Explainable and Bias-Aware AI Models for Clinical Decision Support in U.S.Healthcare Systems

Le résumé fourni par la source

Although AI-enabled clinical decision support systems (CDSS) are becoming more prevalent in U.S. healthcare, inequities and opaque models pose a threat to patient safety and clinician trust. This scoping review mapped evidence on explainable and bias-aware clinical AI systems to inform their equitable deployment. Following the PRISMA-ScR guidelines and a PCC framework, we searched MEDLINE/PubMed and Embase for English peer-reviewed studies published between 2015 and 2025. The full texts of eligible studies were charted across eight domains, including data modality, CDSS use case, model approach, documentation bias, explanation technique, implications for trust and outcomes and mitigation and governance actions. Of the 464 records identified, 18 studies met the inclusion criteria. The evidence spanned imaging (predominantly chest radiography), EHR-based risk prediction and emergency department operational and safety models. The evidence was largely retrospective in nature. Explainability was most defensible when used as a safety audit to support reviewable rationales and detect shortcut learning, typically via feature attribution, perturbation tests, or visualisation. Bias reflected demographic signal leakage, temporal or label leakage, proxy targets, subgroup error disparities, and site confounding factors, which can inflate apparent performance. The most common mitigation strategies combined reweighting or fairness-aware selection, data augmentation, and setting-specific recalibration, but post-deployment monitoring was inconsistently reported. A trustworthy CDSS requires explicit equity objectives, multi-site evaluation, standardised documentation, and continuous surveillance for drift and emergent disparities.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.