adjust the decision model to maintain high levels of accuracy.

Methods: Using data from >75,000 ED patients during a 19-month study period we examined the impact of various static and dynamic training strategies on a decision support system designed to predict hospital admission status for ED patients. Training durations ranged from 1 to 12 weeks. During the study period major institutional changes occurred that affected the system's performance level.

Results: The average area under the receiver operating characteristic curve was higher and more stable when longer training periods were used. The system showed higher accuracy when retrained an updated with more recent data as compared to static training period.

Discussion: To adjust for temporal trends the accuracy of decision support systems can benefit from longer training periods and retraining with more recent data.