Interpretable Machine Learning for LPR Risk Estimation: A Single-Center Retrospective Case-Control Study.

Objective: Based on RSI and RFS scores, an interpretable machine learning model was constructed to identify risk factors for LPR, aiming to screen high-risk patients requiring 24-h pH-impedance monitoring and provide reference for clinical decision-making. Methods: A retrospective case-control study included 537 patients who underwent painless gastroscopy (June 2024-June 2025), split into training ( Results: Six independent LPR predictors were identified: arytenoid IPCL dilation, abdominal circumference, reflux esophagitis, alcohol consumption history, right lateral sleeping position, and GEFV grade III/IV. The random forest model performed best (F1 = 0.725, recall = 0.716, accuracy = 0.727, precision = 0.734), with Bootstrap-validated AUC of 0.815 (95% CI: 0.753-0.873). Calibration curve showed good fit, decision curve analysis confirmed clinical net benefit across thresholds. SHAP analysis ranked feature contributions as: arytenoid IPCL dilation, abdominal circumference, reflux esophagitis, alcohol consumption history, right lateral sleeping position, GEFV grade III/IV. A web-based calculator was deployed (URL: http://127.0.0.1:7292). Conclusion: This study constructed and validated an interpretable machine learning model integrating endoscopic and clinical indicators. The model demonstrates good discriminative ability and calibration and can serve as an auxiliary screening tool for patients with suspected LPR to help clinicians identify high-risk individuals who require priority 24-h MII-pH monitoring.