Abstract
Purpose :
Though visual fields remain the gold standard for assessing glaucoma progression, they are prone to fluctuations. Point-wise visual field indices are variable while global indices are insensitive to local losses. We analyzed existing clinical data to assess the performance of a Convolutional Long Short-Term Memory (LSTM) network trained on longitudinal visual field and clinical data in determining glaucoma progression.
Methods :
From two initial datasets of 265,559 visual fields from 55,056 patients and 350,438 samples of clinical data from 23,967 patients, patients at the intersection of both datasets with four or more visual fields and corresponding clinical data, specifically cup-to-disc ratio, central corneal thickness, and intraocular pressure, were included. After additional exclusion criteria were applied to ensure reliable data, 3103 eyes remained. Three commonly used glaucoma progression algorithms (Visual Field Index slope, Mean Deviation slope, and Pointwise Linear Regression) were used to define eyes as stable or progressing. Two machine learning models, one exclusively trained on visual field data and another trained on both visual field and clinical data, were compared to each other using receiving operating characteristic (ROC) curves and mean accuracies from k-fold cross validation (k=3).
Results :
The convolutional LSTM demonstrated 86-90% accuracy with respect to the different conventional progression algorithms defining ground truth given 4 consecutive visual fields for each subject. The model that was trained on both visual field and clinical data (AUROC between 0.82 and .90) had better diagnostic ability than a model exclusively trained on visual field (AUROC between .56 and .75) data as determined by comparing the area under the ROC curve (P<.001).
Conclusions :
A convolutional LSTM architecture captures local and global trends in visual fields over time. It best fits the problem of assessing glaucoma progression because of its unique ability to extract spatio-temporal features other algorithms cannot. Supplementing visual fields with clinical data improves the model’s ability to assess glaucoma progression and should be accounted for in future research.
This is a 2020 ARVO Annual Meeting abstract.