1. Introduction
The Electromyogram (EMG) signals from the remaining muscles after amputation have long been investigated as a source of control for powered prosthetics, to give an opportunity for people with amputations to live and work in a way that was previously difficult [1]. Advanced commercial prostheses, employing pattern recognition (PR) technologies to revolutionize the way muscles’ bioelectrical activity signals are used to control a multifunctional prosthesis, are nowadays available. However, despite the success of EMG driven PR-based prostheses (a.k.a myoelectric prostheses), recent literature has pointed out that limitations arise when exporting such systems from the laboratory to real-life clinical applications, as it is usually found that the clinical accuracy is inferior to that achieved in controlled laboratory environments [2]. For real-time applications, any such controller should theoretically operate under minimal latencies and memory requirements for processing and decision making, all while maintaining the accuracy of movement identification. It has also been reported that the lack of intuitive control renders such a technology to be rejected by amputees given the delays and inaccuracies associated with many systems [3]. Hence, accuracy and reliability are key points to determine the success of such a technology. Several factors contribute to achieving high accuracies for controlling a prosthetic arm and ensure its reliability including for example training and testing the developed algorithms with many datasets, including more subjects (intact-limbed and amputees), selecting more movements, and recording data under different experimental conditions (limb position change, varying contraction force efforts, forearm orientations, etc.). To prove our results, we implement our algorithm using three main databases of surface EMG (sEMG) signals recorded from the forearm: BioPatrec-database, 3DC-database, and Forces- database.