A machine learning-based model for predicting 5-year conditional survival in patients with colon cancer: a development and validation study

Background Most prognostic models for colon cancer are not tailored to the Chinese population or implemented clinically. This study aimed to identify independent risk factors for five-year conditional survival and compare conventional logistic regression with machine learning models to develop an optimal predictive model. Methods This retrospective study included 430 newly diagnosed colon cancer patients from Suzhou Municipal Hospital (2015–2022). Clinicopathological data and dynamic tumor markers—carcinoembryonic antigen (CEA) and carbohydrate antigen 19-9 (CA19-9)—were collected. Patients were randomly split 70:30 into training (n=301) and validation (n=129) sets. This study employed a dual feature selection strategy combining the least absolute shrinkage and selection operator (LASSO) and Boruta algorithms. One conventional logistic regression model and six machine learning models (including XGBoost) were built. Performance was assessed by the area under the receiver operating characteristic curve (AUC), Brier score, and F1 score, with SHapley Additive exPlanations (SHAP) employed for model interpretability. Results Nine core predictors were selected: clinical stage, differentiation, distant metastasis, surgery, targeted therapy, and baseline and 12-month CEA and CA19–9 levels. The XGBoost model performed best, achieving an AUC of 0.955, the highest F1 score (0.904), and the lowest Brier score (0.085). SHAP analysis revealed that 12-month post-treatment CEA is the crucial predictor for survival prognosis. Conclusion The XGBoost and Random Forest models demonstrated superior predictive performance over conventional logistic regression. XGBoost was ultimately selected as the primary model due to its optimal balance of discrimination, calibration, and compatibility with the SHAP framework. SHAP provided quantitative, individualized risk interpretation, demonstrating the value of dynamic monitoring in disease assessment and management for colon cancer patients. Meanwhile, the SHAP interpretability framework was employed to provide interpretation of the risk factors and individualized risk prediction methods for the model, demonstrating the important value of dynamic monitoring in individualized disease assessment.