META-CLASSIFICATION OF BRAIN SIGNALS IN P300 SPELLER

Abstract

P300 Speller is a brain-computer interface (BCI) spelling system, which is used to enable

a human to communicate with the outer world by processing electroencephalography (EEG) sig- nals. Such speller systems are usually used for patients with di

erent neural motor disorders, such as amyotrophic lateral sclerosis (ALS). Human’s reaction to a visual stimuli can be detected using

event-related potential (ERP) components. A positive voltage deflection detected at about approxi- mately 300 ms after the target visual stimulus, which is called P300 component, is usually used for

speller systems. The main problem is that di

erent people have di

erent latency and amplitude of

the P300 component, thus, P300 Speller has to be trained for each subject separately. In order to achieve robust results for di

erent subjects, generic training (GT) approach is proposed and tested on 20 healthy subjects data using linear discriminant analysis (LDA), support vector machine (SVM), k-Nearest Neighbours (kNN) classifiers with di

erent hyperparameters. GT training showed better performance than standard subject-specific training (SST) approach, and is further used for ensemble learning experiments. A set of ensemble models was proposed based on LDA, SVM, kNN, and convolutional neural network (CNN). The proposed models were trained using multi-channel data from healthy subjects and tested on both healthy subjects and ALS patients data. By trying di

erent number of EEG channels, the best option of using four channels of parietal zone was found. The simulation

results show that the proposed ensemble model achieved better results compared to standalone clas- sifiers. In order to achieve better user experience, the number of channels was decreased from 8 to 4

active channels. The obtained results show that using 4-channel EEG data is more e

cient. Single- channel EEG data classification, however, provided bad performance. The fusion of LDA, kNN and

SVM provides the most accurate results, achieving the accuracy of 99.91% for healthy subjects and almost 84.81% for ALS patients.The best results in terms of time complexity were provided by a simple ensemble of LDA and kNN classifiers.