Subjective breast cancer grading - Analyses of reproducibility after application of Bayesian belief networks

OBJECTIVE: To examine the influence of Bayesian belief networks (BBNs) on the reproducibility of subjective breast cancer grading. STUDY DESIGN: Twenty samples were analyzed for intraobserver and 128 samples for interobserver reproducibility using the Bloom-Richardson and Helpap grading systems. The expression of diagnostic features was evaluated subjectively, and for each a decision it was determined to what extent it represented one of the different outcomes. Evidence was then entered, for each diagnostic feature, into four different BBNs, recently described for breast cancer grading, in the form of a relative likelihood ratio vector. RESULTS: With all cases considered, the use of decision support based on the Bloom-Richardson and Helpap grading systems did not improve intraobserver reproducibility. This was found to be 68% and 80% in subjective gradings, respectively, and 60% and 70% in the BBN-supported method. Interobserver reproducibility was not improved (58% and 70% in subjective gradings and 51-59% based on assessment with decision support). However, when only cases associated with high beliefs were considered, both intraobserver reproducibility (agreement rose from 68% to 93%) and interobserver reproducibility (agreement rose from 60% to 87%) of BBN-supported gradings exceeded the results of subjective assessments. CONCLUSION: The results showed that the observers did not reach the same diagnosis (or grade) and that their observational assessment of histologic features lacked agreement. Since BBNs reflected only the data entered, poor agreement existed in the contribution to the final diagnostic belief by the different features and, ultimately, in belief in the final decision.