Assessment of Heart Rhythm and Arrhythmia Detection from Electromyographic Signals Using a Convolutional Neural Network

TL;DR

Researchers developed a neural network to detect heart rhythms and arrhythmias from muscle signals, showing potential for wearable health monitoring.

Contribution

A convolutional autoencoder was developed to extract rhythmograms from EMG signals for arrhythmia detection.

Findings

The convolutional autoencoder achieved high F-scores for extracting rhythmograms from EMG signals at rest and during static exercise.

The model demonstrated strong generalization when tested on data from subjects not included in training.

The system shows potential for cost-effective wearable monitoring of heart rhythm and muscle activity.

Abstract

The aim of the study is to develop and validate an algorithm based on a convolutional autoencoder for assessing heart rhythm from surface electromyography (EMG) signals recorded by the EMG system Myosuit. The use of this system is intended to expand the capabilities for subsequent development of integrated systems for non-invasive monitoring of functional status (fatigue, stress, and cardiac arrhythmias).

The study involved 6 healthy male subjects (mean age 21±2 years). Synchronous recording of EMG signals was performed using the EMG system Myosuit, the VNS-Micro electrocardiograph, and the Polar H10 monitor at rest, as well as during static and dynamic bicep contractions.

A fully convolutional autoencoder, trained on binary R-wave masks, was developed to extract the corresponding components of the cardiac cycles from the EMG recordings. The model’s performance was evaluated using the F-score in three scenarios: (i) on pooled data from all subjects; (ii) with one subject excluded from training and tested on their data (leave-one-subject-out); (iii) on the personal data of each subject individually. Based on EMG signals recorded from the pectoralis major muscle by the EMG system Myosuit sensors, the convolutional autoencoder demonstrated the feasibility of constructing a rhythmogram — a sequence of R–R intervals essential for heart rate variability analysis and arrhythmia screening. In Scenario (i), the maximum F-score was achieved for EMG signals at rest and during static bicep exercise; classification performance remained high during dynamic bicep exercise. The results of Scenario (ii) indicate that the system reliably operates when tested on data from a subject excluded from the training set. Scenario (iii) yielded the worst results. Given the algorithm’s high generalization capability for data not included in the training set, Scenario (ii) represents the most realistic application use case and demonstrates the best overall performance.

The feasibility of using the universal EMG electrodes of the EMG system Myosuit for simultaneous monitoring of muscle activity and heart rate has been confirmed. The developed algorithm demonstrates high performance in the task of extracting a rhythmogram from EMG signals both at rest and during muscle load. This opens prospects for creating cost-effective wearable systems for the comprehensive assessment of functional status and real-time screening for cardiac arrhythmias without the need for a separate ECG channel.

Figures