Characterization of myocardial motion patterns by unsupervised multiple kernel learning.

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning

Background:

• Heart failure with preserved ejection fraction (HFPEF) presents complex challenges in diagnosis and management.
• Current methods for assessing cardiac function during stress may not fully capture the heterogeneous responses in HFPEF.

Purpose of the Study:

• To develop and validate an objective, unsupervised machine learning method for characterizing functional responses to stress in HFPEF.
• To integrate diverse data sources, including myocardial velocity traces and cardiac event timing, for a comprehensive analysis.

Main Methods:

• Utilized multiple kernel learning (MKL) to combine temporally-aligned myocardial velocity data and cardiac event information (valve openings/closures, atrial activation).
• Employed unsupervised learning to explore intrinsic variability in cardiac function without pre-defined labels.
• Reconstructed velocity traces via multiscale kernel regression for enhanced physiological interpretation of the machine learning output space.

Main Results:

• The proposed method successfully characterized distinct patterns of functional responses to stress in a cohort including healthy individuals, HFPEF patients, and breathless subjects.
• Analysis of 2D echocardiography sequences from 55 subjects demonstrated the method's ability to differentiate cardiac responses.
• The joint analysis of multiple features significantly improved the characterization of myocardial functional response in HFPEF.

Conclusions:

• The developed objective method offers a novel approach to understanding the complex pathophysiology of HFPEF during physiological stress.
• Integrating multimodal data with advanced machine learning techniques provides deeper insights into cardiac mechanics.
• This methodology holds potential for improved diagnostic and prognostic capabilities in HFPEF assessment.