Motor Imagery based Brain Computer Interface (BCI) using Artificial Neural Network Classifiers Maitreyee Wairagkar Brain Embodiment Lab, School of Systems Engineering, University of Reading, Reading, U.K. m.n.wairagkar@pgr.reading.ac.uk, maitreyeew@yahoo.co.in Abstract Brain Computer Interface (BCI) facilitates the communication with external devices (Computer) directly by the brain without involvement of any motor pathways. These systems are useful for assisting people with the impaired motor abilities. The project envisaged the development of a reliable motor imagery BCI capable of classifying right and left hand imaginary movements for controlling a computer cursor. The electroencephalogram (EEG) signals were recorded from the sensory-motor region of the brain using wireless recording device. Using the Event Related Desynchronization (ERD), movement related features were extracted from EEG with the spatial and spectral filtering. These features were classified into two classes using the non-liner Radial Basis Function based Artificial Neural Network (ANN) classifiers. A peak accuracy of 80% was achieved for classifying imaginary hand movements of 16 different subjects. Combination of the non-linear ANN classifiers with the signal processing techniques proved to be an effective method for classifying motor imagery for BCI. 1 Introduction Brain Computer Interface (BCI) provides a new mode of the communication and interaction with a device in an external environment directly via brain, without the involvement of any motor pathways. Different mental states lead to the changes in the electro-physiological signals in brain. BCI records the brain signals, extracts the features related to the changes in mental states and translates them to operative command signals to control the external devices. BCIs do not rely on the brain s output motor pathways including peripheral nervous system and muscles and hence, these systems are especially useful for the people with impaired motor abilities suffering from paraplegia, tetraplegia, Spinal Cord Injuries and many more. It could be used for controlling a wheelchair, moving the cursor on screen, using a spelling device, controlling the neuroprosthetics and other similar devices just by thoughts. BCI is the promising system for allowing patients with locked-in-syndrome to communicate with external world directly via brain. Components of BCI system are shown in Figure 1. BCI operates on an EEG (Electroencephalogram), the electrical signals recorded from scalp over the desired regions of the brain. An EEG has very good temporal resolution [1]. BCIs use the different types of paradigms to operate like Visually Evoked Potentials, Event Related Potentials, Readiness Potentials and Motor Imagery. In this project, the Motor Imagery based BCI was created to control the movement of the mobile robot capable of navigating in the environment. Such assistive robots could be used by the patients for performing the simple day to day activities.
Figure 1: Components of BCI Figure 2: ERD for right hand movement in C3 The right and left hand motor imagery were used for generating the commands for robot by an imagination of the movement. The main aim of the project was to classify between the right/left hand imagery and the resting state. Signal processing and filtering was done on the EEG for extracting the distinguishable features from these three mental states. The Artificial Neural Network (ANN) classifiers were used for classifying the features in order to generate the appropriate commands. The results of classification were translated into the control commands which could be then sent to the different user applications. 2 Motor Imagery and Event Related Desnchronization The rhythmic neural oscillations give rise to the different brain rhythms over different areas of the cortex. Mu rhythm (8 to 13Hz) is observed over the motor cortex [1] which is responsible for the planning, preparation, sequencing, coordination and execution of the voluntary movements. Prominent changes in the brain rhythms are observed in the sensory-motor area of brain during the planning and execution of movements. Thus, the mu rhythm is associated with the planning and execution of voluntary movements of the body [2]. According to the homunculus representation of the body in the primary motor cortex, the motor cortex is divided into the different parts controlling the movements of the body part represented by that section. Each side of the body is controlled by the contralateral hemisphere of the brain. Placing electrodes on the location of the representation of the required body part allows recording the brain activities related to the movement of that part. Thus a movement could be detected from the EEG recorded from the appropriate position on the cortex. Imagining a movement or performing an action mentally is known as Motor Imagery. The Motor imagery produces similar effects on the brain rhythm in the sensory-motor cortex as the real executed movement [3]. The Mu rhythm power undergoes attenuation during the preparation and execution of movement. With an appropriate training and feedback, an individual can learn to control this mu power which could be used to produce the commands for BCI [4]. Different imagined movements could be distinguished by considering the power in the brain rhythms on the sensory- motor cortex during the preparation and execution of the real and an imaginary movement. Maximum mu activity is observed during the resting state. Mu power is attenuated or supressed during the preparation and execution of the movement. This phenomenon of suppression of the mu power is known as an Event Related Desynchronisation (ERD) [5]. ERD is observed in the mu band from about 2 seconds before the onset of movement. This is illustrated in Figure 2. The movement of limb produces a prominent ERD in the contralateral side.
