From classical survival statistics to explainable AI: A multi-method prognostic framework for triple negative breast cancer

ABSTRACT

Triple-negative breast cancer (TNBC) is a biologically heterogeneous subtype associated with variable survival outcomes, limiting the performance of conventional prognostic models. This study integrates classical survival analysis, machine-learning methods, and explainable artificial intelligence to identify key prognostic factors in a retrospective cohort of 869 histopathologically confirmed TNBC patients treated between 2018 and 2024. Kaplan–Meier analysis and Cox proportional hazards regression were used to evaluate survival differences and estimate hazard ratios, with assessment of proportional hazards assumptions using Schoenfeld residuals. Machine-learning models were developed for fixed-horizon survival prediction, and model interpretability was explored using SHAP and LIME analyses to characterize feature contributions. Chemotherapy was associated with improved overall survival in unadjusted analysis (log-rank P < 0.001), while radiotherapy showed no significant association after adjustment. Multivariable Cox analysis identified stage (HR =4.23, P < 0.001), absence of lymphovascular invasion (HR =0.15, P < 0.001), surgery (HR =0.77, P = 0.003), and chemotherapy-related variables as independent prognostic factors. Among machine-learning models, the artificial neural network showed the highest discrimination (C-index =0.98 ± 0.01), followed by random forest (0.92 ± 0.02). SHAP analyses consistently identified stage, tumor size, lymphovascular invasion, chemotherapy, and surgery as key contributors to model predictions. Overall, the combined framework demonstrates improved predictive performance and interpretability compared with traditional models; however, findings are based on internal validation and require external validation before clinical application.