Reduced Order Models of Intraventricular Flow for Interpretable Phenotypic Classification of Heart Failure Patients

Multi-dimensional, high-resolution cardiac flow imaging is increasingly available in the clinical setting, but the adoption of flow-related metrics into clinical decision support systems is lagging behind this technological growth. A bottleneck in this adoption is deriving explainable metrics that identify disease states. Heart failure (HF) is no exception.

Our hypotheses to address this challenge are: 1) Left ventricular (LV) flow patterns contain enough data to phenotype patients with different HF subtypes; 2) Applying machine learning (ML) to data-driven reduced-order models (ROMs) of LV flow can discover simple, interpretable metrics for LV flow phenotyping. We analyzed >200 color-Doppler echocardiograms of healthy volunteers (VOL) and patients with dilated (DCM) or hypertrophic (HCM) cardiomyopathy. We applied proper orthogonal decomposition (POD) to build ROMs of LV flow. We tested classifiers using purposely simple ML analyses of the ROMs and requiring different supervision levels. We ran vector flow mapping to visualize 2D flow patterns (e.g., diastolic vortex) and interpret the ML results. To compare data across patients and blind classifiers to chamber geometry, all Doppler sequences were mapped into a common cubic domain using anatomical landmarks and events of the cardiac cycle.

The POD-based ROMs stably represents each cohort's flow patterns through 10-fold cross-validation. The 1st POD mode captures >80% of the flow kinetic energy (KE) and represents the LV filling/emptying jets. The 2nd mode represents the diastolic vortex; its KE contribution ranges from 8% (HCM) to 39% (DCM). Semi-unsupervised classification using patient-specific ROMs reveals that the KE ratio of these two modes, the vortex-to-jet (V2J) ratio, is a simple, interpretable metric that accurately clusters DCM, HCM, and VOL patients. Receiver operating characteristic curves using V2J as threshold have areas under the curve equal to 0.81, 0.91, & 0.95 for the classification problems HCM vs. VOL, DCM vs. VOL, & DCM vs. HCM, respectively.

The favorable classification accuracy, obtained with overtly uninvolved ML methods and using readily available input data, opens possibilities to amplify the role of cardiac flow imaging in the early detection and management of HF and the development of personalized treatment strategies.