Get 20M+ Full-Text Papers For Less Than $1.50/day. Subscribe now for You or Your Team.

Learn More →

Design of Embedded System for Multivariate Classification of Finger and Thumb Movements Using EEG Signals for Control of Upper Limb Prosthesis

Design of Embedded System for Multivariate Classification of Finger and Thumb Movements Using EEG... Hindawi BioMed Research International Volume 2018, Article ID 2695106, 11 pages https://doi.org/10.1155/2018/2695106 Research Article Design of Embedded System for Multivariate Classification of Finger and Thumb Movements Using EEG Signals for Control of Upper Limb Prosthesis Nasir Rashid , Javaid Iqbal , Amna Javed, Mohsin I. Tiwana, and Umar Shahbaz Khan Department of Mechatronics Engineering, National University of Sciences & Technology, H-12, Islamabad, Pakistan Correspondence should be addressed to Nasir Rashid; n.rashid@ceme.nust.edu.pk Received 10 August 2017; Revised 21 January 2018; Accepted 13 February 2018; Published 20 May 2018 Academic Editor: Noman Naseer Copyright © 2018 Nasir Rashid et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Brain Computer Interface (BCI) determines the intent of the user from a variety of electrophysiological signals. These signals, Slow Cortical Potentials, are recorded from scalp, and cortical neuronal activity is recorded by implanted electrodes. This paper is focused on design of an embedded system that is used to control the finger movements of an upper limb prosthesis using Electroencephalogram (EEG) signals. This is a follow-up of our previous research which explored the best method to classify three movements of fingers (thumb movement, index finger movement, a nd first movement). Two-stage logistic regression classifier exhibited the highest classification accuracy while Power Spectral Density (PSD) was used as a feature of the filtered signal. eTh EEG signal data set was recorded using a 14-channel electrode headset (a noninvasive BCI system) from right-handed, neurologically intact volunteers. Mu (commonly known as alpha waves) and Beta Rhythms (8–30 Hz) containing most of the movement data were retained through filtering using “Arduino Uno” microcontroller followed by 2-stage logistic regression to obtain a mean classification accuracy of 70%. 1. Introduction Movements of a prosthesis are commonly controlled through manipulating the motion of rotary actuator (electric motor) A Brain Computer Interface (BCI) provides a communica- in a BCI system. tion system to control external device(s) in which messages or BCI system consists of input signals (electrophysiological commands are sent to external world through brain signals. activity recorded from scalp of user), a signal processor eTh se signals do not pass through the brain’s normal output (filtering the signal for desired frequency and extracting pathways of nerves and muscles. Rather, BCI provides an features for best representation of user intent), a translating alternate method to its user to interact with the world. For algorithmorclassiefi r(thatanticipatesthehumanintentfrom example, Electroencephalogram (EEG) based BCI messages the selected feature), and n fi ally a control algorithm that are encoded in EEG activity of brain. For people with amputa- controls thedeviceattachedtothesystem [1]. tion or severe neuromuscular disability, who may lack normal Mental activity, such as imagination of movement and output channels, BCIs prove to be useful for controlling movement itself or decision making, results in excitation of external devices [1]. The world of BCI is growing day by Neural Networks which cause changes in electrical potentials day, with applications ranging from control of upper/lower that can be recorded by sensors [2]. This electrical potential is limb prosthesis and wheel chairs to control of multimedia recorded using invasive (placement of sensor under the scalp applications and smart phones for people sueff ring from through surgery) or noninvasive (placement of sensors on the stroke [2, 3]. Table 1 shows some researches in which upper scalp) sensors. The invasive method provides a higher signal limb prosthesis or cursor is controlled using motor imagery. to noise ratio; however, it is cost-wise expensive and involves 2 BioMed Research International Table 1: Examples of research for prosthesis or cursor control using motor imagery. Index Year Research Protocol Accuracy Device Control ECoG of three “Real-time control of movements of left aprosthetichand hand (grasping Prosthesis 2011 using human 69.2% motion, hand control electrocorticography opening motion, signals” [4] scissor type motion) “Control of an Control of Between 44% electrical prosthesis Steady-state visual two-axes 2 2008 and 88% of four with an SSVEP-based evoked potentials electrical hand patients BCI” [5] prosthesis Hybrid feature from motor imagery and “Target Selection with the P300 potential. Online control Hybrid Feature for Target selection by 3 2012 93.99% of cursor on a BCI-Based 2-D focusing and monitor screen Cursor Control” [6] direction control by left-right hand motor imagery. Motor Imagery of left or right hand movement for 1D “Quadcopter control cursor movement left in three-dimensional and right. space using a For 2D movement Between 69.1% Quadcopter 2013 noninvasive motor move cursor up by and 90.5% for 5 control imagery-based imagining squeezing subjects brain–computer or curling both hands interface” [7] and to move the cursor down through the use of a volitional rest. Independently executed overt reach and grasp movements for Intracranial elec- “Simultaneous Neural (Subject 1, troencephalographic Control of Simple Subject 2) were (iEEG) signals of Reaching and (0.85, 0.81) and Dexterous subject who made 5 (0.80, 0.96), robotic 2014 Grasping with the reaching and grasping Modular Prosthetic respectively, prosthetic arm movements to Limb Using during identify task-selective Intracranial EEG” [8] simultaneous electrodes execution they were (0.83, 0.88) and (0.58, 0.88), respectively “Decoding human Average motor activity from Four motor tasks accuracy of 85.5 EEG single trials for a (sustain or cease to 2D cursor 2009 ± 4.65% with discrete move right or left movement physical motor two-dimensional hand) movement cursor control” [9] risk due to surgery. er Th e is a variety of changes in electrical Visual Evoked Potentials, and Mu and Beta Rhythms over the potentials that can be extracted from real time recorded sensorimotor cortex [12]. EEG signals [10] which can be either evoked potentials or In this research, noninvasive electrode equipment is used induced potentials. This includes Event Related Potentials for recording EEG signals form scalp. eTh EEG signal data set (ERP), P300 Evoked Potentials [11], Slow Cortical Potentials, was recorded using a 14-channel electrode headset (Emotiv BioMed Research International 3 AF4 AF3 F4 F3 F7 F8 FC6 FC5 T8 T7 P7 P8 O1 O2 Figure 1: Electrode placement [13]. headset) from right-handed, neurologically intact volunteers. three nge fi r movements. In this research, we are using three This research is a follow-up of our previous research in which targeted finger movements (thumb, index finger, and sfi t) Mu and Beta Rhythms (8–30 Hz) were used. Power Spectral instead of four (as were in our previous research). eTh Density (PSD) was used for analysis of filtered data followed reason to restrict ourselves to three movements only is that by logistic regression for classicfi ation of n fi ger movements embedded system is not able to distinguish between the [13]. PSD describes the distribution of power of signal over index finger movement and index-middle finger combined its frequency. eTh band power of Power Spectral Density [14] movement. of n fi ger movements of one hand occurring over the motor cortex is used as a feature to classify them. eTh Mu and Beta 2. Materials and Methods Rhythms that occur over the motor cortex provide us with information related to the movement [15, 16]. 