Study applicability of a BCI system based on EEG signals for the recognition of imagined movements.

Brain-computer interfaces (BCI) are systems capable of interpreting the neural activity of a subject and translating it into a digital output signal. BCIs are recognized by the scientific community as a potential remedy for restoring the motor functions of physically impaired patients. The long-term goal of this study is to restore tetraplegic patients' motor abilities. As they have lost the ability to correctly execute movements, only imagined or attempted movements can be performed. ElectroEncephaloGram (EEG) based BCIs allow the decoding of imagined movements through a non-invasive approach. The drawback of these systems is their limited clinical application. While they achieve good performance within a laboratory, they are rarely built to work in a real-life setting and when they do, they can hardly achieve acceptable performances. The study aims to identify what limitations prevent an EEG-dependent BCI system from being applied in a real-life situation and tries to propose how these limitations can be overcome both from the point of view of the configuration of the experimental protocol and from the point of view of the models applied. The work is therefore divided into two: an initial study of a classification model, applied to an already existing dataset, and secondly, the development of our own experimental protocol, the data acquisition, and the application of the previously tested models on the new dataset. The online dataset used is the BCI Competition IV dataset 1, containing EEG signals taken using a 59-electrode helmet from seven healthy subjects executing a Motor Imager (MI) task. To overcome the limitations of laboratory BCIs, two points have been identified to be applied to the protocol: firstly, to make the BCI subject-independent, so that it does not need to be calibrated on each subject. Second: make the BCI self-paced, thus allowing the subject to freely execute movements asynchronously, without a cue. The acquired dataset consists of the EEG signals taken using a 21-electrode helmet from four healthy subjects executing a MI task. The models applied are based on different combinations of processing steps and are validated by cross-validation using a fixed training window and a sliding testing window. The signals have been pre-processed with passband time filters in the alpha and beta range, and a Common Average Reference (CAR) filter. To extract useful features a CSP filter has been applied. Two different types of classification algorithms have been used, (LDA and SVM). On the subject-dependent calibration phase of dataset I of BCI competition IV, the applied models show an accuracy that varies between 76% and 92% when classifying the two Intentional Control (IC) classes. When classifying between IC and No Control (NC), the accuracy varies between 73% and 84%. Slightly modified models have been applied to the acquired dataset. The performance results of these models have been in line with the other dataset when classifying between NC and IC while the results dropped when classifying between IC classes. Only the first subject achieved acceptable results, with an accuracy of 67%. Possible causes are the different kinds of MI tasks performed and errors or inaccuracies in data acquisition and protocol definition. Future continuations of the study could focus on finding better-performing classification models and revising the experimental protocol by trying to identify its limitations and flaws.