Prediction of Glaucoma Development from Glaucoma Suspect with OCT Structural Measurements using Artificial Intelligence

Abstract

Purpose : To investigate the ability of macular and circumpapillary retinal nerve fiber (cpRNFL) change rates and clinical/demographic factors to predict development of glaucoma from glaucoma suspects (GS) through elastic net regression (ENR) and machine learning (ML) models.

Methods : 162 eyes of 146 patients with GS were included. The eyes were required to have characteristic glaucomatous optic nerve changes without visual field (VF) defects. Development of glaucoma was defined as having VF glaucoma hemifield test outside normal limit and pattern standard deviation 5% in 3 consecutive exams. Global and sectoral rates of change of macular ganglion cell/inner plexiform layer (GCIPL) and cpRNFL OCT were calculated with linear mixed-effect models. The ENR and ML models, gradient boost model (GBM), support vector machine (SVM) and Naïve Bayes (NB) were fit for structural rates of change and clinical/demographic parameters in order to predict the development of glaucoma in GS. Parameter tuning with Grid Search Method and 5-fold cross validation were applied for each ML model.

Results : 47 eyes developed glaucoma based on the study criteria. The average follow-up time was 2.3 and 2.1 years for glaucoma and GS eyes, respectively. The rates of average GCIPL were significantly faster in the eyes that converted to glaucoma (= –0.10, P<0.001) and the average cpRNFL rates was not significant (= –0.15, P=0.152). Predictive factors that were selected by ENR were average, minimum, superior and temporal superior GCIPL change rates and average intraocular pressure (IOP) during follow-up (AUC (95% confidence interval (CI)=0.71 (0.62, 0.80)). Parameters with highest relative importance from GBM were average GCIPL rates, average IOP during follow-up, age, axial length temporal superior and temporal inferior GCIPL and temporal cpRNFL change rates (Figure 1). The AUC (95% CI) for prediction of glaucoma development was 0.91 (0.86, 0.97), 0.87 (0.81, 0.92) and 0.82 (0.75, 0.89) for GBM, SVM and NB, respectively (Figure 2).

Conclusions : Development of glaucoma in glaucoma suspect eyes can be predicted from longitudinal macular and cpRNFL OCT data and average IOP during follow-up with clinically relevant accuracy. The proposed models may assist clinicians to predict better the development of glaucoma.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.