Decoding firings of a large population of human motor units from high-density surface electromyogram in response to transcranial magnetic stimulation.

We describe a novel application of methodology for high-density surface electromyography (HDsEMG) decomposition to identify motor unit (MU) firings in response to transcranial magnetic stimulation (TMS). The method is based on the MU filter estimation from HDsEMG decomposition with convolution kernel compensation during voluntary isometric contractions and its application to contractions elicited by TMS. First, we simulated synthetic HDsEMG signals during voluntary contractions followed by simulated motor evoked potentials (MEPs) recruiting an increasing proportion of the motor pool. The estimation of MU filters from voluntary contractions and their application to elicited contractions resulted in high (>90%) precision and sensitivity of MU firings during MEPs. Subsequently, we conducted three experiments in humans. From HDsEMG recordings in first dorsal interosseous and tibialis anterior muscles, we demonstrated an increase in the number of identified MUs during MEPs evoked with increasing stimulation intensity, low variability in the MU firing latency and a proportion of MEP energy accounted for by decomposition similar to voluntary contractions. A negative relationship between the MU recruitment threshold and the number of identified MU firings was exhibited during the MEP recruitment curve, suggesting orderly MU recruitment. During isometric dorsiflexion we also showed a negative association between voluntary MU firing rate and the number of firings of the identified MUs during MEPs, suggesting a decrease in the probability of MU firing during MEPs with increased background MU firing rate. We demonstrate accurate identification of a large population of MU firings in a broad recruitment range in response to TMS via non-invasive HDsEMG recordings. KEY POINTS: Transcranial magnetic stimulation (TMS) of the scalp produces multiple descending volleys, exciting motor pools in a diffuse manner. The characteristics of a motor pool response to TMS have been previously investigated with intramuscular electromyography (EMG), but this is limited in its capacity to detect many motor units (MUs) that constitute a motor evoked potential (MEP) in response to TMS. By simulating synthetic signals with known MU firing patterns, and recording high-density EMG signals from two human muscles, we show the feasibility of identifying firings of many MUs that comprise a MEP. We demonstrate the identification of firings of a large population of MUs in the broad recruitment range, up to maximal MEP amplitude, with fewer required stimuli compared to intramuscular EMG recordings. The methodology demonstrates an emerging possibility to study responses to TMS on a level of individual MUs in a non-invasive manner.

Motor Unit Identification in the M Waves Recorded by High-Density Electromyogram.

OBJECTIVE: We describe and test the methodology supporting the identification of individual motor unit (MU) firings in the motor response (M wave) to percutaneous nerve stimulation recorded by surface high-density electromyography (HD-EMG) on synthetic and experimental data. METHODS: A set of simulated voluntary contractions followed by 100 simulated M waves with a normal distribution (MU mean firing latency: 10 ms, Standard Deviation - SDLAT 0.1-1.3 ms) constituted the synthetic signals. In experimental condition, at least 52 progressively increasing M waves were elicited in the soleus muscle of 12 males, at rest (REST), and at 10% (C10) and 20% (C20) of maximal voluntary contraction (MVC). The MU decomposition filters were identified from 15-20 s long isometric plantar flexions performed at 10-70% of MVC and, afterwards, applied to M waves. RESULTS: Synthetic signal analysis demonstrated high accuracy of MU identification in M waves (precision ≥ 85%). In experimental conditions 42.6 ± 11.2 MUs per participant were identified from voluntary contractions. When the MU filters were applied to the M wave recordings, 28.4 ± 14.3, 23.7 ± 14.9 and 20.2 ± 13.5 MU firings were identified in the maximal M waves, with individual MU firing latencies of 10.0 ± 2.8 (SDLAT: 1.2 ± 1.2), 9.6 ± 3.0 (SDLAT: 1.5 ± 1.3) and 10.1 ± 3.7 (SDLAT: 1.7 ± 1.6) ms in REST, C10 and C20 conditions, respectively. CONCLUSION AND SIGNIFICANCE: We present evidence that supports the feasibility of identifying MU firings in M waves recorded by HD-EMG.

