AI Narrows the Gene Search for Inherited Retinal Diseases: Randomized Trial Improves Diagnostic Hit Rate

Inherited retinal diseases can be caused by variants in numerous genes, and even when patients experience similar vision deterioration, the underlying causes may differ. Now, a randomized trial published in *Nature Medicine* shows that if AI first identifies disease features from fundus images, it may help ophthalmology experts more accurately prioritize the disease-causing genes to investigate.

The study enrolled 300 participants and was conducted jointly by seven centers in China. Experts were randomly assigned either to use the Retina4IRD decision-support system or to make assessments without AI assistance; their predictions were then compared with whole-exome sequencing. In the AI-assisted group, the correct gene was included among the top five candidates in 88.5% of cases, compared with 67.3% in the control group—a difference of 21.2 percentage points.

Retina4IRD integrates color fundus photographs and optical coherence tomography (OCT) scans to predict 17 genotype categories before genetic testing. The practical purpose of this design is not to directly declare which variant a patient carries, but to first narrow the broad gene search space, enabling clinicians to arrange subsequent testing, interpret results, and provide genetic counseling on a more informed basis.

According to conference materials released by the Association for Research in Vision and Ophthalmology in the United States, the system uses a Vision Transformer architecture, and its visual foundation model was pretrained on more than 700,000 OCT images and 900,000 color fundus photographs. Subsequent training and validation covered 1,843 patients at nine centers in China, South Korea, and Poland. When only color fundus photographs were used, the top-five prediction accuracy was 84.6%, suggesting that more widely available imaging equipment may also provide diagnostic clues.

Clinical trial registry data show that this was a randomized, parallel-group, double-blind interventional study. Before the trial began, its primary endpoint was specified as the accuracy of “top-five gene mutation predictions” relative to whole-exome sequencing. The study was sponsored by Shanghai General Hospital, Shanghai Jiao Tong University School of Medicine, and was completed in July 2025. The research team also publicly released a 6.5 MB sample dataset, but this is not equivalent to the complete training data.

The findings still have clear limitations. A top-five hit does not mean the first-ranked candidate is correct, nor can it replace genetic sequencing used to confirm specific variants. The system currently focuses on 17 genotype categories, and whether it can maintain its performance for rare phenotypes or those underrepresented in the training data remains to be tested more broadly. Although the randomized trial was conducted across multiple centers, all participants in the clinical phase were from China, and transferability across populations, devices, and healthcare systems has not yet been fully demonstrated.

The key next step is not merely to push accuracy higher. Research must still answer whether AI can shorten the time required for patients to receive a confirmed diagnosis, reduce unnecessary testing, improve referral and treatment decisions, and what consequences incorrect candidate genes might cause. Until these benefits and risks have been prospectively validated and regulatory standards established, Retina4IRD is better positioned as a prioritization tool for physicians rather than an independent diagnostician.