Robust classification of a neuromuscular signal for real-time control of a prosthetic hand

New technological processes that allowed the development of ever smaller and more performing electrodes and simultaneously the advancement of research in AI field has made possible to integrate machine with human body, leading to significant progress in neuroscience field. This new technology is expected to allow neuroscientists to make great strides working on brain signals decoding where large number of channels are generally required as in the case of EEG and in particular of ECoG. Eventually, this progress is expected to lead to a new generation of brain-computer Interfaces (BCIs) that have the potential to restore lost sensory motor function, caused for example by spinal cord injury or neurodegenerative disease. This Master Thesis aims at developing a strategy for Continuous Control (CC) of a prosthetic hand to evaluate how a Machine Learning algorithm can enable a Non-Human-Primate to control a number of DOF. Before focusing on continuous DOF control, I focused on a discrete classification task based on a dataset collected by researchers of DPZ (Deutsches Primatenzentrum), who recorded brain activity via implanted microelectrode arrays (FMAs) from NHPs trained to grasp different objects that vary in size and shape for a total of 48 different objects. Classification task were two: the first was a binary classification with purpose of finding movement activation while the second consist of detection of graspable object. For these tasks I investigated different NN-based approaches, focusing on Bidirectional Recurrent Neural Newtork (BRNN). Eventually, in order to test the capabilities of BRNNs for Continuous Control, due to the lack of proper data readily available, with support of DPZ researchers I created a new virtual kinematics. To do so I grouped objects into 7 macro-classes, based on their shape and the typical grasp behaviour of NHPs. For each object group, a synthetic dataset is recreated containing the values of two coordinates describing the virtual kinematics. As final step I implemented a regression task on this kinematics by means of a BRNN. The overall results are promising, classification task reached an accuracy of 99% for binary classification while for the multi-class classification of 72% of accuracy was reached. For the regression task a global MSE of 0.05 was reached.