Criticizing Conditional Probabilities in Belief Networks

Abstract In order to constmct aBayesian belief network for amedical domain, alarge numberofconditional proba- bilities mustbeobtained. Weinvwtigatedthefollow-ing issues regarding these probabilities: (1) Howac-curate are subjective probabilities provided byphysi- cians? (2) Howcan weuse imprecision in subjectiveprobabilitiestoouradvantage? (3) Howcantheprob-abilities be improved a weobserve newcases ofthe diseas being studied? (4) Howimportant are the probabilities, as compared with the actual structureofthe network? Weconducted prliminary exeriments in the do-main of congenitalheartdiseasetoaddresstheseques-tions. We foundthatcombiningphysician' subjective probabilities with datafromactualcases can improvepredictive ability, but it is likely that the success ofadiagnostic programbasedonbeliefnetworks is ulti- matelylimite bythestructure of the network. Introduction Many computer programs for medical diagnosis are based on Bayesian og (e.g., [5], [10]). Theseprograms tend to haveseveral