A deep CNN framework for neural drive estimation from HD-EMG across contraction intensities and joint angles

Abstract

Objective: Previous studies have demonstrated promising results in estimating the neural drive to muscles, the net output of all motor neurons that innervate the muscle, using high-density electromyography (HD-EMG) for the purpose of interfacing with assistive technologies. Despite the high estimation accuracy, current methods based on neural networks need to be trained with specific motor unit action potential (MUAP) shapes updated for each condition (i.e., varying muscle contraction intensities or joint angles). This preliminary step dramatically limits the potential generalization of these algorithms across tasks. We propose a novel approach to estimate the neural drive using a deep convolutional neural network (CNN), which can identify the cumulative spike train (CST) through general features of MUAPs from a pool of motor units. Methods: We recorded HD-EMG signals from the gastrocnemius medialis muscle under three isometric contraction scenarios: 1) trapezoidal contraction tasks with different intensities, 2) contraction tasks with a trapezoidal or sinusoidal torque target, and 3) trapezoidal contraction tasks at different ankle angles. We applied a convolutive blind source separation (BSS) method to decompose HD-EMG signals to CST and segmented both signals into windows to train and validate the deep CNN. Then, we optimized the structure of the deep CNN and validated its generalizability across contraction tasks within each scenario. Results: With the optimal configuration for the HD-EMG data window (overlap of 20 data points and window length of 40 data points), the deep CNNestimated the CST close to that from BSS, with a correlation coefficient higher than 0.96 and normalized root-mean-square-error lower than 7% with respect to the BSS (golden standard) within each scenario. Conclusion: The proposed deep CNN framework can utilize data from different contraction tasks (e.g., different intensities), learn general features of MUAP variants, and estimate the neural drive for other contraction tasks. Significance: With the proposed deep CNN, we could potentially build a neural-drive-based human-machine interface that is generalizable to different contraction tasks without retraining.

Toward a generalizable deep CNN for neural drive estimation across muscles and participants

“…Then, spike indices were adjusted to align with the center of the MUAP. After pre-processing, we used the same sliding window method as in [34] to segment the HD-EMG signals and CST into small windows. For the HD-EMG windowing, we used a window length of 40 data points and sliding step of 20 data points with an overlap of 20 data points (Fig.…”

“…We have previously developed a deep CNN framework to directly estimate the neural drive in the form of the CST [34]. The deep CNN identified the number of MU spikes in a given window of HD-EMG signals.…”

“…However, if there were more than 4 spikes in the window, we set all four nodes to 1. Therefore, we might lose some of the spikes in the training and testing data, but this is the best choice when comparing the output of the deep CNN with the full CST extracted from the BSS [34].…”

“…The full structural design of the deep CNN is presented in [34]. In short, the deep CNN consists of six parametric layers (four convolutional layers and two fully connected layers) and three non-parametric layers (two max pooling layers and one flatten layer) between the input and output layers.…”

“…A deep CNN is chosen because it demonstrated success 1) in many complex image pattern recognition problems regardless of the orientation and distortion of the objects [30, 31]; 2) in EMG and electroencephalography analysis as well as MU identification using HD-EMG [32, 33]. In addition, we have demonstrated that the deep CNN is feasible for neural drive estimation across different contraction intensities and joint angles in our previous study [34]. As the deep CNN is a non-linear complex feature extraction approach, it could handle more challenging situations and store much more information than a linear separation matrix.…”