Predicting glaucoma interventions with deep learning networks

Abstract

Purpose : Glaucoma is characterised by progressive irreversible vision loss which can be prevented or delayed through timely interventions aimed at lowering intraocular pressure (IOP).Visual field (VF) testing, optic nerve assessment and IOP are the most important indicators used for determining the time of interventions.We investigated how VF and IOP can be used with a deep learning algorithm to predict the timing of a glaucoma intervention.

Methods : In this retrospective database study, patients with primary open-angle, pseudoexfoliation, pigmentary and normal pressure glaucoma seen at the Glaucoma Clinic at Toronto Western Hospital with reliable VFs and a minimum of 10 reliable VFs were included. Data collected for each visit
included age,VFs (Humphrey SITA-Standard 24-2) and IOP. Laser trabeculoplasty and glaucoma surgical interventions were recorded. IOP and VF data from four consecutive visits were used to train a convolutional neural network (CNN) in a 3-fold cross validation scheme to generate the probability of an intervention following the 4th visit(the most recent visit).Three different CNN networks were evaluated: a) CNN-IOP - using only IOP, b) CNN-VF - using only VF, c) CNN-all using IOP+VF+age.

Results : Data from 2743 visits of 84 patients were collected with a mean follow-up of 14 years (range 5-30 years).There were a total of 115 interventions;29.6% laser trabeculoplasties,21.7% trabeculectomies,17.4% phacoemulsification with trabeculectomy and 6.9% Ahmed glaucoma valve surgeries.The CNN-IOP network predicted the timing of intervention with high sensitivity (0.96) and low specificity (0.36).The CNN-VF network predicted the timing of intervention with low sensitivity (0.45) and relatively high specificity (0.79).When the CNN-all was used, the timing of intervention could be predicted with relatively high sensitivity (0.83) and specificity (0.75).The area under the curve of the predictor was 0.86.It suggests that both VF assessments and IOP measurements are essential for the model predictions.The performance of the predictor might be improved with more training examples and by expanding the training data to include parameters from optic nerve assessments.

Conclusions : To our knowledge, this study is the first attempt to apply deep learning to predict the timing of glaucoma procedural interventions.This approach can provide elective supplementary information to guide glaucoma specialists in complex cases.

This is a 2020 ARVO Annual Meeting abstract.