Deriving phenotype-representative left ventricular flow patterns by reduced-order modeling and classification.

Area of Science:

• Cardiovascular Imaging
• Biomedical Engineering
• Computational Fluid Dynamics

Background:

• Clinical translation of advanced cardiac flow imaging is hindered by challenges in extracting representative flow patterns and metrics.
• Reduced-order models (ROMs) offer a promising strategy for deriving simple, interpretable intraventricular flow metrics.
• Integrating ROMs with machine learning (ML) can enhance diagnosis and risk stratification in cardiac patients.

Purpose of the Study:

• To investigate the utility of ROMs derived from 2D color-Doppler echocardiograms for classifying cardiac conditions.
• To develop and validate a simple, interpretable metric for differentiating between non-ischemic dilated cardiomyopathy (DCM), hypertrophic cardiomyopathy (HCM), and healthy controls.
• To explore the potential of ML-based analysis of ROMs for clinical applications in cardiac flow assessment.

Main Methods:

• Proper Orthogonal Decomposition (POD) was applied to 2D color-Doppler echocardiograms from 81 DCM patients, 51 HCM patients, and 77 controls to build patient- and cohort-specific ROMs.
• Three ML classifiers were tested on ROMs, with hyperparameter optimization used to maximize classification power in supervised models.
• Vector flow mapping was employed for visualization and interpretation of flow patterns and ML results.

Main Results:

• POD-based ROMs effectively represented all cohorts, with the principal mode capturing over 80% of flow kinetic energy.
• The ratio of kinetic energy between the second (vortex) and first (jet) POD modes, termed the vortex-to-jet (V2J) energy ratio, emerged as a key discriminating metric.
• The V2J ratio achieved high accuracy in differentiating between DCM, HCM, and control groups, with areas under the ROC curve ranging from 0.81 to 0.95.

Conclusions:

• Modal decomposition using POD can generate ROMs that capture essential cardiac flow dynamics.
• Simple, interpretable flow metrics, such as the V2J energy ratio, can be derived from these ROMs.
• These metrics demonstrate significant potential for discriminating between cardiac disease states and are well-suited for ML-based analysis.