Single-trial EEG Discrimination between Wrist and Finger Movement Imagery and Execution in a Sensorimotor BCI

1  Abstract —A brain-com puter interface (BCI) m ay be used to control a prosth etic or orthotic ha nd using neu ral activity from the brain . The core of this sen sorimoto r B CI lies in the interpre tation of the n eural inform ation e x tracted from electroen cephalogram (EEG). It is desired to imp rove on the interpre tation of EEG to allow peop le with neurom uscular disorde rs to perform daily activities. This paper investigates the possibility of discrimina ting betw een the EEG associated w ith wrist an d finger mo vemen ts. The EEG was recorded fr om test subjects as they executed an d im agined fiv e essential hand m ovem ents using both hands. Indep endent com ponent ana ly sis (ICA) and time-freq uency techniq ues w ere used to extract spectral features bas ed on event-related (de)synch ronisation (ERD/ERS), while the Bhattachar yy a distan ce (BD) w as used for feature reduction. M ahala nobis distance (M D) clustering and artificial neural netw ork s (ANN) w ere used as classifiers and obtain ed average ac curacies of 65 % and 71 % resp ectively. This shows that EEG discrim ination between wrist and finger m ovem ents is possible. The research introd uces a new com bination o f motor tas ks to BCI research. Index Terms — Brain-co mpu ter Interfa ce (BCI), electroen cephalogram (EEG), event-related (de)synchronisa tion (ERD/ERS), im agined hand move ment, Independ ent Com ponent Analysis (ICA) I. I NTRODUCTI ON EOPL E wh o suffer from motor impairm ents can benefit greatly f ro m a system that can return some of the essential functionality of the human hand [1 ]. Such people may have h ad an arm amputated or have suf fered a stroke or spinal cord injury [ 1]. T he lost hand of an amputee can be replaced b y a robotic p rosthetic hand, w hile the n on-functional hand of a victim of a stroke or spinal cord injury can be suppo rted by a robotic exoskeletal o rthotic hand [1]. T hese external devices can then be controlled using the user’s though ts w ith the help of a br ain-com puter interface (B CI) to reroute the signals directly from the brain to actuators in the p rosthetic/o rthotic hand [1, 2] . This solution can be used to allo w m otor -impaired individuals to perform essential hand m ovements that facilitate the performance of d aily activities [1, 3]. Considering the movem ents that patients learn during motor rehabilitation [ 4, 5], five basic hand movem ents are co nsidered i.e. w rist extension (WE), wrist flexion ( WF), f inger extension (FE), Manuscrip t receive d August 13, 2 010. A.K . M ohamed is at the University of W it water srand , T. Marwala is at the University o f Johan nesburg, L. John is at the University of Cape To wn. finger flexion ( FF) and the tri pod p inch (T R). Occupational therapists consider these to b e the most essential hand movem ents [4, 5 , 6]. The co re of an effective BCI solution will re quire that the neural information associated with the essential hand movem ents be extracted and translated from neural signals, such as electroencephalogram (EEG), in real -tim e [7, 8] . The combination of these five essential hand movem ents has not yet been explored in EEG -based BCI literature [9] . It is thus necessary to first investigate the p ossibility of interpreting the EEG for the five hand movemen ts offline on a single-trial basis since this serves as a first step tow ard real-time B CI functionality [1 , 9, 1 0]. BCI literature has show n that discriminating the E EG for different movem ents in a mutliclass problem and EEG discrimination of movem ents on the same limb are challenging tasks [9, 11]. Howev er, success has been show n in the classification of binary com binations of four types of wrist m ovement tasks on the same hand [12, 13 ]. This suggests that the binary classification of other types of unilateral hand movements may be possible. To date, a study has not been conducted to differentiate b etw een m aj or parts of the hand i.e. the wrist and fingers [9, 12, 1 3, 14] . Hence, a s an