Algorithmic Bias in AI-Based Diagnostic Models: A Systematic Review and Lifecycle Framework for Fairness Auditing and Governance in Healthcare

Abstract

Background: Artificial-intelligence-based diagnostic systems increasingly influence healthcare delivery, yet strong aggregate performance can coexist with clinically meaningful disparities across demographic subgroups. Such disparities raise not only model-development concerns but also health-equity, implementation, and governance challenges for healthcare organizations. Methods: We conducted a systematic review reported in accordance with PRISMA 2020 of literature published from January 2019 through August 2026. The search identified 1842 records; 87 records were retained in the overall evidence corpus and 46 underwent structured quantitative extraction. The corpus included primary empirical/model-evaluation evidence together with secondary or adjacent methodological evidence used for qualitative contextualization. We also developed FRCS-10, a ten-item descriptive reporting-completeness index derived from established reporting and governance guidance, and applied it to the full corpus. Results: Peer-reviewed primary sources reported baseline true-positive-rate gaps of 0.062–0.281 across sex, age, and race on CheXpert and MIMIC-CXR, while a contextual dermatology meta-analysis reported AUROC values of 0.89 for Fitzpatrick I–III and 0.82 for Fitzpatrick IV–VI. These values were synthesized as source-reported estimates and were not independently re-analysed from patient-level data. Data representation was the most frequently coded bias source (38%), followed by algorithmic design (21%) and label or annotation bias (19%). The descriptive FRCS-10 mean was 4.6/10; only 7 of 87 records described post-deployment fairness monitoring. Five primary empirical studies with sufficiently explicit estimands were assessed for compatibility; their disparity definitions were not commensurable, so no pooled effect size was computed. Conclusions: We propose the Equity-Aware Bias Detection and Mitigation Framework (EA-BDMF), a literature-derived, implementation-oriented lifecycle specification for healthcare AI. It links data auditing, prevalence-aware preprocessing, fairness-constrained training, calibration and threshold governance, and post-deployment monitoring. The framework is accompanied by a Lifecycle Equity Debt (LED) accounting measure, checkpoint documentation, a Bias Provenance–Mitigation Traceability Matrix, and descriptive regulatory mapping. The complete framework has not yet been validated end to end, and its clinical and organizational utility requires prospective evaluation in real-world healthcare settings.