movement intention detection from autocorrelation of eeg

Movement Intention Detectionfrom Autocorrelation of EEG for BCI

Maitreyee Wairagkar(B), Yoshikatsu Hayashi, and Slawomir Nasuto

University of Reading, Reading, [email protected],

{y.hayashi,s.j.nasuto}@reading.ac.uk

Abstract. Movement intention detection is important for developmentof intuitive movement based Brain Computer Interfaces (BCI). Variouscomplex oscillatory processes are involved in producing voluntary move-ment intention. In this paper, temporal dynamics of electroencephalog-raphy (EEG) associated with movement intention and execution werestudied using autocorrelation. It was observed that the trend of decay ofautocorrelation of EEG changes before and during the voluntary move-ment. A novel feature for movement intention detection was developedbased on relaxation time of autocorrelation obtained by fitting expo-nential decay curve to the autocorrelation. This new single trial featurewas used to classify voluntary finger tapping trials from resting statetrials with peak accuracy of 76.7%. The performance of autocorrelationanalysis was compared with Motor-Related Cortical Potentials (MRCP).

Keywords: Electroencephalography · Autocorrelation · Voluntarymovement intention · Motor-Related Cortical Potentials · BCI

1 Introduction

Brain Computer Interface (BCI) provides a direct mode of interaction with com-puter and other external devices without utilising any motor pathways. Move-ment based BCI has a great potential for the use by patients with severe motordisabilities for operating robotic rehabilitation devices and for performing simpletasks intuitively. Thus, it is important to study underlying neural mechanisms ofvoluntary movement intention. This paper explores the fundamentals of move-ment intention by studying the temporal dynamics of EEG using novel feature.

Motor Related Cortical Potential (MRCP) and Event Related (De) synchro-nization (ERD/S) [1] are widely used movement correlates for movement detec-tion from EEG. Although ERD can detect movement reliably [2,3] , it relies onthe most responsive frequency band that vary from individual to individual. It ischallenging to compute accurate instantaneous frequency distributions withoutcompromising the temporal resolution and inducing delays. MRCP is a slow neg-ative potential occurring from about 2s prior to the onset of the human voluntarymovement [4], observed in frequencies lower than 1Hz [5]. Amplitude of MRCPc© Springer International Publishing Switzerland 2015Y. Guo et al. (Eds.): BIH 2015, LNAI 9250, pp. 212–221, 2015.DOI: 10.1007/978-3-319-23344-4 21

Movement Intention Detection from Autocorrelation of EEG for BCI 213

is extremely small (less than 8-10 μV )and hence, an average about 40-50 trials ofrepeated voluntary movements is used [5]. Single trial analysis is important forpractical online BCI implementation. Although single trial variants of ERD [6]and MRCP [7,8] have been developed, these traditional principles are best suitedfor analysing averaged EEG over several trials [1], [9]. This creates a need for arobust feature for movement detection that can solely be obtained from a singletrial and is independent of any particular frequency band besides providing dif-ferent information related to movement. In this paper, we have proposed a newprocess related to motor command generation in the brain that compliments theinformation that is obtained from conventional ERD and MCRP processes. Thenew autocorrelation relaxation time feature was successfully used for classify-ing movement and resting state trials. The performance of the new single trialmovement feature was compared rigorously with MRCP.

2 Materials and Methods

2.1 Experimental Procedure

Fourteen healthy participants (7 males and 7 females, ages 26±4) participatedin this EEG experiment. Ethical approval for EEG experimentation on humanswas obtained from School of Systems Engineering, University of Reading, UK.Written consent was obtained from all the participants. 12 participants wereright handed and 2 were left handed. Participants did not have any previousexperience of EEG experimentation.

This study was conducted to understand the neural correlates for detection ofmovement intention. A self-paced, asynchronous single finger tapping paradigmwas developed to study EEG corresponding to motor command generation. Sim-ple movement of index finger tapping was chosen as the task because it does notinvolve any complex hand gestures, directional reaching, grasping or trajectoryplanning.

The EEG experimental paradigm was developed in Simulink using the toolsprovided by BioSig toolbox [10]. A customised tapping device was developedusing a programmable micro-controller for recording the finger taps from bothindex fingers. One channel of binary finger tapping signal was recorded simul-taneously with EEG for each hand. EEG and finger tapping signals were co-registered using tools provided by TOBI framework [11].

