AI for Health Equity: A Sociological Framework Linking Social Determinants of Health to Transformative U.S. Public Health Action

While artificial intelligence (AI) and machine learning are increasingly used in public health to forecast population health outcomes and inform resource allocation, little empirical work has examined how structural inequality shapes both health outcomes and the performance of the predictive systems built to model them. This study addresses that gap through a multi-method analysis of the structural determinants of preventable hospitalization in the United States, grounded in Fundamental Cause Theory and the structural racism perspective. Using the 2020 county file of the Agency for Healthcare Research and Quality (AHRQ) Community-Level Health Database, we model the county rate of preventable hospital stays per 100,000 Medicare enrollees and ask three questions: which social determinants of health (SDOH) predict county rates; whether rates and overdose mortality follow a socioeconomic gradient when states are grouped into poverty quartiles; and whether a random forest model trained on these data errs systematically across structurally disadvantaged groups of counties. County-level regression (N = 2,189; R² = 0.296) identifies poverty as the largest positive predictor, with uninsurance, PM2.5 exposure, drug overdose mortality, and percent Black population also positively associated with preventable hospitalization. States in the highest poverty quartile average 3,405.0 preventable stays per 100,000 Medicare enrollees, 44.7% more than states in the lowest quartile (2,352.8). Evaluated by five-fold cross-validation, the random forest explains 39.2% of out-of-fold variance (RMSE = 843.3, compared with an outcome standard deviation of 1,082.1). Its errors are not evenly distributed: the model overpredicts rates for counties in the three lower state poverty quartiles and underpredicts them for counties in highest-poverty-quartile states (mean signed error = −51.1 per 100,000), where its absolute error is also largest; this underprediction appeared in four of five folds and under five alternative random seeds. By contrast, counties with above-median Black population share were not underpredicted. The findings show that aggregate accuracy can conceal directional, socially patterned error, and they support auditing public health AI tools across multiple equity-relevant strata before use in resource allocation.