Mortality prediction with adjuvant tamoxifen in breast cancer: Machine learning-integrated explainable artificial intelligence and Bayesian model results

Background and purpose: Tamoxifen is a cornerstone of adjuvant endocrine therapy for breast cancer, yet significant inter-individual variability in treatment response and mortality exists. Identifying robust predictors of outcomes remains a critical need. This study integrated machine learning, explainable artificial intelligence (XAI) and Bayesian modelling to predict mortality and identify key prognostic factors in breast cancer patients receiving adjuvant tamoxifen. Experimental approach: We analysed data from 568 patients from the International Tamoxifen Pharmacogenomics Consortium database. The outcome was all-cause mortality, with predictors including age, race, menopausal status, tumour size, estrogen receptor status, radiation treatment, and CYP2D6 metabolizer status. Four algorithms, logistic regression, random forest, eXtreme Gradient Boosting (XGBoost) and support vector machine, were developed and validated. Model performance was assessed using accuracy and area under the receiver operating characteristic curve (AUC). SHapley Additive exPlanations (SHAP) analysis provided interpretability for the XGBoost model, and Bayesian logistic regression with weakly informative priors was employed for probabilistic inference. Key results: The overall mortality rate was 19.4 %. XGBoost demonstrated the highest discriminative ability (AUC 0.833; 95 % confidence interval: 0.725 to 0.941), while random forest exhibited superior sensitivity for identifying deceased patients (83.3 %). SHAP analysis revealed that white race, increased age, absence of radiation treatment, larger tumour size and the CYP2D6 poor metabolizer (PM/PM) genotype were associated with elevated mortality risk, whereas the extensive metabolizer (EM/EM) genotype was protective. Significant variability was observed in exploratory subgroup analyses, with the model achieving excellent discrimination in patients without radiation treatment (AUC 0.901) and those with the EM/PM genotype (AUC 0.956) but failing to identify any mortality events in the Caucasian subgroup. Bayesian logistic regression yielded comparable performance to frequentist methods (AUC 0.820), with tumour size emerging as a consistently strong predictor in partial dependence plots. Conclusion: Integrating machine learning with XAI and Bayesian approaches effectively identified key predictors of mortality in tamoxifen-treated breast cancer patients. However, marked heterogeneity in model performance across subgroups highlights the critical need for external validation and careful evaluation of algorithmic fairness before clinical implementation.