AHP1 Task 1 AI Tool Evaluation Example

A Camera, an Algorithm and a Referral: Evaluating an Autonomous AI Diabetic Eye Exam for a Composite Primary Care Clinic, With Human Oversight, Scope of Practice and Patient Communication Defined

Student Name

Leavitt School of Health, Western Governors University

AHP1: AI for Healthcare Professionals, Task 1

Course Instructor

Month Day, Year

A Camera, an Algorithm and a Referral: Evaluating an Autonomous AI Diabetic Eye Exam for a Composite Primary Care Clinic, With Human Oversight, Scope of Practice and Patient Communication Defined

The Clinical Problem

Diabetic retinopathy, damage to the blood vessels of the retina caused by diabetes, is a leading cause of preventable vision loss in working-age adults. It causes no symptoms until it is advanced, so people with diabetes are advised to have a retinal exam every year or two. Many do not. At Northside Family Health, a composite primary care clinic serving about 2,400 adults with diabetes, only 38% had a documented eye exam in the past year. Patients cited the cost of a separate eye appointment, the drive to the nearest eye care practice 40 minutes away, time off work and not understanding why an eye exam mattered when they could see fine. The clinic's leadership is considering an autonomous artificial intelligence system that would perform the exam during routine primary care visits.

How the Tool Works

In the system under consideration, a medical assistant uses a retinal camera to photograph the back of each eye, usually without dilating the pupils. The images are sent to an AI algorithm that analyzes them and returns a result within minutes: negative for more than mild diabetic retinopathy, with a recommendation to rescreen in 12 months; positive, with a recommendation to see an eye care professional; or insufficient image quality. The system is autonomous, meaning it makes the diagnostic determination itself rather than flagging images for a specialist to read. The first such device, IDx-DR, was classified by the U.S. Food and Drug Administration through its De Novo pathway in 2018 for use by health care providers in adults with diabetes not previously diagnosed with retinopathy. The FDA's summary states plainly that the device detects only diabetic retinopathy, does not screen for glaucoma or other conditions and does not treat anything (U.S. Food and Drug Administration, 2018).

The Evidence

Accuracy: In the clinical study the FDA reviewed, 900 adults with diabetes and no history of retinopathy were enrolled at 10 primary care sites. Compared with a rigorous reference standard graded by an expert reading center, the system identified more than mild retinopathy with an enrichment-corrected sensitivity of 87.2% and specificity of 90.7%, and it produced a usable result for 96.1% of participants; the FDA found no significant differences in sensitivity by age, sex, race or ethnicity (U.S. Food and Drug Administration, 2018). Put simply, it caught about 87 of every 100 people with significant disease and correctly cleared about 91 of every 100 without it. Operators received standardized training before the study.

Outcomes: Accuracy matters only if it changes what happens to patients. In a randomized trial among youth aged 8 to 21 with diabetes, 100% of those offered the AI exam at their diabetes visit completed an eye exam within six months, compared with 22% of those given a referral to an eye care provider; among those with abnormal results, 64% in the AI group followed through with an eye care provider, compared with 22% in the control group (Wolf et al., 2024). The trial was conducted in one academic pediatric center, so results in adults at a community clinic may differ, but it shows that bringing the exam to the visit removes the main barrier.

Risks

Missed disease: a sensitivity of 87% means some patients with significant retinopathy will be told their exam is negative. The clinic must explain that no screening test is perfect and keep annual rescreening. Scope of detection: patients may assume a negative result means their eyes are healthy, when the system does not look for glaucoma, cataracts or other conditions.

Bias: AI systems can perform differently across groups. A widely used commercial algorithm for identifying patients with complex health needs was found to underestimate the needs of Black patients because it used past spending as a stand-in for need, and less money is spent on Black patients' care; fixing that single design choice would have lifted Black patients' share of those flagged for extra help from 17.7% to 46.5% (Obermeyer et al., 2019). That example involved a different kind of algorithm, but it shows why the clinic should ask the vendor for performance data by race, ethnicity, age and sex, and should monitor its own results, including the rate of insufficient images, which can be higher in patients with cataracts or small pupils.

Privacy and security: retinal images are protected health information and are sent to the vendor's cloud for analysis. The clinic will require a business associate agreement, encryption and a clear statement that images will not be used for other purposes without consent. Workflow: if positive results are not followed up, screening adds no value, and it may create liability.

Human Oversight and Scope of Practice

Autonomy in diagnosis does not remove human responsibility. The clinic will assign it by role. Medical assistants, trained and competency-checked by the vendor and the clinic's nurse manager, will capture images and repeat them if quality is insufficient; they will not interpret or explain results beyond reading the standard patient script, which is within their scope. The primary care clinician remains responsible for the patient's diabetes care: reviewing each result, discussing it with the patient, placing the referral for positive or insufficient results and documenting the plan. A registered nurse care coordinator will track every positive and insufficient result until the patient has seen an eye care professional, and will call patients who have not scheduled within 30 days. An optometrist partner will accept referrals and share findings back with the clinic. A quarterly review by the medical director will examine completion rates, referral follow-through, insufficient-image rates and results by patient group. The clinic will not use the tool for patients with known retinopathy, who need ongoing specialist care.

Communicating With Patients

Patients will be told plainly, in English or Spanish, that a computer program will read their eye photos, that it has been tested and authorized by the FDA, that it checks only for diabetes-related eye damage, and that a clinician reviews the result with them. Patients may decline and be referred to an eye care practice instead. Positive results will be explained as a reason to see an eye specialist soon, not as a diagnosis of blindness, and the care coordinator will help with scheduling and transportation. Negative results will be explained as reassuring but not a substitute for next year's exam or for reporting vision changes.

Recommendation

The clinic should adopt the system, with conditions: vendor performance data by patient group before purchase, a signed business associate agreement, trained and competency-checked medical assistants, a care coordinator responsible for follow-through, clear patient scripts and quarterly review of results. With these safeguards, the tool addresses the clinic's real barrier, getting patients an eye exam at all, while keeping clinicians accountable for decisions and patients informed about how their care is delivered.

References

Obermeyer, Z., Powers, B., Vogeli, C., & Mullainathan, S. (2019). Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464), 447-453. https://doi.org/10.1126/science.aax2342

U.S. Food and Drug Administration. (2018). De Novo classification request for IDx-DR (DEN180001). https://www.accessdata.fda.gov/cdrh_docs/reviews/DEN180001.pdf

Wolf, R. M., Channa, R., Liu, T. Y. A., Zehra, A., Bromberger, L., Patel, D., Ananthakrishnan, A., Brown, E. A., Prichett, L., Lehmann, H. P., & Abramoff, M. D. (2024). Autonomous artificial intelligence increases screening and follow-up for diabetic retinopathy in youth: The ACCESS randomized control trial. Nature Communications, 15, 421. https://doi.org/10.1038/s41467-023-44676-z