Participants were seated on a chair with palms placed on the table in front.Index fingers of both the hands were placed in finger caps of the tapping device.A fixation cross was displayed on the screen for 2s at the beginning of each trial.It was followed by the instruction for right or left finger tap or resting state. Awindow of 10s was given to perform the instructed task voluntarily at the timeof the participant’s choice. Each trial was followed by a random break of 1s to1.5s. 40 trials for each of the three conditions (right tap, left tap and rest)wererecorded. 19 EEG channels in accordance with 10-20 international system wererecorded using TruScan Deymed EEG amplifier. Ground and reference electrodeswere placed at the centre and all the impedances were kept below 7kΩ.

214 M. Wairagkar et al.

2.2 EEG Pre-processing, Artefacts Removal and Segmentation

EEG was prepared for further analysis by performing pre-processing and arte-facts removal on all the EEG channels. DC offset was removed by subtractingthe mean of each channel from the signal. All the EEG filtering was done usingfourth order Butterworth filter. A notch filter at 50Hz was used to remove thepower line noise. EEG was low-pass filtered with the cut-off frequency of 60Hzto eliminate high frequency noise.

Eye blinks and some movement artefacts were removed using IndependentComponent Analysis [12]. Independent components with artefacts were identifiedmanually and were eliminated. EEG was segmented into time locked trials oflength 6s by extracting 3s prior to the onset of finger tap and 3s after theonset of the finger tap. Channels F3, Fz, F4, C3, Cz, C4, P3, Pz and P4 oversensorimotor cortex were used for movement intention analysis.

2.3 Analysis of Movement-Related Cortical Potentials

MRCPs are obtained from lower frequencies by averaging several trials ofEEG [5]. EEG was filtered between 0.1Hz to 0.5Hz to obtain movement relatedslow cortical potentials. Grand average MRCP was computed by averaging theall the trials from all the participants for nine channels. MRCP was also obtainedfor single trial.

2.4 Movement Intention Analysis Based on Exponential Decay ofAutocorrelation

Autocorrelation gives an estimation of how EEG is related to itself over time. Pre-vious research shows that the autocorrelation changes during movement [13,14].Six virtual channels viz. F3-C3, Fz-Cz, F4-C4, C3-P3, Cz-Pz and C4-P4 were cre-ated using a longitudinal bipolar montage to enhance the movement related signal.EEG was band-pass filtered with cut-off frequencies 0.5Hz and 30Hz. Continuousautocorrelation was computed for each trial to determine the time developmentof the relaxation time of brain activity before, during and after the movement.Normalized autocorrelation was computed on 1s window shifted by 100ms from6s trial.

Let the single window of a signal be represented by A(t), the autocorrelationof A(t) is defined by C(Δt) = 〈A(t)A(t−Δt)〉, where 〈...〉 represents the averageover time. At zero lag, the signal is perfectly correlated, giving C(0) = 〈A2〉,and as lag approaches infinity, the signal becomes completely uncorrelated, giv-ing C(∞) = 〈A〉2. The trend of relaxation process of autocorrelation could bedescribed by C(t) = 〈A2〉e−t

τ . Here, τ (decay constant) represents the relaxationtime of autocorrelation.When autocorrelation is normalised, 〈A2〉 = 1.

To get the relaxation time of autocorrelation by capturing the exponentiallydecaying trend of autocorrelation, the exponential decay curve y = Ke

−tτ was

fitted to the local maxima points of positive lags of autocorrelation of eachwindow in the trial (see Fig. 1). The relaxation time τ was extracted as the


Fig. 1. Exponential curve fitting representing decay of autocorrelation (autocorrelationrelaxation) for right finger tapping trial in F3-C3

feature at every 100ms in the 6s trial. The constant K was set to 1 as theautocorrelation was normalised. Changes in autocorrelation occurring duringmotor command generation were observed by studying the time progression ofτ values.

