We describe the method for identification of motor unit (MU) firings from high-density surface electromyograms (hdEMG), recorded during repeated dynamic muscle contractions. New convolutive data model for dynamic hdEMG is presented, along with Pulse-to-Noise Ratio (PNR) metric for assessment of MU identification accuracy and analysis of the impact of MU action potential (MUAP) changes in dynamic muscle contractions on MU identification. We tested the presented methodology on signals from biceps brachii, vastus lateralis and rectus famoris muscles, all during different speeds of dynamic contractions. In synthetic signals with excitation levels of 10%, 30% and 50% and MUAPs experimentally recorded from biceps brachii muscle, the presented method identified 15 ± 1, 18 ± 1 and 20 ± 1 MUs per contraction, respectively, all with average sensitivity and precision > 90% and PNR > 30 dB. In experimental signals acquired during low force contractions of vastus lateralis and rectus femoris muscle, the method identified 9.4 ± 1.9 and 7.8 ± 1.4 MUs with PNR values of 35.4 ± 3.6 and 34.1 ± 2.7 dB. Comparison with previously introduced Convolution Kernel Compensation (CKC) method confirmed the capability of the new method to follow dynamic MUAP changes, also in relatively fast muscle contractions.

Blind source separation (BSS) algorithms, such as the gradient convolution kernel compensation (gCKC), can efciently and accurately decompose high-density surface electromyography (HD-sEMG) signals into the 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.

Neural interfacing is essential for advancing our fundamental understanding of movement neurophysiology and for developing human-machine interaction systems. This can be achieved at different levels of the central nervous system (CNS) and peripheral nervous system (PNS); however, direct neural interfaces with brain and nerve tissues face important challenges and are currently limited to clinical cases of severe motor impairment. Recent advances in electronics and signal processing for recording and analyzing surface electromyographic (sEMG) signals allow for a radically new way of establishing human interfaces by reverse engineering the neural information embedded in the electrical activity of skeletal muscles. This approach provides a window into the spiking activity of motor neurons in the spinal cord. In this article, we present a brief overview of neural interfaces and discuss the properties of multichannel sEMG in comparison to other CNS and PNS recording modalities. We then describe signal processing approaches for neural interfacing from sEMG, with a focus on recent breakthroughs in convolutive blind source separation (BSS) methods and deep learning techniques. When combined, these approaches establish unique noninvasive human-machine interfaces for neurotechnologies, with applications in medical devices and large-scale consumer electronics.

We analyzed the efficiency of Motor Unit (MU) tracking across different experimentally recorded contraction levels of Biceps Brachii (BB), Tibialis Anterior (TA), First Dorsal Interosseous (FDI) and Abductor Digiti Minimi (ADM) muscles and in simulated conditions. We used the Convolution Kernel Compensation (CKC) algorithm to estimate the MU filters from high-density electromyograms (HDEMGs) at contraction levels ranging from 5 % to 70 % of Maximum Voluntary Contraction (MVC), and applied these MU filters to all the recorded contractions levels of the same muscle. For each MU filter estimation-application pair we assessed the number of identified MUs, Pulse-to-Noise Ratio (PNR), Mean Discharge Rate (MDR) and MUAP peak-to-peak (P2P) amplitude, normalized by HDEMG P2P amplitude. In simulated conditions, we also calculated the Precision ( $Pr$ ) and Sensitivity ( $Se$ ) of MU firing identification. We also reported the number of jointly identified MUs for all possible contraction level pairs. The results in simulated conditions confirmed the efficiency of MU filter transfer from low to high contraction levels. The large majority of MUs identified at lower contraction levels were tracked successfully at higher contraction levels, though many of them were not directly identified at higher contraction levels. Sensitivities of identified MU firings were above 97 %, decreasing only by about 1 % when transferring the MU filters from low to high contraction levels. In the experimental conditions the efficiency of the MU tracking depended significantly on the investigated muscle. The highest efficiency was demonstrated in the FDI muscle, where about 90 % and 70 % of MUs identified at the contraction level of MU filter estimation were also tracked at 10 % and 20 % higher contraction level, respectively. These figures decreased to 47 % and 20 % of MUs in TA, and to 21 % and 18 % of MUs in the BB muscle. The observed differences between the MU filter transfer efficiencies are likely caused by the differences in the size and anatomy of the investigated muscles.

