Wavelet-Based Feature Extraction for the Analysis of EEG Signals Associated with Imagined Fists and Feet Movements

,


Introduction
Electroencephalography (EEG) is defined as the process of measuring the electrical voltage fluctuations along the scalp as a result of the current flows in brain's neurons and the brain's neural activity (Niedermeyer & da Silva, 2005).In typical EEG tests the brain's electrical activity is monitored and recorded using electrodes that are fixed on the scalp (Sleight et al., 2009).Brain-Computer Interface (BCI) is a combination of hardware and software systems that enables the use of the brain's neural activity to communicate with others or to control machines, artificial limbs, or robots without direct physical movements (Levine et al., 1999;Donoghue, 2002;Wolpaw et al., 2002;Vallabhaneni et al., 2005).BCI captures EEG signals in conjunction with a specific user activity then uses different signal processing algorithms to translate these records into control commands for different machine and computer applications (Graimann et al., 2010).It is proved in (Vidal, 1973) that a user's intent could be effectively represented by signals recorded from brain activity.
During the last few years, BCI has become an attractive field of research and applications specially in helping disabled individuals by providing a new channel of communication with the external environment and offering a feasible tool to control artificial limbs (Selim et al., 2008).A variety of BCI applications were described in (Grabianowski, 2007).BCI is a highly interdisciplinary research topic that combines medicine, neurology, psychology, rehabilitation engineering, Human-Computer Interaction (HCI), signal processing and machine learning (Smith et al., 2007).
It can be noted from the literature that the strength of any BCI application depends on the translation approach used to transform EEG signal patterns into machine commands.In (Pfurtscheller et al., 1997), the authors recorded EEG signals for three subjects while imagining either right or left hand movement based on a visual cue stimulus.They were able to classify EEG signals into right and left hand movements using a neural network classifier with an accuracy of 80% and concluded that this accuracy did not improve with increasing number of sessions.Sepulveda (2011) used features produced by Motor Imagery (MI) to control a robot arm.Features such as the band power in specific frequency bands (alpha: 8-12 Hz and beta: 13-30 Hz) were mapped into right and left limb movements.In addition, they used similar features with MI, which are the Event Related Desynchronization and Synchronization (ERD/ERS) comparing the signal's energy in specific frequency bands with respect to the mentally relaxed state.
It was shown in (Mohamed, 2011;Alomari et al., 2013) that the combination of ERD/ERS and Movement-Related Cortical Potentials (MRCP) improves EEG classification as this offers an independent and complimentary information.The authors of (Farina et al., 2007) presented an approach for the classification of single trial MRCP using a discrete dyadic wavelet transform and Support Vector Machines (SVMs) and they provided a promising classification performance.A single trial right/left hand movement classification is reported in (Kim et al., 2003).The authors analyzed both executed and imagined hand movement EEG signals and created a feature vector consisting of the ERD/ERS patterns of the mu and beta rhythms and the coefficients of the autoregressive model.Artificial Neural Networks (ANNs) is applied to two kinds of testing datasets and an average recognition rate of 93% is achieved.
A three-class BCI system was presented in (Wang et al., 2007) for the translation of imagined left/right hands and foot movements into commands that operates a wheelchair.This work used many spatial patterns of ERD on mu rhythms along the sensory-motor cortex and the resulting classification accuracy for online and offline tests was 79.48% and 85.00%, respectively.The authors of (Guger et al., 1999) proposed an EEG-based BCI system that controls hand prosthesis of paralyzed people by movement thoughts of left and right hands.They reported an accuracy of about 90%.
In Su et al. (2011), a hybrid BCI control strategy is presented.The authors expanded the control functions of a P300 potential based BCI for virtual devices and MI related sensorimotor rhythms to navigate in a virtual environment.Imagined left/right hand movements were translated into movement commands in a virtual apartment and an extremely high testing accuracy results were reached.Homri et al. (2012) applied the Daubechies, Coiflet and Symmlet wavelet families to a dataset of MI to extract features and describe right and left hand movement imagery.The authors reported that the use of Linear Discriminate Analysis (LDA) and Multilayer Perceptron (MLP) Neural Networks (NNs) provided good classification results and that LDA classifier achieved higher classification results of up to 88% for different Symmlet wavelets.Tolić and Jović (2013) used the discrete wavelet transform (DWT) to create inputs for a NNs classifier and the authors reported a very high classification accuracy of 99.87% for the recognition of some mental tasks.We proposed a system that could efficiently discriminate between executed left and right fist movements in our previous study (Alomari et al., 2013).The current work is an extension for our studies to classify both imagined fists and feet movements by analyzing EEG signals recorded during a large number of experiments for 100 different subjects.Many wavelet families were used to calculate wavelet coefficients and then all the possible feature candidates were extracted and used in the training/testing and optimization experiments of a NNs classifier.

