Machine Learning algorithms to classify voluntary and involuntary movements from EEG in a BCI for post-coma non-responsive patients

This work is part of a larger project that aims to investigate the level of consciousness in non-responsive patients (Vegetative State or Minimal Conscious State). The ratio of misdiagnosis is indeed very high at the moment. We are developing a Brain Computer Interface able to support the diagnosis. Our BCI uses Readiness Potentials (RP), first identified by Kornuber, which arise just before the start of a voluntary movement. Therefore, the presence of an RP is a marker of presence of consciousness. Within this project, the aim of this thesis was to create a robust machine learning algorithm that can predict whether a movement is voluntary or not. We firstly investigated the characteristics of RP useful for this scope (feature extraction), then we selected the most informative ones (feature selection). Finally, we compared different classifiers model. More in detail, after the EEG pre-processing (i.e. the noise reduction and the correction of artifacts), a feature extraction has been performed in order to find the best RP's attributes for classification; this first part was the most critical one because the extraction of feature influence the performance of the classifiers. The section of feature selection was aimed at detecting which attributes were most informative, thus discarding the redundant or irrelevant ones, that could lead to misclassifications. The selected features were finally used to train and validate three different classifiers, one based on the K-Nearest Neighbour (K-NN), one Decision Tree (DT) and the other using Support Vector Machines (SVM). Significant results were obtained from the binary classification, and in particular: 73.3% specificity and 63.3% sensitivity for linear SVM, 73.3% specificity and 80% sensitivity for cubic SVM, 93.3% specificity and 73.3% sensitivity for K-NN and finally 73.3% specificity and 100% sensitivity for DT.