2018 18th International Conference on Control, Automation and Systems (ICCAS 2018) Oct. 17~20, 2018; YongPyong Resort, PyeongChang, GangWon, Korea
Unorthodox Approach to Classify EEG Signals for Upper Limb Prosthesis Application Nasir Rashid1, Fahad Mahmood1, Javaid Iqbal1 1
National University of Sciences and Technology, H-12 Islamabad, Pakistan
Corresponding author: Nasir Rashid (e-mail:
[email protected]).
Abstract: Brain computer interface (BCI) is targeted for decoding the EEG (Electroencephalogram) signals that the human brain generates which are beneficial for the paraplegic patients. These EEG signals are slow cortical potentials that are directly recorded from scalp thus cortical neuronal activity is explored via non-invasive electrodes. These EEG signals are then further utilized for performing various operations which the paraplegic patients are unable to perform. This research article presents a novel architecture of classification of four finger movements (thumb movement, index finger movement, middle and index finger combined movement and fist movement) of the right hand on the basis of EEG (Electroencephalogram) data of the movements. The presented architecture utilizes Guided filter for reduction of noise (artefacts) from EEG signals alpha and beta band (8-30 Hz). As this band contains the maximum information of movement in terms of motor imagery. Rank Transform is employed as feature extraction approach for further enhancement of processed EEG signals. Two stage Logistic Regression classifier is finally employed for classification of movements using processed EEG signals. The experimental results demonstrate the accuracy, robustness and computational complexity of the proposed approach and have significant improvement as compared with recent architectures for EEG classification. Keywords: Brain Computer Interface (BCI), Rank Transform (RT), Guided Filtering, EEG, Logistics Regression Wang et al. [5] classified left and right finger movements using the architecture of common spatial decomposition as feature extraction approach and support vector clustering as classification algorithm. Literature survey of brain computer interface reveals that Linear Discriminant Analysis, Quadratic Discriminant Analysis, Support vector Machines and most variants of Neural Networks are employed for the classification of most of the movements utilizing EEG (Electroencephalographic) signals [6]. J. Lehtonen [7] presented the online classification architecture of single trial EEG signals for the purpose of finger movements. Muller-Putz et.al [8] proposed the control scheme of electrical prosthesis with a steady-state visual evoked potentials. He controlled the two axis electrical hand prosthesis and performed experiments on four patients. After performing experiments, he got the accuracy between 44% to 88%. Long et.al [9] controlled the 2-D mouse cursor from EEG signals based on hybrid feature selection. Hybrid features from brain computer interface are finally controlled by left-right hand motor imagery. LaFleur et.al [10] controlled the three dimensional motion of Quadcopter using a non-invasive brain computer interface. He got 90.5% accuracy with five subjects. In our previous research [11], we employed power spectral density for recognition of finger movements using EEG data and two stage logistics regression [12] for the classification of EEG signals. Kalaivani et.al [13] explored the EEG signals for detection of abnormalities of brain signals. Liao et.al [14] performed Brain computer interface BCI using individual finger movements from a single hand using EEG signals coming from brain. Guger et.al [15] performed rapid prototyping of an EEG-based brain computer interface.
1. INTRODUCTION Brain computer interface (BCI) decodes the signals of brain through noninvasive electrodes to perform critical tasks that the paraplegic patients are unable to perform. Brain computer interface is a pathway to communicate with the external world bypassing normal neuromuscular pathways [1]. The patients having permanent muscular disability such as ALS (Amyotrophic Lateral Sclerosis) or a spinal cord injury will be able to communicate with the external world. The EEG signals coming from noninvasive electrodes directly attached to the scalp of brain will be decoded and perform the most critical actions which a paraplegic patient is unable to perform. The EEG signals coming from brain will be transformed into digital domain and will be processed through signal processing. The signal processing architecture consists of filtration scheme to remove the artifacts, feature extraction approach for further enhancement of signals and finally the machine learning approach for classifying the incoming signals [2]. The chosen feature extraction should have large signal to noise ratio, have least computation and should possess highest accuracy. RanXiao et al. [3] classified ten classes of movements of finger pairs utilizing Power spectral density for feature extraction and decomposed it using principal component analysis. Classification was done using support vector machines. Vuckovic [1] at al. used Gabor coefficients as features to classify the flexion and extension of left and right wrists. They employed 2-stage 4-class Elman’s neural network for classification purposes. Mohamed at al. [4] classified the finger and wrists movements using Artificial neural network, Bhattacharya distance for feature extraction and independent component analysis for filtration purposes.
