Prediction of visual field progression based on initial OCT structure measurements using Artificial Intelligence

Abstract

Purpose : To investigate the ability of macular and circumpapillary retinal nerve fiber (cpRNFL) baseline and longitudinal optical coherence tomography (OCT) structural measures during an initial follow-up period to predict visual field (VF) progression using machine learning (ML) approach.

Methods : 204 eyes of 126 patients with glaucoma were included. Rates of change of the macular ganglion cell complex (GCC) layer superpixels and regional cpRNFL thickness from OCT were calculated from initial 5 OCT visits. VF progression was evaluated with event-based analysis; two consecutive significant negative slopes from VF mean deviation during the entire follow-up period. 4-fold cross-validation evaluation was applied to each ML model. Different ML models including ridge regression (RR), random forests (RF), support vector machine (SVM), gradient boosting machine (GBM), and Naïve Bayes (NB) were fitted for structural baseline and rates of change; and clinical/demographic parameters to predict visual field (VF) progression and compared with logistic regression (LR) model using paired bootstrap test.

Results : 20 eyes developed VF progression in an average 7.3 years. Average rates of GCC and cpRNFL rates –) were significantly faster in eyes with VF progression than those that never progressed [– 2.09 μm/year (95% CI, −3.40 to −0.80, P=0.001and 0.79μm/year (95% CI, −1.5 to −0.1, P=0.030, respectively)]. The AUC (95% CI) for the prediction of VF progression was 0.81(0.62, 0.98), 0.77(0.57, 0.97), 0.76(0.53, 0.96), 0.74 (0.42, 0.98), 0.66 (0.45, 0.90), and 0.58(0.23,-0.92) for RR, RF, SVM, NB, GBM, and LR respectively (Figure 1). RR (P=0.001), RF(P=0.001), SVM (P=0.009), and NB(P=0.04) outperformed LR for predicting VF progression. Predictive factors that were selected by RR were sectors 10,11, 14, 15, 20, and 21 from cpRNFL, race, and spherical equivalent. Parameters with the highest relative importance from RF were shown in Figure 2.

Conclusions : VF progression in glaucoma eyes can be predicted from baseline, longitudinal macular, and cpRNFL OCT data along with clinical/demographic data during an initial follow-up period using ML models with good accuracy. The proposed longitudinal ML models may assist clinicians in predicting patients at high risk of developing functional disability.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.