2.1. Section I: Experimental Protocol and Data Acquisition. A variety of classification techniques are used in BCI The data was acquired from four subjects (one female and 3 systems such as Neural Networks (NN), Support Vector male) who volunteered to undergo data recording protocol. Machines (SVM), Discriminant Analysis, and Bayesian Clas- One of the male subjects (described as category I in Results) sifiers. As an extension of our previous research logistic has a habit of high involuntary eye blinking frequency. eTh regression is used as a classifier, and output of the classifier is other three subjects are described as category II in Results. used for generating command signals to control upper limb ea Th geofsubjectsisbetween 22 and45years.Thepro- prosthesis [17–19]. cess of data acquisition from the subjects is approved by Our previous research [13] compared different classifiers, the departmental ethical review board. eTh volunteers are namely, Multilayer Perceptron, Linear Discriminant Analysis healthywithnoknown neurological disabilities.Thedatawas (LDA), Quadratic Discriminant Analysis (QDA), and logistic acquired using Emotiv headset at 128 Hz sampling frequency. regression to achieve highest classification accuracy [20, 21]. Emotiv has 14 noninvasive electrodes placed according to Two-stage logistic regression gave the highest classicfi ation the international 10-20 system shown in Figure 1 [13]. The accuracy of 74% for four finger movements (thumb, index data was acquired for four movements, i.e., thumb, sfi t, index finger, index and middle finger combined, and fist). Weka finger, and index-middle finger combined movements. eTh 3.6.9 (data mining software with collection of machine movements are shown in Figure 2 [13]. Three out of these learning algorithms) and Matlab were used to process the acquired movements (thumb, index finger, and st fi move- signals in earlier research. ment) data were used for this research. In the current research our emphasis is on the use During data acquisition, the subjects were comfortably of embedded system to process EEG data for generating sitting in a chair and were asked to perform the movements command signals for upper limb prosthesis. Arduino Uno shownonthe computer screen.For each subjectthe acquired is used as the embedded system to filter signals (between 8 data of one trial contained 10 seconds of data for each and 30 Hz), extract features (PSD), and differentiate between movement, which makes 1280 samples per movement. The 4 BioMed Research International (a) (b) (c) (d) Figure 2: Finger movements that were recorded. (a) u Th mb movement. (b) Fist movement. (c) Index finger movement. (d) Two-finger (index and middle) combined movement [13]. starting note movement movement movement movement ending note 0 1013 2326 3841 5255 6568 75 Figure 3: Data acquisition protocol. total samples in one recorded trial are1280×4 = 5120. signal that controls motors connected to upper limb pros- eTh re were a total of 60 recorded trials for each subject, out thesis n fi gers. Stages of the system from data input to device control are shown in Figure 4. of which 13 trials for each subject were rejected on the basis Data processing steps included digital filtering with a high of visual inspection. Rest of the 47 data trials for each subject pass and low pass filter to retain 8–30 Hz of frequencies. were used. For this research samples of only three movements Filtration was followed by feature extraction (calculation of (thumb, index finger, and sfi t) were used instead of four band power from PSD of the remaining frequencies) from the movements. u Th s, for this research 47×1280×3=180480 data. eTh feature vector was then given as input to a logistic samples have been used for one subject. eTh data acquisition regression classifier network for classification of three finger protocol is shown in Figure 3. movements. Based on the classification, a command signal is generatedandsenttoamotor drivecircuitry (H-Bridgein 2.2. Section II: Embedded System. The aim of this research this case) to actuate the respective motor to start the n fi ger was to design an embedded system that can be used to movement of upper limb prosthesis. classify and control upper limb prosthesis n fi ger movements This research was carried out using data already acquired using acquired EEG signals. “Arduino Uno” is the embedded from the subjects. Data set of each trial consisted of 10 system used to fulfill the aim of this research. eTh attributes of secondsofdataof14channelsatthe sampling rateof 128Hz. embedded system (Arduino Uno) are given in Table 2. From each of the 47 trials, 250 ms of data was extracted The data was given as input to Arduino Uno, which andconverted totext files.Thedatawas savedofflineonan SDcardin text fileandgivenasinputtoArduinoUno.TheSD was programmed to process the input (filtering and classi- fication). Basing upon the result of classification, generate a card was interfaced with the controller using Serial Peripheral BioMed Research International 5 Table 2: Attributes of embedded system. Attribute Specification (1) Memory 32 kB (i) In-System Programming by On-chip (2) Debug ability Boot Program (ii) True Read-While-Write Operation Data Retention: 20 years at 85 C/100 (3) Reliability years at 25 C (4) Throughput Up to 20 MIPS Throughput at 20 MHz (i) In-System Programming by On-chip (5) Testability Boot Program (ii) True Read-While-Write Operation Speed Grade: (i) 0–4 MHz @ 1.8–5.5 V (6) Response (ii) 0–10MHz @ 2.7–5.5 V (iii) 0–20 MHz @ 4.5–5.5 V SD Card Arduino Uno H-Bridge Upper Limb containing data in Data Processing Control of prosthesis Prosthesis form of text file according to logic given by Arduino Figure 4: Stages of system from data input to device control. Table 3: Connection between SD card and Arduino. Table 4: High pass filter coefficients. SD card (Pin) Arduino Uno (Pin) Vector Index 1 Index 2 Index 3 5V 5V a 1 −1.4542 0.5741 Ground Ground b 0.7571 −1.5142 0.7571 CS Pin 10 MOSI Pin 11 Table 5: Low pass filter coefficients. MISO Pin 12 SCK Pin 13 Vector Index 1 Index 2 Index 3 a 1 −0.1151 0.1739 b 0.2647 0.5294 0.2647 Interface (SPI). SPI operates in full duplex mode with a Master Out Slave In Pin, Master In Slave Out Pin, Serial Clock Pin,andChipSelectPin.TheArduino actedasMaster, while using Butterworth filter was its flat response with zero ripples. the SD card acted as Slave. The Arduino rst fi enabled the SD The coefficients of the Butterworth filter were taken from card through Chip Select. eTh clock was set at a baud rate of Matlab “butter” command and are shown in Tables 4 and 5. 9600. eTh Pin congfi uration of connection between SD card Filtration was done using these coecffi ients in the filter andArduino Unoisshownin Table3. difference equation defined by [21] As discussed earlier in this section, 250 ms of data was extracted from each trial, converted, and saved in text file. a(1)∗ y(n)= b(1)∗ x(n)+ b(2)∗ x(n− 1)+ b(3) During processing, rst fi 250 ms of data is read by the embed- ded system, processed, and classified. en, Th next 250 ms of ∗ x(n− 2)− a(2)∗ y(n− 1)− a(3) (1) data of next trial is read and processed and the loop continues for classica fi tion until the data reaches its end. ∗ y n− 2 , ( ) 2.3. Section III: Filtration Techniques. Filtering the data to where y is the output, x is the input, and n is the𝑛th element extract Mu and Beta band of frequencies (8–30 Hz) is carried of the output. out as this band contains maximum information related Before passing the data through filter, it was padded. to finger movements. To execute this through embedded Later, after the forward and reverse filtering, the data was system, 250 ms of EEG data was digitally filtered using a But- truncated back to its original number of samples. Filtration terworth filter between 8 and 30 Hz of order 2. eTh reason for was done in both forward and reverse direction. The data was 6 BioMed Research International rfi st passed through high pass filter of order 2 and 8 Hz cut- 2.5. Section V: Classification. As discussed earlier, our pre- off frequency and the resultant was passed through a low pass vious research had shown highest classification accuracy by filteroforder2and 30Hzcut-off frequency. using linear regression classifier. er Th efore, for this research we used two-stage logistic regression classiefi r to calculate classification accuracy for three n fi ger movements. For the 2.4. Section IV: Feature Extraction. Power Spectral Density logistic regression classifier, the probability of the first class of each channel of the filtered signal was calculated. Each is given by [23] channel of filtered signal was divided into 4 windows of 62.5mseach.Ahamming windowwascreated.Theformula exp(𝐵 ∗𝐹) for hamming window is given in [22] (6) 𝑃 (𝐺=1 )= , (exp(𝐵 ∗𝐹)+1) cos(2𝜋𝑛 ) (2) 𝑤 (𝑛 )=0.54−0.46∗ , 𝑇 where𝐹 is the feature vector and𝐵 are the coefficients of 𝑁−1 logistic regression. where𝑁 is the maximum number of points of the sampling The criterion for selection of class is [24, 25] window. 