978-89-93215-15-1/18/$31.00 • •ICROS
508
EEG DATA ACQUISITION
GUIDED FILTER
RANK TRANSFORM
Logistics Regression
PROSTHESIS ARM
Figure 1: Proposed Scheme Architecture Wolpaw et.al [2] presented the brain computer interface architecture for communication and control purposes, which is the main purpose of brain computer interface (BCI) research. BCI also finds its application in playing games [16]. Yom-Tov et.al [17] proposed the feature selection of the extracted EEG signals for classification of movements from single movement-related potentials. Lee.Set.al [18] proposed an architecture for the removal of high-voltage brain stimulation artefacts from the EEG recordings. Wang, D et.al [19] proposed an accurate and automatic detection of epileptic seizures in long term EEG signals using wavelet decomposition and the directed transfer function algorithm. Jacobs et.al [20] proposed a machine learning based non-invasive EEG that flags an alarm in advance of a clinical seizure onset. Vukovi [21] explored the EEG signals by proposing a two class and four stage brain computer interface system for imaginary right and left wrists movements. The main contribution of this paper is to present a novel, accurate and robust scheme of recognition of EEG signals. Guided filtering approach [22] is employed for removing of unwanted noisy data. EEG signals are filtered to acquire alpha and beta band (8-30 Hz) to retain the maximum information of the movement signals. Rank Transform [23] is applied on the filtered EEG signals to enhance the raw signals data. Finally, the two stage logistic regression [24] is applied to accurately classify the movements using processed EEG signals. These movements are finally targeted so they can be further utilized for the control of upper limb prosthesis.
Figure 1 shows the proposed scheme architecture presented in this research article. 2.1 Data Acquisition The EEG (Electroencephalogram) signals was acquired on various patients in a hospital. Four classes representing thumb movement, fist movement, index movement and two (index and middle) finger movement. The movements of 10 seconds each are recorded on the right hand. The data is recorded with Emotive electrode headset having 14 channels each. The EEG signals are recorded at 128Hz sampling rate. The EEG (Electroencephalogram) topography of the raw data of each movement is shown in Figure 2.
Figure 2: Topology of raw data of finger movements of right hand (A) Thumb movement. (B) Index finger movement. (C) Index and middle finger movement. (D) Fist movement [11]
2. PROPOSED ARCHITECTURE
2.2 Guided Filter
The proposed architecture of classifying EEG (Electroencephalogram) signals consists of Guided filtering approach [22] for removing the unwanted noise from the signals, i.e. to suppress the unwanted information and to enhance the data which is suitable for Brain Computer Interface. Guided filter is effective in a variety of the image processing, computer vision and computer graphics applications. For further enhancement of EEG signals, we employed Rank Transform [24]. Rank Transform is a feature extraction technique which is successfully applied in face recognition, signal processing and others. Two stage logistics regression [25] is employed for accurate classification of EEG signals. To compute the accuracy for our architecture, we performed experiment on prosthesis arm. Experiment results demonstrate the accuracy, robustness and computational complexity of our architecture. Each block of our architecture is explained in detail.
