Hereditary retinal diseases: assisted by Retina4IRD, specialists go from 67 to 88% genetic accuracy

In a multicenter randomized trial published by Nature Medicine, a model that predicts 17 categories of genotypes from retinal images saves ophthalmologists twenty points.

Nature Medicine has published a multicenter randomized trial evaluating Retina4IRD, a decision support system that predicts 17 genotype categories from retinal images. The model, a Vision Transformer pre-trained with RETFound, was trained and validated on multimodal data — color fundus photographs and optical coherence tomography — from 1,843 genetically confirmed patients, or 3,376 eyes, in China, South Korea and Poland.

In validation, the accuracy of the first five predictions was 0.904 internally (95% CI: 0.896-0.912) and 0.856 externally (95% CI: 0.850-0.863). The clinical trial itself included 295 participants with a high-throughput sequencing report, with a median age of 33 years, of whom 114 were women (38.6%).

The measured gap

In the arm where the specialists had Retina4IRD, the genetic accuracy of the first five hypotheses was 88.5%, compared with 67.3% in the specialist-alone arm (P < 0.001). For pathologies where the identification of the causative gene conditions access to available gene therapies, the twenty-point difference relates less to the speed of diagnosis than to the correctly referred patients.