Classification of EEG Signals
EEG is defined as the electroencephalography, which involves the recording of the electrical activity through the sensors sourced on the scalp of the body. The EEG are measured through the various methodologies. The signals are collected with the sensors and the data are collected. The EEG signals are ranging at different frequencies that are collected with the involvement of the easy procedure and no pain. There are various types of the normal waves which denote activity levels at distinct states. The abnormal waves lead to denote medical issues especially related to the disorders of the brain. These signals are widely utilized for the medical diagnosis and establishing the brain computer interfaces for making the assistance to the disabled people for routine tasks. The signals are measured for developing the brain computer interfaces and the diagnosis of the brain related disorders such as the sleeping disorders. For a recording of EEG signals, the process needs the multiple sensors to measure the signals all over the scalp based on the locations. For measuring the feature extractions,
Only collecting the EEG signal data is not enough for the analysis or diagnosis. The main step involves the classification and identification of the EEG signal data based on their features. Normally the EEG signal frequency ranges from 0 to 30 Hertz. Based on the frequency ranges, the EEG signals are classified as alpha, beta, theta, and gamma waves. The frequency variations of the signals normally determined through the power spectral density, which indicate the signal variations and identify the differences among the frequency shifts [2].
The processes involved in the EEG signal classification are given with the different steps such as follows:
- Signal acquisition
- Signal pre-processing
- Feature extraction
- Classification
- Performance Evaluation
EEG as earlier discussed deals with the electrical signals for denoting the electrical activity of the brain. Various researchers identified different technique to process the signals.
Signal acquisition
The signal acquisition involves the acquisition of the data in terms of the signals with the use of the sensors over the scalp. The signals are collected and recorded for the analysis. The sensors are placed in the different locations as shown in the following figure [3].
Signal pre-processing
The signal pre-processing involves the processing of the collected EEG signals [5]. The pre-processing mostly comprises the steps of clearing the signal noises and the irrelevant, invalid inputs [4], [20].
Feature extraction
This step remains the important approach among all the steps. The feature extraction enables to obtain certain features or relationship from the pre-process signals that may help to diagnose the pattern or characteristics of the signals. If the features are inappropriate, then they cannot support the classifier to recognise or to identify the related signal disorders, which creates the brain-computer interface. Features for the Event related potentials [ERPs] are obtained with the suitable feature extraction methods [5].
Event Related Potentials and EEG Signal Classification
There are various methods adopted to extract the features in the signals such as the recognition of the amplitude values, power spectral density, and band powers [22]. During the feature extractions, certain factors are validated such as the noise, outliers, dimensionality, information about the time, non-stationary characteristics, and training sets. Different methods are present in the feature extraction such as Power Spectral Entropy, Wavelet Transformations. Adaptive Auto Regressive parameters, and Fast Fourier Transformations [23].
Fourier transform helps to transform the domain signal into frequency signal. As EEG signal being a time domain signal, it is analyzed with the Fourier transformation and their energy is determined by using the Power Spectral Density. Fourier transformation works well in this method as EEG is a time domain signal. As per Suleimann, Fourier transformation is used for the analysis with the division of signal at 1 sec duration. But this transformation not suits as EEG being short frequency signal [1].
It is alternate method for the transformation of the time-frequency signals and suits well for the short-frequency methods [10]. Normally these transformations involve the use of the low pass and high pass filters to filter the irrelevant and unwanted noises [7]. The low pass and high pass filters provide the approximation and detail coefficients respectively. The coefficients denote the specific ERP activity [8]. There are several methods to classify the signals further [9]. For features, mostly DWT coefficients, wavelet band energy, entropy, and PSD are used [6].
The features of the EEG signals are extracted with the use of the power spectral entropy. PSE supports the new feature for the uncertain system through the information. In this method, PSE is determined followed by the Fourier Transform. Zhang used this technique for classification [11].
It is one of the essential methods in feature extraction in which the random data are converted into the components, which are independent of each other which have the objective to find the linear representation of the non-Gaussian data which denotes the independency among the variables [19]. This method can be used for feature extraction in the ERP signals [12], [20], [21].
