In a study exploring how an AI-assisted diagnostic tool shaped care for underserved populations at multiple community-based primary care sites, investigators at the Wilmer Eye Institute, Johns Hopkins Medicine found that African American patients with diabetes were more likely to receive a diabetic eye exam referral if screened by an AI tool.
The final, exploratory peer-reviewed findings were published April 13 in npj Digital Medicine.
The researchers say their results potentially identify one way the integration of AI-assisted tools could address known healthcare disparities for people with diabetes.
Diabetic retinopathy — the most common diabetes-associated eye disease — is the leading cause of blindness globally. People may not experience symptoms early on, making annual diabetic eye exams essential for timely diagnosis and treatment.
Led by T.Y. Alvin Liu, M.D., principal investigator and founding director of the James P. Gills Jr., M.D., & Heather Gills Artificial Intelligence Innovation Center, the study examined whether referrals made by a U.S. Food and Drug Administration-approved AI-assisted diagnostic screening program increased patient adherence to recommended annual diabetic eye exams. Two historically disadvantaged patient groups, African American patients and patients covered under Medicaid, were focused on in the study.
The current study builds on previous work from Liu and his team, which found that, on a population level, the number of eye exam referrals for people with diabetic retinopathy increased when the AI tool was used.
"A referral [from a primary care provider] doesn't guarantee people will attend a diabetic eye exam, even if it's needed," says Liu. Focusing on African American and Medicaid patients — two groups at high risk for poor visual health outcomes — let the researchers clearly determine if the AI tool's use translated to positive changes in patient care, Liu says.
In their retrospective analysis, the researchers identified 3,745 adult patients with diabetes who visited the Wilmer Eye Institute for a diabetic retinopathy evaluation between August 2020 and September 2022. Of this group, 3,352 patients (mean age 60.6 years) received referrals from their primary care providers (PCPs) and 393 patients (mean age 61.6 years) received a recommendation from an AI-assisted screening tool.
Patients evaluated with the AI tool had retinal images taken using a specialized camera and analyzed in real time during their primary care appointment. If diabetic retinopathy was detected, they were informed and given a referral to the Wilmer Eye Institute or another eye care specialist of their choice that same day.
Comparing the PCP and AI tool referral methods, the researchers observed that a higher percentage of African American patients received an eye exam referral when the AI diagnostic tool was used (64.9% vs. 44.4%) versus when it was not. The number of referrals for patients insured by Medicaid were comparable (0.8% vs. 0.6%) regardless of how they received their referral.
Additionally, they found that patients with hypertension (89.6% vs. 82.6%) and chronic kidney disease (26.2% vs. 20.9%) were also more likely to receive an eye exam referral versus people who did not have either condition when the AI-assisted tool was used.
Investigating how the referral methods translated to changes in patient care, Liu's team found that people who both opted for the AI-assisted tool and attended their diabetic retinopathy evaluation were 15% more likely to be African American. Medicaid coverage did not impact patient appointment attendance whether a referral was received from a PCP or the AI tool.
Liu notes that African Americans and other racial and ethnic minorities are disproportionately affected by diabetic retinopathy and other diabetes-related eye outcomes. Despite this, they are also less likely to receive an annual eye exam for the disease.
"With the AI tool, the patient is evaluated on the spot and given a test result. They're not being asked to attend an appointment because they may have something wrong," says Liu regarding the study results. "Other obstacles may limit whether patients can attend the screenings. But we were able to see that they are more convinced they need care if they're given immediate results with clear instructions on what to do."
Liu says that while the findings are encouraging, further work is needed to evaluate whether improved test access translates to changes in long-term patient vision health outcomes.
"Ultimately, AI tools are not meaningful unless you can demonstrate that their real-world deployment positively impacts patient lives. With future work, we want to examine how patients continue to interact with these AI tools over time and how that translates to specific eye health outcomes."
Researchers who contributed to this study include Michael D. Abramoff, Roomasa Channa, Harold Lehmann, Ariel Leong, Jiangxia Wang and Risa M. Wolf.
Michael D. Abramoff, M.D., Ph.D., shares the following competing interests: patents and patent applications assigned to the University of Iowa and Digital Diagnostics relevant to the subject matter of this manuscript; Digital Diagnostics, Inc, Coralville, Iowa: investor, director, consultant; executive secretary, Healthcare AI Coalition, Washington, D.C.; Treasurer, Collaborative Community on Ophthalmic Imaging, Washington, D.C.; member, American Academy of Ophthalmology (AAO) AI Committee; member, AI Workgroup Digital Medicine Payment Advisory Group (DMPAG) of the American Medical Association. R.M.W. declares the following competing interest: research support from Novo Nordisk, unrelated to this work. The other authors declare no competing interests.
Support for the study was provided by the Gills Artificial Intelligence Innovation Center at the Wilmer Eye Institute and a Research to Prevent Blindness Career Development Award received by Alvin Liu.
