Estimating visual acuity with habitual correction in clinical practice settings from fundus photos using artificial intelligence in eyes with diabetic macular edema

Abstract

Purpose : Determining visual acuity (VA) with habitual (e.g., glasses) correction is essential when managing many ophthalmic diseases. If artificial intelligence (AI) evaluations of a macular image could estimate VA with habitual correction on an ETDRS chart, these techniques might save time and technician costs, as well as potentially simplify home monitoring of habitual VA. The objective of this investigation was to estimate VA with habitual correction on an ETDRS chart among patients with diabetic macular edema (DME) in the clinical practice setting using AI evaluations of fundus photographs.

Methods : Retrospective evaluation was performed using de-identified fundus photographs matched to habitual VA determined by technicians on an ETDRS chart. among patients with a history of DME and at least two visits within 1-6 months at a university-based clinic. Fundus photograph evaluation was performed using a previously developed AI algorithm for determining best-corrected VA from fundus photographs among study participants in a randomized clinical trial evaluating aflibercept for DME. AI-determined habitual VA mean absolute error (MAE) was compared with actual habitual VA.

Results : Among 141 patients at visit 1, mean age (SD) was 73.9 (12.3) years, 41 had non-proliferative diabetic retinopathy (NPDR) and DME based on OCT central subfield thickness (CST), 17 had PDR and DME, 44 had NPDR and no DME, and 39 had PDR and no DME. The MAE (SD) for eyes with NPDR, with or without CI-DME was 1.16 (1.00) lines for habitual VA between 20/10 and 20/20 (n=67), and 1.44 (1.15) lines for habitual VA 20/25 to 20/80 (n=231). The MAE (SD) for eyes with PDR, with or without CI-DME was 1.92 (1.08) lines for habitual VA between 20/10 and 20/20 (n=50), and 1.42 (0.97) lines for habitual VA 20/25 to 20/80 (n=150). There were too few eyes with habitual VA 20/100 or worse for meaningful analyses.

Conclusions : AI evaluations of fundus photographs among patients with DME were able to determine habitual VA with a MAE of approximately 1 to 1.5 ETDRS lines of actual habitual VA among eyes with habitual VA of 20/80 or better. These results provide additional support for pursuing AI evaluation of fundus photographs to determine habitual VA among DME patients if an acceptable proportion of evaluations are within 1 to 2 lines of actual habitual VA.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.