The principal component analysis clearly specifies about the suitability of the feature extraction in the EEG signals. The main reason for the suitability of this method is dimensionality reduction. In this method, d-dimensions are indicated in low dimensions which lead to reduce the time, space complexities. This method is used for the segmentation of the signals sourced in several sources [24].
Classification
Different Classifiers Used for EEG Signal Classification
The features extracted from the signals support classifier for the classification of the disorders and brain-computer interface activities for enhancing the connections.
Various researchers tend to adopt different approaches for the classification of the EEG signals. These approaches drive the researches to find appropriate categorizations that enable further identification of the disorders related to sleep and brain. However, these methods have unique benefits and demerits. In this study, various approaches are followed for the classification of the event related potentials.
The event-related potentials are the small voltages, which are created in the brain as the responses to the stimulus. The classifications of the EEG signals in the ERPs are mandatory as they relate the diagnosis of the disorders. The signals of EEG in event related potentials lead to distinct waveforms consisting of components identifying the number denoting the latency measured in seconds. The negative peak in the ERP waveform for the latency at 100ms after stimulus indicates N100 or N1 similarly the positive peak in the ERP waveform for the latency at 200ms indicates P2 or P200.
The researchers adopted several classifiers for the classification of the EEG signals in the event related potentials. The different types of classifiers are Neural network based classifiers, linear classifiers, Bayesian classifiers, nearest neighbour classifiers, and Gaussian classifiers. The classifiers are algorithms, which consider the features are their inputs and provide the output with a label or confidence values.
Linear Classifiers
These types of the classifiers are classified into support vector machines and linear discriminant analysis [15].
In this method, hyperplane is used for discrimination of the data into several classes. The performance of the LDA gets affected when labelled data points are small. This method gives the measurement of sensitivity and specificity. The pair-wise classification is performed by using k-NN classifier [13].
In this method, SVM adopts the use of the hyper plane for the data separation. The data separation leads to categorize the data easily. The maximization of the margin which denotes the distance between the hyper plane and the points nearest from the classes. This margin is known as the support vectors. SVM classifier is used mostly for the classification of two EEG signals or more [14].
In this method, artificial neural networks are adopted for the non linear classifiers that have interconnected elements known as neurons. As per this method, the neuron consists of biological neuron, which performs the computation. The EEG classification mostly follows the method of multi-layer perceptron neural network. The network consists of three layers with the advantages such as the fast operation, implementation, and suitability with the small training sets.
In this method, the nearest neighbour classifiers will allot a feature vector to specific class based on the nearest neighbour such as k-nearest neighbour classifiers. These methods enable to allot the nearest neighbour classification only with the high value of k. The limitations in the model includes the need of enough training to the samples, high k values, approximation of the functions to produce nonlinear decision boundaries.
In order to learn about the Gaussian Bayes Classifier method, the attributes, which are continuous, are transformed into the discrete variables as these variables will get information such as invalid, missed, noise, and low sensitivity. The discrete variables usually help to receive about the issues. In this study, this model is adopted that helps to classify the EEG signals for the EPRs, which involves the use of the Gaussian function with smoothing for density estimation. The Bayesian classifier along with the continuous attributes is providing the information to a specific class for attributes. The datasets on analysis indicate that the Bayesian classifiers with Kernel function lead to the best classification accuracy as compared to the other methods of classification [17]. This model is used in the study of EEG signal processing for event related potentials. This model suits for the EEG signals especially for ERPs [16].
In this method, the Gaussian Mixture model consists of the several components, which serves the weighted sum of Gaussian probability density functions or Gaussian components. The main objective is to use the Bayesian classifier with flexible handling of acceptable number of variables, Gaussian components, and classes in the mixture model. For this model, two components are considered such as a training function and a classification function. The training methods are then evaluated by terms of testing methods for classification.
Based on the results of the model, the accuracy is determined by performance evaluation of the model. The accuracy results are determined usually by comparing the target variables with the predicted output. The performance about the classification of the variables with model will return the percentage of the accuracy that further helps to enhance the performance of the model.
