Automated Detection of Glaucoma With Interpretable Machine Learning Using Clinical Data and Multimodal Retinal Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness worldwide.
• Early and accurate detection is crucial for effective management and vision preservation.

Purpose of the Study:

• To develop and evaluate a multimodal machine learning model for automated glaucoma detection.
• To identify key features contributing to glaucoma diagnosis and progression.

Main Methods:

• A multimodal deep learning model was trained using macular optical coherence tomography (OCT) volumes, color fundus photographs, and demographic/clinical data from the UK Biobank.
• The study included 863 healthy subjects and 771 subjects with glaucoma.
• Interpretability analysis was performed to understand model predictions and identify important features, including evaluation on subjects who later developed glaucoma (progress-to-glaucoma [PTG]).

Main Results:

• The multimodal model achieved high accuracy in glaucoma detection (area under the curve 0.97).
• Model interpretation highlighted known glaucoma-related features (age, intraocular pressure, optic disc morphology) and suggested novel associations (pulmonary function, retinal outer layers).
• Age and pulmonary function were identified as key predictors for glaucoma progression in PTG subjects.

Conclusions:

• Multimodal data integration significantly enhances glaucoma detection accuracy.
• Interpretable machine learning provides insights into glaucoma pathophysiology and aids in identifying predictive biomarkers.
• The model's ability to predict progression suggests its potential utility in monitoring disease development.