EMG signal decomposition

To enable computation of motor neuron synergies, we applied convolution kernel compensation (CKC) motor unit decomposition (Fig. 3B; see below, Synergy extraction by NMF) to obtain the correspondent neural command received by each motor unit from the respective innervating motor neuron. The HD-sEMG signals were decomposed separately for each grid into the constituent motor unit activities by the CKC algorithm (Holobar and Zazula, 2007) to obtain the firing output of the respective motor neurons. To assess the accuracy of motor unit identification from HD-sEMG, we adopted the pulse-to-noise ratio (PNR; Holobar et al., 2014) as a signal-based metrics. The motor unit innervation pulse trains (Holobar and Zazula, 2007) extracted by the algorithm were manually inspected by experts. This inspection led to rejecting all the motor units with PNR lower than 25 dB, corresponding to a confidence interval of 0–70%, discarding the false-positive peaks over the physiological firing rate (35–40 Hz), and including false-negative undetected peaks (Del Vecchio et al., 2020). To identify the same motor neuron across different grip types, motor neurons were tracked across all grip types by decomposing the concatenated EMG signals from all tasks for each subject. By doing so, the decomposition algorithm identified the activity of the same motor unit (thus, the same motor neuron) when it fired in one or more grip types. To evaluate the capacity of the decomposition algorithm to identify the same motor units across different grip types despite the variability in the level of muscle contraction, an analysis based on spike-triggered averaging (STA) was conducted. For each subject, we performed a STA for each motor unit and for each grip type to estimate the motor unit action potential waveforms. We then computed the 2D cross-correlation between these estimated action potential waveforms (Martinez-Valdes et al., 2017) both for the same motor unit (across different grip types) and for different motor units (for each grip type separately). The aim of this analysis was to prove the stability of the estimated action potential waveforms across different tasks and the fact that motor unit action potentials assigned to the same motor unit were more similar than action potentials assigned to different motor units. Similar analyses have been proposed in previous work for validating the tracking of motor units by EMG decomposition across a variety of conditions (see Del Vecchio et al., 2019).

A binary sequence was created to represent each spike train, with 1 indicating the occurrence of a spike and 0 otherwise, having the length of the original EMG signals. The binary sequence of spikes for each detected motor unit was smoothed by a fourth-order Butterworth low-pass filter with cutoff frequency of 2.5 Hz and then normalized between 0 and 1. This frequency bandwidth corresponds to the one obtained using a 400-ms Hanning window, as done in previous work (De Luca et al., 1982; Negro et al., 2009). The smoothed discharge rates (SDRs) provide an estimation of the instantaneous firing rate (De Luca et al., 1982) for each motor unit. We used sinusoidal contractions to focus on the modulation of motor neuron discharge rates rather than the average baseline value. We note that the baseline of the sinusoidal SDR signals would be determined by the average discharge rates, whereas motor neuron discharge rate oscillations around the average value would be determined by the modulation in discharge rates.

A motor neuron was considered active for a given grip type when discharging at least 20 firings in the 10-s interval analyzed. The percentage of motor neurons active in two tasks was quantified for each task pair.