Development of an Explainable Machine Learning Model for Predicting Reflux Esophagitis Among Candidates for Metabolic Bariatric Surgery.

BACKGROUND: Reflux esophagitis (RE) is a common gastroesophageal disorder among candidates for metabolic bariatric surgery (MBS). This study aimed to develop and internally validate an explainable machine learning (ML) model for predicting preoperative RE risk among patients undergoing MBS. METHODS: This retrospective study included 1,068 MBS candidates who were randomly divided into training and internal test cohorts (7:3). Feature selection was performed using correlation analysis, multicollinearity assessment, LASSO regression, and the Boruta algorithm. Six ML algorithms were developed and evaluated based on AUC, calibration, and classification metrics. SHAP analysis was used for model interpretation, and an online risk assessment tool was developed. RESULTS: Among the 1,068 patients, 280 (26.2%) were diagnosed with RE. Six predictors were ultimately selected, including Sex, Hp, BMI, TG, HGB, and HC. Among the six ML algorithms, the XGB model demonstrated the best overall predictive performance, achieving AUC values of 0.881 and 0.825 in the training and internal test cohorts, respectively. SHAP analysis identified TG, BMI, Hp, HC, HGB, and Sex as the major contributors to model predictions. The online tool developed based on the final XGB model enabled individualized RE risk prediction. CONCLUSIONS: We developed an explainable XGB model based on routinely available preoperative clinical variables to predict RE risk among MBS candidates. The model demonstrated acceptable predictive performance and interpretability and may assist in preoperative RE risk stratification and perioperative risk management. However, external validation in independent cohorts is required to further evaluate its robustness and generalizability.