Development and validation of EHR-based machine learning models for predicting survival and recurrence after radical surgery for colorectal cancer

Accurate prediction of early recurrence and survival in patients with colorectal cancer (CRC) is essential for personalized treatment. Conventional TNM staging has limited accuracy. We aimed to develop machine learning (ML) models for early recurrence and mortality using routine electronic health record (EHR) data. In a retrospective multicenter cohort of 5,136 stage I-III CRC patients after radical surgery, we compared three prognostic models for overall (OS) and disease-free survival (DFS): TNM stage only, multivariate Cox regression, and ML (gradient boosting survival analysis (GBSA) for survival, CatBoost classifier for recurrence risk). The dataset was split 80:20 into training ( n = 4,108) and split-sample validation ( n = 1,028) datasets via stratified sampling by event status and time intervals. Missing values were imputed or handled natively. TNM stage demonstrated acceptable prognostic performance, with good discrimination for both OS (tdAUC 0.71) and DFS (tdAUC 0.70). Cox regression showed only modest discrimination for survival outcomes, with split‑sample validation AUCs of 0.76 for OS and 0.69 for DFS. The GBSA model achieved split-sample validation tdAUC of 0.82 for OS and 0.80 for DFS. CatBoost achieved a split-sample validation AUC of 0.75 for recurrence. Machine learning models based on routine EHR data provided substantially better discrimination for both survival and early recurrence than TNM staging and traditional Cox regression in stage I–III colorectal cancer.