intermediate step, the differentiation b etween EEG for w rist and finger movem ents is investigated in this paper by grouping WE and WF into one class and FE, FF and the TR into another. T his forms part of the effort to improve on the incomplete understanding between central neural signals a nd hand movem ents [2, 1 5]. II. B ACK GROUND A. Electroen cephalog ram and I CA Electrical potentials or iginating from multiple sources i.e . neuron clusters, combine to form a superpo sition of topographical maps on the scalp , which can be measured b y scalp ele ctrodes to for m EEG [17] . There are several challenges associated with the extractio n of relevent information from EEG. T he signals a re small (in the µV range), difficult to measure and are easily contaminated by artifacts [ 1]. EEG also presents a large inter -trial and inter- subject variability [1]. T he billions of simu ltaneously-active neural pro cesses are measu red from a limited number of EEG electrodes (even with high resolution EEG e.g 128 electrod es) [1, 18 ]. T his r esults in a considerable mixing of information sources fro m a ll over the head at each electro de [1, 17, 19]. However, clinical resear ch has increased the understanding of EEG signals and numerous studies have show n relationships Single-trial EEG Discrimination between Wrist and Finger Movement Imagery and Execution in a Sensorimotor BCI A.K. Mohamed, T. Marwala, and L.R. J ohn P 2 between EEG and imagined movemen ts [ 1, 8, 20 , 21]. Inexpensive computer equipment now supports the required computational dem ands for EEG sign al pro cessing [ 1]. The latter factors mak e it po ssible to use EEG as a signal source for basic prosthetic/orthotic hand control [1] in a contro lled laborator y env iro nm ent. Using independent component analysis (ICA), measured signals consisting of a linear mixture of statistically independent source signals, such as EEG, can b e decomposed into their fun da men tal underlying independent components (IC) thus extracting the original source signals [16, 2 2]. ICA is comm only used in BCI research to remove ar tifacts, but has also proven useful in separating bio logically plausible brain components whose activity patterns r elate to behavioural occurrences [18] . In so me studies, ICA has show n su per ior performance over other methods of spatial filtering [23, 2 4] and has aid ed the discrimination of EEG for d ifferent un ilateral wrist movement tasks [ 13]. This sugg ests that it may be beneficial for isolating rhythm ic a ctivity fro m the sensorimotor cortex for other types of hand m ovments [1, 16] . B. Brain -computer I nterface By using EEG or other electrophysiological methods, a B CI provides a comm unication channel from the brain to the external world, circ um venting the natural neuro -mus cular pathway [1, 7]. They can improve the quality of life for those wh o suff er fro m m otor impairm ents [1, 25]. The main co mponents of a BCI are shown in Fig 1. T hey enable execution of the external device accor ding to the user’s intent [1 , 2 5]. BCIs that deal with motor functions or sensory inputs of the bod y d eal with the s ensorimotor cortex of the brain. They a re thus called sensorimotor BCI s and are ideal for the control o f a prosthetic/o rthotic hand. Prominent electrophysiological features associated with the brain’s n o rmal motor output channels are m u (8 –12 Hz) and beta (13– 30 Hz) rhyth ms [1 , 25]. T he rhythm s are syn chronised w hen no sensory inputs or motor outputs are being pr ocessed [1, 2 5]. Movement or preparatio n for movem ent results in a desynchronisation (decr ease in a mplitude) of the mu and beta rhyth ms, r eferred to as event-related desyn chronisation (ERD) [1, 25, 26 ]. Event-related synchronisation ( ERS) o ccurs after movem ent when the rhyth ms synch ro nise ( increase in amplitude) again [1 , 25 , 26 ]. ERD and ERS o ccur during imagin ed movem ents as well, making them suit able for paralysed individuals [1, 3]. Fea tures based on E RD/ERS have been used successf ully to differentiate the EEG for so me types of wrist m ovements [12, 13] . III. M E