High density EMG signal decomposition for force prediction: perspective for man-machine interfacing.

The surface electromyogram (EMG) is used to infer the neural drive sent by the central nervous system to the muscles by associating its amplitude to the strength of the neural activation. In addition, the surface EMG is adopted in the design of technologies for motor rehabilitation with the aim to restore the lost limb functionality of amputees. In this perspective, it is possible to translate this biologic signal into commands for controlling intuitively the prosthetic device. Since the surface EMG amplitude constitutes a poor indicator of the neural information, more advanced methods were developed such as the high-density surface EMG signal decomposition. With this work, I focused my efforts on understanding how the central nervous system (CNS) controls the muscle force, by exploiting the discharge timings of the motor units. This might help to lay the foundations for an innovative human-machine interface directly based on the spinal cord circuitry information. In this context, the aim of this thesis is to compare the use of the traditional surface EMG signal versus the offline high-density surface electromyography (HDsEMG) signal decomposition. For this study, I used a Graphical User Interface (GUI )which implement Convolutive Kernel Convolution (CKC). This graphical tool allows the operator to decompose the EMG signal and to perform a manual correction of the possible algorithm errors. Thanks to the decomposition approach, the neural activation sent by the spinal cord to the muscles was explored. The entire analysis is focus on the activity of the tibialis anterior muscle (TA) which was recorded by using a grid of 64 electrodes (HDsEMG). In each experiment, the participants performed 25 isometric ankle dorsiflexion at three different levels of speed (i.e.%MVC/s). In particular, the study consists of two main part: 1. the investigation of the motor units recruitment and the discharge rates modulations to understand how the CNS responds to the different conditions, such as the changing in the contraction speed; 2. the design of two different linear regressor, based on the neural drive and the EMG envelope respectivelly, to estimate the force of the contraction. In particular three single linear regressor (SLR) and the global linear regressor (GLR) were used. Both the regressions were done considering a time domain feature extracted from the recording force, i.e. the contractile impulse (IC). This was calculated as the integral of the force along time. Thus for both the EMG envelope and the neural drive, the regressions were evaluated with the integral of the signals in each contraction, instead of the signal itself. In order to investigate the performance of the regressors, the R2 value was evaluated and statistically compared among the different cases. A higher R2 mean values was obtained with the single linear ND model based regressor and the global ND model-bases regressor. Results suggest that using the HDEMG decomposition similar or even better performance can be achieved for estimating the contractile force of a muscle. Limitsof this work, due to time and facilities constraints, are mainly related to the numerosity of subjects involved for the analysis. Future development should also investigate dynamic contractions and different muscles.