Which of the five anthropometric and lipid-based indexes best predicts hypertension? A machine learning approach

Hypertension is associated with anthropometric and lipid-based indexes. This study aimed to examine and compare the predictive performance of these indexes, individually and in combination, using conventional statistical methods and modern machine learning approaches. A 10-year prospective cohort study was conducted using data from the Yazd Healthy Heart Project, involving 786 normotensive adults aged 20-74 years at baseline (2005-2006). Anthropometric, biochemical, and lifestyle data were collected, and incident hypertension was evaluated after follow-up. Associations between adiposity indexes and hypertension were examined using Cox proportional hazards regression and receiver operating characteristic (ROC) analysis. Machine learning models, including support vector machines, random forests, and logistic regression, were also employed for feature selection and predictive modeling. The lipid accumulation product (LAP) emerged as the strongest predictor, particularly in males, with the highest hazard ratio (HR: 4.53; 95% CI 2.63-7.82) and ROC area (AUC: 0.73). The triglyceride-glucose index (TyG) also showed predictive power, especially among females (HR: 2.05; 95% CI 1.05-4.01). XGBoost achieved high classification accuracy (79.14% in males), and feature selection identified LAP as the most influential variable. LAP appears to be the most effective anthropometric indicator for predicting hypertension, underscoring the role of visceral adiposity in cardiovascular risk. Incorporating machine learning models may further improve hypertension risk assessment and support their use in clinical settings.