THOD OLOG Y Fig 1 summ arises the m aj or p rocesses that make up the method in order to classify between u nilateral wrist and finger movem ents. The process is applied to rea l and imagined movem ents. A. Data Acquisition Subsequent to ethics approval from the University of Cap e Town, data was captured from five right- handed, healthy, male, untrained volunteers in their early twenties. The subjects were seated in a comfortable chair , resting their forearm on an ar m rest [12, 13]. A co mputer screen was used along with custom Eprime software [27] to queue the movem ents w hile the subjects’ EEG were measured. An EGI system th at co nsisted of 128 high-im ped ance scalp electr odes (forming the GSN 128 ) along with the Geodesic EEG Sys tem and Net Station Software was used [28]. The electrodes were Ag/Ag-C l electrod es with sponge attachmen ts soaked in an electro lyte solution of potassium chloride [29] . Each subj ect was asked to perform real and imagined repetitions of the 5 movem ent sets for each han d (starting with the right hand). Therefore, for each hand, the subjects performed 10 sets of movements: 5 for re al movemen ts and 5 for imag ined movem ents. Each set consisted o f 20 repetitions/trials o f one type of movem ent [ 13]. T he or der of the sets w as randomised and thus d iffered for each subject so that no movem ent type was preferred [1 4]. I n sum mary e ach test subject performed: movement set ( 5) × L/R hand (2) × real/imagined (2) × rep etitions (20) = 400 trials. The type o f movement for each set was shown to the subj ects on the computer screen prior to the co mm encement of the set and a brief practice session w as allow ed. There w ere short breaks b etween sets and the rep etitions for each set w ere performed continually. The trials were queued by instructions show n on the computer screen, th e timeline o f w hich is shown in Fig 2 [13, 3 0]. Subjects were asked not to blink, sw allow, m ove their ey es, adjust their bodies or clear their throats during S1 and S2 , but Fig. 1. Model of a sensorimo tor BCI used for communication to a prosthetic hand. Pre-processing a nd Signal Enhancement: Filter in g, bad dat a removal , arti fact removal , ICA and best I C selection Feature Extraction & Selection: Extract time-frequency spectral features based on ERD/ERS. BD feature sel ection Feature Tra nslation: Classification u sing Mahalanobis d istance clustering and Artificial Neural Netwo rks Data Acquisition Device Controller User intention Assistive Device actuation Visual feedback Real-time solution for co ntrol of a prosthetic/o rthotic hand by a BCI Scope of this s tudy, such that data is captured and proc essed offli ne 3 rather during S3, so as to reduce artifact contamination [30] . Any undesired movem ents or behaviour b y the subjects w as noted. B. Pre-Pro cessing EEGLAB w as used to handle the pre-processing [18]. Noisy channels w ere removed and a bandpass filter between 0.5 Hz and 100 Hz was ap plied to the data [13, 30 ], which w as sampled at 2 00 Hz by the EGI system [29]. A 50 Hz no tch filter was also applied [2 4]. Data w as then divided into 7 s trials, from t = -1s to t = 6 s, placing t = 0 at the Get Re ady event (p re-movem ent stimu lus) show n in Fig 2. T his was done so that the co ntinuous signals were not split in the cr ucial are as of S1 and S2. B ad trials were removed after m anual inspection for voltage spikes and severe distortions acro ss multiple channels. The left hand data f or subjects 1 and 4 was u nusable and thus discarded . The Automatic Artifact Removal ( AA R) toolbo x f or EEGLAB [31] w as used to remove artifacts, which included electro-oculogram from eye-blink s and eye m ovements, an d electromyogram from tongu e, face, neck an d shoulder movem ents [ 1]. Artifacts were removed using spacial filtering and blind source separation [ 31]. A bandpass filter between 8 – 30 Hz was then applied to isolate and mu and beta data [ 12]. C. ICA and Source Lo calisation ICA was run using the infomax algor ithm on the individual hands of each subject [18 ]. T his decomposed the EEG in to separable