Do AI Health Interventions Reduce Disparities? a Scoping Review Based on the Social Determinants of Health and Equity Frameworks

Background
Artificial intelligence (AI) is increasingly integrated into health decision-making, shaping access to diagnosis, treatment, and care services. These systems influence resource allocation and clinical prioritization, and they can reinforce inequities when social determinants of health (SDoH) are inconsistently documented or reflect structural constraints. Thus, as AI health interventions accelerate, there is a need to examine how often—and how well—equity is measured in real-world implementation. This scoping review assesses how AI-based health interventions operationalize inequity and whether equity is treated as a measurable outcome (e.g., subgroup gap reduction) versus a contextual rationale (e.g., deployment in disadvantaged settings).

Methods
Following Arksey and O’Malley’s (2005) framework and PRISMA-ScR guidance (Tricco et al., 2018), we applied a Population–Concept–Context approach. Seven databases were searched (January 2010–September 2025). Of 1,212 records, 260 duplicates were removed; 952 titles/abstracts were screened; 83 full texts were reviewed; and 25 studies met inclusion criteria. Eligible studies examined AI-enabled systems linked to actionable clinical or service-delivery decisions and explicitly engaged SDoH-related inequities. Eligible interventions included machine learning, deep learning, natural language processing (NLP), or multimodal approaches in which algorithmic outputs informed clinical, behavioral, or service decisions and were designed, deployed, or interpreted in relation to inequities. Findings were synthesized descriptively to identify patterns in how equity was defined and evaluated.

Results
Among 25 studies, only 5 (20%) directly quantified subgroup performance or disparity-relevant change; most framed equity through predictive performance in disadvantaged settings without testing whether disparities narrowed. Studies that conducted subgroup analyses demonstrated measurable effects. An imaging-based pain algorithm explained 43% of the racial pain gap and increased Black patients’ surgical eligibility from 11% to 22% (Pierson et al., 2021). A de-biased pulmonary embolism survival model reduced race-, ethnicity-, and gender-based performance gaps while improving overall discrimination (Zhong et al., 2024). Federated learning approaches improved diagnostic accuracy for under-resourced sites by up to 47%, narrowing inter-site disparities. In contrast, many studies reported overall performance metrics without explicit subgroup evaluation (e.g., pediatric mortality prediction across lower-income sites: AUC = 0.945 internal, 0.864 external; rural telehealth: 30% higher consultation accuracy and 32% higher follow-up). NLP models performed well overall, but were less accurate for domains such as housing and medication insecurity, suggesting uneven capture of structurally patterned needs. Across studies, documentation quality, representational choices, and generalizability were recurring mechanisms through which inequities were reproduced or mitigated.

Conclusions
Equity in AI-enabled health interventions is more often operationalized as expanded access or improved accuracy than evaluated as measurable reduction in group-based disparities. Advancing equity-centered AI requires standardized SDoH documentation, routine subgroup auditing, and evaluation frameworks that link algorithmic outputs to downstream decisions and inequity-relevant outcomes. For social work research, education, and policy, findings support (1) integrating AI literacy and equity auditing into professional training, (2) expanding interdisciplinary implementation research that tests whether AI tools reduce disparities in real-world service pathways, and (3) strengthening governance of AI procurement and deployment (transparency, subgroup reporting requirements, and monitoring of impacts on access and service quality for structurally disadvantaged populations).