Predicting atherosclerotic cardiovascular disease risk in retired civil aviation employees using the triglyceride–glucose index and machine learning models: a retrospective cross sectional study with explainable analysis

Abstract Background: Atherosclerotic cardiovascular disease (ASCVD) is the leading cause of death and disability worldwide. Retired civil aviation employees carry substantial accumulated cardiometabolic risk, yet ASCVD risk prediction is understudied in this group. The triglyceride– glucose (TyG) index is a reliable surrogate for insulin resistance, but its predictive value in this population and its incremental benefit over the simple combination of triglyceride (TG) and glucose (GLU) remain unverified. This study evaluated the TyG index for ASCVD risk prediction in retired civil aviation employees and compared machine-learning models. Methods: A retrospective cross-sectional study was conducted using physical examination data of 1,047 retired civil aviation employees from Civil Aviation General Hospital (January 2023– June 2025). Participants were categorized into high-risk and non-high-risk ASCVD groups. Significant variables were screened by univariate logistic regression (P<0.05) and multivariable analysis after variance inflation factor (VIF) based collinearity exclusion. Three models were built—basic, basic + TyG, and basic + TG + GLU—and the incremental value of TyG was assessed by AUC, the DeLong test, the likelihood-ratio test, NRI, and IDI. Six machine-learning algorithms were used to build predictive models, evaluated by 5-fold cross-validation and an independent validation set, and the best model was interpreted with SHapley Additive exPlanations (SHAP). Results: A total of 458 participants were in the high-risk group and 589 in the non-high-risk group. Univariate regression identified 17 significant variables (all VIF <3). In multivariable regression, the TyG model achieved an AUC of 0.8786, significantly higher than the basic model (0.8034) and the TG + GLU model (0.8636; DeLong test P<0.001). Among the six algorithms, binary logistic regression yielded the highest validation-set AUC (0.8495, 95% CI 0.8080–0.8898) and the most stable cross-validation performance (5-fold CV AUC = 0.8788 ± 0.0124). SHAP analysis identified the TyG index as the most important feature (mean |SHAP| = 1.1143), with higher values corresponding to greater ASCVD risk. Conclusions: The TyG index is a core predictor of ASCVD risk in retired civil aviation employees and outperforms the simple combination of TG and GLU. Binary logistic regression showed robust performance with good discrimination and interpretability, making it suitable for clinical application.