The methodology adopted in the present study is given as follows:
- Collection of Raw EEG data
- Filtering process
- Feature transformation using GMM and GAUSS
- Feature classification using GMM and GAUSS
- Usage of BCILAB for Classification
- Performance Evaluation
The EEG data for collection of the data in terms of the event related potentials [ERPs]. The EEG signals are collected by connecting the electrodes with the arrangement of 10-20 system over the scalp. The electrodes are connected over the scalp in the lobes such as frontal [F], central [C], temporal [T], and parietal lobe [P]. The data are collected with the response of the event related potentials. The responses are collected with the application of the stimulus to the individuals. The events are the stimulus in response to the sensory-associated operations like finding colours, shapes, or any type of visual cortex; cognitive control operations such as the selection of response or suppression of action; affective operations such as the responses associated with the positive or negative feelings or emotions; memory related responses such as the recalling of an incident or remembering any new actions, and so on. The ERPs are the potentials same as EEG and hence the similar type of amplifiers is widely applied in the experimentation.
The filters are used to remove the outliers and noises in the raw EEG data, which is acquired. The raw data usually consists of the outliers and noises, which will cause inconsistent results for the feature extraction and classification. In order to derive appropriate results, the invalid data are removed by the use of the filters. Several artefacts in EEG occur, which can be caused by electro-oculography, electrocardiography, and any other sources too. The artefact removal is made in this method with the use of the GMM and GAUSS model itself involving the feature extraction followed by transformation and classification.
In case of Gaussian Bayes Classifier, the Bayesian classifier along with the continuous attributes is leading the information to a specific class for attributes. The datasets on analysis indicate the Bayesian classifiers with Kernel function leading to the best classification accuracy as compared to the other methods of classification
GAUSS
In this method, the Bayesian classifier is used along with kernel functions, dimensional reduction functions to determine the probabilities. The Gaussian parameters can provide the determination of the signals. The discretization for the continuous variables help to find the classification.
Gaussian Mixture models [GMM]
Gaussian mixture models are utilized for Gaussian distributions and subclasses for a class. Probability density functions are defined as the sum of Gaussians.
Here, is weight of the component, ranges from 0 to 1 for all the components.
Algorithm works based on the following steps:
- The feature extraction for filtering the noises and extracting the features through common spatial patterns.
- Collect the data and use the training data
- Apply the expectation maximization method or any other methods for clustering such as variational Bayes, greedy expectation maximization methods.
- Find the essential modelling parameters that may include maximum Gaussian components, dimensions allowed
- For each part of the data: initialize the expectation maximization parameters with Gaussian components and dimensions with iteration time and begin the iterations
- Find the output and develop the mixture model within the dimension in labelled data
Find the probability for each features obtained in the feature extraction process with the iterations involving test datasets
Use of BCILAB
BCILAB being a toolbox in MATLAB, facilitates design, testing, prototyping, experimentation, and evaluation for the brain-computer interfaces and any other EEG systems. The tool allows the user to process data for the EEG system with the different functionalities such as signal processing, feature extraction, machine learning, and even makes use of online plugins. The interfaces for the tool consist of user interface, scripting, and plugins.
Inputs to BCILAB
For this toolbox BCILAB, the inputs include the data collected from the EEG system for the event related potentials. The data input consists of the trials, which are training data and target variable. The other parameters such as the dimensions, blend are also needed as input for BCILAB.
In case of Gaussian Mixture models, the data such as the variants like conventional approaches with expectation maximization, variational bayes, and all other methods are used. These methods are selected in BCILAB. The non-parametric methods are also opted if conventional methods are not required such as variational Drichlet process, collapsed Dirichlet prior, collapsed stick breaking, and much more.
- Collect the raw data from the EEG system
- Use the BCILAB module in MATLAB
- Load the data into the BCILAB by providing the inputs
- Declare the training data, target variable, options and the related values such as dimensions, scaling
- Collect the raw data from the EEG system
- Use the BCILAB module in MATLAB
- Load the data into the BCILAB by providing the inputs
- Declare the training data, target variable, clusters and their number, optional values such as the variants, which may be non-parametric or conventional parametric methods
The performance evaluation of the models is very essential as they support to find the best model for the determination. The accuracy for each model is found by using the results of both models. The accuracy determined will provide the results about the performance of the model. The accuracy can be found by the comparison of the predicted output and the target variables in the study.
