Machine Learning-Based Survival Time Prediction in Colorectal Cancer with Peritoneal Metastasis: A Multi-Institutional Registry-Based Study

Background: Prognostic prediction for colorectal cancer patients with peritoneal metastasis remains challenging due to clinical and biological heterogeneity. We aimed to evaluate the utility of machine learning, comparing high-performance boosting models with interpretable regression approaches for overall survival (OS) prediction. Methods: We analyzed a multi-institutional registry cohort of 150 colorectal cancer patients with synchronous peritoneal metastases. A total of 124 variables were included; continuous variables were standardized, categorical variables were one-hot encoded, and missing values were imputed using the median. Models included XGBoost, LightGBM, Ridge regression, Lasso regression, and linear regression. Training was performed with 3-fold cross-validation, and hyperparameters were optimized using Optuna. Evaluation metrics included mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²). Model interpretability was assessed using SHAP values, LIME local explanations, and regression coefficients. Results: Boosting models consistently outperformed linear models. XGBoost achieved the best performance (MAE 424, RMSE 526, R² = 0.04), while LightGBM showed comparable accuracy. In contrast, Ridge, Lasso, and Linear regression yielded high errors (MAE > 900, RMSE > 1200) with negative R² values, indicating poor predictive ability. SHAP analysis highlighted systemic inflammation markers (CRP, BUN), surgical assessment of tumor depth, operative factors (time, bleeding), and peritoneal metastasis characteristics as major determinants of OS. LIME analyses further provided case-specific interpretability, identifying feature contributions in long-, intermediate-, and short-term survivors. Conclusion: Boosting models, particularly XGBoost, demonstrated superior performance compared with traditional regression models in predicting OS for colorectal cancer patients with peritoneal metastasis, although absolute predictive accuracy remains modest. Integration of SHAP and LIME linked model outputs with clinically plausible prognostic factors, enhancing interpretability. Ensemble learning may provide a promising adjunct for prognostic assessment and should be validated in larger, genomically enriched cohorts.