COMBINING ELECTROENCEPHALOGRAPHY AND FUNCTIONAL NEAR INFRARED SPECTROSCOPY TO IMPROVE ACCURACIES IN INTERPRETING INTENT DURING MOTOR IMAGERY TASKS FOR USE IN BRAIN COMPUTER INTERFACES

Abstract

A Brain Computer Interface (BCI) system can bypass impaired systems in the brain, nervous system, and muscles to enable alternate function and allow the user to reconnect with the outside environment. Measurements for these systems usually include the collection of brain signal activity from sensors mounted on the surface of the scalp. Electroencephalography (EEG) and functional near infrared spectroscopy (fNIRS) are two commonly used non-invasive brain imaging modalities in BCI systems. EEG records the electrical activity produced by neuronal activations whereas fNIRS measures the concentration changes of oxy- (HbO) and deoxyhemoglobin (HbR) molecules in the brain cortex. In this thesis, both EEG and fNIRS data collected during a bilateral right- and left-hand motor imagery task were used to detect brain signals that suggest intent to move. Feature extraction and classification are two important components of a BCI in which discriminant features are extracted from the brain signals and then decoded to interpret the user’s intent. Using the fNIRS data in the first part of this thesis, different features were extracted (mean, peak, minimum, skewness, and kurtosis) and classification algorithms (linear (LDA) and quadratic discriminant analysis (QDA), support vector machine (SVM), Logistic Regression, and Naïve Bayes) were compared to find the set of features and classifiers with the highest accuracy. The mean, peak, and minimum of HbO, as well as the mean of HbR and mean of difference between HbO and HbR produced the highest accuracies among features, whereas skewness and kurtosis of HbO resulted in the lowest accuracies. Furthermore, QDA and SVM with polynomial and Gaussian kernel functions resulted in the highest accuracies compared to other classifiers. Using QDA and SVM, this study assessed a channel selection algorithm to reduce the number of sensors in the BCI examining a strategy to use fNIRS results to target the placement of EEG electrodes. Lastly, the feasibility of hybrid EEG and fNIRS system was investigated by comparing corresponding classification accuracies to EEG and fNIRS alone. The combined EEG and fNIRS system resulted in significant improvements in classification accuracies compared to EEG or fNIRS alone.