This study aimed to determine whether neural drive is redistributed between muscles during a fatiguing isometric contraction, and if so, whether the initial level of common synaptic input between these muscles constrains this redistribution. We studied two muscle groups: triceps surae (14 participants) and quadriceps (15 participants). Participants performed a series of submaximal isometric contractions and a torque-matched contraction maintained until task failure. We used high-density surface electromyography to identify the behavior of 1,874 motor units from the soleus, gastrocnemius medialis (GM), gastrocnemius lateralis (GL), rectus femoris, vastus lateralis (VL), and vastus medialis (VM). We assessed the level of common drive between muscles in the absence of fatigue using a coherence analysis. We also assessed the redistribution of neural drive between muscles during the fatiguing contraction through the correlation between their cumulative spike trains (index of neural drive). The level of common drive between VL and VM was significantly higher than that observed for the other muscle pairs, including GL-GM. The level of common drive increased during the fatiguing contraction, but the differences between muscle pairs persisted. We also observed a strong positive correlation of neural drive between VL and VM during the fatiguing contraction (r = 0.82). This was not observed for the other muscle pairs, including GL-GM, which exhibited differential changes in neural drive. These results suggest that less common synaptic input between muscles allows for more flexible coordination strategies during a fatiguing task, i.e., differential changes in neural drive across muscles. The role of this flexibility on performance remains to be elucidated. NEW & NOTEWORTHY Redundancy of the neuromuscular system theoretically allows for a redistribution of the neural drive across muscles (i.e., between-muscle compensation) during a fatiguing contraction. Our results suggest that a high level of common input between muscles (e.g., vastus lateralis and medialis) represents a neural constraint making it less likely to redistribute the neural drive across these muscles. In this way, redistribution was only observed across muscles that share little common synaptic input (e.g., gastrocnemius lateralis and medialis).

Persistent inward currents (PICs) are crucial for initiation, acceleration, and maintenance of motoneuron firing. As PICs are highly sensitive to synaptic inhibition and facilitated by serotonin and noradrenaline, we hypothesised that both reciprocal inhibition (RI) induced by antagonist nerve stimulation and whole‐body relaxation (WBR) would reduce PICs in humans. To test this, we estimated PICs using the well‐established paired motor unit (MU) technique. High‐density surface electromyograms were recorded from gastrocnemius medialis during voluntary, isometric 20‐s ramp, plantarflexor contractions and decomposed into MU discharges to calculate delta frequency (ΔF). Moreover, another technique (VibStim), which evokes involuntary contractions proposed to result from PIC activation, was used. Plantarflexion torque and soleus activity were recorded during 33‐s Achilles tendon vibration and simultaneous 20‐Hz bouts of neuromuscular electrical stimulation (NMES) of triceps surae. ΔF was decreased by RI (n = 15, 5 females) and WBR (n = 15, 7 females). In VibStim, torque during vibration at the end of NMES and sustained post‐vibration torque were reduced by WBR (n = 19, 10 females), while other variables remained unchanged. All VibStim variables remained unaltered in RI (n = 20, 10 females). Analysis of multiple human MUs in this study demonstrates the ability of local, focused inhibition to attenuate the effects of PICs on motoneuron output during voluntary motor control. Moreover, it shows the potential to reduce PICs through non‐pharmacological, neuromodulatory interventions such as WBR. The absence of a consistent effect in VibStim might be explained by a floor effect resulting from low‐magnitude involuntary torque combined with the negative effects of the interventions.Key points: Spinal motoneurons transmit signals to skeletal muscles to regulate their contraction. Motoneuron firing partly depends on their intrinsic properties such as the strength of persistent (long‐lasting) inward currents (PICs) that make motoneurons more responsive to excitatory input. In this study, we demonstrate that both reciprocal inhibition onto motoneurons and whole‐body relaxation reduce the contribution of PICs to human motoneuron firing. This was observed through analysis of the firing of single motor units during voluntary contractions. However, an alternative technique that involves tendon vibration and neuromuscular electrical stimulation to evoke involuntary contractions showed less effect. Thus, it remains unclear whether this alternative technique can be used to estimate PICs under all physiological conditions. These results improve our understanding of the mechanisms of PIC depression in human motoneurons. Potentially, non‐pharmacological interventions such as electrical stimulation or relaxation could attenuate unwanted PIC‐induced muscle contractions in conditions characterised by motoneuron hyperexcitability.