Elizabeth M. Cherry

Research

Prediction of physiological states using machine learning

Predicting the future dynamics of physiological systems may allow intervention to terminate undesirable states or prevent their occurrence. In several recent papers, our group uses machine-learning techniques to predict complex time series from cardiac action potentials including experimental datasets with good accuracy and over long times compared to previous approaches. We use echo state networks, a type of reservoir computing, that is more efficient to train that other recurrent neural networks because only the output weights are trained. In the cardiac setting, we can predict several seconds (~20 action potentials) into the future.

Cardiac voltage prediction

Prediction of other systems

Estimating physiological system states, dynamics, and model parameter values

Cardiac electrical dynamics within the thickness of the myocardium is difficult to observe directly. Toward this end, our group has developed techniques for estimating these dynamics along with the behavior of unobservable variables using data assimilation. In this approach, we update an estimate of the system and state iteratively by combining observations when available with a prior model forecast from a numerical model, which itself depends on earlier observations. We have shown that our approach using an ensemble Kalman filter is effective and computationally efficient for reconstructing known system states even in the presence of significant model error. We have also applied the technique to predicting the transition of cells from the epithelial to mesenchymal state. In addition, we have used Bayesian approaches, including Hamilton Monte Carlo and Approximate Bayesian Computing Sequential Monte Carlo, to estimate distributions of parameter values of cardiac action potential models fitted to experimental data. Finally, we have also used methods of control theory to analyze the controllability of cardiac cellular electrical dynamics.

Data assimilation: cardiac dynamics

Data assimilation: epithelial-mesenchymal transition

Bayesian approaches for estimating parameter distributions of action potential models

Controllability of cardiac alternans

Efficient numerical methods for cardiac tissue modeling and simulation

Our group has developed several techniques for improved numerical methods for simulating cardiac cells and tissue. A recent approach involves using GPU-accelerated computations in a browser-based setting to allow near-real-time simulations. Some previous methods have focused on efficient techniques to use for three-dimensional complex geometries, and others focus on using space-time adaptivity in the computational mesh representing the solution to improve performance.

GPU computing applications

Complex geometries

Adaptive mesh refinement

Mathematical models of cardiac cellular and tissue electrophysiology

Our group has contributed to the development of several mathematical models of the electrical behavior of cardiac cells and tissue and tools for parameterizing these models. In addition, we have analyzed the behavior of many different cardiac cell models, including comparisons of models for the same cell types and species. We have also studied the use of delay-differential equations for cardiac cell and tissue modeling.

Model development

Model parameterization

Model analysis

Delay-differential equations modeling

Mechanisms for cardiac arrhythmias

In our group. we have used computational models together with experimental data to study multiple mechanisms that can give rise to cardiac arrhythmias. Some recent work in this area has focused on dynamical mechanisms that can give rise to alternans and has shown that cardiac tissue can experience multiple transitions as well as significant spatial heterogeneity and memory effects during the transition to fibrillation. We have also studied the effects of complex structures like the cardiac Purkinje system on arrhythmia development.

Arrhythmia mechanisms

Purkinje modeling and role in arrhythmias

Termination of fibrillation

Our group has worked on multiple methods for terminating fibrillation experimentally and computationally. On the experimental side, we have applied low-energy techniques to terminate fibrillation effectively and efficiently. We have also developed a new method for guiding a catheter to locate atrial fibrillation sources iteratively. In addition, we have shown how an antiarrhythmic agent terminates fibrillation in equine atria and have quantified mechanisms for defibrillation in a computational model of the ventricles.

Low-energy defibrillation

Catheter guidance algorithm

Computational models of defibrillation

Defibrillation via anti-arrhythmic drugs