Methods Figure 3: Experimental Setup The development of the BCI involved four major steps data acquisition, noise and artefacts removal, signal processing for feature extraction, classification of the extracted features and converting them into commands for application. This section details the BCI development process. 2.1 Experimental Procedure and EEG acquisition The EEG data recorded from 16 right handed participants, sampled at 256Hz was used. Data was obtained from the Brain Embodiment Lab of the University of Reading. This contained EEG for the right and the left hand imaginary as well as the real finger taps and the resting state. The data was recorded using the TruScan Deymed EEG system. The participants were seated in a comfortable chair. The cue was displayed on the screen for indicating which hand finger tapping was to be performed. The cues for right, left and resting state were displayed randomly to avoid the pattern learning. The cue was turned off when participant pressed key with required hand finger, thus the finger tapping movement was performed. Two seconds trial after the displaying of the cue was extracted for an analysis. Thus, in each real movement trial, the subject performed single right or left hand finger tap. During the imaginary movement trials, subject imagined single right or left hand finger tap according to the cue. During the resting state trials, subject did not do anything and remained relaxed. This was used as the baseline for identifying movements. Total hundred trials for each condition were recorded. The experimental setup is illustrated in Figure 3. Deymed amplifier was used for the EEG recording. Wireless EPOC headset was also tested for the EEG acquisitions; however, these data were not used due to high noise contamination. EEG from two channels C3 and C4 was used as these channels are located over the sensory-motor area of cortex. 2.2 EEG pre-processing and artefacts removal Figure 4: EEG before (red) and after (blue) artefacts removal The aim of pre-processing was to enhance the required information content from the signal and to increase the quality of features to be extracted. Pre-processing was done using a Butterworth filter to eliminate noise in EEG. The DC offset in the signal was removed using a high-pass filter with a cutoff frequency of 0.5Hz. The power line noise was filtered using a notch filter at 50Hz. Finally, the high frequency noise was eliminated using a low-pass filter with a cut-off frequency of 60Hz.
The Artefacts can occur in EEG due to eye movements, blinks, breathing and other body movements. It was important to remove these artefacts because they cause the distortion in the signal leading to the loss of required information. An Independent Component Analysis (ICA) [6] was used to remove artefacts from the recorded signals. ICA decomposed the signal into its independent sources. Artefacts often have a different source than EEG. The Independent components (ICs) with artefacts were identified manually. An Artefact-free EEG was reconstructed by eliminating these ICs. An EEG sample channel before and after artefacts removal is shown in Figure 4. 2.3 Signal processing and feature extraction EEG signal contained the information of various processes going on in the brain and hence it was essential to extract the features from EEG that were only representative of motor imagery. In order to extract these motor imagery features, various signal processing operations were performed. The movement and motor imagery gives rise to an ERD as explained in section 2 which means that the power of the mu band decreases during the real as well as an imaginary movement. To extract the power of mu band, first, the mu band was segregated from EEG by band-pass filtering between 8Hz to 13Hz. EEG before and after band-pass filtering is shown in Figure 5. The Fourier transform of this band-pass filtered signal was computed to obtain the power of the signal at different frequencies (see Figure 6). The maximum power in the mu band was selected as the feature. The features were extracted from EEG channels C3 over the left motor cortex and C4 over right motor cortex. Thus, from each trial, two features were extracted. The C3 vs. C4 feature plot was made for all the trials of right and, left hand imagery and the resting state. The features from right and left hand motor imagery trials are shown in Figure 8. These features were then sent to the classifiers for classification. Figure 5: Raw and filtered EEG Figure 6: Fourier transform of filtered EEG. Extracted feature is encircled 2.4 Classification using artificial neural networks In order to classify the extracted features, the artificial neural network (ANN) classifiers were used which demonstrated higher classification accuracy due to their ability of learning from the training data. The two stage classification process was used for the classification of the features to first detect the presence of movement and then if the movement is present, classify right and left hand movement. For the first stage of classification, Probabilistic Neural Network (PNN) was used. The PNN is linear classifier and was able to classify the linearly classifiable features of movement vs. rest state. If a feature was classified as a movement, second stage classification was done. The Radial Basis Function (RBF) classifier (see Figure 9) was used to classify the feature identified as the movement feature in the first stage further as right or left hand imagery. The RBF, a non-linear classifier was chosen because the right and left hand features were non-linearly classifiable