localised sources of potentials. T he po tentials or ICs emanating from the motor cor tex were visually selected and isolated. Several ICs representing motor activity were selected p er subject and per hand. This approach is advantag eo us since the inter-subject variab ility of E EG makes it difficult to p redict wh ich electrod es provide relevant information [32 ]. I t also helps to capture the information from different regions of the motor areas, which may activate during d ifferent stages of movem ent [32]. Furtherm o re, it reduces the dimensionality of the data and filters co ntam i nation from non-sensorimotor neural potentials, such as the visual alpha rhyth m [25] . The num ber of selected ICs varied between test subjects, ranging between 8 and 12. T he criteria for selection are based on: 1. View ing lo calised activity mainly in the region o f the primary motor cortex that co ntrols the hand, but activity in the supplementary motor ar ea and premotor area is also considered [ 32, 33 ]. 2. The presence of ERD just prior to and/o r during S2 as well as ERS after S2 [34 ]. T his is calcula ted using the inter- trial variance method [34]. D. Featu re Extraction an d Selection A time-f req uency technique was used to extract power spectral features from the selected ICs due to the non- stationary nature of EEG [2 5]. T he time range from t = 1 s to t = 4 s was considered ( see Fig 2) in order to include pre- movem ent and movem ent e xecution/im agination p hases. An overlapping sliding window of 300 m s w as then applied in increments of 10 0 ms [13, 24]. T he power spectrum for each win do w was calculated us ing an FFT. The freq uency spectrum was then split into 7 bands of 3 Hz each [30 ] and the sum of the powers within each band formed a feature. 28 tim e win do w s were extracted o ver the time r ange considered, with 7 power b and features each. T his was d one for ea ch I C, r esulting in a total number of features ranging betw een 156 8 and 235 2. The B hattacharyy a d istance (BD) was used to select the best features accor ding to how well each feature separ ated the classes [2 4, 3 0]. Hence the BD w as calculated for each feature and the 1 8 features w ith the largest BD were selected . This provided low dimensionality and was found to be the op timu m num ber of features during iterative testing. E. Classification A clustering classifier b ased o n the Mahalanobis distance (MD) is simple and rob ust and has shown good performance in BCI resear ch [7]. The MD measures the dissim ilarity between feature vectors fro m d ifferent classes and can also be used to remove outliers [ 35]. Multilayer pe rceptron artificial neural netw or ks are used widely in BCI r esearch [ 7] and are used to verify an d possibly improve on the MD classification results. The squared MD d i 2 between the i th vector of dataset x and the mean of dataset y can be calc ulated using (1) , where  Y is the m ean of d ataset y a nd C Y -1 is the inv erse covariance matrix of dataset y [ 36].     Y i Y T Y i i x C x d       1 2 (1) The MD is then used to calculate the distance between each trial in a given class to its own mean and to the mean of the other class [ 36]. If the distance between a single-trial feature vector x i and the mean o f its class  x is smaller than the MD between that single-trial vector and the mean of the other cla ss, then it can b e co ncluded that x i belongs to class x . T he trial being tested is removed from the calculations of the means and covariances of the classes/clusters allowing all tr ials to be used for testing. Alternatively , for classification usin g artif icial neural netw or ks (ANNs), the d ata is divided into training and testing data in a 7:3 ratio. The number of hidden nodes is iteratively varied to select that wh ich yields the smallest average error for all subjects. H ence, MLPs each co nsisting of 1 8 input nodes, 24 hidden nodes and 1 output node ar e trained per subject per hand. In clinical ap plications, sensitivity and specificity are often used to evaluate the accuracy of diagnostic tests [3 7]. Sensitivity describes the