A consensus approach to train deep learning models to predict glaucoma progression from visual field data

Abstract

Purpose : The purpose of this study is to construct and examine a deep learning model to predict the progression of glaucoma using minimal visual fields. Additionally, it aims to compare the consensus derived from state-of-the-art progression algorithms with the judgments made by a glaucoma specialist.

Methods : A retrospective analysis of 24-2 visual field (VF) data was conducted on 8,051 glaucoma patients (14,089 eyes), each with at least 5 VF assessments using HFA™ II-i and HFA3 Model 850 (ZEISS, Dublin, CA). The data was split at patient level into train (12,026 eyes) and test (2,063 eyes) set. The patient demographics at baseline are mean age (±SD) 52 ± 15 years, Mean Deviation (MD) -9.7 ± 8.3 dB and Visual Field Index (VFI) of 77.4 ± 26.3 %. The mean (±SD) inter-visit duration was 1.5 ± 1.0 years. We employed 5 methods — MD slope (slope<0, P<0.05), Guided Progression Analysis (GPA) trend (VFI) slope (slope<0, P<0.05), GPA event (“Likely Progression” alert), AGIS, and CIGTS — to evaluate worsening. Training data was labelled as progressing when at least 3 out of 5 methods reported progression. Additionally, these methods were compared against clinical assessment on test set. A simple supervised deep learning model was developed to predict progression at the 5^th visit, using VF data from up to 3^rd/4^th/5^th visits.

Results : The upset plot in Figure 1A shows the relationship between 5 different methods in reporting progression on training set. A consensus among all algorithms was observed in 132 eyes. Trend-based algorithms consistently identified more progression cases (533 eyes). Figure 1B compares algorithm outputs with clinically verified progression, highlighting that 137 eyes noted as progressing clinically were not identified by any algorithm. The architecture of the binary classification model that predicts glaucoma progression from multiple visits of VF data (54 Threshold, 52 Total Deviation, and 52 Pattern Deviation values) is shown in Figure 2A. The AUC scores of the three models on test set are 0.84 (3 visits), 0.91 (4 visits), and 0.92 (5 visits) (Figure 2B).

Conclusions : The model demonstrated good diagnostic performance in identifying glaucoma progression. Training the model with clinical judgment as target might reveal any hidden features in VF data to predict progression.

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