𝐺 𝑥 = class1 if Pr>0.5 ( ) The window function was then multiplied with the signal (7) toshapeitintohamming window.FastFourier transformof 𝐺 (𝑥 )= class1 if Pr<0.5. the windowed signal was then calculated. Formula for Fast Fourier transform is given in [22] In the two-stage model, the first classifier (referred to as network I) distinguished between class 1, which is thumb 𝑁−1 and n fi ger movements, and class 2 which is fist movement. −𝑖2𝜋𝑘(𝑛/𝑁) 𝑋 = ∑𝑥 𝑒 𝑘=0,...,𝑁−1. (3) 𝑘 𝑛 In the 2nd stage a second classifier (referred to as network 𝑛=0 II) distinguished between thumb and finger movement. The classifier model is shown in Figure 9. Training of the classifier eTh absolute value of the resultant is computed and was done using data set of all subjects (75% for training and divided by the normalization factor of the window. 25% for testing) in “Weka” and coefficients of logistic regres- The normalization factor of the window is given by [22] sionwerecalculatedforfurther useinclassicfi ationusing 𝑁−1 the embedded system. 31 randomly chosen samples for each (4) movement were tested for classification in embedded system 𝑈= ∑|𝑤 (𝑛 )| . 𝑛=0 keeping in mind the data handling capability [26]. This gives us the Power Spectral Density of the window, 2.6. Section VI: Device Control. Upper limb prosthesis used which can be represented as in (5) [3]. After the PSD of for this research was developed in the department for carry- each window is calculated, the corresponding values of all ing out the experiments. Figure 10 shows a picture of upper windows are added and averaged, leaving us with a vector of 8 limb prosthesis. constituents. Each value of this vector is then further divided Prosthesis contains two motors connected to two nge fi rs by 2𝜋 to scale the values. andplacedatthepalm.Fingerjointsareconnectedwith 󵄨 󵄨2 each other with the help of a flexible metal wire which is 󵄨 󵄨 𝑋(𝑓) 󵄨 󵄨 󵄨 󵄨 (5) connected with a motor. Motor rotation will cause winding 𝑥= , Fs𝐿𝑈 or unwinding of the flexible metal wire resulting in opening or closing of nge fi rs. Both motors were connected to motor whereFsissamplingfrequency,𝐿 is length of segment,𝑈 is drive which was taking command signal from Arduino Uno. window normalization constant given by (4), and𝑋(𝑓) is data The embedded system generated a control signal based on aer ft FFT. the classification of finger movements. This control signal was The power values are averaged to give the band power sent to the motor drive circuitry to actuate the motor for of one channel of data. eTh process is repeated for all the desired motion. One of the motors is attached to Output Pins 14 channels. In the end, we are left with a feature vector of 4 and 6 (for thumb movement) and the other is attached to 14 values, each representing the band power of 8–30 Hz fre- Output Pins 2 and 3 (for nger fi movement). Motion of motors quency of the channel. Figure 5 shows the topography plots of according to classification is shown in Table 6. the raw data of randomly selected data samples of each move- ment. It can be seen that in each plot the electrodes F3 and FC5contributetoriseincontours.These channels arebasi- 3. Results cally located above the sensorimotor cortex. eTh contours due to these two electrodes have been magnified to show the To train the two-stage logistic regression classiefi r data set of all subjects (category I and category II) is used as discussed difference in the topographies of the movements. eTh differ- in Section 2.5. 75% data is used for training and 25% data is ence is also highlighted in the periodogram graphs that are shown from Figures 6–8 on channels F3 and FC5 of different used for testing. “Weka” (data mining software) was used and coefficients of logistic regression were calculated for further movements. These graphs show the power concentration, in the 8–30 Hz band, of different movements. use in embedded system. 𝑝𝑥 BioMed Research International 7 Table 6: Motion of motors according to classification. Classification of finger State of motor attached to State of motor attached to movement Output Pins 4 and 6 Output Pins 2 and 3 u Th mb movement On Off Finger movement Off On Fist movement On On (a) um Th b movement (b) Finger movement (c) Fist movement Figure 5: Topography plots of movements. FC5 and F3 electrodes have been magnified to show the difference in the topographies of the movements. ×10 Periodogram -F3 0 10 20 30 40 50 60 70 (Hz) ×10 Periodogram -FC5 1.5 0.5 0 10 20 30 40 50 60 70 (Hz) Figure 6: Finger movement periodogram of channels F3 and FC5. 8 BioMed Research International Table 7: Network classification accuracy of a two-stage logistic classifier network. Network number Classification accuracy Network 1 (Class 1-u Th mb + Index Finger and Class 2- Fist) 74% Network 2 (Class 1-u Th mb and Class 2- Index Finger) 76% ×10 Periodogram -F3 0 10 20 30 40 50 60 70 (Hz) ×10 Periodogram -FC5 0 10 20 30 40 50 60 70 (Hz) Figure 7: u Th mb movement periodogram of channels F3 and FC5. 4 Periodogram -F3 ×10 0 10 20 30 40 50 60 70 (Hz) 4 Periodogram -FC5 ×10 0 10 20 30 40 50 60 70 (Hz) Figure 8: Fist movement periodogram of channels F3 and FC5. Table 8: Confusion matrix of category I data set. Results of our research comprise two categories. In cate- gory I (subject having a habit of high involuntary eye blinking Class predicted by 2-stage logistic regression frequency), 31 randomly chosen data samples from each Class classifier movement were used for testing using embedded system. u Th mb Index Finger Fist In category II (subjects other than category I), 31 randomly Thumb 13 12 6 chosen data samples from each movement were used for 816 7 Index Finger testing using embedded system. Fist 89 14 Table 7 shows the classification accuracy of each stage of classifier (network I and network II) using data set of all subjects (category I and category II) as discussed in and Table 9 shows the category II data set tested over 31 Section 2.5. randomly chosen samples for each movement. Table8showsthe confusionmatrixofcategoryIdata set Table 10 shows the per class classification accuracy of tested over 31 randomly chosen samples for each movement randomly chosen samples from category I and category II. BioMed Research International 9 Class 1 Class 2 (Thumb & Fist Finger) Network - 1 Logistic Regression Classifier 1 Thumb Finger Fist Logistic Regression Network - 2 Classifier 2 Thumb Finger Figure 9: Two-stage logistic regression classifier used for the system. Figure 10: Prosthesis controlled by the embedded system. Table 9: Confusion matrix of category II data set. 4. Discussion Class predicted by 2-stage logistic regression The aim of this research was to investigate the design of an Class classifier embedded system for control of upper limb prosthesis as an u Th mb Index finger Fist extension of our previous research. As evident from Table 1, Thumb 20 9 2 research using BCI system for control of prosthesis is focused Index finger 424 3 on spatially distant motor movements. Our focus in this research was to control prosthesis with finger movements Fist 55 21 which have less spatial distance as compared to earlier researches. Table 10: Per class accuracy. Finger movements have the same origin in brain leading to extremely small spatial difference between them. Our Classification Classification accuracy endeavor was to pick up the small difference of brain activity Movement class accuracy of category of category I II recorded in the form of electrical potential and classify it Thumb 42% 65% with higher accuracy. It was seen from the topography plots Index finger 51% 77% showninFigure5thatnfi germovementshavethesameorigin Fist 45% 68% in brain. The minor differences in the topographies were highlighted when the data under electrodes F3 and FC5 was interpolated and plotted in a magnified manner. Band powerofPower Spectral DensityofMuand Beta Percentage accuracies are calculated on the basis of confusion Rhythms was chosen as feature vectors. PSD is used as a matrices showninTables8and9. 