Explainability, Auditability, and Knowledge Integration in Medical AI Using Structural MRI

AI Ethics, Fairness in Machine Learning, Responsible AI, Explainable AI, Machine ethics

This research aims to establish a medical AI framework for structural MRI that extends beyond predictive performance to incorporate explainability, reproducibility, auditability, and external generalizability. Using multi-site structural MRI data for attention-deficit/hyperactivity disorder (ADHD), the first stage systematically audits multiple saliency methods in terms of fidelity, stability, site robustness, and anatomical plausibility. The second stage integrates ROI-level evidence independently extracted from systematic reviews and meta-analyses by GPT, Gemini, and Claude. The extracted findings are validated, normalized to the Harvard–Oxford atlas, and aggregated into consensus literature priors for fusion with image-based deep learning models. The study further evaluates whether literature-informed priors improve prediction, calibration, explanation alignment, faithfulness, and site- or dataset-level generalization, while comparing them with negative-control priors such as uniform, permuted, reversed, random, and zero priors. Ultimately, this research seeks to reposition model explanations from post hoc visualizations to reproducible and auditable components of trustworthy medical AI and clinical decision-support systems.