Comprehensive survival analysis of breast cancer patients: a bayesian network approach.

Background: Breast cancer is recognized as one of the leading causes of cancer-related deaths globally. A deeper understanding of the complex interactions between clinical, pathological, and treatment-related factors is essential for improving patient outcomes.

Methods: Following comprehensive data cleaning and preprocessing, an analysis was performed on a cohort of 1,980 primary breast cancer samples from the METABRIC database. The dataset was divided into a 75/25 training-testing split, and five-fold cross-validation was applied to the training set to mitigate overfitting. Overall and relapse-free survival were then modeled using four fully parametric distributions: Weibull, Exponential, Log-Normal, and Log-Logistic, along with their corresponding Accelerated Failure Time (AFT) forms, to identify significant prognostic features. Competing models were ranked by the Akaike Information Criterion (AIC) and further validated through Quantile-Quantile (QQ) plots. Finally, the probabilistic relationships among the significant factors selected by the optimal AFT models were explored using a Bayesian Belief Network (BBN), whose structure was learned from the training data using multiple score-based algorithms and refined through expert-driven judgment; all conditional probability parameters were estimated using maximum likelihood.

Results: The Weibull model provided the best fit for overall survival, whereas the Log-Normal form was optimal for relapse-free survival, each satisfying their respective error-distribution diagnostics. In the hold-out test set, the Bayesian network achieved an Area Under the Curve (AUC) of 0.880 and an F1-score of 0.779. Age at diagnosis, menopausal status, tumor stage, lymph-node burden, and treatment modality were identified as the most influential predictors, and the learned network clarified their direct and mediated effects on both survival endpoints.

Conclusion: Through the integration of validated parametric survival models with a data-driven BBN, this study delivers a comprehensive framework for estimating individualized survival probabilities and visualizing the complex probabilistic relationships that characterize high-risk cancer patient profiles. This approach supports evidence-based, personalized breast cancer management and demonstrates the potential for guiding clinical decision-making and adapting to diverse external patient cohorts.