Development and validation of an interpretable predictive model for mismatch between endoscopic ultrasonography and histopathological diagnosis in colorectal subepithelial lesions: a retrospective single-center study.

Background: Endoscopic ultrasonography (EUS) is important for evaluating colorectal subepithelial lesions (SELs), but EUS-histopathology mismatch may complicate preoperative management. Evidence for predicting mismatch risk in colorectal SELs remains limited. Methods: We retrospectively included patients with colorectal SELs who underwent EUS and had histopathological confirmation. Candidate predictors were selected using LASSO, and seven machine-learning algorithms were developed and compared. Performance was assessed by discrimination, calibration, decision curve analysis, and repeated stratified 10-fold cross-validation. SHAP was used for model interpretation, and the final model was deployed as a web-based calculator. Results: Among 301 SELs, 222 were concordant and 79 showed mismatch (26.2%). LASSO selected lesion location, echogenicity, and EUS layer of origin. The SVM model achieved the best overall test-set performance, with an AUC of 0.818 (95% CI, 0.673-0.964), accuracy of 0.883, sensitivity of 0.538, specificity of 0.979, PPV of 0.875, NPV of 0.885, F1-score of 0.667, and Brier score of 0.109. Repeated cross-validation yielded an SVM AUC of 0.713 (95% interval, 0.537-0.884). SHAP identified lesion location as the strongest contributor to mismatch risk, followed by echogenicity and layer of origin. Conclusion: An interpretable SVM model based on routinely available EUS features may help identify colorectal SELs at higher risk of EUS-histopathology mismatch and support post-EUS risk alerting. Prospective multicenter validation is required before broader clinical implementation.