Prediction of site-specific recurrence and overall survival for colorectal cancer with liver metastases

Up to 70% of patients with colorectal cancer liver metastases (CRLM) experience recurrence after curative-intent hepatectomy, yet recurrence patterns are heterogeneous across anatomic sites, and no tools currently exist to predict site-specific recurrence dynamically over time. We developed and validated machine learning models to predict overall survival and site-specific recurrence (liver, lung, lymph node, peritoneum) after hepatectomy for CRLM using baseline and dynamic prediction frameworks. In a retrospective cohort of 730 patients who underwent curative-intent hepatectomy for CRLM at the Johns Hopkins Hospital (2000–2024), with a median follow-up of 10.9 years, over 63% of patients experienced recurrence. XGBoost models were trained on clinical, pathologic, and molecular features using nested cross-validation with bootstrap optimism correction. Baseline models achieved C-indices of 0.54–0.67, while dynamic models incorporating surveillance covariates, including adjuvant therapy and interval recurrence at other sites, substantially improved discrimination (C-indices 0.67–0.78; time-dependent AUC 0.68–0.90). Calibration was confirmed with Brier scores of 0.03–0.19. These models enable risk stratification at the time of surgery to inform adjuvant therapy decisions and allow real-time updating of site-specific recurrence risk during follow-up, providing a framework for personalized surveillance strategies in patients with CRLM undergoing curative-intent resection.