2.5 Classification of Movement and Resting State Trials

Classification of MRCP Features. MRCP for each 6s trial was divided into0.5s windows by shifting it by 100ms. Features for classification were obtainedby averaging all the samples in each 0.5s window. Feature from first 0.5s window(-3s to -2.5s) was used as resting state feature for training the classifier. Lineardiscriminant analysis (LDA) classifier was trained for every window in the trialwith two class corresponding features from all the trials. 10x10 cross-fold valida-tion scheme was for this binary classification and percent classification accuracywas obtained. The threshold for classifier outcome was obtained from 95% con-fidence level (p < 0.05) for binary classification. The best performing channelwas selected manually for each participant.

Classification of Autocorrelation Decay Features. Autocorrelation decayfeatures (τ)were obtained for three classes viz. right finger tap, left finger tap andresting state. Classification was performed on each 1s window shifted by 100ms in6s trial. LDA classifier was trained for each window with tap features and restingstate features from the corresponding windows in all the trials. 10x10 cross-foldvalidation scheme was used for binary classification and percent classificationaccuracies were obtained. The classification accuracies for 6s trials for all the sixvirtual channels were plotted for all the participants (see Fig. 2).


Fig. 2. Classification accuracies along the trial for left hand, participant 1 in virtualchannel F3-C3. The horizontal line indicates statistical significance (p < 0.05)

3 Results

3.1 MRCP

Grand Average of MRCP. Grand average of MRCP is shown in the Fig. 3 forright and left hand trials. Negative potential peaks just before the actual onsetof the finger tap in most of the channels. Thus, movement could be predictedfrom the grand average MRCP before its actual occurrence. The peak of theMRCP has small amplitude as expected.

Single Trial Classification of MRCP. (Table. 1) The classification accuracycrosses the threshold prior to the onset of movement indicating that movementcould be predicted before its occurrence.

3.2 Decay of Autocorrelation Related to the Movement Intention

The trend of decay of autocorrelation which represents the relaxation time τ ,changes during the voluntary movement. The τ value starts building-up priorto the onset of the motion. τ features in a single trial for right finger tap illus-trated in Fig. 4 clearly show the increase prior to the movement onset. Thereis no such build-up of τ values in the resting state trials. Two sample t-testswere performed on finger tapping trials and resting state trials on six bipolarchannels (with (p < 0.05)) to confirm the that difference around movement onsetwas statistically significant. Autocorrelation decays slower during intention andexecution of movement and it decays faster otherwise. Table. 2 shows single trialclassification results.


Fig. 3. Grand average of MRCP for nine channels. Red: right finger tapping, Blue: leftfinger tapping.

Table 1. Classification results for MRCP. Statistical significance (p < 0.05) is indicatedin bold.

Right Finger Tapping Left Finger Tapping

ParticipantID

MRCPClassificationAccuracy (%)

ChannelTime ofthresholdcrossing (s)

MRCPClassificationAccuracy (%)


1 70.25 Pz 0.25 68.63 F3 -0.752 64.08 P3 0.30 65.08 P3 -0.803 82.95 C4 -1.60 79.13 P4 -1.604 63.96 P4 -1.35 76.25 P3 -1.755 64.63 F3 -0.90 70.25 F4 -0.506 76.29 P4 -0.90 68.71 F3 -0.207 70.13 F4 -0.95 70.75 Cz -1.208 66.25 C4 -0.20 77.50 P4 -0.659 66.25 P3 -0.95 64.63 F4 -0.4010 70.38 F3 -1.00 67.75 F3 -0.9511 64.50 Cz -1.20 59.88 C4 -12 67.13 Fz -1.10 67.89 Cz 0.6013 69.00 C3 -1.95 71.37 P4 -0.9014 69.50 F4 -0.25 80.63 C3 -0.25

Mean 68.95 -0.84 70.60 -0.72SD 5.26 0.65 5.93 0.62


Fig. 4. Plot of changes in τ in virtual channel F3-C3 in a single trial for right fingertapping (red), left finger tapping (blue) and resting state (black) for participant 1

Table 2. Classification results for autocorrelation analysis. Statistical significance (p <0.05) is indicated in bold.

