From Visual Field to Visual Field: 10-2 Visual Field Map Prediction from 24-2 Visual Field Using Machine Learning and Detection of Central Visual Field Progression

Abstract

Purpose : To investigate whether predictions of 10-2 visual field (VF) maps obtained from a machine learning algorithm, called visual field to visual field (V2V), applied to 24-2 VF can detect central VF progression over time.

Methods : This study included 4645 pairs of 10-2 and 24-2 VF tests from 684 participants (1274 eyes). Data was split on the patient level into training (50%), validation (10%) and test sets (40%). Deep learning (DL) algorithm (Multilayer perceptron) and linear regression (LR) models were trained separately on 24-2 VF maps to estimate 10-2 VF 68 total deviation (TD) values and mean deviation (MD) and were compared with each other (Figure1). An independent sample of eyes from the test that had a longitudinal follow-up of at least 5 VFs was selected. The ability of the slope of predicted 10-2 VF MD to detect central VF progression using clustered pointwise linear regression (PLR) was assessed by Receiver operating characteristic (ROC) curves and was compared with the slope of the mean central 12 points of 24-2 VF.

Results : Our LR model predicting 10-2 VF MD achieved R of 0.94 (95% confidence interval [CI], 0.93, 0.95) and MAEs of 1.3 dB (95% CI, 1.2, 1.3) and was comparable to DL estimates of 10-2 MD. The LR model had an average MAE of 2.47 dB (95% CI: 2.39, 2.54) and R of 0.85 (95% CI: 0.84, 0.86) for pointwise TD values estimations and had comparable performance with DL estimates. 193 eyes of 115 patients in the test sample were followed up for an average of 5.2 years and 8 visits. A significant correlation was found between the change over time in V2V predicted and actual 10-2 VF MDs, Figure 2A. The slope of V2V predicted 10-2 VF MD showed an ROC curve area (95% CI) of 0.87 (0.74, 0.94) to discriminate central VF progressors from non-progressors using PLR method which was comparable to actual 10-2 MD slope 0.93 (0.87, 0.97), P=0.19 and was statistically higher than the slope of mean of central 12 points of 24-2 VF 0.77 (0.64, 0.86), P=0.005, Figure2B.

Conclusions : The V2V algorithm estimates 10-2 functional loss from 24-2 VF with high accuracy and potentially could be used to monitor central glaucomatous progression over time. Given the widespread availability of the 24-2 test, this algorithm may lead to improved individualized patient care and risk stratification of patients who are at risk for central VF damage.

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