Abstract / Summary
Abstract Objective To develop an artificial intelligence (AI) model based on ResNet and Hill classification of gastroesophageal flap valve (GEFV) for endoscopic diagnosis of gastroesophageal reflux disease (GERD) and reduce the subjectivity of manual grading, and provide decision support for clinical GERD risk stratification and anti-reflux therapy selection. Methods In this single-center retrospective diagnostic accuracy study, 7,652 endoscopic images from patients with gastroesophageal reflux disease (GERD) were split into training, validation, and test sets at a patient-level 7:2:1 ratio. A ResNet-152 model was trained with class-weighted cross-entropy loss. Diagnostic performance was evaluated using accuracy, precision, recall, specificity, and F1-score. Correlation with mean lower esophageal sphincter (LES) pressure from high-resolution manometry (HRM) was assessed in an independent cohort of 92 patients. Results Inter-observer agreement among expert endoscopists yielded a weighted Kappa of 0.73. The final model achieved an overall test-set accuracy of 85.02%, with per-grade accuracies of 84.76%, 80.95%, 76.19%, and 72.38% for Hill grades I–IV, respectively. All global performance metrics exceeded 0.8. And the model significantly outperformed junior endoscopists on a random test subset ( P < 0.01). AI-predicted Hill grades showed a significant negative correlation with mean resting LES pressure (Spearman's ρ=−0.501, P < 0.001). Conclusions The ResNet-152 model demonstrated reliable performance for automated GEFV Hill classification and correlates significantly with objective manometric parameters. This AI tool has potential as a clinical decision-support system to standardize esophagogastric junction morphological assessment in GERD.