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

Area of Science:

• Gastroenterology
• Medical Informatics
• Machine Learning

Background:

• Laryngopharyngeal reflux (LPR) diagnosis often requires invasive 24-hour pH-impedance monitoring.
• Identifying high-risk patients for LPR can optimize diagnostic resource allocation.

Purpose of the Study:

• To develop and validate an interpretable machine learning model for LPR risk stratification.
• To identify key predictors of LPR using clinical and endoscopic data.

Main Methods:

• A retrospective case-control study of 537 patients undergoing gastroscopy.
• Machine learning models were trained and validated, with performance assessed using metrics like F1 score and AUC.
• SHAP analysis and decision curve analysis were used for model interpretation and clinical utility assessment.

Main 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 achieved an F1 score of 0.725 and a Bootstrap-validated AUC of 0.815.
• An interpretable model with good discriminative ability and calibration was developed, along with a web-based risk calculator.

Conclusions:

• An interpretable machine learning model integrating endoscopic and clinical indicators can effectively screen patients with suspected LPR.
• The model aids clinicians in identifying high-risk individuals who may benefit from priority 24-hour MII-pH monitoring.
• This approach offers a valuable auxiliary tool for LPR diagnosis and management.