Using Quantitative Retinal SD-OCT Features to Predict Eyes at Risk of Fast Expansion of Geographic Atrophy in Dry Age-related Macular Degeneration

Abstract

Purpose : Geographic atrophy (GA) is an advanced stage of dry age-related macular degeneration (AMD). With the recent advances in therapies aiming at slowing disease progression, predicting patients at risk of fast GA expansion could be particularly useful, not only for clinical prognostication, but also for targeting patients who would benefit most from treatment. The goal of this study was to develop a prediction model for patients at risk of fast GA expansion within 1 year, based on retinal quantitative features.

Methods : This was an IRB-approved retrospective longitudinal study, that included baseline and 1-year follow up SD-OCT scans from 311 eyes of 311 patients with geographic atrophy. A previously validated deep learning-based platform was used for automated segmentation of multiple retinal layers with segmentation validation performed by certified readers. This allowed for extraction of several quantitative features describing photoreceptor integrity and disease burden. Patients were then split based on their 1-year GA expansion rate into fast (> 2 mm²/year) and normal/slow (< 2 mm²/year) progressors. To predict risk of fast expansion, a random forest (RF) classification model was trained on 70% of the data and tested on the remaining 30%. A Shapley additive explanations (SHAP) analysis was then performed to explain the output of the RF model.

Results : The mean GA growth rate in the slow/normal progressor group was 0.5 ± 0.5 mm²/year vs 2.9 ± 0.9 mm²/year for the fast progressor group (p-value < 0.001). The preliminary prediction model had a sensitivity of 89%, specificity of 88%, positive predictive value of 82%, negative predictive value of 93%, and accuracy of 88%. The mean GA growth rate in the misclassified patients was 1.9 ± 0.5 mm²/year which was at the boundary that defined the 2 groups. The SHAP analysis revealed that larger GA area, larger area of partial photoreceptor attenuation (EZ-RPE thickness ≤ 20 microns), and high drusen burden were the most significant contributors to the model’s predictions.

Conclusions : This work shows that OCT-extracted quantitative metrics carry significant value for predicting patients at risk of 1-year fast GA expansion. Future work will focus on fine tuning this model and external dataset validation.

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