Identification of Motor Unit Firings in H-Reflex of Soleus Muscle Recorded by High-Density Surface Electromyography.

We developed and tested the methodology that supports the identification of individual motor unit (MU) firings from the Hoffman (or H) reflex recorded by surface high-density EMG (HD-EMG). Synthetic HD-EMG signals were constructed from simulated 10% to 90% of maximum voluntary contraction (MVC), followed by 100 simulated H-reflexes. In each H-reflex the MU firings were normally distributed with mean latency of 20 ms and standard deviations (SDLAT) ranging from 0.1 to 1.3 ms. Experimental H-reflexes were recorded from the soleus muscle of 12 men (33.6 ± 5.8 years) using HD-EMG array of 5×13 surface electrodes. Participants performed 15 to 20 s long voluntary plantarflexions with contraction levels ranging from 10% to 70% MVC. Afterwards, at least 60 H-reflexes were electrically elicited at three levels of background muscle activity: rest, 10% and 20% MVC. HD-EMGs of voluntary contractions were decomposed using the Convolution Kernel Compensation method to estimate the MU filters. When applied to HD-EMG signals with synthetic H reflexes, MU filters demonstrated high MU identification accuracy, especially for [Formula: see text] ms. When applied to experimental H-reflex recordings, the MU filters identified 14.1 ± 12.1, 18.2 ± 12.1 and 20.8 ± 8.7 firings per H-reflex, with individual MU firing latencies of 35.9 ± 3.3, 35.1 ± 3.0 and 34.6 ± 3.3 ms for rest, 10% and 20% MVC background muscle activity, respectively. Standard deviation of MU latencies across experimental H-reflexes were 1.0 ± 0.8, 1.3 ± 1.1 and 1.5 ± 1.2 ms, in agreement with intramuscular EMG studies.

High-density electromyography biofeedback during robotic wrist exercises for reducing co-activation of antagonist muscles: a case report.

High-density (HD) electrodes have been introduced in research and diagnostic electromyography. Recent advances in technology offer an opportunity for using the HDEMG signal as biofeedback in stroke rehabilitation. The purpose of this case study was to test the feasibility of using two 5 × 13 electrode arrays for providing real-time HDEMG biofeedback and the preliminary outcome of combining HDEMG biofeedback with robotic wrist exercises over 4 weeks in a person who suffered a stroke 26 months earlier. The isometric wrist flexion/extension task required to keep the paretic agonist activity within variable preset limits with minimal activation of the antagonists. The participant was able to utilize the provided biofeedback interface and after eight sessions significantly decreased co-activation in the antagonist wrist extensor muscles during isometric wrist flexion. The HDEMG biofeedback seems feasible and may be used alone or in combination with robotic therapy for increasing the selectivity of muscle activation after stroke.

Deep Learning for Robust Decomposition of High-Density Surface EMG Signals.

Blind source separation (BSS) algorithms, such as gradient convolution kernel compensation (gCKC), can efficiently and accurately decompose high-density surface electromyography (HD-sEMG) signals into constituent motor unit (MU) action potential trains. Once the separation matrix is blindly estimated on a signal interval, it is also possible to apply the same matrix to subsequent signal segments. Nonetheless, the trained separation matrices are sub-optimal in noisy conditions and require that incoming data undergo computationally expensive whitening. One unexplored alternative is to instead use the paired HD-sEMG signal and BSS output to train a model to predict MU activations within a supervised learning framework. A gated recurrent unit (GRU) network was trained to decompose both simulated and experimental unwhitened HD-sEMG signal using the output of the gCKC algorithm. The results on the experimental data were validated by comparison with the decomposition of concurrently recorded intramuscular EMG signals. The GRU network outperformed gCKC at low signal-to-noise ratios, proving superior performance in generalising to new data. Using 12 seconds of experimental data per recording, the GRU performed similarly to gCKC, at rates of agreement of 92.5% (84.5%-97.5%) and 94.9% (88.8%-100.0%) respectively for GRU and gCKC against matched intramuscular sources.