This article recommends to employ Guided Image Filtering mechanism [22] to remove the unwanted signals and noise from the EEG (Electroencephalogram) signals for BCI (brain computer interface) applications. The main purpose of applying Guided filter in BCI application lies in its fast computation and less complexity which is independent of filtering kernel size. ܹǡ ሺܫሻ ൌ
ͳ ȁݓȁଶ
ǣሺǡሻఢ௪ೖ
ቆͳ
ሺܫ െ ߤ ሻ ൫ܫ െ ߤ ൯ ߪଶ ߳
In Guided image filtering, the weights independent of the guidance image ܫand explained as: In the above mathematical equation, ߪ represents the variance and mean of the guidance image.
509
ቇ
(1)
ܹǡ are is further and ߤ computed
Classifier 2 Thumb & Finger EEG SIGNALS
Classifier 1 Fist & Two finger Classifier 2
Figure 3: Classification Scheme of Logistics Regression 2.3 Rank Transform Figure 3 demonstrates the 2-stage and 4-class model logistics regression. In this research article EEG signals are mainly classified in four classes that is thumb movement, finger movement, fist movement and two fingers movement.
In this research article, we compute the features of EEG (Electroencephalogram) signals from Rank Transform [23]. The Rank Transform encodes for each pixel the position of its gray value in the ranking of all the gray values in its neighborhood. The main purpose of employing Rank Transform in BCI applications because of its more information and exhibits strong invariance. It is computationally efficient, parameter free algorithm, robust to noise and clearly leads to improved results. Rank transform employed for each input matrix as
2.5 Prosthesis Arm The efficiency proposed scheme of classifying EEG signals is tested on embedded system based Prosthesis arm. It contains two motors connected between the fingers and palm. Arduino Uno is employed for the control of prosthesis arm. EEG signals are mainly classified in four classes i.e. thumb movement, finger movement, fist movement and two fingers movement. The presented scheme is able to classify the EEG signals coming from brain in real time. Figure 4 demonstrates the prosthesis arm employed for testing the accuracy and efficiency of EEG signals for Brain Computer Interface application. After employing our EEG classification architecture in prosthesis arm, we got much better accuracy in real time. The paraplegic patient can easily perform his real world duties in real time with our presented scheme. This 3-D printed prosthesis arm is better than metallic or aluminum because of light weight, flexible and water proof structure and insulation medium of electricity and heat. Figure 5 demonstrates the control topology of the upper limb prosthesis arm. Arduino Uno was employed for the control of the motors installed in Upper Limb Prosthesis. H-Bridge was employed for providing suitable current to the motors of Upper limb Prosthesis arm.
݂ோ் ሺݔǡ ݕሻ ൌ ܵሺܫሺݔǡ ݕሻ െ ܫ െ ͳሻ
(2)
ୀ
ͳ݂݅ ݔ Ͳ ܵሺݔǡ ݕሻ ൌ ቄ Ͳ݁ݏ݅ݓݎ݄݁ݐ ܨሺݔǡ ݕሻ ൌ
݂ோ் ሺ݅ǡ ݆ሻ
ሺǡሻאஐሺ௫ǡ௬ሻ
(3)
(4)
The Rank Transform is a 2-D filter based on spatial domain having the input as ݂ோ் ሺ݅ǡ ݆ሻ and the output as ܨሺݔǡ ݕሻ. This transform greatly enhances the input EEG (Electroencephalogram) signals which not only enhances the information of the input signals but also increases the accuracy of the mechanism. 2.4 Logistics Regression A 2-stage 4-class model of logistics regression [11] was developed in our previous research. Logistic Regression is a probabilistic statistical classification architecture that estimates the probabilistic statistical classification model that estimates the probability of an event occurring. Unlike linear regression, it does not assume linear relationship between the dependent and independent variable. For logistic regression classifier, the probability of the first class is given by ܲሺ ܩൌ ͳሻ ൌ
݁ݔሺܨ כ ்ܤሻ ሺ݁ݔሺܨ כ ்ܤሻ ͳሻ
(5)
ٻ
where ܨis a feature vector and ܤis the coefficients of logistic regression.
Figure 4: 3-D printed Prosthesis Arm Employed for classifying EEG signals coming from brain.
