Fusing Structural Phenotypes with Functional Data for Early Prediction of Primary Angle-Closure Glaucoma Progression

To classify eyes as slow or fast glaucoma progressors in patients with primary angle-closure glaucoma (PACG) using an integrated approach combining optic nerve head (ONH) structural features and sector-based visual field (VF) functional parameters.

Retrospective longitudinal study.

Patients with PACG from glaucoma clinics.

Patients with PACG with ≥5 reliable VF tests over ≥5 years were included. Visual field progression was assessed using ZEISS FORUM Software, with the baseline VF test selected within 6 months of the OCT scan. Fast progression was defined as a VF index (VFI) decline of < -2.0% per year; slow progression as ≥ -2.0% per year. OCT volumes were processed using artificial intelligence-based segmentation networks to extract 31 ONH structural parameters. The Glaucoma Hemifield test defined 5 regions in each hemifield, reflecting retinal nerve fiber layer distribution. Mean raw sensitivity values were computed for each region and combined with the structural parameters to train machine learning (ML) classifiers for progression classification. Several ML models were evaluated. Shapley Additive Explanations identified the most influential predictors.

Classification of slow versus fast progressors using combined structural and functional data.

A total of 250 eyes were analyzed after excluding severe glaucoma. The mean VFI progression rate was -0.80% per year; 82.4% showed slow progression, and 17.6% fast progression. The Random Forest multimodal model integrating structural and functional features achieved the highest classification accuracy with an area under the receiver operating characteristic curve (AUC) of 0.85 ± 0.02 over 2000 Monte Carlo cross-validation iterations, outperforming the conventional multimodal model based on routinely available clinical features (AUC, 0.77 ± 0.02), as well as structure-only (AUC, 0.79 ± 0.03) and function-only models (AUC, 0.75 ± 0.04). Adding clinical features to the proposed structure-function model did not significantly improve model performance (AUC, 0.86 ± 0.02; P = 0.29). Shapley Additive Explanation analysis identified key predictive features, which included inferior minimum rim width, retinal nerve fiber layer thicknesses, and average VF sensitivity in the superior-nasal sector.

Integration of ONH structural and VF functional parameters improves the classification of progression risk in patients with PACG beyond conventional clinical models. These findings support the utility of multimodal approaches for enhanced risk stratification.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.