A study based in Mumbai tested the performance of a smartphone-based automated system that analyzes retinal images, taken by a minimally trained health care worker, to detect referable diabetic retinopathy (DR). Using this artificial intelligence (AI) platform could allow screening in remote areas where such services are often unavailable, possibly giving results to patients immediately.

Researchers found the AI algorithm’s sensitivity and specificity to diagnose referable DR in 213 participants were 100.0% and 88.4%, respectively, and its sensitivity and specificity for any DR were 85.2% and 92.0%, respectively.

The automated analysis, while promising, comes with some limitations. For one, it was designed only to provide binary results—referable DR or no DR—and did not grade the stages of the disease. In addition, the researchers noted that the offline AI was purposely not trained on mild nonproliferative DR images to ensure high specificity in the binary diagnoses. The AI also tended to over-diagnose retinal lesions, such as retinitis pigmentosa, drusen and retinal pigment epithelium changes, as referable DR. Although incorrectly labeled, these cases likely would warrant a referral to a practitioner anyway, the researchers said in the study.

The study concluded that a smartphone-based fundus camera and AI algorithm could help address the lack of specialist access through automatic, instant analysis of retinal images and offer a possible solution for implementing large-scale DR screening.

Natarajan S, Jain A, Krishnan R, et al. Diagnostic accuracy of community-based diabetic retinopathy screening with an offline artificial intelligence system on a smartphone. JAMA Ophthalmol. August 8, 2019. [Epub ahead of print].