Interpretable Multimodal Learning for Cardiovascular Hemodynamics Assessment

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence
• Medical Imaging

Background:

• Pulmonary Arterial Wedge Pressure (PAWP) is a key indicator for heart failure detection.
• Current gold standard for PAWP assessment is invasive Right Heart Catheterization.
• Non-invasive screening methods are crucial for large-scale patient populations.

Purpose of the Study:

• To develop a multimodal learning pipeline for predicting PAWP.
• To integrate data from Cardiac Magnetic Resonance Imaging (CMR) and Electronic Health Records (EHRs).
• To create an interpretable and effective tool for cardiovascular hemodynamics assessment.

Main Methods:

• Utilized tensor-based learning for spatio-temporal feature extraction from CMR scans.
• Employed a graph attention network to select significant EHR features.
• Implemented early, intermediate, late, and hybrid feature fusion strategies.
• Used a linear classifier and linear fusion for interpretability.

Main Results:

• The multimodal pipeline demonstrated superior performance compared to state-of-the-art methods.
• Validated on a large dataset of 2,641 subjects from the ASPIRE registry.
• Decision curve analysis confirmed the pipeline's utility for population screening.

Conclusions:

• The proposed multimodal learning pipeline offers a superior, interpretable method for PAWP prediction.
• This approach enhances non-invasive screening for cardiovascular conditions like heart failure.
• The pipeline shows significant potential for clinical application in large-scale patient management.