10 BioMed Research International feature vector in 70% of the research which focuses on motor [2] A. Guillot and C. Collet, “eTh neurophysiological foundations of mental and motor imagery,” eTh Neurophysiological Founda- controls. eTh periodogram of the movements was plotted tions of Mental and Motor Imagery,pp. 1–320,2012. for the channels above sensorimotor cortex to visualize the [3] D.J.McFarland,L.A.Miner,T.M.Vaughan,andJ. R.Wolpaw, differences. “Mu and beta rhythm topographies during motor imagery and The development of the embedded system was focused on actual movements,” Brain Topography,vol.12, no.3,pp. 177–186, to design a control for upper limb prosthesis that has small size and light weight and is easy to carry onboard system [4] T.Yanagisawa, M. Hirata,Y.Saitohetal.,“Real-timecontrol for prosthesis user. The acquired data was saved on SD card of aprosthetichandusing humanelectrocorticography signals: rather than on the controller to depict a real time data pro- Technical note,” Journal of Neurosurgery,vol.114,no.6,pp.1715– cessing and classification. 70% mean classification accuracy 1722, 2011. was achieved with 2-stage logistic regression classifier using [5] G. R. Muller ¨ -Putz and G. Pfurtscheller, “Control of an electrical an Arduino Uno based embedded system. It should also prosthesis with an SSVEP-based BCI,” IEEE Transactions on be noted that the index and middle n fi ger combined move- Biomedical Engineering,vol.55,no.1,pp. 361–364,2008. ment could not be classified with higher accuracies since [6] J.Long, Y.Li,T.Yu,andZ.Gu,“Targetselection withhybrid the spatial distance is very less. This accuracy needs to be feature for BCI-based 2-D cursor control,” IEEE Transactions increased further for developing better control of prosthesis on Biomedical Engineering,vol.59, no.1,pp.132–140, 2012. and practical implementation. [7] K. Lafleur, K. Cassady, A. Doud, K. Shades, E. Rogin, and B. ec Th hallengeistocreateanonlinemodelthatcanprocess He, “Quadcopter control in three-dimensional space using a andclassifythe real time data.However,implementation noninvasive motor imagery-based brain-computer interface,” of the model on patients requires a more robust system Journal of Neural Engineering,vol.10, no.4,pp.711–726,2013. suggested subsequently. A headset with greater number of [8] M. S. Fifer, G. Hotson, B. A. Wester et al., “Simultaneous neural channels especially above the motor cortex is required for control of simple reaching and grasping with the modular recording more comprehensive signals. It is also recom- prosthetic limb using intracranial EEG,” IEEE Transactions on mended to use an embedded system with higher computa- Neural Systems and Rehabilitation Engineering,vol.22,no.3,pp. 695–705, 2014. tional speed which can process signals in real time. It canalsobeseenfromourresultsthateye blinking [9] D. Huang, P. Lin, D. Y. Fei, X. Chen, and O. Bai, “Decoding Human Motor Activity from EEG Single Trials for A Discrete during data acquisition induces ocular artefacts which result Two-Dimensional Cursor Control,” Journal of Neural Engineer- in lower classification accuracies. For a better and higher clas- ing,vol.6,no.4, ArticleID046005, 2009. sification accuracies a signal with higher signal to noise ratio [10] K. Liao, R. Xiao, J. Gonzalez, and L. Ding, “Decoding individual is required. Different techniques for removing ocular artefact finger movements from one hand using human EEG signals,” may be used for future work. PLoS ONE,vol.9,no.1, ArticleIDe85192, 2014. Overall the research shows that upper limb prosthesis [11] B. Dal Seno, M. Matteucci, and L. Mainardi, “Online detection control can be achieved even with signals that are taken from of P300 and error potentials in a BCI speller,” Computational closelysituatedbodypartwithanaverageaccuracyof 70% Intelligence and Neuroscience, vol. 2010, Article ID 307254, 5 (calculatedonthe basisofclassicfi ation accuracy ofcategory pages, 2010. II randomly chosen samples as mentioned in Table 10). [12] D. J. McFarland et al., “Mu and beta rhythm topographies dur- ing motor imagery and actual movements,” Brain Topography, 5. Conclusion vol.12,no. 3,pp.177–186,2000. [13] A.Javed,M.I.Tiwana,M. I. Tiwana,N.Rashid,J.Iqbal,and U. The designed embedded system in this research is capable of S. Khan, “Recognition of finger movements using EEG signals controllingthe prosthesis basedonthemodeldeveloped for control of upper limb prosthesis using logistic regression,” earlier. A two-stage classifier has been designed and imple- Journal of Biomedical Research,vol.28, no.17, pp.7361–7369, mented over the embedded systems. eTh classifier is capable of distinguishing between three movements of nge fi r, thumb, [14] E. Yom-Tov and G. F. Inbar, “Feature selection for the classifica- andfist.Themeanclassicfi ation accuracyof 70%isattained tion of movements from single movement-related potentials,” by the developed system. Further work to improve the IEEE Transactions on Neural Systems and Rehabilitation Engi- classification accuracy using advanced embedded system can neering,vol.10, no.3,pp.170–177,2002. be undertaken for enhanced prosthesis control. [15] R. Xiao and L. Ding, “Evaluation of EEG features in decoding individual finger movements from one hand,” Computational Conflicts of Interest and Mathematical Methods in Medicine,vol.2013, ArticleID 243257, 2013. eTh authors declare that they have no conflicts of interest. [16] A. Vuckovic and F. Sepulveda, “A four-class BCI based on motor imagination of the right and the left hand wrist,” in Proceedings of the First International Symposium on Sciences on References Biomedical and Communication Technologies, (ISABEL ’08),pp. 1–4, October 2008. [1] J. R. Wolpaw,N.Birbaumer,D.J.McFarland,G.Pfurtscheller, and T. M. Vaughan, “Brain-computer interfaces for communi- [17] A. K. Mohamed, T. Marwala, and L. R. John, “Single-trial EEG discrimination between wrist and finger movement imagery cation and control,” Clinical Neurophysiology, vol. 113, no. 6, pp. 767–791, 2002. and execution in a sensorimotor BCI,” in Proceedings of the BioMed Research International 11 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, (EMBS ’11), pp. 6289–6293, IEEE, Boston, Mass, USA, September 2011. [18] A. Vuckovic and F. Sepulveda, “A two-stage four-class BCI based on imaginary movements of the left and the right wrist,” Medical Engineering & Physics, vol. 34, no. 7, pp. 964–971, 2012. [19] D. W. Hosmer Jr., S. Lemeshow, and R. X. Sturdivant, Applied Logistic Regression, vol. 398, John Wiley & Sons, 2013. [20] J. Lehtonen, P. Jylan ¨ ki, L. Kauhanen, and M. Sams, “Online classification of single EEG trials during finger movements,” IEEE Transactions on Biomedical Engineering,vol.55,no.2,pp. 713–720, 2008. [21] B. Wang and F. Wan, “Classification of Single-Trial EEG based on support vector clustering during finger movement,” Advances in Neural Networks–ISNN,pp.354–363, 2009. [22] K. M. M. Prabhu, Window Functions and Their Applications in Signal Processing,CRCPress, 2013. [23] F. Lotte, M. Congedo, A. Le´cuyer, F. Lamarche, and B. Arnaldi, “A review of classification algorithms for EEG-based brain–computer interfaces,” Journal of Neural Engineering,vol. 4, no. 2, pp. R1–R13, 2007. [24] N. Nedjah, “Evolvable machines: theory practice,” in Evolvable machines: theory practice, p. 161, Springer Science Business Media, 2005. [25] S. W. Smith, The Scientist And Engineer’s Guide to Digital Signal Processing, Elsevier, 1997. [26] A.T. Campbell,T.Choudhury,S.Huetal.,“Neuro-phone: brain-mobile phone interface using a wireless EEG headset,” in Proceedings of the 2nd ACM SIGCOMM Workshop on Network- ing, Systems, and Applications on Mobile Handhelds (MobiHeld ’10),pp.3–8,NewDelhi,India,September 2010. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png BioMed Research International Pubmed Central

