Machine Learning-Based Risk Prediction Models for Peripheral Artery Disease: A Nationally Representative Analysis of NHANES 1999-2004.

BackgroundWe evaluated the association between small dense low-density lipoprotein (sdLDL) and peripheral artery disease (PAD) and developed prediction models using NHANES 1999-2004.MethodsThis weighted cross-sectional study included 3,415 participants (3,227 without PAD and 188 with PAD). PAD was defined by ankle-brachial index (ABI) <0.90; participants with ABI ≥1.40 or missing ABI were excluded. Survey-weighted logistic regression assessed the association between sdLDL and PAD. For machine learning, complete-case data (n=2,828) were split into training (70%) and test (30%) sets before feature selection and preprocessing. Correlation screening, LASSO, encoding, standardization, and SMOTE were restricted to the training set. Nine algorithms were evaluated using discrimination and calibration metrics.ResultssdLDL was not independently associated with PAD in the fully adjusted model (OR=1.01, 95% CI 0.94-1.10, P=0.69). LASSO retained eight predictor domains: age, body mass index, C-reactive protein, education, estimated glomerular filtration rate, hypertension, poverty-income ratio, and smoking. In the test set, gradient boosting machine achieved the highest AUC (0.833, 95% CI 0.777-0.889), followed by logistic regression (0.819) and XGBoost (0.811). SHAP analysis highlighted age, eGFR, and smoking-related variables as important contributors.Conclusions: sdLDL was not independently associated with PAD. Several models showed acceptable discrimination, with GBM performing numerically best and logistic regression remaining competitive. External validation in contemporary cohorts is required.