FORGE: Functionally-guided OCT Representation for Glaucoma Endophenotyping

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma phenotyping is complex due to disease variability and reliance on single data types.
• Current methods struggle to capture the full spectrum of glaucoma progression.
• Integrating structural and functional data is crucial for accurate disease characterization.

Purpose of the Study:

• To develop a visual field (VF)-guided entangled learning framework for glaucoma phenotyping.
• To integrate structural (macular retinal nerve fiber layer - mRNFL) and functional (VF) signals for improved phenotyping.
• To enable OCT-only inference after training with multimodal data.

Main Methods:

• Developed a VF-guided entangled learning framework.
• Integrated structural mRNFL and functional VF data during training.
• Validated the framework on MEEI (n=5,372) and UK Biobank (n=74,077) datasets.
• Performed genetic analyses to compare phenotyping approaches.

Main Results:

• Identified 9 distinct mRNFL phenotypes with varying progression rates (MD slopes -0.2 to -1.8 dB/year).
• Achieved superior clustering accuracy compared to OCT-only methods (22% FCM, 11% GMM).
• Demonstrated generalizability and improved risk association in external validation (r=-0.33).
• Discovered 12 additional glaucoma loci through genetic analysis.

Conclusions:

• VF-guided entangled learning significantly enhances the clinical and genetic coherence of mRNFL phenotyping.
• This multimodal approach improves glaucoma classification and understanding of disease progression.
• The framework shows broad applicability for multimodal medical imaging analysis in ophthalmology.