From body composition to reflux esophagitis: an interpretable machine learning model based on CT-derived features

Background The prevalence of gastroesophageal reflux disease (GERD) has been increasing in China. Previous studies link sarcopenia and visceral adiposity to GERD, but most models lack CT-based body composition data. This study aims to improve the identification of Reflux esophagitis (RE) by applying machine learning (ML) to third lumbar vertebra cross-sectional CT (L3-CT) images for quantitative analysis of muscle and fat mass. Methods Participants underwent comprehensive abdominal CT and gastroscopy. Body composition parameters, including skeletal muscle mass (SM), total fat mass (FM), fat-free mass (FFM), visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and intermuscular adipose tissue (IMAT), were derived from the L3-CT images. Six ML models were developed: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Logistic Regression (LR), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), and Artificial Neural Network (ANN). Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). Shapley Additive exPlanations (SHAP) values were used to interpret the RF model. Results A total of 324 subjects were included, with 135 diagnosed with RE. The prevalence of low skeletal muscle mass index (low SMI) was significantly higher in the RE group compared to controls (52.6% vs. 36.5%, P 0.01). The top six variables in the RF importance matrix were IMAT, visceral-to-subcutaneous fat ratio (VSR), age, VAT, SAT, and hiatal hernia. In the validation set, RF (AUC = 0.829, 95% CI: 0.731–0.905) and LR (AUC = 0.829, 95% CI: 0.736–0.909) demonstrated the best discriminative performance for RE. SHAP summary plots illustrated the positive and negative contributions of the top 20 features, while SHAP dependence plots explained the impact of individual variables on the RF model output. Conclusion This study reveals a significant association between low SMI and RE. ML models identified key body composition factors, providing insights for targeted screening and clinical assessment.