Multi-channel integration and analysis of an sEMG-based hand gestures recognition systems for rehabilitation applications

Hand rehabilitation represents a significant priority for individuals with tetraplegia and stroke survivors, given the debilitating impacts of these conditions that persist in 45% of cases after 18 months. However, current rehabilitation practices, e.g. repetitive Transcranial Magnetic Stimulation (rTMS), show limited evidence of their effectiveness. Therefore, the need for new approaches, such as using Functional Electrical Stimulation (FES) systems, emerges. This active rehabilitation technique uses low-intensity electrical pulses to stimulate skeletal muscles, acting on the nervous system by promoting new synaptic connections. The purpose of the thesis is the development of an event-driven sEMG-based system for real-time FES control, aimed at recovering hand functionalities. The project focused on identifying the muscles required for hand motor control, defining the integration of the devices into the acquisition system, and their placement. For synchronous operation of the acquisition units, a control unit, from an existing version, was developed that employs Object-Oriented Programming (OOP) in Python programming language and controls a Graphical User Interface (GUI). The surface ElectroMyoGraphic (sEMG) signal is extracted using two types of devices, made available by the eLiONS Lab. research group. The first is a wearable device, and four of them were used, while the second is a circular ring made up of seven units designed to fit around the forearm. The acquisition units acquire the sEMG signals, which go through direct processing using the Average Threshold Crossing (ATC) technique. Subsequently, ATC data are wirelessly transmitted to the control unit where they are used to compute the stimulation parameters. The ATC paradigm is considered optimal for the sEMG-based FES system as it maintains a high correlation with muscle contraction. The FES system operates with therapist-patient as the main usage mode. This modality enables patients to replicate rehabilitative movements performed by the therapist, as the stimulation patterns applied to the patient’s muscles are defined based on the muscle activation of the therapist. To select the stimulation channels based on the movement performed by the therapist, a classifier was developed. At this stage, an artificial neural network was developed for the multi-class classification of executed movements. These movements were systematically organized within a carefully defined acquisition protocol, delineating the precise timing and execution mode. This protocol includes a selection of nine movements chosen to enhance motor recovery and hand functionality. Data collection was required for the training and testing phases of the classifier. Two types of datasets were then generated: the first comprised signals acquired from 12 subjects, while the second contained data from a single subject, a scenario that could simulate the use of the system personalized to a specific therapist. The classifier’s performance was evaluated in terms of average accuracy, reaching 59.3% and 86.1% for multi-subject and single-subject datasets, respectively. In the first case, accuracy shows high variability in individual movements and overall does not reach an appropriate level for the use of the model. Instead, the results on the single subject enable the integration of the classifier into the control unit. Therefore, hand movement recognition was tested in real-time, showing promising prospects for integration into the FES system.