Design of Embedded System for Multivariate Classification of Finger and Thumb Movements Using EEG Signals for Control of Upper Limb Prosthesis

BioMed Research International , Volume 2018 – May 20, 2018

Loading next page...
 
/lp/pubmed-central/design-of-embedded-system-for-multivariate-classification-of-finger-6OiCjPUaeu

References

References for this paper are not available at this time. We will be adding them shortly, thank you for your patience.

Publisher
Pubmed Central
Copyright
Copyright © 2018 Nasir Rashid et al.
ISSN
2314-6133
eISSN
2314-6141
DOI
10.1155/2018/2695106
Publisher site
See Article on Publisher Site

Abstract

Hindawi BioMed Research International Volume 2018, Article ID 2695106, 11 pages https://doi.org/10.1155/2018/2695106 Research Article Design of Embedded System for Multivariate Classification of Finger and Thumb Movements Using EEG Signals for Control of Upper Limb Prosthesis Nasir Rashid , Javaid Iqbal , Amna Javed, Mohsin I. Tiwana, and Umar Shahbaz Khan Department of Mechatronics Engineering, National University of Sciences & Technology, H-12, Islamabad, Pakistan Correspondence should be addressed to Nasir Rashid; n.rashid@ceme.nust.edu.pk Received 10 August 2017; Revised 21 January 2018; Accepted 13 February 2018; Published 20 May 2018 Academic Editor: Noman Naseer Copyright © 2018 Nasir Rashid et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Brain Computer Interface (BCI) determines the intent of the user from a variety of electrophysiological signals. These signals, Slow Cortical Potentials, are recorded from scalp, and cortical neuronal activity is recorded by implanted electrodes. This paper is focused on design of an embedded system that is used to control the finger movements of an upper limb prosthesis using Electroencephalogram (EEG) signals. This is a follow-up of our previous research which explored the best method to classify three movements of fingers (thumb movement, index finger movement, a nd first movement). Two-stage logistic regression classifier exhibited the highest classification accuracy while Power Spectral Density (PSD) was used as a feature of the filtered signal. eTh EEG signal data set was recorded using a 14-channel electrode headset (a noninvasive BCI system) from right-handed, neurologically intact volunteers. Mu (commonly known as alpha waves) and Beta Rhythms (8–30 Hz) containing most of the movement data were retained through filtering using “Arduino Uno” microcontroller followed by 2-stage logistic regression to obtain a mean classification accuracy of 70%. 1. Introduction Movements of a prosthesis are commonly controlled through manipulating the motion of rotary actuator (electric motor) A Brain Computer Interface (BCI) provides a communica- in a BCI system. tion system to control external device(s) in which messages or BCI system consists of input signals (electrophysiological commands are sent to external world through brain signals. activity recorded from scalp of user), a signal processor eTh se signals do not pass through the brain’s normal output (filtering the signal for desired frequency and extracting pathways of nerves and muscles. Rather, BCI provides an features for best representation of user intent), a translating alternate method to its user to interact with the world. For algorithmorclassiefi r(thatanticipatesthehumanintentfrom example, Electroencephalogram (EEG) based BCI messages the selected feature), and n fi ally a control algorithm that are encoded in EEG activity of brain. For people with amputa- controls thedeviceattachedtothesystem [1]. tion or severe neuromuscular disability, who may lack normal Mental activity, such as imagination of movement and output channels, BCIs prove to be useful for controlling movement itself or decision making, results in excitation of external devices [1]. The world of BCI is growing day by Neural Networks which cause changes in electrical potentials day, with applications ranging from control of upper/lower that can be recorded by sensors [2]. This electrical potential is limb prosthesis and wheel chairs to control of multimedia recorded using invasive (placement of sensor under the scalp applications and smart phones for people sueff ring from through surgery) or noninvasive (placement of sensors on the stroke [2, 3]. Table 1 shows some researches in which upper scalp) sensors. The invasive method provides a higher signal limb prosthesis or cursor is controlled using motor imagery. to noise ratio; however, it is cost-wise expensive and involves 2 BioMed Research International Table 1: Examples of research for prosthesis or cursor control using motor imagery. Index Year Research Protocol Accuracy Device Control ECoG of three “Real-time control of movements of left aprosthetichand hand (grasping Prosthesis 2011 using human 69.2% motion, hand control electrocorticography opening motion, signals” [4] scissor type motion) “Control of an Control of Between 44% electrical prosthesis Steady-state visual two-axes 2 2008 and 88% of four with an SSVEP-based evoked potentials electrical hand patients BCI” [5] prosthesis Hybrid feature from motor imagery and “Target Selection with the P300 potential. Online control Hybrid Feature for Target selection by 3 2012 93.99% of cursor on a BCI-Based 2-D focusing and monitor screen Cursor Control” [6] direction control by left-right hand motor imagery. Motor Imagery of left or right hand movement for 1D “Quadcopter control cursor movement left in three-dimensional and right. space using a For 2D movement Between 69.1% Quadcopter 2013 noninvasive motor move cursor up by and 90.5% for 5 control imagery-based imagining squeezing subjects brain–computer or curling both hands interface” [7] and to move the cursor down through the use of a volitional rest. Independently executed overt reach and grasp movements for Intracranial elec- “Simultaneous Neural (Subject 1, troencephalographic Control of Simple Subject 2) were (iEEG) signals of Reaching and (0.85, 0.81) and Dexterous subject who made 5 (0.80, 0.96), robotic 2014 Grasping with the reaching and grasping Modular Prosthetic respectively, prosthetic arm movements to Limb Using during identify task-selective Intracranial EEG” [8] simultaneous electrodes execution they were (0.83, 0.88) and (0.58, 0.88), respectively “Decoding human Average motor activity from Four motor tasks accuracy of 85.5 EEG single trials for a (sustain or cease to 2D cursor 2009 ± 4.65% with discrete move right or left movement physical motor two-dimensional hand) movement cursor control” [9] risk due to surgery. er Th e is a variety of changes in electrical Visual Evoked Potentials, and Mu and Beta Rhythms over the potentials that can be extracted from real time recorded sensorimotor cortex [12]. EEG signals [10] which can be either evoked potentials or In this research, noninvasive electrode equipment is used induced potentials. This includes Event Related Potentials for recording EEG signals form scalp. eTh EEG signal data set (ERP), P300 Evoked Potentials [11], Slow Cortical Potentials, was recorded using a 14-channel electrode headset (Emotiv BioMed Research International 3 AF4 AF3 F4 F3 F7 F8 FC6 FC5 T8 T7 P7 P8 O1 O2 Figure 1: Electrode placement [13]. headset) from right-handed, neurologically intact volunteers. three nge fi r movements. In this research, we are using three This research is a follow-up of our previous research in which targeted finger movements (thumb, index finger, and sfi t) Mu and Beta Rhythms (8–30 Hz) were used. Power Spectral instead of four (as were in our previous research). eTh Density (PSD) was used for analysis of filtered data followed reason to restrict ourselves to three movements only is that by logistic regression for classicfi ation of n fi ger movements embedded system is not able to distinguish between the [13]. PSD describes the distribution of power of signal over index finger movement and index-middle finger combined its frequency. eTh band power of Power Spectral Density [14] movement. of n fi ger movements of one hand occurring over the motor cortex is used as a feature to classify them. eTh Mu and Beta 2. Materials and Methods Rhythms that occur over the motor cortex provide us with information related to the movement [15, 16]. 