Artificial intelligence analysis of baseline fundus photos to predict effects of aflibercept treatment on best-corrected visual acuity

Abstract

Purpose : Best-corrected visual acuity (BCVA) is an important predictor of BCVA outcomes following anti-vascular endothelial growth factor (anti-VEGF) treatment of diabetic macular edema (DME) or certain other macular diseases. Most ophthalmologists likely do not obtain a BCVA, including a protocol refraction, prior to initiating anti-VEGF therapy, and therefore, may not necessarily have precise BCVA to help inform patients accurately of the likelihood of VA changes. Recent studies showed AI could estimate BCVA from fundus photos in eyes with DME; thus, this investigation evaluated the ability of AI analyses of fundus images to predict changes in BCVA at 1, 2, and 3 years after initiating aflibercept for DME.

Methods : Deidentified baseline color fundus images of study eyes after dilation were used post hoc to train AI systems to determine changes of 10 or more letters of BCVA on an ETDRS chart at 1, 2 and 3 years after baseline. Participants were patients enrolled in the VISTA randomized clinical trial wherein study eyes were treated with aflibercept or laser. Training, validation, and testing partitions were created by an 80%/10%/10% split of participants, respectively. Baseline image only Method (IMG Only) was a Resnet50 architecture taking images of resolution 224 by 224 pixels. Baseline BCVA only Method (BLV Only) was a logistic regression using only baseline value to predict changes of >10 letters. Combined baseline image and value Method (IMG+BLV) took output of the linear regression and corrected that output towards the true value by using the image only Method.

Results : Analysis included 164 participants. Baseline BCVA letter score ranged from 73 to 24 (Snellen equivalent 20/40 to 20/320). IMG Only Method (Table) resulted in accuracies, sensitivities and F1 Scores that appeared similar to values obtained with BLV Only, with little improvement noted when combining IMG and BLV. Please see attached table.

Conclusions : This suggests AI analysis of fundus images can predict likelihood of a change in BCVA of at least 10 letters from baseline at 1, 2 and 3 years after initiating aflibercept for DME similar to that obtained using baseline BCVA from a protocol refraction and protocol VA measurement on an ETDRS chart, but without refraction or subjective VA measurements by trained VA examiners.

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