Autonomous artificial intelligence for diabetic eye disease increases access and health equity in underserved populations

Diabetic eye disease (DED) affects about one-third of people with diabetes and is a leading cause of blindness among working-age adults in developed countries. Because early-stage DED is often asymptomatic, professional guidelines from the American Academy of Ophthalmology and the American Diabetes Association recommend annual eye examinations for patients with diabetes to enable early detection and treatment. Despite this, adherence to annual DED testing has remained low: only 50–60% of Medicare beneficiaries with diabetes receive annual eye exams, and adherence can be even lower (30–40%) in smaller health systems and low-income metropolitan populations. Reported barriers include misinformation about the importance of screening, logistical challenges with scheduling, and anxiety about exams. In 2018, the FDA De Novo authorized an autonomous AI system (Luminetics Core, formerly IDx-DR) for diagnosing DED at the point of care, demonstrating 87.2% sensitivity and 90.7% specificity in a pivotal trial. Subsequent studies have further validated performance. This study evaluates whether deployment of autonomous AI within a large health system is associated with improved adherence to annual DED testing overall and across patient subgroups, with attention to access and health equity.

Prior U.S. studies report suboptimal adherence to annual eye examinations among patients with diabetes: 50–60% among Medicare beneficiaries and 30–40% in smaller systems and low-income metropolitan populations. Qualitative research identifies misconceptions about screening, logistical barriers, and anxiety as key drivers of non-attendance. Autonomous AI for DED (Luminetics Core/IDx-DR) was de novo authorized by the FDA in 2018 and showed strong diagnostic performance (sensitivity 87.2%, specificity 90.7%) against outcome-based standards. Subsequent validations corroborated accuracy, sensitivity, and specificity in diverse settings. The literature also documents disparities in eye care use, with Black/African American populations utilizing services at lower rates than White patients despite higher risk for diabetic retinopathy, and broader disparities across language, insurance, and socioeconomic indicators. These findings motivate evaluating autonomous AI as a potential tool to improve screening adherence and reduce inequities.

Design: Retrospective observational study approved by the Johns Hopkins School of Medicine IRB, adhering to the Declaration of Helsinki. Setting and population: Adult patients with diabetes mellitus managed at Johns Hopkins Health System primary care sites during calendar years 2019 (pre-AI deployment) and 2021 (post-AI deployment). Data sources and variables: Electronic health records provided demographics and social determinants, including gender, age, race, ethnicity, preferred language, insurance status, ZIP code-based geography (metropolitan, micropolitan, small town, rural per 2010 RUCA codes), national Area Deprivation Index (ADI; 1–100, higher indicates greater socioeconomic disadvantage), and inflation-adjusted median household income. Exposure: Clinic sites categorized as “AI-switched” if they adopted the autonomous AI system (Luminetics Core, formerly IDx-DR) by 2021 but not in 2019; “non-AI” sites never adopted and referred patients externally for DED testing. Outcome: Adherence to annual DED testing. Statistical analysis: Baseline characteristics were summarized as counts and percentages; chi-squared tests compared AI-switched vs non-AI within years (Table 1). Primary analysis used a difference-in-differences framework with inverse-probability-weighted regression adjustment to estimate average treatment effect among four groups (non-AI 2019, AI-switched 2019, non-AI 2021, AI-switched 2021). Covariates in propensity weighting and adjustment included age, gender, race, ethnicity, preferred language, insurance, ADI quartile, family income, and geographic region. Covariate balance was checked post-weighting. Robust standard errors clustered on participants accounted for repeated measures (patients present in both years). Subgroup analyses of adherence changes (2019–2021) used Poisson regression with a generalized estimating equations framework to estimate percentage-point changes with 95% confidence intervals (Table 2). Software: SAS for descriptive statistics and chi-squared tests; R version 4.3.1 and Stata 17.0 for propensity score weighting and outcome modeling. Statistical significance threshold p < 0.05.

