Foundation model-driven distributed learning for enhanced retinal age prediction.

Area of Science:

• Ophthalmology and Medical Imaging
• Machine Learning and Artificial Intelligence
• Biomarker Discovery

Background:

• The retinal age gap (RAG) shows promise as a biomarker for systemic diseases.
• Accurate RAG prediction from fundus images requires robust machine learning models.
• Data diversity limitations hinder the development of generalizable RAG prediction models.

Purpose of the Study:

• To develop a computationally efficient distributed learning framework for retinal age prediction.
• To enable accurate RAG prediction using diverse and potentially limited datasets.
• To enhance the clinical utility of RAG as a disease biomarker.

Main Methods:

• Utilized an 8-bit quantized foundation model (RETFound) for feature extraction from fundus images.
• Employed federated learning (FL) and traveling model (TM) for distributed training of a linear regression head.
• Evaluated the framework using UK Biobank and BRSET datasets, including patients with type 1 diabetes.

Main Results:

• The distributed learning framework achieved performance comparable to centralized methods (MAE ~3.6 years).
• Traveling model (TM) demonstrated faster convergence than federated learning (FL).
• Significantly higher RAG values were observed in type 1 diabetes patients compared to controls (P < .001).

Conclusions:

• The developed framework is computationally and memory efficient, suitable for resource-constrained environments.
• Distributed learning enhances RAG model generalizability by integrating data from diverse populations.
• This approach improves accessibility and clinical utility of RAG as a disease biomarker.