The study explores the correlation of surrogate measures of insulin resistance (IR) with coronary artery disease (CAD) among patients with type 2 diabetes (T2D). Since direct measures of IR are inconvenient and impractical in clinical settings, the study investigates simpler substitutes, which are triglyceride-glucose index (TyG), TyG-body mass index (TyG-BMI), TyG-waist circumference (TyG-WC), TyG-waist-to-height ratio (TyG-WHtR), metabolic score for insulin resistance (METS-IR), and the ratio of triglycerides/high-density lipoprotein cholesterol (TG/HDL-C). Among 2,702 T2D patients, the patients were divided into groups based on CAD status.
Logistic regression, restricted cubic spline, and receiver operating characteristic (ROC) curves were applied to compare predictive performance of these measures. Machine learning algorithms, Random Forest and XGBoost, were also applied to enhance predictive accuracy.
The findings indicated that all surrogate IR indices correlated strongly with CAD and, importantly, had a U-shaped relationship for the TyG index in which low and high values were both associated with increased CAD risk. Among the indices, the TG/HDL-C ratio had the highest prediction for CAD with an area under the curve (AUC) of 0.721, 71% sensitivity, and 64% specificity. METS-IR also had a very high predictive value with an AUC of 0.713. TyG-derived indices such as TyG-WC had the highest specificity of 78%, with TyG-BMI and TyG-WHtR also having significant correlations with CAD risk. Machine learning algorithms were superior to single IR indices in prediction accuracy, with XGBoost having the highest AUC of 0.792, followed by Random Forest at 0.791.
The research concludes that surrogate IR indices are convenient and useful markers for the estimation of cardiometabolic risk in T2D patients. The ratio of TG/HDL-C was the strongest predictor and can be a useful marker for CAD screening. Machine learning models greatly improved prediction, and their potential to enhance cardiovascular risk estimation in clinical practice was demonstrated. The finding indicates that the combination of AI-based models with conventional diagnostic methods may result in more accurate and earlier CAD detection in T2D patients. Further longitudinal studies are needed to determine optimal risk thresholds and confirm these findings in larger populations.
Source: lipidworld.biomedcentral.com/articles/10.1186/s12944-025-02526-5
