AMD-Mamba: A Phenotype-Aware Multi-modal Framework for Robust AMD Prognosis

Area of Science:

• Ophthalmology and Artificial Intelligence
• Biomedical Informatics

Background:

• Age-related macular degeneration (AMD) is a primary cause of irreversible vision loss, necessitating accurate prognosis for timely intervention.
• Current prognostic methods often focus on limited local features, potentially missing crucial disease progression patterns.

Purpose of the Study:

• To introduce AMD-Mamba, a novel multi-modal framework for AMD prognosis.
• To develop and validate a new AMD biomarker for improved early detection and risk stratification.

Main Methods:

• Developed AMD-Mamba, a multi-modal framework integrating color fundus images, genetic variants, and socio-demographic data.
• Employed a novel metric learning strategy using AMD severity scales for richer feature representation.
• Utilized Vision Mamba to fuse local and global image information, enhancing analysis beyond traditional CNNs.
• Implemented multi-scale fusion combining imaging and clinical variables at different resolutions.

Main Results:

• The proposed AMD biomarker demonstrated significant predictive power for AMD progression.
• AMD-Mamba achieved improved detection of high-risk AMD patients in early stages when combined with existing variables.
• Experimental validation on the AREDS dataset (45,818 images, 52 genetic variants, 3 socio-demographic variables from 2,741 subjects) confirmed the framework's efficacy.

Conclusions:

• AMD-Mamba offers a promising multi-modal approach for more precise AMD prognosis.
• The novel biomarker significantly contributes to identifying individuals at high risk for AMD progression.
• This framework facilitates proactive and personalized management strategies for AMD patients.