2.1. Section I: Experimental Protocol and Data Acquisition. A variety of classification techniques are used in BCI The data was acquired from four subjects (one female and 3 systems such as Neural Networks (NN), Support Vector male) who volunteered to undergo data recording protocol. Machines (SVM), Discriminant Analysis, and Bayesian Clas- One of the male subjects (described as category I in Results) sifiers. As an extension of our previous research logistic has a habit of high involuntary eye blinking frequency. eTh regression is used as a classifier, and output of the classifier is other three subjects are described as category II in Results. used for generating command signals to control upper limb ea Th geofsubjectsisbetween 22 and45years.Thepro- prosthesis [17–19]. cess of data acquisition from the subjects is approved by Our previous research [13] compared different classifiers, the departmental ethical review board. eTh volunteers are namely, Multilayer Perceptron, Linear Discriminant Analysis healthywithnoknown neurological disabilities.Thedatawas (LDA), Quadratic Discriminant Analysis (QDA), and logistic acquired using Emotiv headset at 128 Hz sampling frequency. regression to achieve highest classification accuracy [20, 21]. Emotiv has 14 noninvasive electrodes placed according to Two-stage logistic regression gave the highest classicfi ation the international 10-20 system shown in Figure 1 [13]. The accuracy of 74% for four finger movements (thumb, index data was acquired for four movements, i.e., thumb, sfi t, index finger, index and middle finger combined, and fist). Weka finger, and index-middle finger combined movements. eTh 3.6.9 (data mining software with collection of machine movements are shown in Figure 2 [13]. Three out of these learning algorithms) and Matlab were used to process the acquired movements (thumb, index finger, and st fi move- signals in earlier research. ment) data were used for this research. In the current research our emphasis is on the use During data acquisition, the subjects were comfortably of embedded system to process EEG data for generating sitting in a chair and were asked to perform the movements command signals for upper limb prosthesis. Arduino Uno shownonthe computer screen.For each subjectthe acquired is used as the embedded system to filter signals (between 8 data of one trial contained 10 seconds of data for each and 30 Hz), extract features (PSD), and differentiate between movement, which makes 1280 samples per movement. The 4 BioMed Research International (a) (b) (c) (d) Figure 2: Finger movements that were recorded. (a) u Th mb movement. (b) Fist movement. (c) Index finger movement. (d) Two-finger (index and middle) combined movement [13]. starting note movement movement movement movement ending note 0 1013 2326 3841 5255 6568 75 Figure 3: Data acquisition protocol. total samples in one recorded trial are1280×4 = 5120. signal that controls motors connected to upper limb pros- eTh re were a total of 60 recorded trials for each subject, out thesis n fi gers. Stages of the system from data input to device control are shown in Figure 4. of which 13 trials for each subject were rejected on the basis Data processing steps included digital filtering with a high of visual inspection. Rest of the 47 data trials for each subject pass and low pass filter to retain 8–30 Hz of frequencies. were used. For this research samples of only three movements Filtration was followed by feature extraction (calculation of (thumb, index finger, and sfi t) were used instead of four band power from PSD of the remaining frequencies) from the movements. u Th s, for this research 47×1280×3=180480 data. eTh feature vector was then given as input to a logistic samples have been used for one subject. eTh data acquisition regression classifier network for classification of three finger protocol is shown in Figure 3. movements. Based on the classification, a command signal is generatedandsenttoamotor drivecircuitry (H-Bridgein 2.2. Section II: Embedded System. The aim of this research this case) to actuate the respective motor to start the n fi ger was to design an embedded system that can be used to movement of upper limb prosthesis. classify and control upper limb prosthesis n fi ger movements This research was carried out using data already acquired using acquired EEG signals. “Arduino Uno” is the embedded from the subjects. Data set of each trial consisted of 10 system used to fulfill the aim of this research. eTh attributes of secondsofdataof14channelsatthe sampling rateof 128Hz. embedded system (Arduino Uno) are given in Table 2. From each of the 47 trials, 250 ms of data was extracted The data was given as input to Arduino Uno, which andconverted totext files.Thedatawas savedofflineonan SDcardin text fileandgivenasinputtoArduinoUno.TheSD was programmed to process the input (filtering and classi- fication). Basing upon the result of classification, generate a card was interfaced with the controller using Serial Peripheral BioMed Research International 5 Table 2: Attributes of embedded system. Attribute Specification (1) Memory 32 kB (i) In-System Programming by On-chip (2) Debug ability Boot Program (ii) True Read-While-Write Operation Data Retention: 20 years at 85 C/100 (3) Reliability years at 25 C (4) Throughput Up to 20 MIPS Throughput at 20 MHz (i) In-System Programming by On-chip (5) Testability Boot Program (ii) True Read-While-Write Operation Speed Grade: (i) 0–4 MHz @ 1.8–5.5 V (6) Response (ii) 0–10MHz @ 2.7–5.5 V (iii) 0–20 MHz @ 4.5–5.5 V SD Card Arduino Uno H-Bridge Upper Limb containing data in Data Processing Control of prosthesis Prosthesis form of text file according to logic given by Arduino Figure 4: Stages of system from data input to device control. Table 3: Connection between SD card and Arduino. Table 4: High pass filter coefficients. SD card (Pin) Arduino Uno (Pin) Vector Index 1 Index 2 Index 3 5V 5V a 1 −1.4542 0.5741 Ground Ground b 0.7571 −1.5142 0.7571 CS Pin 10 MOSI Pin 11 Table 5: Low pass filter coefficients. MISO Pin 12 SCK Pin 13 Vector Index 1 Index 2 Index 3 a 1 −0.1151 0.1739 b 0.2647 0.5294 0.2647 Interface (SPI). SPI operates in full duplex mode with a Master Out Slave In Pin, Master In Slave Out Pin, Serial Clock Pin,andChipSelectPin.TheArduino actedasMaster, while using Butterworth filter was its flat response with zero ripples. the SD card acted as Slave. The Arduino rst fi enabled the SD The coefficients of the Butterworth filter were taken from card through Chip Select. eTh clock was set at a baud rate of Matlab “butter” command and are shown in Tables 4 and 5. 9600. eTh Pin congfi uration of connection between SD card Filtration was done using these coecffi ients in the filter andArduino Unoisshownin Table3. difference equation defined by [21] As discussed earlier in this section, 250 ms of data was extracted from each trial, converted, and saved in text file. a(1)∗ y(n)= b(1)∗ x(n)+ b(2)∗ x(n− 1)+ b(3) During processing, rst fi 250 ms of data is read by the embed- ded system, processed, and classified. en, Th next 250 ms of ∗ x(n− 2)− a(2)∗ y(n− 1)− a(3) (1) data of next trial is read and processed and the loop continues for classica fi tion until the data reaches its end. ∗ y n− 2 , ( ) 2.3. Section III: Filtration Techniques. Filtering the data to where y is the output, x is the input, and n is the𝑛th element extract Mu and Beta band of frequencies (8–30 Hz) is carried of the output. out as this band contains maximum information related Before passing the data through filter, it was padded. to finger movements. To execute this through embedded Later, after the forward and reverse filtering, the data was system, 250 ms of EEG data was digitally filtered using a But- truncated back to its original number of samples. Filtration terworth filter between 8 and 30 Hz of order 2. eTh reason for was done in both forward and reverse direction. The data was 6 BioMed Research International rfi st passed through high pass filter of order 2 and 8 Hz cut- 2.5. Section V: Classification. As discussed earlier, our pre- off frequency and the resultant was passed through a low pass vious research had shown highest classification accuracy by filteroforder2and 30Hzcut-off frequency. using linear regression classifier. er Th efore, for this research we used two-stage logistic regression classiefi r to calculate classification accuracy for three n fi ger movements. For the 2.4. Section IV: Feature Extraction. Power Spectral Density logistic regression classifier, the probability of the first class of each channel of the filtered signal was calculated. Each is given by [23] channel of filtered signal was divided into 4 windows of 62.5mseach.Ahamming windowwascreated.Theformula exp(𝐵 ∗𝐹) for hamming window is given in [22] (6) 𝑃 (𝐺=1 )= , (exp(𝐵 ∗𝐹)+1) cos(2𝜋𝑛 ) (2) 𝑤 (𝑛 )=0.54−0.46∗ , 𝑇 where𝐹 is the feature vector and𝐵 are the coefficients of 𝑁−1 logistic regression. where𝑁 is the maximum number of points of the sampling The criterion for selection of class is [24, 25] window. 