肥満修飾TyG指数を用いた胆嚢摘出術の予測精度向上：機械学習およびSHAPベースの研究

This study investigates the association between TyG-related composite indices and the risk of gallstones and the history of cholecystectomy, using logistic regression and machine learning models to evaluate predictive performance and clinical utility. Additionally, the study explores the relationship between TyG-derived obesity indices and the age at the history of cholecystectomy.

A total of 3737 participants were analyzed. Logistic regression models were used to assess the relationship between TyG, TyG.BMI, TyG.WC, and TyG.WHtR with gallstone prevalence and the history of cholecystectomy. Performance metrics for 11 machine learning models, including XGB, logistic regression (LR), and gradient boosting machine (GBM), were evaluated using AUC-ROC, accuracy, sensitivity, and specificity. Decision curve analysis (DCA), calibration plots, and SHAP (Shapley additive explanations) analysis were used to assess clinical utility and interpretability. Additionally, De Long test was applied to compare the AUC-ROC values of the machine learning models to identify statistically significant differences.

Among 3737 participants, 395 (10.6%) had gallstones. Individuals with gallstones were older (median 59 vs. 51 years, P < 0.01), predominantly female, and had higher levels of TyG and TyG-related indices (all P < 0.01). Logistic regression analyses revealed that while TyG was not significantly associated with gallstones after full adjustment, composite indices incorporating obesity measures-TyG.BMI, TyG.WC, and TyG.WHtR-remained robustly associated with gallstone risk in the fully adjusted model. Participants in the highest quartile (Q4) of these indices had higher odds of gallstones compared to those in the lowest quartile (Q1). Further analysis revealed that TyG.BMI, TyG.WC, and TyG.WHtR were associated with younger age at the history of cholecystectomy, with threshold effects identified at TyG-BMI = 184.35 and TyG-WC = 776.69, above which the association with younger cholecystectomy age became significant. In predicting the history of cholecystectomy, XGB outperformed other models with an AUC-ROC of 0.83, accuracy of 0.89, and F1-score of 0.73, showing balanced sensitivity (0.72) and specificity (0.82). The De Long test indicated that XGB demonstrated statistically significant superior performance compared to all other models (P < 0.01 for pairwise comparisons), reaffirming its high predictive capability. Supplementary Fig. 2 presents ROC curves for all models, where XGB achieved the highest AUC-ROC (0.827), outperforming other models such as LR (AUC-ROC = 0.746) and GBM (AUC-ROC = 0.742).

TyG-derived composite indices, particularly TyG.WHtR, are strong predictors of both gallstone prevalence and the history of cholecystectomy. The XGB model demonstrated the best performance in predicting cholecystectomy risk, with the De Long test confirming its superior AUC-ROC compared to other models. The combination of strong predictive performance, good calibration, and high interpretability makes XGB a valuable tool for clinical decision-making in managing gallbladder disease risk.