AI decision support lifts diagnostic accuracy for inherited eye diseases in a randomized trial

A trial in 295 patients found AI-assisted specialists reached the correct genetic category far more often than specialists alone.

A study published in Nature Medicine on 24 July 2026 describes Retina4IRD, an AI decision support system for diagnosing inherited retinal diseases from retinal images. The system predicts 17 genotype categories using a Vision Transformer model pretrained with RETFound, trained and validated on color fundus photographs and optical coherence tomography scans from 1,843 genetically confirmed patients across China, South Korea and Poland.

The researchers ran a randomized controlled trial (NCT06839170) enrolling 300 participants with suspected disease, split evenly between an AI-assisted specialist arm and a specialist-only arm. Among the 295 with sequencing reports analyzed, top-5 genetic accuracy reached 88.5% with AI assistance versus 67.3% without. Top-1 accuracy was 37.8% versus 22.4%. A composite downstream management score also favored the AI group, 37.7 versus 28.5.

Inherited retinal diseases are hard to diagnose and often require genetic sequencing that is slow or unavailable. A tool that helps a specialist narrow the genetic cause from an image could shorten the path to diagnosis and appropriate care, especially where sequencing capacity is scarce. That makes accurate diagnosis more available rather than gated behind specialist genetics labs.

Worth watching: whether the system holds up beyond the three countries studied, and whether the management benefit survives scrutiny. The management findings came from post hoc analyses, external validation accuracy (0.856) trailed internal (0.904), and the published text available is an abstract preview with the full paper paywalled.

Source: Nature Medicine