The PhysioNet EEG Dataset
In this work, we used the EEG dataset that was created and contributed to PhysioNet (Goldberger et al., 2000) by the developers of the BCI2000 (Schalk et al., 2004) instrumentation system.The dataset is publically available online at http://www.physionet.org/pn4/eegmmidb.It consists of more than 1500 one or two minutes-duration EEG records obtained from 109 healthy subjects.Subjects were asked to execute and imagine different tasks while 64 channels of EEG signals were recorded from the electrodes that were fitted along the scalp.
In the records of the dataset that are related to the current research, each subject performed three experimental runs of imagining the movement of both fists or both feet.During each two-minute run, either the top or the bottom of a computer screen shows a target.The subject imagines opening and closing either both fists (if the target is on top) or both feet (if the target is on the bottom) until the target disappears where he relaxes.This was repeated 15 times during each two-minute run.Then the obtained EEG signals were recorded according to the international 10-20 system as seen in Figure 1.For this work, we created a subset of three two-minute runs for 100 subjects for a total of 4500 events (45 imagined events per subject).

Wavele
The Wave (Riouel & 2013) (1)  Integrated EEG (IEEG) (3) The Daubechies, Symlets, and Coiflets wavelets were used to analyze the channels C 3 , C 4 , and C z of each EEG record.Then, as depicted in Figure 3, the features RMS, MAV, IEEG, SSI, VAR, and AAC were calculated for the wavelet coefficients using Equations 1 through 6.This process was repeated for each event in our dataset of 4500 vectors.At the end of these calculations, 9 RMS features (3 channels  3 details), 9 MAV features, 9 IEEG features, 9 SSI features, 9 VAR features, and 9 AAC features were generated for each wavelet.These features were numerically represented in a format that is suitable for use with NN algorithms (Qahwaji et al., 2008;Al-Omari et al., 2010) as described in next section.

Neural Networks Experiments
Neural networks learning algorithms were used in (Pfurtscheller et al., 1997;Homri et al., 2012;Kharat & Dudul, 2012;Tolić & Jović, 2013) and provided good classification performance.A detailed description of NN can be found in (Qahwaji et al., 2008).The MATLAB NN toolbox was used for all the training and testing experiments.
The training experiments were handled with the aid of the back-propagation learning algorithm (Fahlmann & Lebiere, 1989).
The numb layers, 80% for testing number of A huge n performan the same g the neural  By comparing the results, it was found that the optimum classification accuracy that can be achieved using our system is 89.11%.This performance was achieved by inputting the MAV feature of a Coif4 wavelet into a neural network of 20 hidden layers.This result is consistent with the conclusions reported in (Phinyomark et al., 2013) where it was shown that both MAV and RMS were accurate inputs for recognition and classification systems.
If we compare the highest accuracies in all tables, we can note that the Symlets wavelet outperforms the other wavelet families in most cases.The VAR feature for the Sym2 wavelet provided an accuracy of 87.44% using a NN of 14 hidden layers.In addition, the AAC feature for the same wavelet provided a 85.49% performance using a NN of 17 hidden layers.
On the other hand, it can be concluded from all tables that the MAV feature provides the best overall performance using any wavelet family.It is associated with the performances of 89.11%, 84.03%, and 84.51 while applying the Coiflets, Daubechies, and Symlets, respectively.The next best feature is the VAR then the IEEG.It can be mentioned here that The IEEG feature for the Db8 wavelet provided an accuracy of 86.78% using a NN of 17 hidden layers.

Conclusions
This work describes a classification system that can classify imagined EEG signals into fists and feet movements.Symlets, Daubechies, Coiflets wavelet families were compared for their abilities to decompose EEG signals and extract features that can be used as inputs to neural networks.Extensive experiments were carried out and the neural networks were optimized.The optimum classification performance of 89.11% was achieved with a NN classifier of 20 hidden layers while using the mean absolute value of the Coiflets wavelet coefficients as inputs to NN.It is believed that this work is one of the best to achieve such classification performance while working on imagined fists and feet activities.Real-time applications of this work can be implemented in the near future. Figur Figure 4.

Table 1 .
Phinyomark et al. (2013)decomposed details and approximationPhinyomark et al. (2013)provided the mathematical definitions of many amplitude estimators for neurological activities.If we assume that the n th sample of a wavelet decomposed detail at level i is D i (n), then we can define the following features: Root Mean Square (RMS)

Table 3 .
Best Average Accuracy (Avg Acc) results achieved using different Daubechies functions with different features and a variable number of Hidden Layers (HL) for the NN classifier

Table 4 .
Best Average Accuracy (Avg Acc) results achieved using different Symlets functions with different features and a variable number of Hidden Layers (HL) for the NN classifier