Beyond unimodal analysis: Multimodal ensemble learning for enhanced assessment of atherosclerotic disease progression

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence
• Medical Imaging

Background:

• Atherosclerosis is a major cause of cardiovascular disease, leading to plaque rupture and stroke.
• Current risk scores (e.g., Framingham) and ultrasound markers (e.g., carotid intima-media thickness) have limited sensitivity for early detection.
• Existing models often analyze clinical or imaging data separately, lacking a comprehensive multimodal approach.

Purpose of the Study:

• To develop and evaluate a multimodal ensemble learning framework for assessing atherosclerosis severity, particularly in its sub-clinical stages.
• To integrate features from clinical risk factors and ultrasound imaging for a more accurate cardiovascular risk assessment.
• To measure the efficacy of multimodal models in evaluating vascular aging (plaque presence, vascular age) over six years.

Main Methods:

• Development of a multimodal ensemble learning framework using multi-objective optimization for performance and diversity.
• Integration of clinical risk factors and ultrasound image-derived features.
• Application of eXplainable Artificial Intelligence (XAI) techniques to understand model drivers and identify subject subgroups.

Main Results:

• The multimodal framework effectively integrates diverse data sources for improved atherosclerosis assessment.
• The study demonstrates the efficacy of combining clinical and imaging data for evaluating vascular aging.
• XAI techniques provided insights into key predictive features and distinct patient subgroups.

Conclusions:

• Multimodal ensemble learning offers a superior approach to assessing early-stage atherosclerosis compared to single-modality methods.
• Combining clinical and imaging data enhances the prediction of vascular aging and cardiovascular risk.
• The developed framework and XAI analysis provide valuable tools for personalized cardiovascular risk stratification.