Long-Time Prediction of Arrhythmic Cardiac Action Potentials Using Recurrent Neural Networks and Reservoir Computing.

Shahrokh Shahi,Christopher D Marcotte,Elizabeth M Cherry,Yohannes Shiferaw,Flavio H Fenton,Conner J Herndon

doi:10.3389/fphys.2021.734178

Abstract

The electrical signals triggering the heart's contraction are governed by non-linear processes that can produce complex irregular activity, especially during or preceding the onset of cardiac arrhythmias. Forecasts of cardiac voltage time series in such conditions could allow new opportunities for intervention and control but would require efficient computation of highly accurate predictions. Although machine-learning (ML) approaches hold promise for delivering such results, non-linear time-series forecasting poses significant challenges. In this manuscript, we study the performance of two recurrent neural network (RNN) approaches along with echo state networks (ESNs) from the reservoir computing (RC) paradigm in predicting cardiac voltage data in terms of accuracy, efficiency, and robustness. We show that these ML time-series prediction methods can forecast synthetic and experimental cardiac action potentials for at least 15–20 beats with a high degree of accuracy, with ESNs typically two orders of magnitude faster than RNN approaches for the same network size.

Highlights

Cardiac electrical signals, known as action potentials, exhibit complex non-linear dynamics, including period-doubling bifurcations in their duration (Guevara et al, 1984; Watanabe et al, 2001) and amplitude (Chen et al, 2017), along with higher-order period-doublings (Gizzi et al, 2013) and chaotic behavior (Chialvo et al, 1990)
The hybrid echo state networks (ESNs) does the best job of matching voltage values during these phases; the ESN approaches produce good results during the plateau but show depolarization preceding each action potential rather than remaining at a stable rest potential
The long short-term memory (LSTM) and gated recurrent units (GRUs) methods show the largest discrepancies, including plateau height mismatches and significant slowing in repolarization leading to elevated resting potentials

Summary

Introduction

Known as action potentials, exhibit complex non-linear dynamics, including period-doubling bifurcations in their duration (Guevara et al, 1984; Watanabe et al, 2001) and amplitude (Chen et al, 2017), along with higher-order period-doublings (Gizzi et al, 2013) and chaotic behavior (Chialvo et al, 1990). Data-driven approaches can be used to forecast systems like cardiac action potentials by inferring the dynamics from observed data represented as time series (Kutz, 2013). Gated RNNs can help overcome some of these problems; for example, to overcome the vanishing gradient problem, gated RNNs take advantage of memory cell architecture and a gating mechanism allowing the network to select which information should be kept and which forgotten (Hochreiter and Schmidhuber, 1997). This process enables the network to learn the long-term dependencies in sequential temporal data. Two widely used gated RNN approaches include long short-term memory (LSTM) networks and gated recurrent units (GRUs)

Methods

Results

Conclusion