Development and Temporal Validation of an Explainable Machine Learning Model for Identifying Colorectal Adenoma in Patients With Metabolic Syndrome: A Retrospective Study

Abstract Background Colorectal adenoma (CRA) is a precursor to colorectal cancer. Metabolic syndrome (MetS) is associated with an increased likelihood of colorectal neoplasia. We aimed to develop and temporally validate an explainable machine learning model to estimate the probability of CRA in patients with MetS. Methods Demographic, clinical, and laboratory data were retrospectively collected from patients with MetS who underwent colonoscopy between January 2024 and June 2026. Based on colonoscopic and histopathological findings, patients were classified into a CRA group and a normal colonoscopy group. Patients from January 2024 to December 2025 were assigned to the training set, whereas those from January to June 2026 formed the temporal test set. All preprocessing, feature selection, and model development were restricted to the training set. Mutual information was used for preliminary screening, followed by a grid-search-optimized GBDT and five-fold cross-validated permutation importance. Three base models were developed: Elastic Net Logistic Regression, XGBoost, and Extra Trees. Calibration and ensemble modeling yielded 11 candidate configurations. The final model and classification threshold were determined using five-fold cross-validation in the training set. Model performance was evaluated using AUC, sensitivity, specificity, Brier score, decision curve analysis, and SHAP. Results A total of 774 patients were included: 467 (60.3%) with CRA and 307 (39.7%) with normal colonoscopy findings. Two-stage feature selection identified 10 predictors. Prognostic nutritional index (PNI), hemoglobin, and age had the highest mean permutation importance. ExtraTrees_Isotonic was selected as the final model with a classification threshold of 0.5563. In the temporal test set, this model achieved an AUC of 0.811 (95% CI: 0.740–0.875), a sensitivity of 0.819, a specificity of 0.705, and a Brier score of 0.176. Decision curve analysis showed that the model provided greater net benefit than the treat-all and treat-none strategies across most evaluated thresholds. SHAP analysis showed that sex, history of constipation, and PNI contributed most to model predictions. Conclusions The ExtraTrees_Isotonic model showed good discrimination and potential clinical net benefit across a range of threshold probabilities. The model's generalizability and clinical utility require further validation in independent external cohorts and prospective studies.