Predicting post-ESD bleeding in elderly patients with early gastrointestinal cancer: an explainable model based on endoscopic features.

Background: Endoscopic submucosal dissection (ESD) has emerged as the treatment of choice for early gastrointestinal cancers due to its minimally invasive nature, short recovery period, and brief hospitalization. However, postprocedural bleeding remains a frequent complication that may adversely affect clinical outcomes, particularly in elderly patients. Accurate prediction of bleeding risk after ESD is therefore essential to guide timely preventive interventions, mitigate complication rates, and ultimately enhance patient prognosis. Objective: To develop and interpret a machine learning model based on endoscopic features for predicting post-endoscopic submucosal dissection bleeding in elderly patients with early gastrointestinal cancer, using SHAP (SHapley Additive exPlanations) to provide explainable insights. This model is intended to support clinicians in early risk stratification and to facilitate individualized perioperative management aimed at improving outcomes. Methods: We conducted a retrospective study of 236 elderly patients with early gastrointestinal cancer who were treated at the Department of Gastroenterology and Emergency of Lu'an Hospital Affiliated with Anhui Medical University between January 2022 and June 2025. Clinical data were systematically collected. Using stratified random sampling, patients were allocated to training and testing datasets in a 7:3 ratio. Feature selection was carried out via least absolute shrinkage and selection operator (LASSO) regression. Nine machine learning models were developed, and the model demonstrating the highest performance underwent SHAP (SHapley Additive exPlanations) analysis to enhance interpretability. Results: Among the 236 patients included in the analysis, 35 met the criteria for delayed post-ESD bleeding, resulting in an incidence of 14.8%. The remaining 201 patients did not experience delayed bleeding. LASSO regression identified seven variables associated with bleeding risk: large submucosal vessels, submucosal fibrosis, lesion diameter, lesion depth, procedure duration, use of anticoagulant medication, and hemoglobin level. Using these features, we constructed nine machine-learning models: K-nearest neighbors, support vector machine, extreme gradient boosting, approximate nearest neighbors, decision tree, light gradient boosting machine, random forest, extra trees, and gradient boosting machine. On the test set, the extra trees model performed best, with a sensitivity of 0.182, specificity of 0.983, F1-score of 0.286, and an area under the receiver-operating-characteristic curve of 0.879. SHAP analysis indicated that the features contributing most to the model's predictions were, in descending order: large submucosal vessels, lesion diameter, anticoagulant use, submucosal fibrosis, lesion depth, procedure duration, and hemoglobin level. Conclusion: We developed an interpretable Extra Trees model that effectively predicts the risk of post-ESD bleeding in elderly patients with early gastrointestinal cancer. This tool may aid clinicians in early identification of high-risk patients, thereby supporting targeted perioperative management and potentially improving clinical outcomes.