Machine-Learning-Based Prediction and Interpretation of Non-Erosive Reflux Disease Risk

Objective To develop a machine learning (ML) model for predicting non-erosive reflux disease (NERD) risk, interpret the optimal model using Shapley Additive Explanations (SHAP), and create an online prediction tool. Methods This single-center retrospective cohort study enrolled 556 patients undergoing sedated gastroscopy at Chengde Central Hospital (June 1, 2024–June 1, 2025). Stratified random sampling allocated participants to training (n = 390) and validation (n = 166) sets (7:3 ratio). Clinical characteristics were analyzed using LASSO regression with 10-fold cross-validation to identify predictors. Nine ML models were developed and compared: elastic net GLM, random forest, support vector machine, gradient boosting machine, XGBoost, artificial neural network, K-nearest neighbors, linear discriminant analysis, and elastic net regression. Performance was evaluated by F1-score, AUC, Brier score, recall, precision, and accuracy. Bootstrap resampling (1000 iterations) and calibration curves compared predictive efficacy, with the optimal model selected by highest calibrated AUC. Decision curve analysis (DCA) quantified clinical utility. SHAP interpreted the optimal model (via bar/summary plots), and an online calculator was deployed. Results LASSO identified five predictors: Dilation of capillary loops in the epithelial papillae of the arytenoid cartilage, waistline, non-exposed cardia glands, cardia polyps, and Hill grade III/IV gastro-oesophageal flap valve (GEFV). All models achieved AUCs > 0.770 in training and validation sets. After internal validation, random forest demonstrated optimal performance (validation set calibrated AUC: 0.805, 95% CI: 0.741–0.866). Brier scores were 0.178 (training) and 0.227 (validation). DCA confirmed net clinical benefit across 0.01–0.99 threshold probabilities. SHAP analysis ranked predictor contributions: Dilation of capillary loops in the epithelial papillae of the arytenoid cartilage, Waistline, non-exposed cardia glands, cardia polyps, Hill grade III/IV GEFV, all positively associated with NERD risk. The online calculator was validated locally. Conclusion : Five key NERD predictors were identified. The SHAP-interpretable random forest model demonstrates robust performance and clinical utility. The deployed calculator may enable early prevention, personalized management, and targeted interventions for NERD.