Activation of Environment Control Units (ECU's) by Mental Prosthesis

J. Santhosh, Computer Centre, IIT Delhi
S. Anand, Centre for Biomedical Engineering, IIT Delhi and AIIMS, New Delhi
M. Bhatia, Department of Neurology, AIIMS, New Delhi

Abstract: 

The paper describes a novel approach to establish a hemispheric asymmetry between two mental states as 'normal Relaxed' and 'movement Imagery' of a specific motor activity, from raw electroencephalogram (EEG) data. This characteristic difference was later used as control signals for the activation of environment control units (ECU's) like bell, light, fan or TV, as a communication device for severely handicapped. Brain controls muscular movement and a change in EEG potential reflecting a subject's readiness to move (Readiness Potential, RP) appears as a negative voltage deflection, at least one second before initiating any movement. This negative voltage deflection known as Event Related Desynchronisation (ERD), occurring in the contra-lateral hemisphere during a movement imagery, is due to changes in EEG power from a normal relaxed state. This attenuation of power is extracted by suitable electrode placement in and around the motor cortex on the scalp. A study on 10 healthy right handed subjects and on 5 neuro-muscular disorder patients, successfully detected the change in EEG pattern between 'normal Relaxed' and 'movement Imagery' states. Advanced feature extraction and pattern recognition methods like wavelet packets and artificial neural networks are used for the classification of these two states from raw electroencephalogram data in to binary output 0/1 which then used as control signals to operate an electronic switch.

Keywords: Brain Computer Interface (BCI), Wavelet Packets, Artificial Neural Networks, Electroencephalogram (EEG), Severely Handicapped.

1. Introduction
Electroencephalogram (EEG) reflects neuronal activity of a person [1]. A correlation of neural activity and mental experience could provide an alternate channel for sending messages to the external world through the help of a computer interface, which is called a Brain Computer Interface (BCI) [2-5]. Communication is a severe handicap in all affected by diseases like amyotrophic lateral sclerosis (ALS), spinal cord injuries of cervical level, spastics and other such neuromuscular disorders [6]. Their healthy thoughts cannot be communicated to external world due to their inability to twitch the muscles. Some minimal quality of life can be offered to them by converting their healthy thoughts into activation of controls for movement or selection of letters to compose words for communication through a computer [7-10]. Movement related potential changes preceding voluntary movement have been reported earlier [11-15]. During planning stage of a movement, at least one second before, the relaxed neuronal activity from the motor area known as Mu rhythm gets disrupted over the contra-lateral hemisphere there by attenuate the EEG power. This change of activity in Mu rhythm is called Event Related Desynchronization (ERD). Mu rhythm, occurring at 7-13 Hz, returns to baseline levels within a second after the initiation of movement and may increase briefly above the baseline, which is known as Event Related Synchronization (ERS) [16]. Neural centers like supplementary motor area (SMA) and premotor area (PMA) initiates the readiness potential and primary motor cortex is responsible for later initiating the movement activity [17-19]. Hence the relevant brain signals are obtained by placing EEG electrodes in and around C3 & C4 positions as per the 10-20 system of electrode placement [20]. 

The attenuation of EEG power during planning of movement when compared to relaxed state can be extracted after averaging the relax state which is taken as a baseline task. In the present study, activity related brain responses reflected in Electroencephalogram (EEG) were recorded at specific frequencies and time to examine relevant information through sophisticated signal processing methods [21]. Earlier attempts made by scientists had reported that application of raw EEG data or data parameterized, gave poor reliability even with long computation time. Latest attempts recommend wavelet transform as more efficient and faster method, providing both time and frequency characteristics of signal [22]. 

Compared to earlier attempts, the present study used a different paradigm for the experiment to obtain the relevant EEG signal and also used wavelet packets to get wavelet coefficients corresponding to the required frequency components. Artificial Neural Network (ANN) have been already established to give best results for input parameters evaluated on the basis of wavelet coefficients there by offering a powerful frame work for temporal processing [23-26], hence ANN was used as the classifier for the present study. 

2. Methodology
2.1 Subject Selection and Data Acquisition

For the current study EEG data was collected for 15 times on various days from 10 healthy right-handed subjects of 20-63 years of age and 5 ALS (Amyotrophic Lateral Sclerosis) patients of 25-65 years of age. The data was recorded on a Medelec Profile Digital EEG machine available in EEG laboratory, Department of Neurology, All India Institute of Medical Sciences, New Delhi. The settings of: High Frequency filter 50 Hz, Low Frequency filter 1.6 Hz, Notch filter 50 Hz, Sensitivity 70 microVolts/mm and a Sampling rate of 256 Hz, were used for the basic signal processing.