On the Selection of Neural Network Architecture for Supervised Motor Unit Identification from High-Density Surface EMG.

In the last decade, accurate identification of motor unit (MU) firings received a lot of research interest. Different decomposition methods have been developed, each with its advantages and disadvantages. In this study, we evaluated the capability of three different types of neural networks (NNs), namely dense NN, long short-term memory (LSTM) NN and convolutional NN, to identify MU firings from high-density surface electromyograms (HDsEMG). Each type of NN was evaluated on simulated HDsEMG signals with a known MU firing pattern and high variety of MU characteristics. Compared to dense NN, LSTM and convolutional NN yielded significantly higher precision and significantly lower miss rate of MU identification. LSTM NN demonstrated higher sensitivity to noise than convolutional NN.Clinical Relevance-MU identification from HDsEMG signals offers valuable insight into neurophysiology of motor system but requires relatively high level of expert knowledge. This study assesses the capability of self-learning artificial neural networks to cope with this problem.

Estimation of Muscle Co-Activations in Wrist Rehabilitation After Stroke is Sensitive to Motor Unit Distribution and Action Potential Shapes.

We evaluated different muscle excitation estimation techniques, and their sensitivity to Motor Unit (MU) distribution in muscle tissue. For this purpose, the Convolution Kernel Compensation (CKC) method was used to identify the MU spike trains from High-Density ElectroMyoGrams (HDEMG). Afterwards, Cumulative MU Spike Train (CST) was calculated by summing up the identified MU spike trains. Muscle excitation estimation from CST was compared to the recently introduced Cumulative Motor Unit Activity Index (CAI) and classically used Root-Mean-Square (RMS) amplitude envelop of EMG. To emphasize their dependence on the MU distribution further, all three muscle excitation estimates were used to calculate the agonist-antagonist co-activation index. We showed on synthetic HDEMG that RMS envelopes are the most sensitive to MU distribution (10 % dispersion around the real value), followed by the CST (7 % dispersion) and CAI (5 % dispersion). In experimental HDEMG from wrist extensors and flexors of post-stroke subjects, RMS envelopes yielded significantly smaller excitations of antagonistic muscles than CST and CAI. As a result, RMS-based co-activation estimates differed significantly from the ones produced by CST and CAI, illuminating the problem of large diversity of muscle excitation estimates when multiple muscles are studied in pathological conditions. Similar results were also observed in experimental HDEMG of six intact young males.

Automatic Identification of Individual Motor Unit Firing Accuracy From High-Density Surface Electromyograms.

We introduce an algorithm for automatic identification of true positive (TP) and false positive (FP) spikes in the motor unit spike train, identified by blind source separation (BSS) of high-density surface electromyograms (HDsEMG). The algorithm selects predefined number of spikes, so called witnesses, from identified spike train. The other spikes in the spike train are called test spikes and are classified into TP or FP spikes by our algorithm. For this purpose, the algorithm constructs as many motor unit filters as there are test spikes, using the information from all the witnesses and each individual test spike. Afterwards, it applies each motor unit filter to HDsEMG to get new estimate of MU spike train for each selected test spike and calculates previously introduced Pulse-to-Noise Ratio (PNR) on preselected witnesses in this new spike train. When accumulated over all the test spikes, these PNR values exhibit bimodal distribution with the peak at lower PNR values representing FPs and the peak at higher PNR values representing TPs. Therefore, FPs and TPs can be discriminated by applying computationally efficient segmentation algorithm to corresponding PNR values. We also propose and mutually compare different witness selection strategies and show that selection of about 40 spikes with maximal amplitude in the identified spike train minimizes the selection of FPs as witnesses and maximizes the TP vs. FP discrimination power. In our tests on 20 s long experimental HDsEMG signals from biceps brachii muscle the number of FPs decreased from 23.9 ± 4.7 to 4.1 ± 4.4 when the proposed algorithm was used.