An algorithm is presented for making diagnostic inferences on the basis of a causal network model of medical knowledge. The algorithm is based on Bayes Rule, but is unique in the way that it accounts for the presence of conditional non-independence of observations and for the presence of multiple diseases in the same patient. An evaluation of the system is performed on a database of patients with chest pain. In this evaluation, the diagnostic accuracy of the system is found to be inferior to that of a logistic regression model and comparable to that of a linear discriminant function. In a review of selected cases from this database, the system can be shown to provide inferences that are not possible with other simpler statistical models. The practicality of this and other computer aids to medical diagnosis is discussed.
Boosting backpropagation algorithm by stimulus-sampling: Application in computer-aided medical diagnosis