likelihood of a po sitive test result if a patient has a disease, while sp ecificity indicates the likelihood of a negative result if the p atient d oes not have the disease [37]. Sensitivity and specificity c an be generalized to 2 class datasets, for example: wrist movem ents = po sitive test result and finger movements = negative test result. Classification accuracy is thus m easured b y calculating the average of the S1 Preparation for m ovem ent S2 Sustain ed M ovem ent t = 2 t = 5 t = 7 Get Ready Stop Movement t = 0 S3 Rest Period Single trial Start Movement Fig 2. Time sequence and instructions for a sin gle t rial 4 sensitivity and specificity measures ( SSA ) as shown in ( 2), wh ere T and F respectively rep resent the nu mber of corr ectly and falsely classified trials for each class. Subscripts W and F denote wrist and finger classes respectively .             F F F W W W F T T F T T SSA 2 1 (2) IV. R ESULT S AND D I SCUSSIO N The MD and ANN r esults are sum marised in Ta ble I and Table II respectively. Classification is shown per subject for real and imaginary movemen ts . The results show reasonable classification accuracies, which are consistent across most test subjects for both hands. ANNs perfor med b etter than MD clustering. This is pro bably due to the ANNs man aging to capture the hidden patter ns amongst the features more accurately than the simple distance-based approach of the MD method. Classification is slightly more successful for imagin ed movem ents than for real movements. T his is contrary to the findings of other BCI studies [30 ], wh ere classification results for real m ovements are superior due to real movem ents generating stronger motor neural activity [30, 39]. Howev er, some studies have show n similar results for r eal and imag ined movem ents [ 13]. The superio r results for imagined movem ents in this study could be due to the fact that all the test subjects were university stu d ents w ho were familiar w ith motor imagery . Consequen tly their co ncentration levels and imagin ative skills may have been abo ve average, which may have increased the classification accuracy for imagined movem ents [40 ]. Subjects who par ticipated in the study in [12 ] reported an ease of im agining m ovements such as WE since it is us ed in everyday life. In this study , the use of WE, W F, FE, FF and the TR in everyday life may have made the motor imagery tasks easier for the test subjects, thus enhancing their sensorimotor EEG patterns, desp ite having n o training. The success of this research is important since it shows that the discrimination of neural signals from neighbouring areas of the motor cortex is possible using EEG. T his allows the real or imagin ed movement of major p arts o f the hand i. e. the wrist and fingers, to be interp reted via EEG. T he use of ICA along with high resolution EEG (128 channels) played an impo rtant role in this regard. Common hand movem ents such as FE and the T R [4. 5], which ar e novel to BCI literature, can be explored in future research involving pr osthetic/orthotic hand control using a BCI [9]. Future work involves working towards accurately classifyin g the individual five essential hand movements ; first offline and thereafer in real-time. V. C ONCLUS ION This p aper focuses on discriminating b etw een u nilateral wrist and finger movem ents in order to improve EEG interpretation to allow a sensorimotor BCI to contro l a prosthetic/orthotic hand. T he average results for the MD a nd ANN classifiers are 65 % and 71 % respectively. These r esults show that the offline discrimination between wrist and finger movem ent EEG, for r eal and imagined movem ents, is p ossible. This is an important step towards allowing a p rosthetic/orthotic hand to perform essential hand movem ents. R EFERENCES [1] Wol pa w J R, Birbau mer N, McFarland D J, Pfurtscheller G, Vaughan T M. Brain -computer in terface s for c ommunicatio n and con trol . Clinical Neurophysiolo gy, Vol 113, 20 02, pp 767 – 791 . [2] Afshar P, Masuoka Y. Neural -Based Contro l o f a Ro botic Hand: Evidence for Distin ct Muscle Strat