Machine-Learning Prediction of Bleeding After Endoscopic Submucosal Dissection for Early Gastric Cancer: A Multicenter Study

Abstract

Background

Endoscopic submucosal dissection (ESD) is a minimally invasive treatment for early gastric cancer (GC); however, post-ESD bleeding remains a serious and unpredictable complication. This study aimed to develop machine-learning (ML) models to predict post-ESD bleeding and identify associated risk factors, ensuring accurate and interpretable predictions.

Methods

A retrospective, multicenter clinical database was constructed for patients who underwent ESD for early GC. An ML model was developed using patient characteristics and perioperative findings to predict bleeding within 28 days post-ESD. Its performance was compared with that of a logistic regression–based non-ML model. Feature importance analysis was performed to aid interpretation.

Results

Among 1084 patients (median age: 75 years), post-ESD bleeding occurred in 63 (5.8%). The area under the curve of the ML model was better than that of the non-ML model (0.80 vs. 0.71, p = 0.03). Furthermore, the ML model demonstrated a trend toward higher sensitivity compared with the non-ML model (0.74 vs. 0.58, p = 0.58). When stratified by ML-estimated bleeding probability, observed bleeding rates were 2.3%, 8.8%, and 28.6% in the low- (< 50%), intermediate- (50%–80%), and high-probability (≥ 80%) groups, respectively. The odds of bleeding were significantly higher in the intermediate- (OR 4.03, p = 0.03) and high-probability (OR 16.7, p < 0.01) groups compared to the low-probability group. Anticoagulant use with atrial fibrillation emerged as a key predictor.

Conclusions

The ML model effectively rules out post-ESD bleeding and identifies clinically meaningful risk factors, supporting its use in personalized prophylactic strategies.