Abstract
Background: Unlike most cancers, breast cancer poses a persistent risk of distant recurrence-often years after initial treatment-making long-term risk stratification uniquely challenging. Current tools fall short in predicting late metastatic events, particularly for early-stage patients. Methods: We present an interpretable machine learning (ML) pipeline to predict distant recurrence-free survival at 5, 10, and 15 years, integrating Bayesian network-based causal feature selection, deep feed-forward neural network models (DNMs), and SHAP-based interpretation. Using electronic health record (EHR)-based clinical data from over 6000 patients, we first applied the Markov blanket and interactive risk factor learner (MBIL) to identify minimally sufficient predictor subsets. These were then used to train optimized DNM classifiers, with hyperparameters tuned via grid search and benchmarked against models from 10 traditional ML methods and models trained using all predictors. Results: Our best models achieved area under the curve (AUC) scores of 0.79, 0.83, and 0.89 for 5-, 10-, and 15-year predictions, respectively-substantially outperforming baselines. MBIL reduced input dimensionality by over 80% without sacrificing accuracy. Importantly, MBIL-selected features (e.g., nodal status, hormone receptor expression, tumor size) overlapped strongly with top SHAP contributors, reinforcing interpretability. Calibration plots further demonstrated close agreement between predicted probabilities and observed recurrence rates. The percentage performance improvement due to grid search ranged from 25.3% to 60%. Conclusions: This study demonstrates that combining causal selection, deep learning, and grid search improves prediction accuracy, transparency, and calibration for long-horizon breast cancer recurrence risk. The proposed framework is well-positioned for clinical use, especially to guide long-term follow-up and therapy decisions in early-stage patients.