# Real-world performance of the AI diagnostic system IDx-DR in the diagnosis of diabetic retinopathy and its main confounders

## TL;DR

This study evaluates how well the AI system IDx-DR detects diabetic retinopathy in real-world settings and identifies factors that affect its performance.

## Contribution

The study provides real-world validation of IDx-DR's performance and identifies key confounders affecting its diagnostic accuracy.

## Key findings

- IDx-DR achieved high sensitivity and specificity in detecting severe diabetic retinopathy when image quality was sufficient.
- Factors like pupil size and patient age significantly impacted image acquisition and analyzability.
- IDx-DR results matched ophthalmologists' diagnoses in about half of the cases with good-quality images.

## Abstract

The escalating prevalence of diabetes mellitus (DM) emphasizes the critical need for early detection of diabetic retinopathy (DR). This study assesses the performance of the autonomous AI-based diagnostic system IDx-DR in detecting DR and its associated confounders in a real-world clinical setting. This prospective cross-sectional study involved 875 diabetic patients with a mean age of 52 years (range: 8–92). Retinal images were captured by trained assistants. IDx-DR results were compared with mydriatic fundus examination (gold standard) and Ophthalmologists’ image analysis. Factors impacting image acquisition or analyzability were examined. Among all patients, 10.5% yielded no image in miosis, and 26.1% were unanalyzable by IDx-DR. Confounders affecting image acquisition were examiner, pupil size, patient age and patients’ visual acuity. When good quality images were achieved, IDx-DR performed well, particularly in detection of severe DR (sensitivity 94.4%; specificity 90.5%). IDx-DR results exactly matched Ophthalmologists’ mydriatic fundoscopy gradings in 54.2% if images of sufficient quality were obtainable. Undergrading of DR severity by IDx-DR was rare (4.8%). IDx-DR shows promise in detecting DR, especially in resource-limited settings and in detecting severe DR. One remaining challenge is good image acquisition in miotic patients.

The online version contains supplementary material available at 10.1038/s41598-026-36970-9.