DeepPressure: Non-invasive estimation of the left ventricle pressure-volume loop using cardiac strains

Introduction: Accurate assessment of the pressure-volume (P-V) loop of the left ventricle (LV) is crucial in diagnosing and managing various cardiac diseases. However, current methods of measuring the P-V loop require invasive cardiac catheterization. Here, we propose a novel, non-invasive approach that estimates the P-V loop of the LV via a machine learning (ML) model using only cardiac strains as inputs. In this approach we demonstrate the remarkable performance of a multi-fidelity ML model that combines in silico training data with in vivo human data to predict the P-V loop. We developed a library of high-fidelity finite-element (FE) rodent computational cardiac models (RCCMs) which simulate the full cardiac cycle. Cardiac strains at end-diastole (ED) and end-systole (ES) from the RCCMs were used as low-fidelity training data. Additionally, we incorporated high-fidelity training data in the form of strains extracted from cardiac magnetic resonance (CMR) imaging from human patients. With this approach, we demonstrate the potential for the use of ML models, trained primarily on synthetic data and augmented with small amounts of human data, in accurately predicting important hemodynamic metrics. The results of this work indicate the feasibility of image-based characterization of the P-V loop, reducing the risk of invasive procedures to patients, thus improving diagnostic capabilities and patient care.

Materials and

Methods: High-fidelity FE RCCMs were developed using in vivo hemodynamic measurements and ex vivo imaging data. A comprehensive ML training dataset of 500 examples of RCCMs was developed using latin hypercube sampling to span the geometry-architecture-loading space. A forward simulation of each example simulated the complete cardiac cycle, generating a unique P-V loop. Cardiac circumferential, radial, and longitudinal (CRL) strains calculated at ED and ES and were mapped to a standard 17-segment AHA representation. The AHA representation was processed by conversion to grayscale and normalization, and then mapped to RGB channels, each of which represented one CRL strain component. The processed strain data was then used to train a ML model based on the UNet architecture. The model was then extended into a multi-fidelity composite neural network via the incorporation of in vivo human strain data that was processed in the same manner as the RCCM data.

Results, Conclusions, and Discussions: The presented approach made use exclusively of cardiac strain data that can be readily derived from routine imaging protocols to predict the full P-V loop of the LV. Results indicated that the augmentation of model-based training data with small amounts of human data can significantly improve the model’s accuracy when applied to human-derived images. Thus, the general difficulty of obtaining and processing large amounts of patient data to train ML models could be alleviated through a multi-fidelity approach similar to that presented in this work. Further development of the model we presented could indeed augment, and potentially replace, the current invasive means of capturing the P-V loop, minimizing potential adverse effects associated with invasive procedures. This advancement hold promise to transform current diagnostic protocols and optimize patient safety.

Acknowledgements (Optional):