Three paediatric cardiologists assessed nearly 1000 imprecise subjectiveconditional probabilities for a simple belief network representing congenitalheart disease, and the quality of the assessments has been measured usingprospective data on 200 babies. Quality has been assessed by a Brier scoringrule, which decomposes into terms measuring lack of discrimination andreliability. The results are displayed for each of 27 diseases and 24questions, and generally the assessments are reliable although there was atendency for the probabilities to be too extreme. The imprecision allows thejudgements to be converted to implicit samples, and by combining with theobserved data the probabilities naturally adapt with experience. This appearsto be a practical procedure even for reasonably large expert systems.
展开▼