510
Processed EEG Data
Arduino UnoController
Upper Limb Prosthesis
H-Bridge for Motor Control
Figure 5: Architecture for the control topology of the Upper Limb Prosthesis Arm
3. RESULTS AND DISCUSSION Table 2: Confusion matrix for 2-stage logistic regression with 5 fold cross validation testing
After extensive experimentation, we conclude that our scheme is robust to noise, accurate and has least computational complexity. Out of our acquired dataset testing data samples were chosen randomly for each movement. In classification stage of test samples, 52 randomly chosen data samples were employed for testing of this embedded system. Table 1 demonstrates the confusion matrix that was obtained through testing on the randomly chosen dataset. Table 2 explains the computational accuracy in the form of confusion matrix for 52 randomly chosen dataset. Confusion matrix is a table that describes the performance of a classification model (or "classifier") on known true values of test data set. The classification accuracy is normally defined as the evaluation of binary classifiers i.e, it compares two classes and assigns them a binary attribute. Table 3 compares the proposed scheme architecture with the past approaches of BCI for motor imagery classification. Previous approaches include the variants of Common Spatial Pattern (CSP) as feature extraction technique combined with Support Vector Machine (SVM) as binary classifier. The results of our proposed scheme shows that this paradigm works in better way with 75% training and 25% test data than 5 fold cross validation data. Implementation of our technique on upper limb prosthesis through an embedded system is shown in figure 5. Arduino Uno is used for the control of H-bridge to actuate upper limb prosthesis motors.
Class
Thumb Finger Fist Two
Table 3: Comparison of presented scheme with previous approaches Features approach and Classifier
Proposed Scheme Multi-CSP2 + SVM [16] Multi-GECSP + SVM [16] Multi-sTRCSP + SVM [16]
Table 1: Confusion matrix for 2 stage model logistic regression with 75% training and 25% test set Class
Thumb Finger Fist Two
Class the feature was classified as by the classifier Thumb Finger Fist Two Fingers 32 4 9 7 9 23 6 14 4 4 29 15 8 13 10 21
Results
Classification Accuracy (%) 71.2
Standard Deviation (%) 3.26
69.07
5.40
68.73
6.35
70.43
6.73
4. CONCLUSION AND FUTURE WORK Brain computer interface (BCI) is targeted for decoding the EEG (Electroencephalogram) signals that human brain generates which can be used for driving an actuator. This can be utilized for the benefit of paraplegic patients to actuate a prosthetic limb for performing gripping tasks. This research article presents a novel architecture of classification of four finger movements (thumb movement, index finger movement, middle and index finger combined movement and fist movement) of the right hand on the basis of EEG (Electroencephalogram) motor imagery data of the
Class the feature was classified as by the classifier Thumb Finger Fist Two 8 1 2 2 3 10 0 0 0 0 10 3 0 1 3 9
511
[10] LaFleur, K., et al. (2013). "Quadcopter control in three-dimensional space using a noninvasive motor imagery-based brain–computer interface." Journal of neural engineering 10(4): 046003. [11] Javed, A., et al. (2017). "Recognition of finger movements using EEG signals for control of upper limb prosthesis using logistic regression." Biomedical Research 28(17). [12] Sturdivant, R. X., et al. (2013). Applied Logistic Regression, New York: Wiley. [13] Kalaivani, M., et al. (2014). "Analysis of EEG Signal for the Detection of Brain Abnormalities." at International Journal of Computer Applications® year. [14] Liao, K., et al. (2014). "Decoding individual finger movements from one hand using human EEG signals." PloS one 9(1): e85192. [15] Guger, C., et al. (2001). "Rapid prototyping of an EEG-based brain-computer interface (BCI)." IEEE Transactions on Neural Systems and Rehabilitation Engineering 9(1): 49-58. [16] Hasan BA, Gan JQ. Hangman BCI: An
movements. The presented architecture utilizes Guided filter for smoothing EEG signals in alpha and beta band (8-30 Hz). This band contains the maximum information of movement recorded from scalp. For feature extraction Rank transform is used as a novel technique in BCI applications. Regression classifier is employed for classification of movements using processed EEG signals. We have obtained an accuracy of 71.2% and also tested our approach on upper limb prosthesis application. In future, we are planning to explore more feature extraction and filtering approaches which are not only suitable for Brain Computer Interface application but also robust and computationally inexpensive.