References
Sulieman, “Feature Extraction Techniques of EEG Signal for BCI Applications”, Computer and Information Engineering Department College Of Electronics Engineering, University of Mosul, Mosul, Iraq
Khorshidtalab and M. Salami, “EEG signal classification for real time brain?computer interface applications: a review”, IEEE International Conferenceon Mechatronics, May 2011, pp. 1?7.
Faris, “Feature Extracted Classifiers Based EEG Signal: A Survey”, Life Science Journal, vol. 11, 2014,pp: 364?375.
Correa, “Artefact removal from EEG signal using Adaptive Filters in Cascode”, IOP, 2007.
Khatwani and A. Tiwari, “A Survey on Different Noise Removal Techniques of EEG Signals”, International Journal of Advanced Research in Computer an Communication Engineering, Vol. 2, Feb 2013, pp: 1091?1095.
Guo, “Classification of EEG Signals Using Relative Wavelet Energy and Artificial Neural Networks”,GEC, Jun 32009, pp: 177?183.
Yohanes, “Discrete Wavelet Transform Coefficients for Emotion Recognition from EEG signal”, IEEE, International Conference of the IEEE EMBS,2012
Dingyin, “Feature Extraction of Motor Imagery EEG Signals Based on Wavelet Packet Decompositions”, IEEE, International Conference on Medical Engineering, 2011.
Hazarika, “Classification of EEG Signals Using the Wavelet Transform”, IEEE
Imran, “Motor Imagery EEG Signal Classification Scheme Based on Wavelet Domain Statistical Features”, IEEE, International Conference on Electrical Engineering and Information & Communication Technology, 2014.
Zhang, “Feature extraction Using Power Spectral Entropy”, IEEE International Conference on Biomedical Engineering and Informatics, 2008
Oveisi, “EEG Signal Classification using Nonlinear Independent Component Analysis”, IEEE, 2009
Lee, “Extending Linear Discriminant Analysis by Using Unlabeled Data”, IEEE, Computer and Information Technology , 2011
Hortal, “Online Classification of Two Mental Tasks Using A SVM based BCI System”, IEEE Neural Engineering,2013.
Gursel, “Discrimination Between Mental and Motor Tasks of EEG Signals Using Different Classification Methods”, IEEE, 2010
A. T. Figueiredo and A. K. Jain, “Unsupervised learning of finite mixture models,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, no. 3, pp. 381–396, 2002.
Ng and M. Jordan. On generative vs. discriminative classifiers: A comparison of logistic regression and naive bayes. In Proceedings of Advances in Neural Information Processing, 2002.
Barbati, C. Porcaro, F. Zappasodi, P. M. Rossini, and F. Tecchio, “Optimization of an independent component analysis approach for artifact identification and removal in magnetoencephalographic signals,” Clinical Neurophysiology, vol. 115, no. 5, pp. 1220–1232, 2004
Lawhern, D. Slayback, D. Wu, and B. J. Lance, “Efficient labeling of EEG signal artifacts using active learning,” in 2015 IEEE International Conference on Systems, Man, and Cybernetics, Kowloon, China, October 2015
Winkler, S. Haufe, and M. Tangermann, “Automatic classification of artifactual ICA-components for artifact removal in EEG signals,” Behavioral and Brain Functions, vol. 7, no. 1, pp. 1–15, 2011. View at Publisher · View at Google Scholar · View at Scopus
Zhou and J. Gotman, “Automatic removal of eye movement artifacts from the EEG using ICA and the dipole model,” Progress in Natural Science, vol. 19, no. 9, pp. 1165–1170, 2009. View at Publisher · View at Google Scholar ·
Vergult, W. de Clercq, A. Palmini et al., “Improving the interpretation of ictal scalp EEG: BSS–CCA algorithm for muscle artifact removal,” Epilepsia, vol. 48, no. 5, pp. 950–958, 2007
Delorme, T. Sejnowski, and S. Makeig, “Enhanced detection of artifacts in EEG data using higher-order statistics and independent component analysis,” NeuroImage, vol. 34, no. 4, pp. 1443–1449, 2007.
Choi, A. Cichocki, H. M. Park, and S. Y. Lee, “Blind source separation and independent component analysis: a review,” Neural Information Processing – Letters and Reviews, vol. 6, pp. 1–57, 2005.
Order an Essay Now & Get These Features For Free:
Turnitin Report
Formatting
Title Page
Citation
Outline
