Survival prediction in colorectal cancer liver metastases using machine learning with SHAP-based interpretation.

Background: Colorectal cancer liver metastasis (CRLM) remains a leading cause of cancer-related mortality worldwide. Accurate prognostic stratification is crucial for making individualized therapeutic decisions. Conventional statistical approaches are limited in capturing complex nonlinear interactions among multidimensional clinical variables. This study aimed to develop, temporally validate, and deploy an interpretable machine learning (ML) model incorporating Traditional Chinese Medicine (TCM) intervention to predict long-term survival in patients with CRLM. Methods: A retrospective cohort of 861 CRLM patients was included following institutional ethical approval. Clinical, pathological, and treatment-related variables, including TCM exposure characteristics, were systematically collected. After data preprocessing and feature selection, six machine learning algorithms-Random Forest (RF), XGBoost, K-Nearest Neighbors (KNN), Support Vector Machine (SVM), LightGBM, and CatBoost-were trained using five-fold cross-validation to predict 36- and 60-month overall survival. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), precision-recall curve (PRC), and confusion matrix metrics. The best-performing model was further validated in an temporal dataset to assess generalizability. Model interpretability was enhanced using SHapley Additive exPlanations (SHAP), and the final optimized model was deployed as a web-based clinical application to facilitate individualized survival prediction and real-time risk stratification. Results: Among candidate models, the optimized XGBoost algorithm demonstrated superior predictive performance. For 36-month survival prediction, the AUC reached 0.891 in the training cohort and 0.833 in the testing cohort, with consistent performance for 60-month survival prediction. Temporal validation confirmed model robustness and stability. SHAP analysis revealed that TNM stage, liver metastasis burden, and TCM intervention intensity were among the most influential prognostic factors. TCM exposure exhibited a protective association with survival probability in a dose-dependent pattern. The web-based tool enables clinicians to input individual patient parameters and obtain dynamic risk estimates with transparent, interpretable outputs. Conclusions: We developed and temporal validated an interpretable ML-based prognostic model for CRLM and successfully translated it into a web-based clinical decision-support tool. By integrating TCM intervention into predictive modeling, this study provides quantitative evidence supporting its potential survival benefit. The deployed model offers a practical and accessible instrument for personalized prognostic assessment and optimized treatment planning for CRLM patients.