Overview of the approach.

During isometric contractions, electromyographic (EMG) signals can be considered as the sum of all action potentials that originate from the muscle fibres of all the active motor units that lie within the electrodes recording zone. Each of these spike trains are the convolution of a series of discharge times with the recorded profile of motor unit action potentials (MUAP). (A) The shape of the recorded action potentials differs across electrodes when recorded with an array of surface or intramuscular electrodes. The EMG signal and each individual MUAP profile depends on the position of the electrode, as highlighted by the different colours. (B) Decomposing EMG signals consists of solving the inverse problem, that is to estimate the discharge times of the active motor units from the EMG signals. Our software uses a fast independent component analysis (fasICA) to optimize a set of separation vectors (i.e., filters) for each motor unit. To this end, each separation vector is iteratively optimized to maximize the sparseness of the source. At the end of this step, the source is refined, and a K-mean classification is applied to separate the high peaks, which represent the targeted motor unit spikes, from the low peaks (other motor units and noise).

Overview of the online electromyographic (EMG) decomposition approach.

(A1) The participants performed submaximal isometric contractions while tracking a visual target that displayed the level of force/torque exerted upon the dynamometer. During this contraction, EMG signals were recorded with either high-density grids of surface electrodes (as shown for quadriceps) or arrays of intramuscular electrodes (not shown). (A2) Offline EMG decomposition was performed to optimize a separation vector for each motor unit and estimate the centroids of the ‘spikes’ and ‘noise’ classes using K-mean classification. (B) During the online EMG decomposition, the extended EMG signals recorded over 125-ms segments were projected on the separation vectors (motor unit filters), and the peaks were detected using the function ‘islocalmax’. Each peak is classified as a spike or as noise depending on the distance separating them from each centroid. At the end of this process, the motor unit firing activity is translated as visual feedback to the participant, in the form of a raster plot of the discharge times for each motor unit of a given muscle, a quadrant displaying the cumulative spike trains (CST) of two groups of motor units for a given muscle, a quadrant displaying the firing rate of two motor units (not shown), and the smoothed discharge rates of all the identified motor units for a given muscle.

Effect of the manual editing on the quality of motor unit (MU) filters.

Once the participants completed the baseline contraction, we ran an automatic offline decomposition over 150 iterations. As described in the method section, we set a threshold for the silhouette value (SIL) at 0.8 to maximize the number of identified motor units for the vastus lateralis (VL), vastus medialis (VM), gastrocnemius lateralis (GL) and gastrocnemius medialis (GM). Then, the operator removed all the motor units for which the spikes were not clearly separated from the noise (red scatters in A). The remaining motor units were manually edited by the operator (black scatters in A). For the tibialis anterior, the SIL value was set at 0.9. (B) The manual editing consisted of removing false positives and adding the false negatives, before updating the motor unit filter. The effect of this step on the SIL value and the coefficient of variation of the inter-spike intervals (CoV of ISI, without units) is shown on the right panel. The red scatters are the motor units before editing and the green scatters are the motor units after editing. These scatters are connected with a grey vector to show the changes in SIL value and CoV of ISI.

Computational time of the online EMG decomposition.

(A) We considered the computational time for the decomposition as the time between the reception of the EMG signals and the identification of the spikes for all the motor units. We computed the linear regression between the number of identified motor units and the computational time and considered the slope as the computational time per motor unit. Each scatter represents one decomposition, and the colour scheme depends on the muscle. (B) As the sampling frequency differed between recordings with high-density grids and intramuscular arrays of electrodes, we compared the computational times for both techniques with the same number of identified motor units (i.e., 9). (C) and (D) After the decomposition, the motor unit discharge activity was translated into visual feedback, either in the form of a raster plot or the smoothed discharge rates (DR) for all the identified motor units. As for the decomposition, we normalized the computational time per motor unit using a linear regression.

Accuracy of the online EMG decomposition.

We compared the motor unit spike trains automatically identified in real-time with their manually edited version. We calculated the sensitivity, the precision, the false negative rate, and the rate of agreement for each motor unit. Each scatter is an individual motor unit, each box represents the 25th and 75th percentiles of the distribution of values, each bar represents the 5th and 95th percentiles of the distribution of values, and each line is the median. VL: vastus lateralis, VM: vastus medialis, GL: gastrocnemius lateralis, GM: gastrocnemius medialis, TAg: tibialis anterior recorded with a high-density grid of electrodes, and TAi: tibialis anterior recorded with an intramuscular array of electrodes.

Accuracy of the online EMG decomposition with variations of contraction intensity.

We compared the motor unit spike trains automatically identified in real-time with their manually edited version while the contraction intensity was 5% below the level of the target (-5), at the level of the target (0) and 5% above the level of the target (+5). We calculated the sensitivity, the precision, the false negative rate, and the rate of agreement for each motor unit. Each scatter is an individual motor unit, each box represents the 25th and 75th percentiles of the distribution of values, each bar represents the 5th and 95th percentiles of the distribution of values, and each line is the median. VL: Vastus Lateralis, VM: Vastus Medialis, GL: Gastrocnemius Lateralis, GM: Gastrocnemius Medialis.

Accuracy of the biofeedback based on the motor unit smoothed discharge rates.

After completing the online EMG decomposition, we converted the motor unit discharge times into biofeedback based on the firing activity of individual motor units. We estimated the accuracy of the biofeedback by calculating the root mean squared error (RMSE) between the smoothed discharge rates estimated from the motor unit spike trains identified in real-time and from their manually edited version at 20% of MVC and 40% of MVC. Each scatter is an individual motor unit, each box represents the 25th and 75th percentiles of the distribution of values, each bar represents the 5th and 95th percentiles of the distribution of values, and each line is the median. VL: vastus lateralis, VM: vastus medialis, GL: gastrocnemius lateralis, GM: gastrocnemius medialis, TAg: tibialis anterior recorded with a high-density grid of surface electrodes, and TAi: tibialis anterior recorded with an intramuscular array of electrodes.