Decoding muscle activity via CNN-LSTM from 3D spatiotemporal EEG

Abstract

Objective. Reconstructing muscle activity from electromyogram (EMG) data using non-invasive electroencephalogram (EEG) signals could lead to significant advancements in brain-computer interfaces (BCIs). However, extracting muscle-related signals from EEG poses considerable challenges due to the mixed nature of signals captured by EEG sensors from various cortical regions. Approach. This study introduces a new method for estimating muscle activity from non-invasive EEG signals while participants performed the grasp and lift (GAL) task. Envelopes of the delta, theta, alpha, beta, and gamma frequency bands were chosen as EEG features for the decoding models, computed similarly to muscle activity (EMG envelopes). These were converted into three-dimensional spatiotemporal matrices based on EEG electrode locations. A deep learning model, combining convolutional neural networks (CNN) for spatial and long short-term memory (LSTM) network for temporal EEG information extraction, was applied. This model was compared with two linear and nonlinear decoding methods: multivariate linear regression (mLR) and multilayer perceptron (MLP). Main Results. The average ± standard deviation of the normalized root mean square error (nRMSE), coefficient of determination (R²), and correlation coefficient (CC) between the estimated and actual muscle activity of two muscles in five participants were 0.21 ± 0.05, 0.54 ± 0.17, and 0.76 ± 0.10, respectively. The CNN-LSTM model outperformed both mLR and MLP approaches (p-value < 0.016), with higher frequencies proving more effective for decoding. Significance. The proposed model effectively captures nonlinear relationships between brain and muscle activities, indicating its potential to enhance the accuracy and reliability of non-invasive BCIs.