EEG electrodes are placed according to the international standard 10-20 system of electrode placement. The placement of electrodes is shown in Figure 1. Bipolar and unipolar EEG was recorded from eight Ag/AgCl scalp electrodes, which were placed 2.5 cm anterior and posterior to the central electrodes C3 and C4 (left and right side of the hemisphere). The reference electrodes are placed on the left and right ears and the ground electrode on the forehead. EOG (Electrooculogram) being a noise artifact, was derived from two electrodes, placed on the outer canthus of left and right eye in order to detect eye movement. These EOG signals are then used to eliminate eye movement artifacts. 

Fig 1. Electrode placement for the present study.

The subject was asked to lie down comfortably in a relaxed position with eyes closed and advised to minimize eye movements. The EEG was recorded for the relaxed state for 5 minutes. Following this, an audio beep of 60 db and 0.91 sec duration was given at the start and end of a 5 second epoch where the subject was asked to mentally plan lifting of the right hand thumb. This activity is collected as a 5 second epoch data corresponding to 'movement Imagery' state. After a gap of 5 minutes the same cue is given to repeat the experiment. The whole experiment lasts for approximately 30 minutes collecting data for 5 sessions of 5 second epoch each for normal relaxed state and 5 sessions of 5 second epoch each for movement imagery. No actual movement is performed during the session. All data sets were visually checked for artifacts before final selection. The data collected, taken as relaxed and movement planning events are then classified using the software MATLAB on a Pentium III PC. The experiment paradigm shown in figure 2 involves the following steps:

Fig 2. Experiment paradigm. (A - start of 1st beep, B - start of 2nd beep)

• Obtain the distinctive parameters of readiness potentials (RPs) associated with the planning of voluntary movements (Right hand thumb) against a baseline (Relax) task.

• Classify these RP patterns to distinguish between the two mental states.

2.2 Feature Extraction and Classification
Wavelet packet analysis has been used for signal decomposition with equal frequency bandwidth at each level of decomposition, which lead to an equal number of the approximation and detail coefficients [27-29]. By applying Wavelet packet analysis on the original signal (for 256 samples at sampling frequency 256 Hz) we obtained the wavelet coefficients in the 7-13 Hz frequency band at the 5th level node (5,3). The signal is reconstructed at node (5,3) and its FFT plot gave the frequency band 7-13 Hz as the most discriminating, in conjunction with the wavelet Daubechies#5 (db5). Daubechies wavelets were chosen because of the physiologically shaped basic wavelet [30].

The artificial neural network used was the standard Multiplayer Perceptron (MLP) with back-propagation for learning. There are four steps involved in the training process: Assembling the training data, Create the network object, Train the network and Simulate the network response to new inputs. The Wavelet coefficients extracted from the 7-13 Hz frequency components were fed to the designed ANN with a 16-35-15-1 structure having sigmoid transfer functions. The network object always got initialized before use. The simulation function takes the network input and network object, and returns the network response towards the new input. The network design and training had been done with the MATLAB's Neural Networking tool kit. Figure 3 shows the EEG analysis circuitry.

Fig. 3. EEG Analysis Circuitry

3. Results
The Fast Fourier Transformation (FFT) revealed the sudden change in mu rhythm, the 7-13 Hz frequency band from the motor area. The hemispheric asymmetry between the two states was established after analyzing the raw EEG data through various processes as described in methodology. The indigenously designed ANN architecture was trained to distinguish the features from the two mental states, which are in the form of wavelet coefficients, and to provide the classification in binary values as either "0" or "1". These binary outputs are later used as control signals to operate an electronic switch as OFF/ON. Classification accuracy up to 87% is achieved between the two mental states for the present study. Figure 4 shows the FFT plots of signals from the left hemispheric electrode C3-A1, for the two states. Figure 5 shows the corresponding wavelet coefficient plots and Figure 6 gives the maximum difference between the two states from the left hemispheric monopolar electrode pairs, and shows that the maximum difference occurs in central electrode C3-A1. All figures correspond to the analysis of one second data, from a 5 second epoch. 

