Bayesian Hierarchical Modeling Framework for Breast Cancer Treatment Outcome Prediction: Integrating Clinical, Pathological, And Treatment Variables

Abstract

Breast cancer represents the most prevalent malignancy among women globally, with approximately 2.3 million new cases annually. In Kenya, it disproportionately affects younger women (35-50 years) and represents the leading cancer diagnosis. Current prediction models inadequately quantify uncertainty in treatment responses, leading to suboptimal clinical decision-making. This study developed a Bayesian hierarchical modeling framework to predict pathological complete response (pCR) in breast cancer patients by systematically integrating clinical, pathological, and treatment variables. We conducted a retrospective analysis of 5,400 patients across 12 Kenyan treatment centers using Bayesian logistic regression with random effects to model hierarchical data structure. The framework incorporated tumor stage, molecular markers (hormone receptor status, HER2), histological grade, patient demographics, and treatment protocols. Markov Chain Monte Carlo (MCMC) methods estimated posterior distributions with multiple imputation addressing missing data. The developed model demonstratedsuperior predictive accuracy (AUC = 0.837) compared to classical approaches, with significant effects identified for tumor stage (Stage IV OR: 3.19, 95% CrI: 1.89-4.54), hormone receptor status (OR: 0.31, 95% CrI: 0.15-0.66), and HER2 positivity (OR: 2.33, 95% CrI: 1.08-4.78). Treatment center heterogeneity accounted for 12.5% of outcome variability. This framework provides the first population-specific Bayesian approach for subSaharan African breast cancer prediction, enabling personalized treatment planning and improved clinical decision-making in resource-constrained settings.