Abstract
Purpose :
To develop and evaluate a deep learning model (DLM) to predict eyes at high risk for future incisional glaucoma surgery from observations early in the disease course.
Methods :
We used the DLM architecture depicted in Figure 1. A vision transformer (VIT) is used for feature extraction. The input of the transformer consists of 3 channels consisting of spatially aligned VF and OCT data. After VIT-based feature extraction, the latent representation of VF and OCT data is concatenated with the normalized clinical data and fed into a linear classifier predicting the occurrence of glaucoma surgery within the pre-specified time horizon. Data were split at the patient level into 76%, 12%, and 12% for training, validation, and testing, respectively.
Results :
DLM performance was evaluated and compared using Area Under the Receiver Operating Characteristic Curve (AUC). Additionally, sensitivity, specificity, and positive predictive value were calculated using an optimal threshold selected using Youden’s J index.
Between 1788 and 2893 eyes were included in the analysis depending on the time horizon of interest. In the test set, the DLM achieved an AUC of 0.92 (95% CI: 0.86, 0.98), a sensitivity of 0.74, and a specificity of 0.92 for predicting glaucoma surgery within 2 years from baseline (Table 1). The predictions for other time intervals achieved clinically useful AUC values (>0.8) except for the 5-year time interval with an AUC of 0.78 (95% CI: 0.71, 0.87).
Conclusions :
It is possible to identify eyes in need of glaucoma surgery with very high sensitivity and specificity. Implementing such prediction models in the clinical setting may help stratify high and low-risk patients. Future studies will need to assess the performance of such models prospectively in a varied clinical setting.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.