Analysis of Electroencephalogram Signal for P300 Based Brain-Computer Interface Speller

Abstract

Abrain-computer interface (BCI) speller is a communication medium with the outer world for the patients suffering with neuro-muscular disorders. A P300 speller which translates the brain signal into machine commands provides such communication inter-face to convey their thought without any motor movement. A P300 speller aims to spell characters by using the electroencephalogram (EEG) signal and its performance can be defined by the number of correctly recognised characters. P300 is an event-related potential (ERP) which is appeared in the EEG signal when random stimuli occur to the subject. Various types of BCI paradigms can be used for character spelling which provide stimuli to the subject. Generally, row-column paradigm is used for P300 based character recognition in BCI application due to its simple graphical user interface (GUI). The row and column of the matrix are intensified randomly and successively which generate the stimuli. P300 appears in the EEG signal when the row or column of the desired char- acter intensified and the character is recognised from the detected P300. To detect the P300, various signal processing algorithms have been introduced which analyse the EEG signal and classified them as P300 target or P300 non-target. A P300 based character recognition system consists of the following stages: preprocessing, feature extraction, feature optimization and classification. The amplitude and latency of the EEG signal vary based on the psycho-physiological condition of the subjects and the EEG signal is easily affected by the surrounding noise which will reduce the character recognition performance. Therefore, there is a requirement of feature extraction technique to rep- resent the EEG signal efficiently and to detect the dynamic changes of the EEG signal. The EEG signals are acquired through a multi-channel system with high sampling rate. High dimension of EEG features can create over-fitting problem, and irrelevant features can reduce the classification performance. Therefore, the need of feature optimization algorithms are constantly increasing to overcome the dimensionality problems. Mo- tivated by the above discussion, this thesis focuses on the feature extraction, feature optimization and classification technique for the improvement of P300 based character recognition performance. At first, a feature extraction technique is developed based on principal component analysis (PCA) which extracts the important features from the dataset and removes the redundant and irrelevant features. With these PCA based fea- tures, ensemble of weighted support vector machine (EWSVM) technique is proposed for classification. In EWSVM, a weight is assigned to each classifier, so that better classifiers get more weightage compared to other classifiers. As a result, when the clas- sifiers’ outcomes are averaged out, best classifier provides more impact on the output for the weight assigned to it. Next, for P300 based character recognition, hand-crafted features are not efficient to represent the signal properly due to amplitude variation and nonlinearity of EEG signal. To overcome the limitation of the hand-crafted features, convolutional neural network (CNN) based automated feature extraction technique has been developed as it extracts hierarchical features from the dataset. In the developed technique, two different convolution layers are used to extract the spatial and temporal features from the dataset, respectively. The deep features are extracted from the fully- connected layer of the trained CNN architecture. After extracting deep features from the EEG signals, Fisher’s ratio (F-ratio) based feature selection technique is adopted to find out the optimal features from the extracted features. Subsequently, to improve the P300 based character recognition performance, a feature fusion framework is devel- oped. Autoencoder (AE) technique is proposed which extracts abstract features from the input data. Unlike PCA, an AE uses non-linear transformation with non-linear acti- vation function which extracts abstract information from the EEG signal and temporal features represent the dynamic information of the EEG signal. Therefore, these two features may be partially complementary in nature. It is seen from the experimental results that their combination helps to improve the P300 based character recognition performance. In AE, a sparsity constraint is imposed on the hidden nodes to overcome the over-fitting problem, and the modified AE is referred as sparse autoencoder (SAE). SAEs are stacked together and it is denoted as stacked sparse autoencoder (SSAE). Finally, multiscale convolutional neural network (MsCNN) is proposed which extracts multi-resolution deep features from the data. These features learn diverse information from the acquired EEG signal. To overcome the limited dataset problem and require-ment of long calibration time, transfer learning (TL) technique is proposed in this work for P300 based character recognition. In this process, a network is trained with suffi- cient number of training data of one subject. After training, this pre-trained network is fine-tuned with a new subject. Less data is required for fine-tuning, and as a result, the requirement of calibration time is also less. Experimental results are conducted on two publicly available datasets, BCI Competition II and III datasets to demonstrate the effectiveness of all the proposed techniques.