𝐺 𝑥 = class1 if Pr>0.5 ( ) The window function was then multiplied with the signal (7) toshapeitintohamming window.FastFourier transformof 𝐺 (𝑥 )= class1 if Pr<0.5. the windowed signal was then calculated. Formula for Fast Fourier transform is given in [22] In the two-stage model, the first classifier (referred to as network I) distinguished between class 1, which is thumb 𝑁−1 and n fi ger movements, and class 2 which is fist movement. −𝑖2𝜋𝑘(𝑛/𝑁) 𝑋 = ∑𝑥 𝑒 𝑘=0,...,𝑁−1. (3) 𝑘 𝑛 In the 2nd stage a second classifier (referred to as network 𝑛=0 II) distinguished between thumb and finger movement. The classifier model is shown in Figure 9. Training of the classifier eTh absolute value of the resultant is computed and was done using data set of all subjects (75% for training and divided by the normalization factor of the window. 25% for testing) in “Weka” and coefficients of logistic regres- The normalization factor of the window is given by [22] sionwerecalculatedforfurther useinclassicfi ationusing 𝑁−1 the embedded system. 31 randomly chosen samples for each (4) movement were tested for classification in embedded system 𝑈= ∑|𝑤 (𝑛 )| . 𝑛=0 keeping in mind the data handling capability [26]. This gives us the Power Spectral Density of the window, 2.6. Section VI: Device Control. Upper limb prosthesis used which can be represented as in (5) [3]. After the PSD of for this research was developed in the department for carry- each window is calculated, the corresponding values of all ing out the experiments. Figure 10 shows a picture of upper windows are added and averaged, leaving us with a vector of 8 limb prosthesis. constituents. Each value of this vector is then further divided Prosthesis contains two motors connected to two nge fi rs by 2𝜋 to scale the values. andplacedatthepalm.Fingerjointsareconnectedwith 󵄨 󵄨2 each other with the help of a flexible metal wire which is 󵄨 󵄨 𝑋(𝑓) 󵄨 󵄨 󵄨 󵄨 (5) connected with a motor. Motor rotation will cause winding 𝑥= , Fs𝐿𝑈 or unwinding of the flexible metal wire resulting in opening or closing of nge fi rs. Both motors were connected to motor whereFsissamplingfrequency,𝐿 is length of segment,𝑈 is drive which was taking command signal from Arduino Uno. window normalization constant given by (4), and𝑋(𝑓) is data The embedded system generated a control signal based on aer ft FFT. the classification of finger movements. This control signal was The power values are averaged to give the band power sent to the motor drive circuitry to actuate the motor for of one channel of data. eTh process is repeated for all the desired motion. One of the motors is attached to Output Pins 14 channels. In the end, we are left with a feature vector of 4 and 6 (for thumb movement) and the other is attached to 14 values, each representing the band power of 8–30 Hz fre- Output Pins 2 and 3 (for nger fi movement). Motion of motors quency of the channel. Figure 5 shows the topography plots of according to classification is shown in Table 6. the raw data of randomly selected data samples of each move- ment. It can be seen that in each plot the electrodes F3 and FC5contributetoriseincontours.These channels arebasi- 3. Results cally located above the sensorimotor cortex. eTh contours due to these two electrodes have been magnified to show the To train the two-stage logistic regression classiefi r data set of all subjects (category I and category II) is used as discussed difference in the topographies of the movements. eTh differ- in Section 2.5. 75% data is used for training and 25% data is ence is also highlighted in the periodogram graphs that are shown from Figures 6–8 on channels F3 and FC5 of different used for testing. “Weka” (data mining software) was used and coefficients of logistic regression were calculated for further movements. These graphs show the power concentration, in the 8–30 Hz band, of different movements. use in embedded system. 𝑝𝑥 BioMed Research International 7 Table 6: Motion of motors according to classification. Classification of finger State of motor attached to State of motor attached to movement Output Pins 4 and 6 Output Pins 2 and 3 u Th mb movement On Off Finger movement Off On Fist movement On On (a) um Th b movement (b) Finger movement (c) Fist movement Figure 5: Topography plots of movements. FC5 and F3 electrodes have been magnified to show the difference in the topographies of the movements. ×10 Periodogram -F3 0 10 20 30 40 50 60 70 (Hz) ×10 Periodogram -FC5 1.5 0.5 0 10 20 30 40 50 60 70 (Hz) Figure 6: Finger movement periodogram of channels F3 and FC5. 8 BioMed Research International Table 7: Network classification accuracy of a two-stage logistic classifier network. Network number Classification accuracy Network 1 (Class 1-u Th mb + Index Finger and Class 2- Fist) 74% Network 2 (Class 1-u Th mb and Class 2- Index Finger) 76% ×10 Periodogram -F3 0 10 20 30 40 50 60 70 (Hz) ×10 Periodogram -FC5 0 10 20 30 40 50 60 70 (Hz) Figure 7: u Th mb movement periodogram of channels F3 and FC5. 4 Periodogram -F3 ×10 0 10 20 30 40 50 60 70 (Hz) 4 Periodogram -FC5 ×10 0 10 20 30 40 50 60 70 (Hz) Figure 8: Fist movement periodogram of channels F3 and FC5. Table 8: Confusion matrix of category I data set. Results of our research comprise two categories. In cate- gory I (subject having a habit of high involuntary eye blinking Class predicted by 2-stage logistic regression frequency), 31 randomly chosen data samples from each Class classifier movement were used for testing using embedded system. u Th mb Index Finger Fist In category II (subjects other than category I), 31 randomly Thumb 13 12 6 chosen data samples from each movement were used for 816 7 Index Finger testing using embedded system. Fist 89 14 Table 7 shows the classification accuracy of each stage of classifier (network I and network II) using data set of all subjects (category I and category II) as discussed in and Table 9 shows the category II data set tested over 31 Section 2.5. randomly chosen samples for each movement. Table8showsthe confusionmatrixofcategoryIdata set Table 10 shows the per class classification accuracy of tested over 31 randomly chosen samples for each movement randomly chosen samples from category I and category II. BioMed Research International 9 Class 1 Class 2 (Thumb & Fist Finger) Network - 1 Logistic Regression Classifier 1 Thumb Finger Fist Logistic Regression Network - 2 Classifier 2 Thumb Finger Figure 9: Two-stage logistic regression classifier used for the system. Figure 10: Prosthesis controlled by the embedded system. Table 9: Confusion matrix of category II data set. 4. Discussion Class predicted by 2-stage logistic regression The aim of this research was to investigate the design of an Class classifier embedded system for control of upper limb prosthesis as an u Th mb Index finger Fist extension of our previous research. As evident from Table 1, Thumb 20 9 2 research using BCI system for control of prosthesis is focused Index finger 424 3 on spatially distant motor movements. Our focus in this research was to control prosthesis with finger movements Fist 55 21 which have less spatial distance as compared to earlier researches. Table 10: Per class accuracy. Finger movements have the same origin in brain leading to extremely small spatial difference between them. Our Classification Classification accuracy endeavor was to pick up the small difference of brain activity Movement class accuracy of category of category I II recorded in the form of electrical potential and classify it Thumb 42% 65% with higher accuracy. It was seen from the topography plots Index finger 51% 77% showninFigure5thatnfi germovementshavethesameorigin Fist 45% 68% in brain. The minor differences in the topographies were highlighted when the data under electrodes F3 and FC5 was interpolated and plotted in a magnified manner. Band powerofPower Spectral DensityofMuand Beta Percentage accuracies are calculated on the basis of confusion Rhythms was chosen as feature vectors. PSD is used as a matrices showninTables8and9. 