Abstract
Purpose :
Early initiation of therapy in glaucoma may prevent irreversible vision loss, yet identification of patients at greatest risk for progression remains challenging. We aimed to evaluate the predictive value of baseline macular optical coherence tomography (OCT) for clinically significant glaucoma progression using deep learning.
Methods :
Data were collected from a deidentified, longitudinal institutional dataset of macular OCT and 24-2 Humphrey visual fields (HVFs). Inclusion criteria were diagnosis of primary open-angle glaucoma; at least 5 reliable HVFs; and a macular OCT study obtained prior to the date HVF progression was first detected in patients with progression, or prior to the fifth-most-recent HVF in patients without progression. Exclusion criteria were presence of cataract with visual acuity 20/40 or worse, treatment with anti-VEGF agent, or incisional glaucoma surgery during the HVF observation period. Clinically significant glaucoma progression was defined as mean deviation rate of change of -0.5 dB/year.
A vision transformer (ViT) based classifier was developed to predict progression (Figure 1a). Using an 80/20 train/test split among unique eyes, k-nearest neighbors classification on framewise embeddings generated by the ViT model was used to predict progression for each OCT tomogram (Figure 1b). The proportion of predictions for each OCT study was used as a prediction score. Accuracy and area under the receiver operating curve (AUC) were assessed by varying thresholds of this score.
Results :
1346 eyes from 785 patients were included, of which 333 (24.7%) eyes progressed. Macular OCT exams were performed a median of 194 days (IQR 119 to 368 days) prior to progression. Mean patient age was 67.3 ± 12.9 years. On the test set of 270 eyes, the model achieved an AUC of 0.73 for prediction of future progression (Figure 2). At a prediction score threshold of 0.32, the model correctly identified 45 of 70 patients with future progression (sensitivity 0.64, specificity 0.45). No significant difference in predictive value between OCT frames by superior-to-inferior quintile was found, suggesting that all parts of the macular OCT contain early markers of progression.
Conclusions :
We present a novel method for analysis of macular OCT using deep learning, which identifies structural features heralding glaucoma progression. This method may aid treatment decisions for glaucoma patients.
This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.