Abstract / Summary
Abstract Background Postoperative recurrence remains a major determinant of long-term survival among patients with colorectal cancer (CRC). Although TNM staging provides the foundation for prognostic stratification, considerable heterogeneity exists within each stage, and accessible perioperative indicators—including inflammatory and nutritional biomarkers—may offer additional prognostic value. Methods This retrospective study included patients who underwent radical colorectal cancer surgery between April 2018 and April 2022, with a temporal split assigning those treated before June 2021 to the training cohort and those treated thereafter to the temporal validation cohort. Perioperative clinical, pathological, and laboratory variables, including both continuous and categorical data, were collected for model development. Feature selection was performed using LASSO regression and Boruta algorithm, and six machine-learning algorithms and logistic regression were compared, with model performance evaluated in terms of discrimination, calibration, and clinical utility, SHAP was used for model interpretation. Results Six variables were consistently selected by LASSO regression and Boruta algorithm: tumor macroscopic types, AJCC stage, T stage, N stage, preoperative albumin (Pre-ALB), and postoperative day-3 white blood cell count (POD3 WBC). Among the evaluated prediction models, logistic regression demonstrated the best overall performance, achieving AUC values of 0.713 in internal cross-validation and 0.726 in temporal validation. SHAP analysis identified N stage and Pre-ALB as the strongest predictors of recurrence. This exploratory, proof-of-concept model showed modest-to-moderate discrimination in internal temporal validation (ROC AUC, 0.726; PR AUC, 0.596) and requires external validation and prospective clinical-impact assessment before clinical use. Conclusion This clinically interpretable logistic regression model, based on routinely available perioperative variables, demonstrated stable performance in temporal validation within a single institution. Its transparent structure supports clinical applicability, while further multicenter external validation is warranted to confirm generalizability across diverse populations.