Development and validation of an interpretable machine learning model for predicting 5-year recurrence in breast cancer.

BACKGROUND

Accurate prediction of breast cancer recurrence is vital for optimizing adjuvant therapy intensity. However, the traditional TNM staging system often lacks precision in capturing individual risk due to biological heterogeneity. This study aimed to develop an interpretable machine learning model to refine risk stratification.

METHODS

We retrospectively analyzed clinical data from 578 breast cancer patients with a median follow-up duration of 62 months. The cohort was randomly partitioned into a training set (n = 405, 190 recurrence events) and an internal validation set (n = 173, 72 recurrence events). An Extreme Gradient Boosting (XGBoost) model was developed using 15 clinicopathological features. Model interpretability was achieved using SHapley Additive exPlanations (SHAP), and a visual nomogram was constructed for clinical application.

RESULTS

In the internal validation cohort, the XGBoost model demonstrated superior discriminative performance with an AUC of 0.877 (95% CI: 0.835-0.918), significantly outperforming logistic regression (AUC = 0.693, 95% CI: 0.612-0.768) and the standard TNM system. SHAP analysis identified the Ki-67 index and positive lymph nodes as the most influential predictors, revealing non-linear risk associations. Crucially, the model successfully stratified patients within TNM Stages II and III into distinct high- and low-risk trajectories (Log-rank p < 0.05), addressing intra stage heterogeneity.

CONCLUSION

The proposed XGBoost-based framework provides a robust and interpretable tool for predicting 5-year recurrence, offering superior prognostic accuracy over standard anatomical staging. This approach holds promise for facilitating personalized clinical decision-making.