A Novel Foundation Model-Based Framework for Multimodal Retinal Age Prediction.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Biomarker Discovery

Background:

• The retinal age gap (RAG) is an emerging biomarker for ocular and non-ocular diseases, traditionally predicted using CNNs on color fundus photography (CFP).
• Multimodal fusion of CFP with optical coherence tomography (OCT) data offers potential for enhanced RAG prediction accuracy and diagnostic utility.

Purpose of the Study:

• To develop a novel foundation model-based framework for multimodal retinal age prediction using CFP and OCT data.
• To investigate the potential of multimodal RAG for non-ocular disease classification.

Main Methods:

• Feature extraction from CFP and OCT images using the RETFound foundation model.
• Innovative fusion strategy to combine features and train a linear regression model for retinal age prediction.
• Utilized UK Biobank data from over 80,000 participants for training and validation.

Main Results:

• The multimodal model achieved a new benchmark in retinal age prediction with a mean absolute error of 2.75 years.
• Outperformed traditional CNN and single-modality approaches in accuracy.
• Multimodal RAG values showed superior performance in classifying diabetes mellitus type 1, multiple sclerosis, and chronic kidney disease.

Conclusions:

• Multimodal fusion of CFP and OCT significantly enhances retinal age prediction and RAG-based analyses.
• The proposed approach improves disease classification accuracy and shows potential for clinical integration.
• Retinal age gap analysis using multimodal imaging can serve as a scalable, non-invasive screening tool.