Undilated Fundus Photography: AI Flags Macular Disease Requiring Referral in Primary Care

Age-related macular degeneration (AMD) may cause no obvious symptoms in its early stages. By the time patients seek care after central vision has been impaired, the opportunity to preserve vision has often diminished. If fundus images can be captured in family medicine or general clinics to identify, on the spot, people who need referral to an ophthalmologist, screening may no longer be limited to specialist clinics. A prospective study published in *Scientific Reports* provides performance data for this clinical pathway from a setting closer to real-world practice.

The research team recruited 845 adults aged 50 or older who had never been diagnosed with AMD from three primary care clinics and three general ophthalmology clinics in New York City. Color images of both eyes were first captured with a nonmydriatic fundus camera and analyzed by iPredict-AMD. Dilated images were then captured and independently graded by three ophthalmologists, with the majority determination used as the reference standard. After excluding people who were ineligible, already had AMD, or did not complete the procedures, 696 participants entered the primary endpoint analysis: 339 from primary care clinics and 357 from ophthalmology clinics.

iPredict-AMD is not intended to provide a complete diagnosis. Instead, it addresses a more practical referral question: Do the images show disease beyond early AMD—that is, intermediate or late disease that should undergo further evaluation by an ophthalmologist? At the participant level, the system had an area under the curve (AUC) of 0.92, sensitivity of 90.27%, and specificity of 83.36%. Of the 113 participants whom experts determined required referral, the system correctly identified 102 but missed 11. The study also reported a negative predictive value of 97.79%, although this figure changes with disease prevalence in the screened population and cannot be applied directly to every setting.

The workflow was deliberately designed to resemble that of a general clinic: images were captured without prior pupil dilation, uploaded to the cloud after passing a quality check, and interpreted jointly by five deep-learning models. The system returned one of three results—“refer,” “do not refer at this time,” or “insufficient image quality”—within about one minute. The models were developed using 116,875 fundus photographs from the AREDS study. This study used prospective enrollment and independent ophthalmologist grading to test their performance outside the existing dataset. The trial registration originally estimated enrollment of 1,000 participants, but the final paper reported 845, and the primary analysis was further reduced to 696—a discrepancy that cannot be overlooked when interpreting the results.

The tool’s intended role also defines its limitations. It was trained only to identify the need for AMD referral. The fact that it may classify other abnormal images as requiring referral does not mean it can diagnose diabetic retinopathy, high myopia, or epiretinal membrane. All centers were also concentrated in the same city and used the same fundus camera and cloud-based workflow. Whether sensitivity and specificity would be maintained across different populations, equipment, network conditions, and healthcare referral systems remains to be answered by external prospective studies.

The study demonstrates screening accuracy but has not yet shown that implementing the system can increase ophthalmology attendance, shorten treatment delays, or reduce vision loss. Publication of the paper also does not mean that regulators have approved the product for market. The work was funded by the U.S. National Institutes of Health Small Business Innovation Research program. The corresponding author is employed by the developer, iHealthScreen, and holds related patents for image-based screening. Before the system can enter routine medical care, its performance must be validated across regions, and questions about image failure rates, the burden of unnecessary referrals, data governance, cost-effectiveness, and regulatory positioning must also be addressed.