Abstract
Purpose :
To evaluate global visual field prediction using a gradient-boosted approach based on optical coherence tomography (OCT) structural measures in primary open angle glaucoma (POAG).
Methods :
721 glaucoma eyes from 506 Asian Chinese participants from a study site in Singapore and 111 glaucoma eyes from 63 Caucasian participants from Bucharest were included in the study. All participants underwent the same study protocol, including OCT imaging and 24-2 visual field testing using standard automated perimetry. Retinal nerve fiber layer (RNFL) thickness measurements and global visual field mean deviation (MD) values were extracted for each study eye. Data from the Singapore site was split using stratified sampling based on glaucoma severity into a training dataset and internal testing dataset using a 80/20 ratio, while an external test dataset was constructed with the Bucharest data. A gradient boosted ensemble tree (GBT) model was trained using 5-fold cross validation with the training data, and evaluated on the internal and external test datasets using mean average errors (MAE). Predictions were compared with a baseline MAE which predicted the mean MD, and with multi-variate linear regression (LR), using Pearson correlation analysis and Wilcoxon signed rank testing. 95% confidence intervals were generated using bootstrapping with 5000 samples.
Results :
Participants in the internal dataset had an average age of 66.7±8.1 years and MD of -5.85±5.05 dB, while those in the external dataset had an average age of 63.6±11.7 years and MD of -4.79±5.79 dB. Differences in the mean MD were not significant (P=0.118). Baseline MAE on the internal and external test datasets was 3.80 dB and 4.09 dB respectively. The gradient boosted approach achieved a MAE of 3.01 dB (95%CI: 2.58-3.49) and Pearson correlation of 0.59 (0.47-0.70) on the internal test dataset, and a MAE of 3.04 dB (95% CI: 2.52-3.68) and Pearson correlation of 0.66 (95%CI: 0.44-0.82) on the external test dataset. Results obtained with GBT were significantly better than LR (P<.001) on both datasets.
Conclusions :
Global visual field prediction using gradient boosted trees performed better than baseline and linear regression on independent internal and external datasets. The results motivate further evaluation of the approach and its applicability in structure-function studies.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.