Motor imagery based mobile brain computer interface design using machine learning techniques

Brain Computer Interface (BCI) is a fast growing area where brain signals are translated into reasonable commands bridging the gap between human and machines. On one hand, it is used as a game equipment or lie detection etc. for healthy people. On the other hand, it is a promising development more than anything for disabled people regaining their movement and spelling ability partially. Most of the motor imagery BCI applications rely on electroencephalography (EEG) due to its compactness, inexpensiveness and high temporal resolution. However, low spatial resolution due to volume conduction effect makes classification task tedious. In this study, mobile brain computer interface is designed using commercially available, portable EEG headset (Emotiv Epoc). The thesis begins with EEG method introduction and its comparison with brain imaging techniques. Next, several BCI types depending on different neurophysiological phenomenon has been presented. Among these methods, BCI based on oscillatory activity (i.e motor imagery) with its specific signals (µ and β rhythms) is described. Spatial filtering techniques, especially common spatial patterns, originally developed for two class problems, are mentioned with its extension technique to multi-class case; one versus one technique and one versus the rest technique. Most popular machine learning algorithms such as Linear discriminant analysis (LDA), Support vector machines (SVM) and K-nearest neighbour classifier (KNN) were elaborated with the main idea behind. Popular competition dataset in BCI academics and its paradigm are carefully examined and dataset is processed with written code in Matlab not only in order to assess the accuracy of the code, but also to develop a baseline for comparison at the end. Similar paradigm is created in Openvibe program, both off-line and on-line classification has been made through acquired signals from mobile Emotiv Epoc headset. Off-line results of Emotiv headset data showed similar performance with BCI competition dataset using the same electrode locations. The same configuration files such as CSP filter and classifier coefficients are used in on-line demonstration for two and three class classification. The importance of electrode location placed on motor cortex is once more understood by comparing the results of Competition dataset using all electrode samples that is covering motor cortex region with the results of Competition dataset using samples from reduced number of electrodes that are around the motor cortex. In the future, new portable headset designs utilizing motor cortex electrodes may exploit the accuracy of the motor imagery classification further.