Machine Learning-Based Predictive Modeling of Factors Associated with Low HDL-C Levels: Insights from a Large-Scale Cohort Study

Abstract Introduction Our study applies machine learning methods to identify determinants of low–high-density lipoprotein cholesterol (HDL-C) in northeastern Iran. Clarifying these risk factors may support earlier diagnosis and treatment of cardiovascular disease (CVD) and inform timely prevention strategies. Methods This analytic cross-sectional study used baseline data from the Mashhad Stroke and Heart Atherosclerotic Disorder (MASHAD) cohort to develop predictive models of factors associated with low HDL-C. Participants were stratified into two groups based on HDL-C cut-off values: 40 mg/dL for men and 50 mg/dL for women. Our objective was to construct and evaluate predictive models to identify key factors associated with low HDL-C using Logistic Regression (LR), Decision Tree (DT), and Bootstrap Forest (BF). Results Among the 7526 participants assessed, 4842 (64.3%) were identified with low HDL-C levels. Logistic regression analysis demonstrated that physical activity level (PAL) was the most influential determinant, followed by sex and hip circumference. In parallel, the Bootstrap Forest model underscored mid-upper arm circumference and demi-span as the principal predictors of HDL-C status. Conclusion PAL, sex, hip circumference, mid-upper arm circumference, and demi-span emerged as potential predictors of HDL-C levels. Moreover, DT and BF models demonstrated robust capabilities in constructing predictive models for HDL-C-related factors.