Right Finger Tapping Left Finger Tapping

ParticipantID

AutocorrelationClassificationAccuracy (%)


AutocorrelationClassificationAccuracy (%)


1 71.00 Cz-Pz -0.20 74.00 C4-P4 -0.252 76.70 F4-C4 -0.25 76.58 F4-C4 -0.903 65.75 Fz-Cz 0.40 66.63 F4-C4 -0.604 66.88 F4-C4 -0.60 69.00 Fz-Cz -0.555 64.50 F4-C4 -0.67 65.38 C3-P3 0.356 69.57 Cz-Pz 0.35 62.71 F4-C4 0.257 69.25 F3-C3 0.30 72.50 F3-C3 0.308 66.13 Cz-Pz -0.20 63.00 F4-C4 0.109 68.25 F4-C4 -1.50 65.38 C3-P3 0.6010 64.38 Cz-Pz -0.75 58.88 Cz-Pz -11 64.25 F3-C3 0.35 65.88 F3-C3 0.1012 63.38 C4-P4 0.75 65.25 C4-P4 -1.0013 61.50 Fz-Cz - 65.38 Fz-Cz -1.0014 70.50 C3-P3 0.40 67.87 C3-P3 -0.60

Mean 67.28 -0.12 67.03 -0.32SD 3.92 0.63 4.71 0.57


Fig. 5. Classification accuracies for autocorrelation and MRCP. Blue: Classificationaccuracies for Autocorrelation, Red: Classification accuracies for MRCP.

3.3 Comparison of Autocorrelation and MRCP Analysis

Classification Accuracies of Single Trial Autocorrelation and MRCP.(Fig. 5) The classification accuracies of the newly proposed autocorrelation anal-ysis method are very similar to the MRCP accuracies. The mean accuracies ofautocorrelation and MRCP for right and left finger tapping are comparable(67.28% and 68.95% for right and 67.03% and 70.60% for left respectively).MRCP has higher maximum classification accuracy than autocorrelation butalso greater SD (see Table 1 and 2).

Timings for Threshold Crossing of Classification Accuracies and Spa-tial Locations for Autocorrelation and MRCP. (Fig. 6) Like MRCP, auto-correlation analysis could also predict the movement before its onset in mostcases and can be used for detection of movement intention. However, MRCPcan predict the movement earlier than autocorrelation decay feature. Spatiallocations of autocorrelation and MRCP analysis are different. Lateralization isnot observed in both the cases. Different spatial locations suggests that the originof the movement related information is distinct in both the cases.

Fig. 6. Timings for threshold crossing for classification accuracies of autocorrelationand MRCP. Blue: Timings for autocorrelation, Red: Timings for MRCP.


4 Discussion and Conclusion

The trend in decay of autocorrelation changes during the process of motor com-mand generation. Autocorrelation decays slower during the preparation and exe-cution of the voluntary movement than in the resting state. Decay constant τ ,representing the relaxation time of autocorrelation, proves to be a robust featurerelated to the movement detection. The change in τ is observed in the single trialand, hence, it could be used for online BCI applications. In this paper, we havevalidated the new movement related process to complement traditional MRCPand ERD processes for movement prediction and detection.

Classification accuracies for autocorrelation (mean 67% with peak accuracyof 76.7%) are similar to the mean accuracies (69.95% and 70.60%) obtainedfrom single trial MRCP analysis. Similar results were obtained with modifiedsingle trial ERD analysis (paper currently in preparation). These classificationaccuracies are also comparable to the other work done for movement predictionand detection using MRCP and ERD [6–9], [15]. The accuracies obtained fromthe offline analysis using both the methods are not very high probably due tosingle finger tapping which involves fewer muscles. The results obtained fromthis task are purely based on movement intention and execution as opposed toother studies that used different full arm directional movements which can resultto inclusion of cognitive information of trajectory planning and coordination.

This new feature is important because the information content on movementintention obtained from this feature is different from MRCP and ERD. Thisfeature gives a third independent process related to movement in brain thatcomplements MRCP and ERD. The advantage of this feature over ERD is that itdoes not require most responsive frequency band selection for individual and thuscan be used more robustly. Unlike MRCP, this new feature provides informationfrom a wide frequency range without eliminating any potential movement relatedinformation from higher frequencies. Further work on this autocorrelation featurecould be used for modelling the oscillatory processes during movement intentiongeneration in brain and thus understand methodically what happens in brainwhen an individual intends to move.

