Utilizing Higher-Order Quantitative SD-OCT Biomarkers in a Machine Learning Prediction Model for the Development of Subfoveal Geographic Atrophy in Age-Related Macular Degeneration

Abstract

Purpose : This study sought to evaluate higher-order SD-OCT features, including ellipsoid zone integrity and drusen burden, as predictive imaging biomarkers for a machine learning (ML) prognostic model for the development of subfoveal geographic atrophy (sfGA) in non-neovascular age-related macular degeneration (NNVAMD) over 5 years.

Methods : This was a retrospective cohort study of eyes with NNVAMD without sfGA with a 5-year follow-up interval. Based on sfGA status at year five, eyes were categorized into two subgroups: non-converter and sfGA converter. The macular scans at baseline were evaluated using an automated machine learning-enhanced multi-layer segmentation platform with expert reader verification for assessment of retinal anatomy, including outer retinal integrity [e.g., ellipsoid zone (EZ)] and the sub-RPE compartment (e.g., drusen burden). The feature complexities were compared using multiple decision trees and random forest ML modeling for a prediction assessment. Model performance was evaluated with a 5-fold cross-validation.

Results : One hundred and thirty-seven were included in this assessment, including 116 non-converters and 21 sfGA converters. Multiple higher-order OCT features, including EZ integrity metrics and sub-RPE compartment metrics, that were associated with sfGA conversion were analyzed. Feature selection for model inclusion was based on univariate analysis significance. Utilizing 7 significant baseline imaging features, the predictive performance of the ML sfGA prediction model was evaluated and achieved an area under the receiver operating characteristic (ROC) curve (AUC) of 0.85. SD-OCT features were ranked on their predictive ability to identify the development of sfGA and demonstrated that the most important feature was associated with EZ integrity and drusen burden.

Conclusions : Utilizing a ML classifier, higher-order SD-OCT quantitative biomarkers (i.e., EZ integrity, sub-RPE compartment) appeared to provide a high-performance model for predicting the development of sfGA in NNVAMD within 5 years of baseline assessment.

This is a 2021 ARVO Annual Meeting abstract.