Bayesian belief network for the Gleason patterns in prostatic adenocarcinoma: development of a diagnostic decision support system for educational purposes.

OBJECTIVE: To develop a Bayesian belief network (BBN) for Gleason grading of prostate adenocarcinoma. STUDY DESIGN: A shallow network was developed for Gleason grading with open-tree topology, with a root node containing 5 subjective diagnostic alternatives and 8 first-level descendant nodes for diagnostic features. Features or diagnostic clues of the descendant nodes were based on architecture of Gleason patterns. Data collected on 20 slides in the first and third slide circulations in the U.K.-based investigation of observer reproducibility of Gleason grading of prostatic biopsies were used. Circulations were called A and B. Level of agreement was studied using kappa statistics. RESULTS: Mean of percentage agreements between subjective Gleason major pattern attributed to slides by pathologists and subjective Gleason major pattern most frequently assigned to each slide was 85% in A and 88% in B. Mean of percentage agreements between BBN reading for slides read by pathologists and BBN reading most frequently seen in each slide was 77% in A and 70% in B. CONCLUSION: The BBN for Gleason grading is readily implemented, allowing use of linguistic variables and descriptive terms and accumulation of evidence presented by morphologic clues. This diagnostic decision support system has potential in pathology education.