i.e. could not be separated using a straight line. The RBF classifier computed the curved boundary that would classify these two classes. Thus using this two stage classification, the feature was identified as rest state, right hand motor imagery or left hand motor imagery. The classified feature was then translated into command for driving an application like mobile robot and a cursor control on the screen. Figure 7: Power spectrum of right (blue), left (red) hand movement and rest (black) Figure 8: Features of right (blue) and left (red) hand movement Figure 9: Structure of RBF neural network classifier 3 Results The clear distinction between the resting state trials, right and left hand trials was observed. Figure 7 shows the Fourier transforms of all the three conditions. It could be seen that the mu power is highest during the resting state, where as the mu power is attenuated for right and left hand imagery. This confirms the occurrence of ERD during motor imagery. Thus, this decrease in the mu power is the clear indication of the motor imagery. The final classification accuracy of the two stage classification using PNN and RBF classifier was determined for 100 trials for each condition. The peak classification accuracy of 82% was obtained for participant E for real left hand movement. Average classification accuracy for right vs. left hand real as well as imagery movement was 63.8%. It was observed that the classification accuracies for an
actual movement and the imaginary movement are in the same range. all the 16 participants are shown in Table 1. Classification accuracies for Participant ID Real Movement classification (%) Imaginary Movement classification (%) Right hand Left hand Right hand Left hand A 49 52 55 53 B 41 61 62 59 C 68 50 74 78 D 57 60 54 59 E 67 82 65 62 F 63 59 68 72 G 69 57 63 66 H 61 65 44 64 I 63 56 66 68 J 68 62 55 62 K 75 76 68 59 L 74 78 71 64 M 73 73 81 67 N 64 56 68 62 O 59 56 57 59 P 71 72 65 74 Average 63.87 63.4375 63.5 64.25 S.D. 9.1497 9.7977 8.9592 6.4343 Table 1: Results for real and imaginary Right and Left hand movements for 16 participants. 4 Conclusions and future work The Signal processing and classification technique was developed in this project for the motor imagery based BCI. The BCI was capable of classifying the rest state, the right hand motor imagery and the left hand motor imagery with the use of ANNs. The property of an ANN to learn the pattern from the training data allowed it to classify the new unseen data efficiently. Thus, using two staged ANN classification process proved to be a very effective in classifying the unknown EEG trials instead of using the fixed thresholding approach.
The spectral features based on the mu power captured the ERD occurrence during real as well as imaginary movements. Thus, the movement imagery was appropriately detected and then converted into the command for an application module. The system developed was modular and hence could be used to drive the different applications. In order to improve the detection accuracy of the system, the advanced signal processing algorithms would be used. The combination of the multiple movement features extracted from EEG would give better classification accuracies. Moreover, attenuation of the mu band is not the only indicator of the movement in EEG. Other movement related information in EEG would be studied in future. References [1] D. McFarland, L. Miner, J. Wolpaw and T. Vaughan, Mu and Beta Rhythm Topologies During Motor Imagery and Actual Movements, Brain Topography, vol. 12, no. 3, pp. 177-186, 2000. [2] H. Jasper and W. Penfield, Electrocorticograms in man: effect of the voluntary movement upon the electrical activity of the precentral gyrus, Arch. Psychiat. Z. Neurol., vol. 138, pp. 163-174, 1949. [3] C. Hema, M. Paulraj, S. Yaacob, A. Adom and R. Nagrajan, An Analysis of the Effect of EEG Frequency Bands on the Classification of Motor Imagery Signals, Biomedical Soft Computing and Human Sciences, vol. 16, no. 1, pp. 121-126, 2010. [4] G. Pfurtscheller and C. Neuper, Motor Imagery and Direct Brain-Computer Communication, IEEE Transactions, vol. 89, no. 7, pp. 1123-1134, 2001. [5] G. Pfurtscheller and A. Aranibar, Event-related cortical desynchronization detected by power measurements of scalp EEG, Clinical Neurophysiology, vol. 42, pp. 817-826, 1977 [6] T. Jung, S. Makeig, C. Humpharies and T. Lee, Removing electroencephalographic artifacts by blind source separation, Psychophysiology, 37, p. 163 178, 2000.