Development of a Bayesian Classifier for Breast Cancer Risk Stratification: A Feasibility Study

Background: Lifetime risk assessment tools are relatively limited in identifying breast cancer risk in younger women. The predictive value of mathematical models to estimate risk varies according to age, menopausal status, race/ethnicity, and family history. Current risk prediction models estimate population, not individual, levels of breast cancer risk; hence, individualized risk prediction models are needed to identify younger at-risk women who could benefit from timely risk reduction interventions. Clinical data collected as part of breast cancer screening studies may be modeled using Bayesian classification. Purpose: To train a proof-of-concept Bayesian classifier for breast cancer risk stratification. Patients and Methods: We trained a Bayesian belief network (BBN) model on cohort data (including risk factors, demographic, electrical impedance scanning (EIS), breast imaging, and biopsy data) from a prospective pilot screening trial in younger women (N = 591). Receiver operating characteristic curve analysis and cross-validation of the model were used to derive preliminary guidance on the robustness of this approach and to gain insights into what a cross-validation exercise could provide in terms of risk stratification in a larger population. Results: Independent predictors of biopsy outcome in the BBN model included personal breast disease history, breast size, EIS (low vs high risk) and imaging results, and Gail cutoff (5-year risk: <1.66% vs ≥1.66%). Area under the receiver operating characteristic curve and positive predictive value for benign and malignant biopsy outcomes were 0.88 and 97% and 0.97 and 42%, respectively. Patient-specific probability of biopsy outcome given positive EIS result and Gail model 5-year risk ≥1.66% indicated that the combined effect of these predictors on likelihood that a biopsy would prove malignant exceeded the sum of the individual effects; breast cancer likelihood is as follows: 3% (EIS negative and Gail model 5-year risk <1.66%) versus 9% (EIS positive and Gail model 5-year risk <1.66%) versus 27% (EIS negative and Gail model 5-year risk ≥1.66%) versus 45% (EIS positive and Gail model 5-year risk ≥1.66%). Conclusion: Clinical data collected as part of breast cancer screening studies can be modeled using Bayesian classification. The BBN model may be predictive and may provide clinically useful incremental risk information for individualized breast cancer risk assessment in younger women.