Automated Medical Diagnosis based on Decision Theory and Learning from Cases

Medical diagnosis is an important but complicated task whose automation would be very useful. A scarce resource could be made less so and diagnosis could be made more accurate and efficient in both monetary terms and in reduced suffering or pain for a patient. We have used decision and probability theory to construct such systems from a database of typical cases. This simplifies the task of knowledgeextraction from physicians, who often do not know how they have come to a certain diagnosis. Probability models are constructed using mixture models that are efficiently learned by the Expectation-Maximization algorithm. Problems with missing data are then solved, both for missingdata in thecase database and during diagnosis (missing data are then observations not yet made). Decision theory is used to find the most informative next question to ask, observation to make, or test to d o in order to minimize the total cost for the diagnosis. It is also used todecide when to stop requesting more information. To automatically find good utility values for the decision theoretic model, temporaldifference reinforcement learning is used to increase the system's accuracy. Results are presented on a case database for heart disease.