Can Artificial Intelligence Predict Glaucomatous Visual Field Progression? A Spatial-Ordinal Convolutional Neural...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucomatous visual field (VF) progression is a leading cause of irreversible blindness.
• Accurate prediction of VF progression is crucial for timely intervention and management.
• Current predictive models have limitations in capturing complex spatial-ordinal patterns of VF changes.

Purpose of the Study:

• To develop and evaluate an artificial neural network (NN) model for predicting glaucomatous VF progression.
• To incorporate both spatial and ordinal information within the NN architecture.
• To compare the predictive performance of NN models against traditional linear models.

Main Methods:

• A cohort study included 9212 eyes from 6047 primary open-angle glaucoma patients with over 4 years of VF data.
• Spatial-ordinal tensors were constructed by stacking 3 consecutive VF tests (VF-blocks).
• Convolutional neural network (NN) models were developed and compared with linear regression and pointwise linear regression (PLR) models using AUROC.

Main Results:

• The best-performing NN model achieved an AUROC of 0.864 (sensitivity 0.42, specificity 0.95).
• Linear models, including PLR, showed significantly lower performance (PLR AUROC 0.611).
• Progression was identified in 5.5% to 10.2% of VF-blocks based on different criteria, with progression groups being older and having worse initial VF indices.

Conclusions:

• NN models integrating spatial-ordinal characteristics significantly outperform linear models in predicting glaucomatous VF progression.
• This approach offers a more robust method for identifying patients at risk of vision loss.
• The findings support the use of advanced AI models for enhanced glaucoma management.