EOCRCPred: An Artificial Intelligence–Based Prognostic Model Integrating Serum CEA and Clinical Factors for Early‐Onset Colorectal Cancer
Guangsheng Wan, Qingjie Wang, Chunwu Zhu, Liyu Wang, Yongbao Zhu, Meiying Li, Sheng Guo, Hongjie Yu, Luyun YuanPurpose
Early‐onset colorectal cancer (EOCRC) is increasing globally and exhibits aggressive biology, yet reliable biomarker‐guided prognostic tools for postoperative patients remain scarce. We aimed to develop and validate an artificial intelligence (AI)–based prognostic model—EOCRCPred—that integrates serum carcinoembryonic antigen (CEA) and routine clinical variables for postoperative risk stratification in EOCRC.
Patients and Methods
We analyzed 7408 Stage I–III EOCRC patients from the SEER database (training/test) and 250 patients from two Chinese hospitals (external validation). After LASSO–Cox feature selection, six survival models (RSF, CoxPH, S‐SVM, XGBSE, Gradient Boosting Survival Analysis [GBSA], and DeepSurv) were constructed. Model performance was assessed via C‐index, time‐dependent AUC, Brier score, calibration, and decision curves. SHAP analysis identified key prognostic drivers, and an online calculator was deployed for clinical use.
Results
The GBSA model showed the most consistent discrimination, with integrated C‐indices of 0.774, 0.749, and 0.741 across training, test, and external validation cohorts, respectively, and mean AUCs of 0.812, 0.783, and 0.779. Calibration and decision curves confirmed its accuracy and net clinical benefit. SHAP analysis revealed serum CEA as the top prognostic driver, followed by chemotherapy, marital status, perineural invasion, and household income. The EOCRCPred risk stratification system effectively separated patients into low‐, intermediate‐, and high‐risk groups (all log‐rank
p
< 0.001). The web‐based tool (
Conclusions
EOCRCPred is an externally validated AI‐based prognostic model that integrates serum CEA with clinical parameters for postoperative survival prediction and risk stratification in EOCRC patients. The model showed consistent predictive performance across the training, test, and external validation cohorts and may have potential utility in personalized clinical management.