OPTIMIZATION OF GENERALIZED SEQUENTIAL DECISION SCHEMES FOR MEDICAL DIAGNOSIS

An improved ability to diagnose arthritic patients on the basis of observed features has been achieved using a novel statistical approach and numerical analysis. This improvement is tested against the diagnoses of both experienced physicians and established computational schemes. Approaches are investigated to derive and then optimize sequential Bayesian decision schemes for medical diagnosis. The assumptions often made with the classical Bayesian approach include (1) the disorders under consideration are mutually exclusive and exhaustive and (2) the features are conditionally independent under each disorder. These assumptions are seldom valid for practical applications.; Development in this dissertation of the Generalized Sequential Decision Scheme (GSDS) extends the classical Bayesian approach and requires assumptions that are less restrictive. By considering explicitly the a priori most likely multiple disorders in the patient population, GSDS can be used to calculate directly their a posteriori probabilities. The disorders need only be assumed to be almost exhaustive. The assumption of conditional independence under each disorder is relaxed by grouping into clusters the set of features relevant for a given disorder. Clustering features is difficult, however, because of the large number of conditional probabilities that must be estimated. Both the concept of irrelevant features and the calculation of consistency bounds on joint probabilities (within feature clusters) are used to reduce the required number of probability estimates.; One stage of the sequential decision process involves observing a feature or diagnosing the patient. Optimizing the sequential decision process based on dynamic programming is impractical because of computational limitations. Computation is reduced by assuming some form of conditional independence for the features. However, the development of limited-look-ahead approximations to the dynamic programming solution almost always is necessary. To effect the limited-look-ahead approximation, the sequential decision process is viewed as having a tree structure. A tree searching technique, called the gamma procedure, is efficient because it avoids searching some branches whose values will not affect decisions at higher levels of the tree.; Implementing GSDS with two-stage-ahead optimization has been demonstrated for diagnoses involving fifteen arthritic disorders and 172 features. Simulation results show that assuming conditional independence between feature clusters provides a small improvement in performance over assuming conditional independence under each disorder. Even such a small improvement is significant, however, because the present scarcity of data makes defining feature clusters difficult. Computation time for the two-stage-ahead strategy is 21 times greater than myopic optimization. However, this increased computation is more than compensated for by the improved diagnostic results and the reduced total expected cost of classification. Under the assumptions of GSDS, the total expected cost for classification of patients was calculated using the actual choices by physicians of laboratory tests. Costs were much less when the computer made the decisions, while the number of correct diagnoses remained unchanged.