PRACTICAL BCI SPELLER BASED ON USER-SPECIFIC FEATURE REPRESENTATION AND CONVOLUTIONAL NEURAL NETWORK

Abstract

Brain-Computer Interface (BCI) systems have a great impact on improving people's lives. One of the popular implementations are BCI spellers which are realized in big sized monitor layouts. However, in real-world such applications are inconvenient and immobile. Therefore, smartphone layouts should be used for real-world BCI spellers as diminutive low-impacted stimuli are suitable to solve such issue. Nevertheless, if the stimuli intensity is diminished it might lead to P300 amplitude reduction, which causes poor classification accuracy. The aim of this thesis research is to propose a state-of-the-art BCI speller with less than 1 mm sized visual stimuli implemented in a smartphone layout. To increase the task performance, subjects were instructed to produce a certain mental task while gazing at the target character. During mental tasks experiments the new event-related potential (ERP) component, the late positive potential (LPP), was discovered between 700-800 ms after the stimuli. The dataset of 14 subjects was collected.

In this study, BCI system is proposed based on classification of the P300 wave which is one of the difficult tasks in processing electroencephalography (EEG) signals as it is affected by surrounded noise and has low signal-to-noise ratio. Before P300 classification, prerpocessing recorded signals is an essential step where some features are extracted and selected. The majority of the previous studies utilizes hand-crafted features to detect P300 wave; however, this technique is inefficient in representing the signals because an environment and subjects vary in each individual experiment. Inspired by this, convolutional neural network (CNN) architecture is suggested for P300 signal detection to extract important features automatically.

Additionally, to characterize the user-specific spatial-temporal features a data-driven optimization approach is utilized. %The spelling accuracy of Linear Discriminant Analysis (LDA) is 96.8\% with an Information Transfer Rate (ITR) of 31.6 [bits/min]. Average spelling accuracy for LDA in individual sequences for passive tasks in normal-speller - 76.3\% and in the proposed dot-speller - 78.1\%, whereas for active tasks in normal - 94.1\% and dot- 96.8\%. Average spelling accuracy for CNN results in individual sequences for active tasks in dot-speller before data augmentation EEGNet - 97\%, ShallowConvNet - 59\%, DeepConvNet - 94\%, while after augmentation EEGNet - 97\%, ShallowConvNet - 88\%, DeepConvNet - 97\%. These results are comparable to traditional BCI spellers. This study represents practicability of creating feasible and useful BCI spellers in the future.