Document Type : Original Article
Authors
1
School of Management and Accounting, Allameh Tabatabai University, Tehran, Iran
2
Department of Marketing and Supply Chain, School of Business and Economics, Maastricht University, Maastricht, The Netherland
3
Department of Public Administration, School of Management and Accounting, Allameh Tabatabai University, Tehran, Iran
10.22060/ajmc.2026.24183.1398
Abstract
Health disparities during epidemics expose structural inequities in health systems: during COVID-19, marginalized groups experienced mortality more than 2.8 times that of privileged populations due to cumulative disadvantages spanning housing, occupational exposure, access to care, and social capital. While machine learning (ML) shows promise for mitigating such disparities, prevailing approaches rarely codesign fairness into the pipeline or engage stakeholders in ways that are viable for resource-constrained settings. Following PRISMA 2020 and PRISMA-AI, we conducted a systematic review of 76 peer-reviewed studies (2019–2024) to identify which established ML methods most reliably support equitable epidemic management. The review synthesized evidence on diagnostic and operational gains as well as recurring barriers—algorithmic bias, infrastructural defcits, and limited community participation—that impede equitable deployment. Guided by these fndings, we propose a three-layer integration framework—prediction, allocation, and governance—that embeds fairness at design time rather than as a post-hoc fx. To aid interdisciplinary readers, we defne key terms at frst use: equity checkpoints are formal decision points in the ML lifecycle at which group-wise error, beneft, and burden gaps are tested against pre-specifed thresholds; if thresholds are exceeded, the system triggers model reweighting, constraint tuning, or data collection updates before proceeding. The framework draws on statistical learning theory, convex optimization, and algorithmic fairness to preserve accuracy and equity guarantees within a single optimization procedure. We then validate the framework via simulation—not as an isolated modeling exercise, but as a direct test of the design principles distilled from the review. Across diverse scenarios, disparity metrics improve by 40–55Policy implications include mandatory algorithmic impact assessments with equity criteria, independent audits, and targeted digital-equity investments. Collectively, the review-informed framework and its validation demonstrate a practical path for deploying established ML methods that measurably reduce health disparities without sacrifcing system performance.
Keywords
Subjects
