Machine Learning Identification and Quantification of Drusen At-Risk on Optical Coherence Tomography for Geographic Atrophy Prediction

Abstract

Purpose : The hallmark lesions of intermediate Age-Related Macular Degeneration (AMD) are macular drusen. As a result, drusen characteristics are used to predict the likelihood of Geographic Atrophy (GA). The goal of this analysis was to evaluate the potential feasibility for identification of “drusen at-risk” for progression to GA through the use of a multi-model machine learning (ML) assessment tool.

Methods : An IRB-approved retrospective cohort analysis of 115 eyes with intermediate AMD with 5 years follow was conducted. A hybrid multi-model system was developed to define drusen at-risk through the integration of a previously described ML-enabled multilayer retinal segmentation system that utilized RPE-BM thickness of 50 um or greater to define significant drusen and an additional ML model that identified At-Risk Ellipsoid Zone regions (At-Risk EZ) for progression. Drusen At-Risk were identified as areas where significant drusen defined by the layer segmentation were concurrent with areas of EZ At-Risk. Assessment of lesion progression was performed on co-registered OCT volumes from baseline and year 5. Mean baseline percentage area of Drusen At-Risk was compared between patients that progressed to develop GA at year 5 vs those that did not develop GA. Probability of conversion to GA was calculated by obtaining a ratio of baseline area of Drusen At-Risk that converted to GA at year 5 and total area of baseline Drusen At-Risk.

Results : Of the 115 eyes included in this analysis, 31 progressed to develop GA at year 5. Mean macular percentage of Drusen At-Risk coverage at baseline was significantly higher in the patients that progressed to develop GA as compared to patients that did not progress (2.7% vs 0.6% respectively, p value: <0.001). The probability of conversion of baseline Drusen At-Risk to GA at year 5 was calculated as 27%. In addition, 26% of the newly developed GA area was directly attributable to conversion of Drusen At-Risk.

Conclusions : This study demonstrates that Drusen At-Risk can be used as a biomarker to predict the development of GA. Nearly 25% of all new GA area developed directly from Drusen At-Risk. Further research will focus on additional validation and comparative performance of this biomarker to other potential features that predict future GA.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.