An explainable multimodal model integrating post-ESD mucosal defect images and clinical features to predict delayed bleeding after colorectal endoscopic submucosal dissection

Existing tools for predicting delayed bleeding (DB) after colorectal endoscopic submucosal dissection (ESD) use only structured clinical variables, leaving information in post-ESD mucosal defect images unused. We developed and externally validated an explainable multimodal model integrating these images with clinical features. We retrospectively included 3,416 patients who underwent colorectal ESD at two hospitals from January 2022 through June 2026 (development cohort, n = 2,778; external validation cohort, n = 638). DB was active bleeding 6 h to 28 days after ESD, confirmed by repeat endoscopy and requiring endoscopic hemostasis. YOLOv11m-seg segmented mucosal defects, ResNet50 generated an image-derived prediction score (E-Score), and the E-Score was fused with clinical features. After LASSO feature selection, five algorithms with and without SMOTE were evaluated by fivefold stratified cross-validation. SHAP and Grad-CAM provided post-hoc feature- and image-level explanations. DB occurred in 224 patients (8.06%) in the development cohort and 53 (8.31%) in the external validation cohort ( P = 0.839). The segmentation model achieved mAP50 and mAP50-95 values of 0.984 and 0.838. Six features were retained. XGBoost without SMOTE (PrismDB) performed best (area under the receiver operating characteristic curve [AUC], 0.804; 95% confidence interval [CI], 0.769–0.837), with 64.73% sensitivity, 85.40% specificity, and 96.50% negative predictive value. In external validation, the AUC was 0.788 (95% CI, 0.724–0.850), sensitivity was 79.25%, specificity was 83.25%, and negative predictive value was 97.79%. Probability calibration in the development cohort under the natural event distribution showed a calibration slope of 1.101 and an intercept of + 0.227 (Brier score, 0.0584). E-Score had the highest mean absolute SHAP contribution, and representative activation maps overlapped with exposed vessels, adherent clots, and eschar. PrismDB showed stable discrimination for DB after colorectal ESD. Low-risk classification did not exclude DB. Probability calibration was acceptable in the development cohort but has not been evaluated externally, and the safety of model-guided changes to postoperative care remains to be established; the companion web application demonstrates the prediction workflow.