Fig. 4. The FFT plots of signals from coefficients C3-A1 electrode for the two states.	Fig. 5. Plots of extracted wavelet for the two states.

Fig. 6. Maximum difference between 'normal Relaxed' and 'Movement Imagery' states, from each of the monopolar electrode pair from left hemisphere. Plots shows that the maximum difference occurs in central electrode, C3-A1.

4. Discussions
Processing with Wavelets and Artificial Neural Network, the study detected the changes in EEG power as Event Related Desynchronisation (ERD) from the raw EEG data corresponding to a movement imagery task with respect to normal relaxed state. A total of 15 experiment each having 5 sessions of 5 second epoch each for the two states were collected for the present study. The analysis proved 87% of accuracy in pattern recognition. The advantage of choosing readiness potential for the study is that, apart from the fact that readiness potentials are present in every normal individual as well as in persons with severe motor deficit, it is easy to conduct experiment without the need of a stimulus. Earlier attempts by scientists were primarily to evaluate RPs for cursor movement on screen, slow cortical potential (SCP) shift for word compilation, occipital alpha wave for hand grasp through functional electrical stimulation (FES) [7-9]. But the present study made use of RPs for the design of a communication device by activating an electronic switch. Also, Wavelet packets were employed for feature extraction instead of simple methods like Fast Fourier Transform (FFT), Auto Regressive (AR), Linear Discriminant Analysis (LDA), and Wavelet Transform as used in earlier experiments. Here FFT was employed to identify the exact frequency bands where attenuation of EEG power occurs. Wavelet Packet analysis was applied to get the wavelet coefficients corresponding to the required frequency components. These coefficients were then fed to the neural network architecture designed for this study, which then classifies the two mental states accordingly in to binary values 0/1. These binary values are then used as control signals to an electronic switch which in turn can operate any environment control units.

The paper integrates well diverse disciplines like computer science, electrical engineering and medical science so as to evolve a clinically acceptable solution and decode human thoughts using wavelets and artificial neural net. Further studies are underway to improve the efficiency of the system and use it for other affected neurological patients. 

5. Conclusions
The present study offers novel approach to obtain binary outputs through Wavelet packet analysis and ANN for classifying mental states. The technique can be easily extended to various other applications where different selection processes can be reduced to a series of binary selections. These binary selections can be used as parameters useful in areas like rehabilitation, bioengineering applications, supplementary motor area related neurological diseases, neuro-scientific research in behavioral sciences and in Man-Machine interface. 

6. Acknowledgements
The authors would like to acknowledge the financial support from the Department of Science and Technology (DST), Government of India, New Delhi for carrying out the study. The authors also thank the technical staff at Rehabilitation laboratory, Indian Institute of Technology, Delhi and Clinical Neurophysiology laboratory, All India Institute of Medical Sciences, New Delhi for the help in carrying out the experiment.

REFERENCES

1. Elman J, Bates E, Johnson M, Karmiloff-Smith A, Parisi D, Plunkett K. Rethinking Innateness: A connectionist perspective on development, Chapter 5, Cambridge, MIT/Bradford Books, 1996.

2. Pfurtscheller G, Flotzinger D, Kalcher, J. Brain Computer Interface - a new communication device for handicapped persons. Journal of Microcomputer Applications,16, 293-299, 1993.

3. Vaughan T M, Guest editorial brain-computer interface technology: a review of the second international meeting, IEEE Trans. Rehab. Engg., 11 (2), 94-109, June 2003.

4. Culpepper B J, Keller R M. Enabling Computer Decisions Based on EEG Input, IEEE Trans. Rehab. Engg., 11 (4), 354-360, Dec 2003.

5. Cincotti F, Mattia D, Babiloni C, Carducci F, Salinari S, Bianchi L, Marciani M G, F. Babiloni. The use of EEG modifications due to motor imagery for brain-computer interfaces, The use of EEG modifications due to motor imagery for brain-computer interfaces, IEEE Trans. Rehab. Engg., 11(2), 131-133, June 2003.

6. Hayashi H, Kato S. Total manifestations of amyotrophic lateral sclerosis-ALS in the totally locked-in state. J. Neuroscience, 93, 19-35, 1989.

7. Neuper C, Muller G R, Kuber A, Birbaumer N, Pfurtscheller G. Clinical application of an EEG based brain-computer interface: a case study in a patient with severe motor impairment, Clinical Neurophysiology 114, 399-409, 2003.

