Alexander E. Olsson

Processing myoelectrical activity in the forearm has for long been considered a promising framework to allow transradial amputees to control motorized prostheses. In spite of expectations, contemporary muscle–computer interfaces built for this purpose typically fail to satisfy one or more important desiderata, such as accuracy, robustness, and/or naturalness of control, in part due to difficulties in acquiring high-quality signals continuously outside laboratory conditions. In light of such problems, surgically implanted electrodes have been made a viable option that allows for long-term acquisition of intramuscular electromyography (iEMG) measurements of spatially precise origin. As it stands, the question of how information embedded in such signals is best extracted and combined across multiple channels remains open. This study presents and evaluates an approach to this end that uses deep neural networks based on the Long Short-Term Memory (LSTMs) architecture to regress forces exerted by multiple degrees of freedom (DoFs) from multichannel iEMG. Three deep learning models, representing three distinct regression strategies, were evaluated: (I) One-to-One, wherein each DoF is separately estimated by an LSTM model processing a single iEMG channel, (II) All-to-One, wherein each DoF is separately estimated by an LSTM model processing all iEMG channels, and (III) All-to-All, wherein a single LSTM model with access to all iEMG channels estimates all DoFs simultaneously. All models operate on raw iEMG, with no preliminary feature extraction required. When evaluated on a dataset comprising six iEMG channels with concurrent force measurements acquired from 14 subjects, all LSTM strategies were found to significantly outperform a baseline feature-based linear control regression method. This finding indicates that recurrent neural networks can learn to transform raw forearm iEMG signals directly into representations that correlate with forces exerted at the level of the hand to a greater degree than simple features do. Furthermore, the All-to-All and All-to-One strategies were found to exhibit better performance than the One-to-One strategy. This finding suggests that, in spite of the spatially local nature of signals, iEMG from muscles not directly actuating the relevant DoF can provide contextual information that aid in decoding motor intent.

By being predicated on supervised machine learning, pattern recognition approaches to myoelectric prosthesis control require electromyography (EMG) training data collected concurrently with every detectable motion. Within this framework, calibration protocols for simultaneous control of multifunctional prosthetic hands rapidly become prohibitively long—the number of unique motions grows geometrically with the number of controllable degrees of freedom (DoFs). This paper proposes a technique intended to circumvent this combinatorial explosion. Using EMG windows from 1-DoF motions as input and EMG windows from 2-DoF motions as targets, we train generative deep learning models to synthesize EMG windows appertaining to multi-DoF motions. Once trained, such models can be used to complete datasets consisting of only 1-DoF motions, enabling simple calibration protocols with durations that scale linearly with the number of DoFs. We evaluated synthetic EMG produced in this way via a classification task using a database of forearm surface EMG collected during 1-DoF and 2-DoF motions. Multi-output classifiers were trained on either (I) real data from 1-DoF and 2-DoF motions, (II) real data from only 1-DoF motions, or (III) real data from 1-DoF motions appended with synthetic EMG from 2-DoF motions. When tested on data containing all possible motions, classifiers trained on synthetic-appended data (III) significantly outperformed classifiers trained on 1-DoF real data (II), although significantly underperformed classifiers trained on both 1- and 2-DoF real data (I) (I < 0.05). These findings suggest that it is feasible to model EMG concurrent with multiarticulate motions as nonlinear combinations of EMG from constituent 1-DoF motions, and that such modelling can be harnessed to synthesize realistic training data.

The use of natural myoelectric interfaces promises great value for a variety of potential applications, clinical and otherwise, provided a computational mapping between measured neuromuscular activity and executed motion can be approximated to a satisfactory degree. However, prevalent methods intended for such decoding of movement intent from the surface electromyogram (sEMG) based on pattern recognition typically do not capitalize on the inherently time series-like nature of the acquired signals. In this paper, we present the results from a comparative study in which the performances of traditional cross-sectional pattern recognition methods were compared with that of a classifier built on the natural assumption of temporal ordering by utilizing a long short-term memory (LSTM) neural network. The resulting evaluation indicate that the LSTM approach outperforms traditional gesture recognition techniques which are based on cross-sectional inference. These findings held both when the LSTM classifier operated on conventional features and on raw sEMG and for both healthy subjects and transradial amputees.

In contemporary muscle-computer interfaces for upper limb prosthetics there is often a trade-off between control robustness and range of executable movements. As a very low movement error rate is necessary in practical applications, this often results in a quite severe limitation of controllability; a problem growing ever more salient as the mechanical sophistication of multifunctional myoelectric prostheses continues to improve. A possible remedy for this could come from the use of multi-label machine learning methods, where complex movements can be expressed as the superposition of several simpler movements. Here, we investigate this claim by applying a multi-labeled classification scheme in the form of a deep convolutional neural network (CNN) to high density surface electromyography (HD-sEMG) recordings. We use 16 independent labels to model the movements of the hand and forearm state, representing its major degrees of freedom. By training the neural network on 16 × 8 sEMG image sequences 24 samples long with a sampling rate of 2048 Hz to detect these labels, we achieved a mean exact match rate of 78.7% and a mean Hamming loss of 2.9% across 14 healthy test subjects. With this, we demonstrate the feasibility of highly versatile and responsive sEMG control interfaces without loss of accuracy.