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

Area of Science:

• Computational biology
• Biomedical engineering
• Artificial intelligence in medicine

Background:

• Cardiac arrhythmias stem from complex, non-linear electrical signal dynamics.
• Accurate forecasting of cardiac voltage time series is crucial for early intervention.
• Machine learning (ML) offers potential for precise cardiac signal prediction, but faces challenges in non-linear time-series forecasting.

Purpose of the Study:

• To evaluate the performance of recurrent neural networks (RNNs) and echo state networks (ESNs) for cardiac voltage time-series prediction.
• To compare the accuracy, efficiency, and robustness of different ML approaches in forecasting cardiac action potentials.
• To assess the potential of reservoir computing (RC) for predicting complex cardiac electrical activity.

Main Methods:

• Utilized two distinct recurrent neural network (RNN) architectures.
• Implemented echo state networks (ESNs) within the reservoir computing (RC) framework.
• Tested prediction models on both synthetic and experimental cardiac voltage data.

Main Results:

• Both RNNs and ESNs demonstrated high accuracy in predicting cardiac action potentials for 15-20 beats.
• Echo state networks (ESNs) were found to be approximately two orders of magnitude faster than RNNs for equivalent network sizes.
• The study confirmed the robustness of the evaluated ML methods for cardiac signal forecasting.

Conclusions:

• ML time-series prediction methods, particularly ESNs, show significant promise for accurate and efficient cardiac voltage forecasting.
• ESNs offer a computationally efficient solution for predicting complex cardiac electrical activity, potentially aiding in arrhythmia detection and management.
• These findings highlight the utility of reservoir computing for advancing predictive capabilities in cardiac electrophysiology.