Update on the Pathfinder Project

We review highlights of our research on Pathfinder, a decision-theoretic expert system for hematopathology diagnosis. We have developed techniques for efficiently acquiring, representing, and reasoning with uncertain biomedical knowledge. Specifically, we have developed a methodology for coping with complex dependencies among findings and disease in pathology. The methodology includes an extension of the belief-network representation called similarity networks. Using this methodology, we have constructed a large probabilistic knowledge base for the domain of lymph-node pathology. We have also developed techniques for improving the clarity of explanations through the use of human-oriented abstractions. Finally, we have conducted a formal evaluation of Pathfinder’s diagnostic accuracy.