The single feature - autocorrelation relaxation time is capable of identifyingthe intention of movement from single trial with equally good performance asMRCP. Thus, this autocorrelation decay feature could be adapted for onlineBCI applications. The SD of autocorrelation was very small as compared to theSDs of other ERD and MRCP studies [7,8]. This is another indicator of therobustness and consistent performance of the autocorrelation. Thus, this paperintroduces a new neural correlate of movement intention based on the temporaldynamics of EEG that is different from MRCP and ERD.


References

1. Pfurtscheller, G., Da Silva, F.L.: Event-related eeg/meg synchronization and desyn-chronization: basic principles. Clinical Neurophysiology 110(11), 1842–1857 (1999)

2. Blankertz, B., Losch, F., Krauledat, M., Dornhege, G., Curio, G., Muller, K.R.:The berlin brain-computer interface: accurate performance from first-session in bci-naive subjects. IEEE Transactions on Biomedical Engineering 55(10), 2452–2462(2008)

3. Guger, C., Edlinger, G., Harkam, W., Niedermayer, I., Pfurtscheller, G.: How manypeople are able to operate an eeg-based brain-computer interface (bci)? IEEETransactions on Neural Systems and Rehabilitation Engineering 11(2), 145–147(2003)

4. Shibasaki, H., Hallett, M.: What is the bereitschaftspotential? Clinical Neurophys-iology 117(11), 2341–2356 (2006)

5. Bai, O., Rathi, V., Lin, P., Huang, D., Battapady, H., Fei, D.Y., Schneider, L.,Houdayer, E., Chen, X., Hallett, M.: Prediction of human voluntary movementbefore it occurs. Clinical Neurophysiology 122(2), 364–372 (2011)

6. Lopez-Larraz, E., Montesano, L., Gil-Agudo, A., Minguez, J.: Continuous decodingof movement intention of upper limb self-initiated analytic movements from pre-movement eeg correlates. Journal of Neuroengineering and Rehabilitation 11(1),153 (2014)

7. Ibanez, J., Serrano, J., Del Castillo, M., Monge-Pereira, E., Molina-Rueda, F.,Alguacil-Diego, I., Pons, J.: Detection of the onset of upper-limb movements basedon the combined analysis of changes in the sensorimotor rhythms and slow corticalpotentials. Journal of Neural Engineering 11(5), 056009 (2014)

8. Xu, R., Jiang, N., Lin, C., Mrachacz-Kersting, N., Dremstrup, K., Farina, D.:Enhanced low-latency detection of motor intention from eeg for closed-loop brain-computer interface applications. IEEE Transactions on Biomedical Engineering61(2), 288–296 (2014)

9. Lew, E.Y., Chavarriaga, R., Silvoni, S., Millan, J.D.R.: Single trial prediction ofself-paced reaching directions from eeg signals. Frontiers in Neuroscience 8 (2014)

10. Vidaurre, C., Sander, T.H., Schlogl, A.: Biosig: the free and open source softwarelibrary for biomedical signal processing. Computational Intelligence and Neuro-science (2011)

11. Breitwieser, C., Daly, I., Neuper, C., Muller-Putz, G.: Proposing a standardizedprotocol for raw biosignal transmission. IEEE Transactions on Biomedical Engi-neering 59(3), 852–859 (2012)

12. Jung, T.P., Makeig, S., Humphries, C., Lee, T.W., Mckeown, M.J., Iragui, V.,Sejnowski, T.J.: Removing electroencephalographic artifacts by blind source sepa-ration. Psychophysiology 37(02), 163–178 (2000)

13. Hayashi, Y., Nagai, K., Ito, K., Nasuto, S.J., Loureiro, R.C., Harwin, W.S.: Analy-sis of eeg signal to detect motor command generation towards stroke rehabilitation.In: Converging Clinical and Engineering Research on Neurorehabilitation, pp. 569–573. Springer (2013)

14. Wairagkar, M., Daly, I., Hayashi, Y., Nauto, S.J.: Novel single trial movement clas-sification based on temporal dynamics of eeg. In: Proceedings of 6th InternationalBrain Computer Interface Conference, Gratz (2014)

15. Lew, E., Chavarriaga, R., Silvoni, S., Millan, J.D.R.: Detection of self-paced reach-ing movement intention from eeg signals. Front. Neuroeng. 5(13) (2012)

movement intention detection from autocorrelation of eeg

Documents