Deep learning system may detect HbA1c levels from retinal photographs

VANCOUVER, British Columbia — Using good-quality macular-centered fundus photographs and serum samples, a newly designed deep learning system could accurately estimate hemoglobin A1c levels, according to a speaker here.

“With the advent of AI, we are hoping to further introduce this as an exponentially paradigm shift for home care of diabetes,” Yih-Chung Tham, PhD, said at the Association for Research in Vision and Ophthalmology annual meeting.

Tham and researchers from Singapore Eye Research Institute included 17,422 participants in the retrospective review; 13,937 individuals were used to train the deep learning system and 3,485 were used to validate the system.

The mean error between the deep learning system and the actual HbA1c was 0.87%, according to Tham.

Overall, the system slightly underestimated HbA1c for patients with diabetes and slightly overestimated for healthy subjects. The results were promising overall, Tham said.

Tham said the deep learning system could eventually be used in a new home care system, possibly with a smartphone or smart device, for diabetes control instead of the typical “finger prick and current paper test.” – by Robert Linnehan

Reference:

Tham YC. Estimation of haemoglobin A1c from retinal photographs via deep learning. Presented at: Association for Research in Vision and Ophthalmology; April 28-May 2, 2019; Vancouver, British Columbia.

Disclosure: Tham reports no relevant financial disclosures.