Date of Award

Summer 7-19-2021

Document Type


Degree Name

Master of Science in Biomedical Engineering



First Advisor

Alan Chiu

Second Advisor

William Weiner

Third Advisor

Jennifer O'Connor


One of the chief contributing factors to slowing down BCI spellers for users with profound disabilities is backtracking to delete mistakes or correct certain selections. The ability to design an EEG-based strategy to identify the desire to make corrections in the BCI speller would enhance the user experience and bit rate of the device. Past efforts suggested that Common Spatial Patterns (CSP) may show promise in helping detect and classify semantic violations in reading despite CSP not being widely used in event-related potential (ERP) applications. Semantic violations in EEG often exhibit deflections in the N400 and P600 region coinciding with the violation. This research aims to create a CSP model that can improve the classification accuracy of semantic violations in reading by incorporating neural oscillation information. Visual stimuli consisting of 150 pairs of nouns from Maguire et al. (2010) and Calvo et al. (2018) were presented, where the first word served as the Primer and the second word served as the Target. EEG signals from 14 channels were parsed to obtain the average N400 potential, the average P600 potential, and signal power in the alpha and theta bands, creating a 56-dimension (14 channels by 4 feature types) feature space. The CSP algorithm was implemented to improve orthogonality in the feature space, and the feature space dimension was reduced to 2. Three types of classification strategies Linear Discriminant Analysis (LDA), Naïve Bayes (LB), and K-Nearest Neighbor (KNN), were implemented. Graphical analysis by CSP showed that while individual features did not appear separable, a higher dimension dataset including all four feature types does demonstrate separability. The 10-fold validation results showed that two-class models optimized for individual subjects achieved accuracies ranging from 50% to 66% with LDA, 78% to 98% with KNN, and 76% to 100% with NB. Examining the mixing matrices during the dimension reduction step in CSP suggested that alpha frequency band and EEG locations P7, F4, F3, AF4, and FC6 consistently play a critical role in the success of these classifiers across the different subjects.