10 BioMed Research International feature vector in 70% of the research which focuses on motor [2] A. Guillot and C. Collet, “eTh neurophysiological foundations of mental and motor imagery,” eTh Neurophysiological Founda- controls. eTh periodogram of the movements was plotted tions of Mental and Motor Imagery,pp. 1–320,2012. for the channels above sensorimotor cortex to visualize the [3] D.J.McFarland,L.A.Miner,T.M.Vaughan,andJ. R.Wolpaw, differences. “Mu and beta rhythm topographies during motor imagery and The development of the embedded system was focused on actual movements,” Brain Topography,vol.12, no.3,pp. 177–186, to design a control for upper limb prosthesis that has small size and light weight and is easy to carry onboard system [4] T.Yanagisawa, M. Hirata,Y.Saitohetal.,“Real-timecontrol for prosthesis user. The acquired data was saved on SD card of aprosthetichandusing humanelectrocorticography signals: rather than on the controller to depict a real time data pro- Technical note,” Journal of Neurosurgery,vol.114,no.6,pp.1715– cessing and classification. 70% mean classification accuracy 1722, 2011. was achieved with 2-stage logistic regression classifier using [5] G. R. Muller ¨ -Putz and G. Pfurtscheller, “Control of an electrical an Arduino Uno based embedded system. It should also prosthesis with an SSVEP-based BCI,” IEEE Transactions on be noted that the index and middle n fi ger combined move- Biomedical Engineering,vol.55,no.1,pp. 361–364,2008. ment could not be classified with higher accuracies since [6] J.Long, Y.Li,T.Yu,andZ.Gu,“Targetselection withhybrid the spatial distance is very less. This accuracy needs to be feature for BCI-based 2-D cursor control,” IEEE Transactions increased further for developing better control of prosthesis on Biomedical Engineering,vol.59, no.1,pp.132–140, 2012. and practical implementation. [7] K. Lafleur, K. Cassady, A. Doud, K. Shades, E. Rogin, and B. ec Th hallengeistocreateanonlinemodelthatcanprocess He, “Quadcopter control in three-dimensional space using a andclassifythe real time data.However,implementation noninvasive motor imagery-based brain-computer interface,” of the model on patients requires a more robust system Journal of Neural Engineering,vol.10, no.4,pp.711–726,2013. suggested subsequently. A headset with greater number of [8] M. S. Fifer, G. Hotson, B. A. Wester et al., “Simultaneous neural channels especially above the motor cortex is required for control of simple reaching and grasping with the modular recording more comprehensive signals. It is also recom- prosthetic limb using intracranial EEG,” IEEE Transactions on mended to use an embedded system with higher computa- Neural Systems and Rehabilitation Engineering,vol.22,no.3,pp. 695–705, 2014. tional speed which can process signals in real time. It canalsobeseenfromourresultsthateye blinking [9] D. Huang, P. Lin, D. Y. Fei, X. Chen, and O. Bai, “Decoding Human Motor Activity from EEG Single Trials for A Discrete during data acquisition induces ocular artefacts which result Two-Dimensional Cursor Control,” Journal of Neural Engineer- in lower classification accuracies. For a better and higher clas- ing,vol.6,no.4, ArticleID046005, 2009. sification accuracies a signal with higher signal to noise ratio [10] K. Liao, R. Xiao, J. Gonzalez, and L. Ding, “Decoding individual is required. Different techniques for removing ocular artefact finger movements from one hand using human EEG signals,” may be used for future work. PLoS ONE,vol.9,no.1, ArticleIDe85192, 2014. Overall the research shows that upper limb prosthesis [11] B. Dal Seno, M. Matteucci, and L. Mainardi, “Online detection control can be achieved even with signals that are taken from of P300 and error potentials in a BCI speller,” Computational closelysituatedbodypartwithanaverageaccuracyof 70% Intelligence and Neuroscience, vol. 2010, Article ID 307254, 5 (calculatedonthe basisofclassicfi ation accuracy ofcategory pages, 2010. II randomly chosen samples as mentioned in Table 10). [12] D. J. McFarland et al., “Mu and beta rhythm topographies dur- ing motor imagery and actual movements,” Brain Topography, 5. Conclusion vol.12,no. 3,pp.177–186,2000. [13] A.Javed,M.I.Tiwana,M. I. Tiwana,N.Rashid,J.Iqbal,and U. The designed embedded system in this research is capable of S. Khan, “Recognition of finger movements using EEG signals controllingthe prosthesis basedonthemodeldeveloped for control of upper limb prosthesis using logistic regression,” earlier. A two-stage classifier has been designed and imple- Journal of Biomedical Research,vol.28, no.17, pp.7361–7369, mented over the embedded systems. eTh classifier is capable of distinguishing between three movements of nge fi r, thumb, [14] E. Yom-Tov and G. F. Inbar, “Feature selection for the classifica- andfist.Themeanclassicfi ation accuracyof 70%isattained tion of movements from single movement-related potentials,” by the developed system. Further work to improve the IEEE Transactions on Neural Systems and Rehabilitation Engi- classification accuracy using advanced embedded system can neering,vol.10, no.3,pp.170–177,2002. be undertaken for enhanced prosthesis control. [15] R. Xiao and L. Ding, “Evaluation of EEG features in decoding individual finger movements from one hand,” Computational Conflicts of Interest and Mathematical Methods in Medicine,vol.2013, ArticleID 243257, 2013. eTh authors declare that they have no conflicts of interest. [16] A. Vuckovic and F. Sepulveda, “A four-class BCI based on motor imagination of the right and the left hand wrist,” in Proceedings of the First International Symposium on Sciences on References Biomedical and Communication Technologies, (ISABEL ’08),pp. 1–4, October 2008. [1] J. R. Wolpaw,N.Birbaumer,D.J.McFarland,G.Pfurtscheller, and T. M. Vaughan, “Brain-computer interfaces for communi- [17] A. K. Mohamed, T. Marwala, and L. R. John, “Single-trial EEG discrimination between wrist and finger movement imagery cation and control,” Clinical Neurophysiology, vol. 113, no. 6, pp. 767–791, 2002. and execution in a sensorimotor BCI,” in Proceedings of the BioMed Research International 11 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, (EMBS ’11), pp. 6289–6293, IEEE, Boston, Mass, USA, September 2011. [18] A. Vuckovic and F. Sepulveda, “A two-stage four-class BCI based on imaginary movements of the left and the right wrist,” Medical Engineering & Physics, vol. 34, no. 7, pp. 964–971, 2012. [19] D. W. Hosmer Jr., S. Lemeshow, and R. X. Sturdivant, Applied Logistic Regression, vol. 398, John Wiley & Sons, 2013. [20] J. Lehtonen, P. Jylan ¨ ki, L. Kauhanen, and M. Sams, “Online classification of single EEG trials during finger movements,” IEEE Transactions on Biomedical Engineering,vol.55,no.2,pp. 713–720, 2008. [21] B. Wang and F. Wan, “Classification of Single-Trial EEG based on support vector clustering during finger movement,” Advances in Neural Networks–ISNN,pp.354–363, 2009. [22] K. M. M. Prabhu, Window Functions and Their Applications in Signal Processing,CRCPress, 2013. [23] F. Lotte, M. Congedo, A. Le´cuyer, F. Lamarche, and B. Arnaldi, “A review of classification algorithms for EEG-based brain–computer interfaces,” Journal of Neural Engineering,vol. 4, no. 2, pp. R1–R13, 2007. [24] N. Nedjah, “Evolvable machines: theory practice,” in Evolvable machines: theory practice, p. 161, Springer Science Business Media, 2005. [25] S. W. Smith, The Scientist And Engineer’s Guide to Digital Signal Processing, Elsevier, 1997. [26] A.T. Campbell,T.Choudhury,S.Huetal.,“Neuro-phone: brain-mobile phone interface using a wireless EEG headset,” in Proceedings of the 2nd ACM SIGCOMM Workshop on Network- ing, Systems, and Applications on Mobile Handhelds (MobiHeld ’10),pp.3–8,NewDelhi,India,September 2010.

Journal

BioMed Research InternationalPubmed Central

Published: May 20, 2018

There are no references for this article.