A deep learning system achieved 98% sensitivity for detecting systemic lupus erythematosus from retinal images alone, significantly outperforming both primary care physicians and rheumatology specialists in head-to-head comparisons, according to research published in Cell Reports Medicine.
The AI system, called DeepSLE, demonstrated 92% sensitivity and 93% specificity for lupus detection in standard internal validation testing. In a separate prospective reader study comparing the system with five primary care physicians and five rheumatology specialists using a curated dataset, DeepSLE achieved 98% sensitivity, exceeding all physicians.
The system was trained and validated on 254,246 retinal fundus images from 91,598 participants across multiple ethnic groups. Performance remained robust across gender, age, ethnicity, and socioeconomic subgroups, suggesting broad clinical applicability.
DeepSLE also detected lupus-related complications, achieving 100% sensitivity for lupus retinopathy in some external datasets and 76% sensitivity for lupus nephritis.
"Our proposed retinal image-based DL system provides a proof-of-concept digital solution to address the current gap in the detection of SLE and related complications, in which delayed diagnosis is highly prevalent," wrote Tingyao Li, PhD, of Shanghai Jiao Tong University, and colleagues.
The findings suggest retinal imaging could enable opportunistic lupus screening during routine eye examinations, with digital retinal photography being widely available in primary care settings for referral to specialists for confirmatory testing.
The study was limited by datasets from China and the United Kingdom, which may affect generalizability to other populations.
Source: Cell Reports Medicine