egies . The proce edings for the 2004 I EEE I nternati onal C onfere nce on rob otics an d au tomation, New Orleans, LA , April 2004 , pp 4 633 – 4638 . [3] Guger C, Harkam W, Hertnaes C, Pfurtscheller G. Prosth etic Control by a n EEG-ba sed Brai n-Compute r Interfa ce (BCI). In Pro ceedings of the 5th European Conference for the Advancement of Assist ive Technolog y (AA ATE), G ermany, 1999. [4] Trombly C, Radomski M . Occupatio nal Therapy for physical dysfunct ion , 5 th edition, 200 2. [5] Smith J C. OT for chil dren , Development o f han d fu nction , 2 nd edition, 200 4. [6] Bulbulia R., Personal communicat ion [7] Lo tt e F, Congedo M, ´ ecuyer A L , Lamarche F, Arnaldi B. A revi ew of classificat ion al gorithms for EEG-b ased bra in-comput er interfac es , Journal of Neural Engineering, Vol 4 , 2007, pp R1 – R13. [8] Babiloni F, Cin cotti F, Bian chi L, Pirri G . MillánJ R, Mouriń o J, Salinari S, Marciani MG. Recogni tion of imagined hand movements with low resolutio n surface La placian and li near classif iers . Medic al Engineering and Physics, Vol 23, 2 001, pp 323 – 328 . [9] Vuckovic , A. Non-in vasive BCI: How fa r can we g et with mo tor imagina tion? . Clinical Neurophysiolo gy , Vol 120, 20 09, pp 1422 - 1423. [10] Hallet M, Bai O, Bonin C. Pred icting Movement: When , Which and Where . IEEE/I CM E I nt ernational Conference on Complex M edical Engineering, Beijing, May 2007, pp 5 – 7. [11] Obermaier B, Neuper C, G u ger C, Pfurtschell er G. Info rmation transfer ra te in a f ive-cl ass brain -computer int erface . I EEE Transactions on Neural Systems and Rehabilitation En gineering , Vol 9, September 2001, p p 283 – 2 88. [12] Gu Y, Dremstrup K, F arina D. Sing le-trial discrimina tion o f type an d speed o f wrist movements from EEG recordi ngs . Clinical Neurophysiol ogy, Vol 20, August 2009, pp 15 96 – 16 00. [13] Vuckovic A, Sepulve d a F. Delta ban d contri bution in cue b ased single t rial classi fication of real and imag ina ry wrist movement. Medical and Bi olog ic al Engineering and Computin g, Vol 46, 2 008, pp 52 9 – 539 . TA BLE II C LAS SIFICA TION ACCURACY (%) FOR E EG DISCRIMI NATION BETWEEN WRIST A ND FINGER MOVEMENTS USING ANN CLAS SIFIERS Real Imaginary RH LH RH LH Subject 1 81 - 70 - Subject 2 73 75 79 68 Subject 3 73 56 61 72 Subject 4 70 - 76 - Subject 5 52 69 82 67 Subject Average 70 67 73 69 Grand Average 71 % TA BLE I C LAS SIFICA TION ACCURACY (%) FOR E EG DISCRIMI NATION BETWEEN WRIST A ND FINGER MOVEMENTS USING MD CLAS SIFIERS Real Imaginary RH LH RH LH Subject 1 68 - 61 - Subject 2 63 84 56 54 Subject 3 62 45 69 63 Subject 4 71 - 76 - Subject 5 49 55 81 70 Subject Average 63 61 69 62 Grand Average 65 % 5 [14] Khan Y U, Sepulveda F . Brain –computer interface for single -tria l EEG classifica tion fo r wrist movement imagery usin g spati al filterin g in th e gamma ban d . IET Signal Proce ssing, Vol 4, 2 010, pp 510 – 517. [15] Matsuoka Y, Afshar P. Neuromuscul ar S trategie s fo r Dyna mic Finger Mov ements: A Rob otic Ap proach . Proceedings of the 26 th Annual I nternati onal C onfere nce of the I EEE EMBS, Vol 2, September 2004, pp 4649 – 46 52. [16] Wang S, James C J. Extract ing Rhyt hmic Brain Ac tivity fo r Brain - Computer In terfacing through Constrai ned Indep endent Component Analysis. Computat ional I nt ell igence and Neuroscie n ce, Vol 2007 , 2007. [17] Ungureanu M, Bigan C, Strungaru R, Lazares cu V. Indepen dent Componen t An alysis A pplied in Biomedi cal Sig nal Processi ng. Measurement Science Revie w, Vol 4, 2 004. [18] Delo rme A, Makeig S. EEGLAB: an o pen source toolbo x for analysi s of singl e-trial EEG dynamics in cluding indepe ndent componen t analy sis . Journal of Neuroscience Methods , Vol 134, 2004, pp 9 – 21. [19] Blankertz B, Tomioka R, Lemm S, K awanabe M, Mü lle r K R. Optimizing Spatia l Filters fo r Robust EEG S ingle -Trial An alysis . I EEE Signa l Proce ssin g Magazine , Vol 25, 20 08, pp 4 1 – 5 6. [20] Tow nsend B, Grainmann B, Pfurtsc helle r G. Con tinuou s EEG Classific ation During Motor Imagery – Simulatio n of an Asynchron ous BCI . IEEE Transactions on Neural Systems