Interpretable machine learning for individualized survival prediction in node-positive, non-metastatic prostate cancer: a population-based study.

Background

Prostate cancer with regional lymph node involvement but no distant metastasis (N1M0) has heterogeneous prognosis. This study aimed to develop and validate an interpretable machine learning model for predicting cancer-specific survival (CSS).

Methods

Data from 18,287 N1M0 patients (2000-2022) in the SEER database were divided into training (n = 4780), internal testing (n = 1193), and two temporal validation cohorts (n = 3149; n = 9165). Cox proportional hazards and four machine learning models were compared using C-index, time-dependent AUC, and Integrated Brier Score. SHAP was used for interpretability, and IPTW for sensitivity analysis.

Results

Random Survival Forest (RSF) outperformed all models, achieving the highest C-index (0.697 internal; 0.740 temporal validation). RSF-stratified risk groups showed significant survival differences (P < 0.001). SHAP revealed radical prostatectomy as the strongest protective factor, followed by lower T stage and PSA. IPTW confirmed survival benefits of aggressive local control. CSS outperformed overall survival in discriminative accuracy.

Conclusion

The RSF model provides accurate, robust, and interpretable prognostication for N1M0 prostate cancer. Deployed as a web-based calculator, it enables precise risk stratification and individualized treatment planning.