Machine Learning Classification of Prevalent Chronic Disease Using Multidimensional Social, Behavioral, and Psychological Determinants: A Cross-Sectional Analysis of the 2024 National Health Interview Survey.
Abstract
Background
Chronic diseases account for approximately 90% of the $4.5 trillion annual healthcare expenditure in the United States. While traditional clinical risk factors have been extensively studied, the predictive utility of multidimensional social determinants of health (SDOH), including psychosocial factors such as loneliness, psychological distress, and social support, remains inadequately characterized within machine learning (ML) prediction frameworks.
Objective
To develop and compare ML models that classify the presence of six major chronic conditions-hypertension, type 2 diabetes, coronary heart disease, COPD, depression, and anxiety-using an integrated framework encompassing demographic, socioeconomic, behavioral, psychosocial, healthcare access, functional status, and COVID-19-related predictors; and to quantify the relative predictive importance of each domain across disease categories.
Methods
We conducted a cross-sectional analysis of 32,614 adults from the 2024 NHIS. Twenty-four predictor variables spanning seven SDOH domains were used to train three ML algorithms: Logistic Regression (LR), Gradient Boosting Machine (GBM), and Random Forest (RF). Model performance was evaluated using 5-fold stratified cross-validation with AUROC, F1 score, and AUPRC. Subgroup analyses were performed by age, sex, and race/ethnicity. Because the design is cross-sectional, the models estimate the likelihood of prevalent disease, not incident risk; robustness was confirmed with alternative imputation (KNN and random-forest MICE), hyperparameter optimization, and three importance methods (Gini, permutation, and SHAP).
Results
GBM achieved the highest AUROC for 5 of 6 outcomes, ranging from 0.793 (anxiety) to 0.857 (COPD). For cardiometabolic outcomes, age and BMI were dominant predictors (hypertension: age importance = 0.641, BMI = 0.117). In contrast, psychological distress (K6) and loneliness emerged as the top predictors for depression (importance = 0.285 and 0.234) and anxiety (0.245 and 0.161). Psychosocial factors collectively contributed 52% and 42% of predictive importance for depression and anxiety, but less than 3% for cardiometabolic diseases. Subgroup analyses showed consistent performance across demographic strata.
Conclusions
Integrating psychosocial determinants substantially enhances prediction of mental health outcomes but contributes minimally to cardiometabolic disease prediction. These findings support disease-specific screening and case-identification strategies that appropriately weight social, behavioral, and psychological dimensions.