REFERENCES [1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
A. Vuckovic, and F. Sepulveda, 'A Four-Class BCI Based on Motor Imagination of the Right and the Left Hand Wrist', in Applied Sciences on Biomedical and Communication Technologies, 2008. ISABEL'08. First International Symposium on (IEEE, 2008), pp. 1-4. Jonathan R. Wolpaw, Niels Birbaumer, Dennis J. McFarland, Gert Pfurtscheller, and Theresa M. Vaughan, 'Brain–Computer Interfaces for Communication and Control', Clinical neurophysiology, 113 (2002), 767-91. Ran Xiao, and Lei Ding, 'Evaluation of EEG Features in Decoding Individual Finger Movements from One Hand', Computational and mathematical methods in medicine, 2013 (2013). A. K. Mohamed, Tshilidzi Marwala, and L. R. John, 'Single-Trial EEG Discrimination between Wrist and Finger Movement Imagery and Execution in a Sensorimotor BCI', in Engineering in Medicine and Biology Society, EMBC, 2011 Annual International Conference of the IEEE (IEEE, 2011), pp. 6289-93. Boyu Wang, and Feng Wan, 'Classification of Single-Trial EEG Based on Support Vector Clustering During Finger Movement', in International Symposium on Neural Networks (Springer, 2009), pp. 354-63. F. Lotte, M. Congedo, A. Lecuyer, F. Lamarch e, and B. Arnaldi, ´ “A review of classificatio n algorithms for EEG-based brain-computer int erfaces,” J Neural Eng, vol. 4, pp. R1–R13, 2 007. J. Lehtonen, P. Jyla nki, L. Kauhanen, and M. Sams, “Online classification of single EEG trials during finger movements,” in press. Muller-Putz, G. R. and G. Pfurtscheller (2008). "Control of an electrical prosthesis with an SSVEPbased BCI." IEEE Transactions on Biomedical Engineering 55(1): 361-364. Long, J., et al. (2012). "Target selection with hybrid feature for BCI-based 2-D cursor control." IEEE Transactions on Biomedical Engineering 59(1): 132-140.
unsupervised adaptive selfǦpaced brainǦ computer interface for playing games. Comput Biol Med 2012;42:598Ǧ606. [17] Yom-Tov, E. and G. F. Inbar (2002). "Feature selection for the classification of movements from single movement-related potentials." IEEE Transactions on Neural Systems and Rehabilitation Engineering 10(3): 170-177. [18] Lee, S., et al. (2018). "Removal of high-voltage brain stimulation artifacts from simultaneous EEG recordings." IEEE Transactions on Biomedical Engineering. [19] Wang, D., et al. (2018). "Epileptic Seizure Detection in Long-Term EEG Recordings by Using Wavelet-Based Directed Transfer Function." IEEE Transactions on Biomedical Engineering. [20] Jacobs, D., et al. (2018). "Classification of PreClinical Seizure States using Scalp EEG CrossFrequency Coupling Features." IEEE Transactions on Biomedical Engineering. [21] Vukovi, A. and F. Sepulveda (2012). "A twostage four-class BCI based on imaginary movements of the left and the right wrist." Medical Engineering and Physics 34(7): 964-971. [22] He, K., et al. (2010). Guided image filtering. European conference on computer vision, Springer. [23] Mahmood, F., et al. (2016). 3-D Shape Recovery from Image Focus Using Rank Transform. International Symposium on Visual Computing, Springer. [24] Subasi, A. and E. Ercelebi (2005). "Classification of EEG signals using neural network and logistic regression." Computer methods and programs in biomedicine 78(2): 87-99.
512
