Abstract
BACKGROUND: The prevalence of Peripheral Artery Disease (PAD) is rising globally, yet early risk stratification remains challenging due to the limitations of traditional obesity metrics. TyG-ABSI, an index combining Triglyceride-Glucose (TyG) with A Body Shape Index (ABSI), is a novel marker reflecting both functional insulin resistance and structural visceral adiposity. However, its predictive value for PAD remains unexplored in large prospective cohorts. METHODS: We included 390,274 adults from the UK Biobank. Baseline characteristics were analyzed across TyG-ABSI quartiles and PAD status. Associations between TyG-related indices and incident PAD were assessed using multivariable-adjusted Cox regression, Kaplan-Meier survival curves, and restricted cubic splines. Robustness was evaluated via Fine-Gray competing risk models, propensity score matching, subgroup analyses, and external validation in the NHANES database. Consensus k-means clustering, integrating biochemical and insulin resistance markers, identified metabolic phenotypes and stratified PAD risk. Feature selection (LASSO, Boruta, and Minimum Redundancy Maximum Relevance [mRMR]) guided the development of six machine learning models (logistic regression, GBM, XGBoost, AdaBoost, LightGBM, and neural network) for PAD prediction, with interpretability assessed via SHAP analysis. RESULTS: Higher TyG-ABSI and related indices were strongly associated with increased PAD incidence (cumulative incidence at 15 years: 4.16% in the top quartile vs. 0.98% in the bottom quartile; fully-adjusted Hazard Ratio [HR] per 1-SD increase for TyG-ABSI: 1.22, 95% Confidence Interval [CI] 1.17-1.27), which were robust in the NHANES external validation cohort. Clustering analysis revealed four distinct metabolic subgroups, with the highest PAD risk in the insulin resistance/glucose dysfunction cluster (HR vs. healthy phenotype: 7.48, 95% CI 6.82-8.21). Feature selection identified 19 key predictors. Logistic regression provided the most stable and generalizable prediction (validation Area Under the Curve [AUC] = 0.788, 95% CI 0.778-0.798), demonstrating superior generalizability compared to complex ensemble methods. SHAP analysis demonstrated TyG-ABSI, age, and neutrophil count as leading predictors for incident PAD and confirmed the interpretability of the model. CONCLUSION: TyG-ABSI is a robust, independent predictor of long-term PAD risk. Data-driven phenotyping and interpretable machine learning facilitate more precise risk stratification. Logistic regression offers optimal performance and interpretability, holding potential clinical utility for individualized PAD risk prediction.