OBJECTIVE: To develop and test a Bayesian belief network (BBN) for the identification of prostatic adenocarcinomas (PACs) with combination endocrine therapy (CET) changes from PACs with poor to no treatment CET effect and from untreated PACs. METHODS: A network was designed with a decision node containing three diagnostic alternatives (PAC with CET effect, PAC with poor to no treatment effect, and untreated PAC) and seven first-level evidence nodes for the diagnostic features: nuclear enlargement; frequency of prominent nucleoli; cell cytoplasm vacuolization; shrunken acini; individual infiltrating tumor cells; WHO prostate cancer pattern recognition, and amount of interstitial tissue stroma. Three prototype cases, one for each diagnostic alternative, were used to develop the BBN. The BBN performance was then evaluated in 40 prostatectomies for PAC, consisting of 20 CET treated and 20 untreated cases. RESULTS: The results obtained with the three prototypes showed that the network can identify the diagnostic alternatives with certainty when seven features are polled. When the performance was evaluated in the 40 PACs, the belief values were 1.0 or close to it in most of the cases (the value range is 0.0-1.0; the closer to 1.0, the greater the belief). Moreover, the BBN allowed an identification with high certainty of PACs with treatment-related changes from those either with poor to no treatment effect or untreated. CONCLUSIONS: A BBN for the evaluation of androgen-deprived PAC offers a descriptive classifier which is readily implemented and allows the use of descriptive, linguistic terms.
OBJECTIVE: To develop and test a Bayesian belief network (BBN) for the identification of prostatic adenocarcinomas (PACs) with combination endocrine therapy (CET) changes from PACs with poor to no treatment CET effect and from untreated PACs. METHODS: A network was designed with a decision node containing three diagnostic alternatives (PAC with CET effect, PAC with poor to no treatment effect, and untreated PAC) and seven first-level evidence nodes for the diagnostic features: nuclear enlargement; frequency of prominent nucleoli; cell cytoplasm vacuolization; shrunken acini; individual infiltrating tumor cells; WHO prostate cancer pattern recognition, and amount of interstitial tissue stroma. Three prototype cases, one for each diagnostic alternative, were used to develop the BBN. The BBN performance was then evaluated in 40 prostatectomies for PAC, consisting of 20 CET treated and 20 untreated cases. RESULTS: The results obtained with the three prototypes showed that the network can identify the diagnostic alternatives with certainty when seven features are polled. When the performance was evaluated in the 40 PACs, the belief values were 1.0 or close to it in most of the cases (the value range is 0.0-1.0; the closer to 1.0, the greater the belief). Moreover, the BBN allowed an identification with high certainty of PACs with treatment-related changes from those either with poor to no treatment effect or untreated. CONCLUSIONS: A BBN for the evaluation of androgen-deprived PAC offers a descriptive classifier which is readily implemented and allows the use of descriptive, linguistic terms.