Fairness over Time: Evaluating How Fairness Objectives Shape Health Disparities in AI-Driven Clinical Decision Support Systems

Abstract

Fairness is not a single concept and may be operationalized using a variety of definitions reflecting different ethical commitments, yet these are rarely compared systematically in healthcare AI. Fairness is often assessed using static metrics, even though clinical decisions are sequential and evolve over time. In chronic disease management, small inequities early in a care pathway may accumulate into long-term disparities. We develop a Markov cohort microsimulation model to examine how different fairness operationalizations align with or diverge from their intended goals. Using our simulation model, we compare five heuristic allocation policies based on Utilitarian, Egalitarian, Sufficientarian, Prioritarian, and Needs-Based principles against a Status Quo baseline. Outcomes are evaluated longitudinally using quality-adjusted life years, between-group inequality, threshold attainment, and cost-effectiveness. No policy performs best across all criteria; most achieve their intended objective while performing worse on metrics aligned with other frameworks, though the policies with the strongest local equity commitments, Egalitarian and Sufficientarian, notably underperform even on their own aligned metrics. These same two policies also produce the largest cumulative disparities over time. We show that this may come from a structural access mechanism, where locally fair allocation rules do not offset the access disparity that determines whether resources are actually received. These findings show that the choice of fairness definition can have consequential and sometimes counterintuitive effects, reinforcing the need for explicit, longitudinal fairness evaluation.