@article{AES9347,
author = {May-Lyn Hu and Nigel Morlet and Wei Liu and David Glance and Viska Mutiawani and Bill Morgan and Siobhan Manners and Jonathon Ng},
title = {Deep learning artificial intelligence models compared to ordinary linear regression for prediction of visual field progression},
journal = {Annals of Eye Science},
volume = {0},
number = {0},
year = {2026},
keywords = {},
abstract = {Background: Loss of field from common conditions such as glaucoma may cause significant irreversible blindness, and predicting visual field loss in clinical practice currently uses ordinary linear regression (OLR). Better predictive methods are desirable to guide management, so this study aims to investigate the utility of a number of well described deep learning (DL) neural networks for prediction of progression.Methods: From a general clinical cohort of 9,569 patients who had visual field testing for a variety of clinical indications, 1,526 patients that had six to fourteen consecutive Humphrey visual field tests were selected. Following filtering to produce six sequential ‘test window’ sets, and mirroring left eye data, the reconfigured cohort was split into a training set (4,304 tests from 2,348 sequences) and a test set (3,576 tests from 596 sequences). We trained both simple and bidirectional models of a recurrent neural network (RNN), a long short-term memory (LSTM) and a gated recurrent unit (GRU), a single and double layer convolutional LSTM (ConvLSTM), and modified three-dimensional convolution (C3D) convolutional neural network (CNN). Those models examined the first three fields in a set to predict the 4th, 5th, and 6th tests. The root mean square error (RMSE) was used to compare the results with OLR.Results: All DL models were significantly better than OLR. The 4th test had the best predictions, with the error increasing for the subsequent tests. Although no one model was clearly superior, the best 4th test result was from the C3D CNN [RMSE 2.376 decibels vs. 3.936 decibel (dB) for OLR], with the single layer conv-LSTM performing slightly better for the 5th and 6th predictions (2.681 vs. 4.155 dB for OLR and 2.838 vs. 4.476 dB for OLR, respectively).Conclusions: Patients may generally have 2 to 3 dB variability in the test performance, and this is more so with glaucoma patients. Relating the Jensen’s inequality to the pointwise prediction error, while taking into account that the theoretical pointwise mean absolute error (MAE) lower limit was 2.32 dB, our C3D CNN result for the 4th sequence was close to the limit of theoretical predictability. These results confirm that DL models are better for predicting visual field loss, however no one model was shown as superior.},
issn = {2520-4122}, url = {https://aes.amegroups.org/article/view/9347}
}