UC San Diego

Extracting explainable flow metrics from cardiac flow imaging modalities for the use of clinical decision making is a challenge especially when our understanding of complex intraventricular hemodynamics is incomplete. We hypothesized that reduced-order models (ROMs) of intraventricular flow are a suitable strategy for deriving simple and interpretable clinical metrics suitable for further assessments. Combined with machine learning (ML) flow-based ROMs could provide new insight to help diagnose, risk-stratify, and monitor patients LV flow. We analyzed color-Doppler echocardiograms of 81 non-ischemic dilated cardiomyopathy (DCM) patients, 51 hypertrophic cardiomyopathy (HCM) patients, and 77 normal volunteers (Control). To test whether flow-based ROMs provide metrics that discriminate phenotypic differences, we used these ROMs to build three LV flow classifiers with different levels of supervision: supervised, weakly supervised, and quasi-unsupervised.

The first study applies proper orthogonal decomposition (POD) and dynamic mode decomposition (DMD) to obtain cohort-specific ROMs of LV flow. Each ROM aggregates a low number of components representing a spatially dependent velocity map. These components/hyperparameter were found using gridsearch to derive supervised ROMs that maximize classification power. POD-based and DMD-based ROMs stably represented each cohort through repeated 10-fold cross-validation. POD cohort-representative models slightly outperformed DMD cohort-representative ROM with POD mode captured >80% of the flow kinetic energy (KE) in all cohorts. This classifier performed favorably, with AUCs exceeding 0.75 and going up to 0.93.

The second part of the study explores whether there are intrinsic properties of patient-represented POD ROM that can differentiate among phenotypes. Semi-unsupervised classification using patient-specific ROMs revealed that KE ratio of these two principal modes, the vortex-to-jet (V2J) energy ratio, is a simple, interpretable metric that discriminates DCM, HCM, and Control patients. Receiver operating characteristic curves using V2J as classifier had areas under the curve ranging from 0.81-0.95 for distinguishing between phenotypes.

Modal decomposition of cardiac flow can be used to create ROMs of normal and pathological flow patterns, uncovering simple interpretable flow metrics with power to discriminate disease states, and particularly suitable for further processing using ML. Future work includes extending ROM for longitudinal studies by considering trajectory inference modeling for hemodynamic monitoring of patients with end-stage DCM.