8. Birbaumer N, Kuber A, Ghanayim N, Hinterberger T, Perelmouter J et al. The thought translation device (TTD) for completely paralysed patients, IEEE Trans. Rehab. Engg., 8, 190-193, 2000.

9. Heasman J M, Scott T R D, Kirkup L, Flynn R V et al. Control of a hand grasp neuroprosthesis usinh an electroencephalogram-triggered switch: demonstration of improvements in performance using wavepacket analysis, Medical & Biological Engineering & Computing, 40, 588-93, 2002.

10. Palaniappan R, Paramesran R, A New Brain-Computer Interface Design using Fuzzy ARTMAP, IEEE Transactions on Neural systems and Rehabilitation Engineering, 10(3), 140-48, 2002.

11. Kristeva R, Keller E, Deecke L, Kornhuber HH. Cerebral Potentials preceding unilateral and simultaneous bilateral finger movements. Electroencephalography and Clin. Neurophysiology, 47(2), 229-38, 1979.

12. Praamstra P, Stegeman DF, Horstink MW, Brunia CH, Cools AR. Movement-related potentials preceding voluntary movement are modulated by the mode of movement selection. Experimental Brain Research, 103(3), 429-39, 1995.

13. Praamstra P, Stegeman DF, Horstink MW, Cools AR. Dipole source analysis suggests selective modulation of the supplementary motor area contribution to he readiness potential. Electroencephalography and Clin. Neurophysiology. 98(6), 468-477, 1996.

14. Dirnberger G, Fiukel U, Lindinger G, Lang W, Jahanshani M. The mode of movement selection. Movement-related cortical potentials prior to freely selected and repetitive movements. Experimental Brain Research, 120(2), 263-72, 1998.

15. Allison BZ, Vankov A, Hughes J L, Arnoldussen A, Pineda JA . Near real time recognition if EEG changes associated with movement. Proceedings of Society for Neuroscience annual meeting, Los Angeles, C.A, 1998.

16. Pfurtscheller G, Neuper C. Simultaneous EEG 10 hz desynchronisation and 40 Hz synchronization during finger movements. Neuroreport, 3 (12), 1057-1060, 1992.

17. Stancak A Jr, Pfurtscheller G. The effects of handedness and type of movement on the contralateral preponderance of m-rhythm desynchronisation. Electroencephalography & Clin. Neurophysiology, 99(2), 174-182, 1996.

18. Deecke L, Scheid P, Kornhuber H H. Distribution of readiness potential, pre-motion positively, and motor potential of the human cerebral cortex preceding voluntary finger movements. Experimental Brain Research, 7(2), 158-168, 1969.

19. Deecke L. Bereitschaftspotential as an indicator of movement preparation in supplementary motor area and motor cortex. Ciba Foundation Symposium, 132(11), 231-50, 1987.

20. Pregenzer M., Pfurtscheller G., Flotzinger D : Selection of electrode positions for for an EEG based Brain Computer Interface (BCI). Biomedizinische Technik, 39, 264-269, 1994.

21. Eugene N Bruce. Biomedical Signal Processing and Signal Modeling. John Wiley & Sons, Canada; 2001.

22. Thilo hinterberger, Kubler A, Kaiser J, Neumann N, Birbaumer N, A brain-computer interface (BCI) for the locked-in: comparison of different EEG classifications for the thought translation device, Clinical Neurophysiology 114, pp, 416-425, 2003.

23. Donna L Hudson, Maurice E Cohen. Neural Networks and Artificial Intelligence for Biomedical Engineering. IEEE Press, NY, 2000.

24. Bishop CM. Neural networks for pattern recognition. Oxford edn: Clarendon Press 1995.

25. Simon Haykin. Neural Networks- A Comprehensive Foundation. 2nd Edn. Pearson Edu, Inc. 1999. 

26. Neural Network Toolbox with MATLAB Version 6.

27. Gilbert Strang, Truong Nguyen. Wavelets and Filter Banks. Wellesley-Cambridge, USA, 1996.

28. Wavelet Toolbox for use with MATLAB Version 6

29. Raghuveer M Rao, Ajit SB. Wavelet Transforms: Introduction to theory and applications. Pearson Education Asia, 2nd Reprint 2001.

30. Dickhaus H, Heinrich H. Classifying biosignals with wavelet networks. IEEE Engg in Med Biology, 15, 103-111, 1996.