Event-related potentials (ERPs) are small voltages generated in the brain in response to specific sensory, cognitive, or motor event. The P300 wave is an ERP component elicited in the process of decision making and surfaces as a positive peak in the ERP signal about 300ms after stimuli. Due to the consistency, reproducibility, and ubiquity of the P300 response and its relevance to unconscious decision-making processes, it has been widely used to develop brain-computer interfaces (BCIs). BCI is a direct communication pathway between the brain and an external device (e.g., computer, wheelchair, prosthesis) that translates brain signals into meaningful commands for controlling the device. The most important part of a P300-based BCI is its processing part which is responsible for detecting the P300 component from the input ERP signal to recognize the mental state/activity. Most state-of-the-art P300 detection methods are based on feature extraction and machine learning which are computationally demanding to be trained, inferred, or both. Moreover, such methods are mostly uninterpretable, making it almost impossible to gain insights into brain functionality. In this thesis, elicitation of P300 in response to visual stimuli is studied from statistical and probabilistic points of view. To this end, the problem is modeled in a generic Bayesian framework, and two probabilistic classifiers, namely naïve Bayes and Markov chain-based Bayes, are presented for P300 detection. The performance of the two classifiers is then eva‎luated on a P300 speller data set, and it is shown that the naïve Bayes model, especially in the multi-channel settings, outperforms Markov chain-based Bayes model. A suboptimal algorithm is then proposed to further improve the performance of the multi-channel naïve Bayes model in P300 detection and character recognition by finding suitable time range and set of channels. The final probabilistic classifier is proposed by extending the naïve Bayes model to weighted naïve Bayes (WNB), optimizing the weights using a proposed optimization method, and incorporating the aforementioned suboptimal algorithm. The proposed WNB-based method achieved 100% character recognition accuracy and 32.72 bit/min information transfer rate using 4 trials of spelling experiments. The performance of the proposed method is also compared with some state-of-the-art methods, and it is shown that the proposed method achieves state-of-the-art accuracy while being interpretable and less computationally demanding.

بهزاد نظري، سعيد صدري

علي يوسفي

رسول اميرفتاحي،‌ فرزانه شايق