and Rehabilitation Engineering, Vol 12, June 20 04, pp 258 – 265 . [21] Pfurtscheller G, C. B runner C, Schlögl A, L op ez da Silva F H. Mu rhythm (de)synchro nization and EEG single -trial classi fication of differen t motor imagery ta sks . NeuroI mage, Vol 3 1, 2006, pp 153 – 159. [22] Hyv ärin en A, Oja E. Indepen dent Comp onent An alysis: Algori thms and Ap plicatio n. Neural Networ k s, Vol 13, 2000 , pp 411 – 430. [23] Brunner C, Naeem M, Lee b R, Graimann B, Pfurtscheller G . S patial filterin g and sel ection o f optimize d compone nts in fo ur class moto r imagery EEG da ta using indepe ndent co mponent a nalysis . Pat tern Recognition Le tt ers, Vol 2 8, June 2 007, 9 57 – 9 64. [24] Bai O, Lin P, Vorbach S, Li J, Furlani S, Halle t M. Explora tion of computat ional met hods for classificat ion of movement int ention during human Vo luntary movement from sin gle tria l EEG . Clinical Neurophysiol ogy, Vol 1 18, Decemeber 2007, pp 263 7–2655 . [25] Bashashati A, F atourechi M , Ward R K, Birch G E, A survey of signal p rocessing algo rithms in brain-co mputer interface s based on electrica l bra in signals . Journal of Ne ural Engin eering, Vol 4, 2 007, pp R32 – R5 7. [26] Neuper C, Pfurtscheller G. Event -related dynamic s o f co rtical rhythms: frequ ency-speci fic featu res and func tional co rrelates . I nt ernational Journal of Psy choph ysiol ogy, Vol 43, 2001 , pp 41 – 5 8. [27] Psychol ogy Softw are Tools I nc. E-Prime 2, http://www .p stnet.com/eprime.cfm, L ast accessed 11 Janau ry 2 011. [28] EGI , http ://ww w.egi.com/rese arch -division-research-products/eeg- systems /1 91-ges300mr, Last accessed 13 April 2009. [29] Electrical Ge odesics Inc. Geodesi c Sensor Ne t Technica l Manua l , Electrical Ge odesics Inc. htt p://ww w.egi.com, January 2007, S - MAN-200-GSN R-00 1. [30] Morash V, Bai O, Furlani S, Lin P, Hallett M. Classifyi ng EEG signals preceding right ha nd, lef t hand, tongue ,and righ t foot movements an d motor imag ery. Clin ical Neurophysio logy, Vol 119, Novembe r 20 08, pp 2570 – 257 8. [31] G´ omez-He rrero G. Automatic Arti fact Remova l (AAR) toolbo x v1.3 (Release 09 .12.200 7) for MATLAB, Tampere University of Technolog y, December 2007. [32] Åberg M CB, Wessberg J. EVolu tionary optimizati on of cl assifiers and fe atures for si ngle -trial EEG Discrimina tion . BioMedical Engineering OnLine, Vol 6, August 2 007. [33] Dornhege G , Blankertz B, Curio G, and Müller K-R. Combining Features f or BCI . Advances in N eural Info rmation Proc essing Systems (NIPS 02), Vol 15, 2003, p p 1115 –1122 . [34] Pfurtscheller G, Lopes da Silva F H. Event -related EEG/MEG synchron ization a nd desyn chronisat ion: basi c princip les . Clinical Neurophysiol ogy, Vol 1 10, 199 9, pp 1 845 – 185 7. [35] De Maesschalck R, Jouan -Rimbaud D, M assart D L. The Mahala nobis di stance . Chemometrics and Intell igent Laboratory Syste ms, Vol 50, 2000 , pp 1 – 18 . [36] Babiloni F, Bianch i L, Semer aro F, del R Millan J, Mourin o J, Cattini A, Salin ari S, Marciani M G, Cincotti F. Ma halan obis Distance-Ba sed Classif iers Are Able t o Recogni ze EEG Patterns b y Using Few EEG El ectrodes . Proceedings of the 23rd Annual I nt ernational Conference o f th e I EEE EMBS, Istanbul, October 200 1. [37] Peat J, Parton B. Me dical St atistics: A Gui de to Data Ana lysis an d Critical Appraisal . Blackwell publishing, 200 5, pp 2 82 – 28 3. [38] Tow nsend B , Grainmann B, Pfurtscheller G . Co ntinuo us EEG Classific ation Duri ng Moto r Imagery – Simu lation of an Asynchron ous BCI . IEEE Transactions on Neural Syste ms an d Rehabilitatio n Engineering, Vol 12, June 20 04, pp 258 – 265 . [39] Blankertz B, Curio G, Müller K R. Cl assifying Single Trial EEG: Towards Bra in-compute r Interfacin g . Advances in Neural I nformation Processing Systems, Vol 1 4, 2002 , pp 15 7 – 164. [40] Erfanian A, Erfani A. ICA-Based Classifica tion Sc heme for EEG - based Bra in-Comput er Interface: The Role of Mental Pra ctice and Concent ration S kills . Proceedings of the 26 th Annual I EEE EM BS Conference, San Fr an cisco CA, USA , September 2004 , pp 23 5 – 238 .