Multi-Dimensional Training Data Bias Detection and Fairness-Aware Augmentation for Equitable AI-Assisted Medical Imaging Diagnosis

Abstract

The rapid expansion of FDA-authorized AI-enabled medical imaging devices has revealed substantial performance disparities across demographic subgroups, with growing evidence indicating that the root cause lies in systematic bias embedded in training datasets rather than in downstream model architecture. Underrepresented minorities, older adults, and lower-income populations consistently appear at reduced frequencies in publicly available chest radiograph cohorts, and conventional debiasing approaches such as simple oversampling or instance reweighting fail to address bias dimensions beyond sample count imbalance. This study proposes a four-dimensional bias detection framework that quantifies sample representation deviation against population benchmarks, feature distribution skew across subgroups in deep embedding spaces, cross-subgroup annotation consistency disparity, and implicit associations between sensitive attributes and labels mediated through non-sensitive image features. A fairness-aware conditional generative augmentation pipeline complements the detection module by synthesizing high-quality samples for underrepresented subgroups under anatomical, radiomic, expert-validation, and anti-drift constraints. Causal mediation analysis identifies hidden indirect pathways through which sensitive attributes influence diagnostic labels. Experimental validation on CheXpert, MIMIC-CXR, and NIH ChestX-ray14 cohorts demonstrates that the integrated framework reduces the worst-group AUC gap from 0.087 to 0.024 while preserving overall diagnostic utility, identifying 23.7% more demographic bias pathways than single-dimension baselines. The framework provides actionable instruments for HHS health equity directives, FDA AI/ML training data representativeness expectations, and NIST AI Risk Management Framework compliance.