- Sample: Approximately 17,000 adult patients with diabetes across Johns Hopkins Medicine primary care sites. - Propensity score weighting analysis: • AI-switched sites: adherence increased by 4.7 percentage points from 2019 to 2021 (95% CI: 2.5 to 9.1; p < 0.001). • Non-AI sites: adherence decreased by 0.3 percentage points (95% CI: −1.5 to 0.9; p = 0.663). • Difference-in-differences: AI-switched sites experienced a 7.6 percentage point higher increase vs non-AI (95% CI: 5.2 to 10.1; p < 0.001). - Overall adherence levels: • 2019 overall adherence: 42.2% (baseline 46.1% at AI-switched; 40.4% at non-AI). • 2021 overall adherence increased to approximately 48% (54.5% at AI-switched; 40.1% at non-AI). - Equity-related outcomes: • Black/African American adherence increased by about 12% at AI-switched sites (e.g., +12.2 percentage points in some analyses) but decreased slightly at non-AI sites (−0.6 percentage points). • Native Hawaiian or Other Pacific Islander adherence at AI-switched sites increased by about 19%, while decreasing by ~1% at non-AI sites. • Racial gap between Asian Americans and Black/African Americans shrank markedly after AI deployment; the gap declined to 3.5% by 2021, with reports indicating a reduction from about 15.6% in 2019 to 3.5% in 2021. • Insurance coverage disparities narrowed: the gap between military insurance and Medicaid diminished substantially after AI implementation. • Socioeconomic disparities improved: patients in higher ADI quartiles showed notable gains in adherence at AI-switched sites. - Subgroups with minimal gains at AI-switched sites included Asian (+0.4%), American Indian or Alaska Native (+2.3%), non-English speakers (+2.8%), and those with military insurance (+2.8%).

The study addressed whether deploying autonomous AI for DED screening in primary care increases adherence to annual eye examinations and whether benefits extend across diverse patient groups. Using a difference-in-differences approach with propensity score weighting to mitigate confounding, AI-switched sites exhibited significantly greater improvements in adherence than non-AI sites, despite the study period spanning the COVID-19 pandemic, which generally suppressed preventive care utilization. The magnitude and consistency of effects across analyses suggest that embedding autonomous AI at point-of-care can reduce logistical barriers and streamline access to retinal evaluation, thereby improving overall adherence. Importantly, improvements were most pronounced among historically disadvantaged groups, including Black/African American patients, Medicaid-covered patients, and those residing in higher ADI areas, indicating enhanced health equity. Narrowing gaps across race, insurance status, and socioeconomic strata supports the hypothesis that autonomous AI can serve as an equalizing force when integrated into primary care workflows. Nevertheless, heterogeneous responses across subgroups (e.g., minimal gains among Asian patients, non-English speakers, and military-insured patients) point to persistent barriers beyond access, such as differential trust in AI, cultural/linguistic factors, or system-level nuances. These findings imply that, alongside AI deployment, targeted patient engagement and education, culturally tailored communication, and careful implementation strategies are needed to maximize equitable benefits. The improvements at AI-switched sites also suggest potential positive impacts on quality metrics (HEDIS, CMS MIPS) and payer reimbursement if scaled system-wide.

Real-world deployment of an autonomous AI system for diabetic eye disease screening in primary care was associated with higher adherence to annual testing and improved access, particularly among historically disadvantaged populations, leading to reduced disparities across race, insurance type, and socioeconomic status. These results indicate that autonomous AI can be an effective tool to improve health equity within integrated healthcare systems. Future work should include prospective randomized trials to establish causality, qualitative studies to understand patient perceptions and trust in AI, evaluations of follow-up rates after positive AI findings, and assessments of downstream clinical outcomes and healthcare costs. Broad implementation coupled with targeted outreach could further close remaining adherence gaps and enhance population-level eye health outcomes.

- Retrospective observational design limits causal inference despite use of difference-in-differences and propensity score weighting. - Study population predominantly metropolitan, potentially limiting generalizability to rural settings. - Potential unmeasured confounders and treatment heterogeneity not captured in EHR data. - The analysis spans the COVID-19 era, which may introduce time-varying system-level effects on preventive care. - Some subgroup results showed limited improvement, suggesting unresolved barriers (e.g., trust, language, cultural factors). - Data and code availability are by request, limiting external reproducibility.