Data-Driven Approaches to Reducing Health Disparities in Underserved United States Communities: A Critical Review of Predictive Analytics and Decision-support Models

Abstract

Predictive analytics and clinical decision-support models are now embedded in routine care delivery across United States health systems, and they are increasingly promoted as instruments for narrowing longstanding inequities affecting low-income, minoritised, rural and publicly insured populations. The empirical record supporting that promise is uneven. Demonstrations that widely deployed algorithms encode and amplify disadvantage have accumulated rapidly, whereas evidence that model-guided programmes measurably reduce disparities in health outcomes remains sparse and methodologically fragile. This critical narrative review examines how predictive and decision-support models interact with the structural, informational and organisational conditions that produce disparities in underserved United States communities, and evaluates what can defensibly be concluded from the present literature. Sources were identified through searching of open scholarly databases and indexes, citation tracking and examination of authoritative federal materials, with critical appraisal focused on target definition, data provenance, subgroup evaluation, deployment context and outcome measurement. Four analytical claims emerge. First, the most consequential equity failures originate in the specification of the prediction target and in the measurement infrastructure, rather than in model architecture. Second, area-based social risk proxies and individual-level social risk data are not interchangeable, and the substitution of one for the other reshapes who is identified as high risk. Third, statistical fairness adjustments cannot resolve conflicts that arise from finite intervention capacity, and subgroup net benefit provides a more decision-relevant frame than parity in error rates. Fourth, the empirical chain from model output to changed outcomes is broken at the intervention stage far more often than at the prediction stage, and safety-net settings face the steepest implementation constraints. Confidence in any claim that predictive analytics reduces disparities remains low, and the field requires prospective, capacity-explicit, outcome-anchored evaluation rather than further retrospective demonstrations of bias.