FIELD OF THE INVENTION

This invention relates to the technical field of computer-implemented systems and data processing for muscle artifacts removal from an EncephaloGram (EG), in particular from an ElectroEncephaloGram (EEG) or MagnetoEncephaloGram (MEG). More particularly the present invention involves a system for muscle artifact removal in EG, preferably a digital EG, based on Canonical Correlation Analysis (CCA) as a Blind Source Separation (BSS) technique. The method is particularly suitable for successfully removing the muscle artifact without altering the recorded underlying brain activity and outperforms a low pass filter with different cut-off frequencies and the Independent Component Analysis (ICA) based technique for muscle artifact removal of the state of the art. Moreover, the fast computational time and the ordering of the CCA components make the system of present invention applicable in clinical and medical practice.

BACKGROUND OF THE INVENTION AND ASSESMENT THEREOF

Description of the Related Art The ElectroEncephaloGram (EEG) is frequently contaminated by electrophysiological potentials associated with muscle contraction due to biting, chewing, frowning or other muscle activities. These muscle artifacts, also known as ElectroMyoGram (EMG) artifacts, obscure the EEG and complicate the interpretation of the EEG or even make the interpretation unfeasible. In the domain of Event Related Potentials (ERP), epochs heavily contaminated with muscle artifacts are typically rejected. When a small number of epochs is available, for instance in complex memory tasks, each rejected epoch will reduce the interpretability of the ERP (M. Stecker, "The effects of automatic artifact rejection on evoked potential recordings." Computers in Biology and Medicine, vol. 32, pp. 247-259, 2002). In automated epileptic seizure detection, muscle artifacts considerably reduce the selectivity of the method (M. E. Saab and J. Gotman "A system to detect the onset of epileptic seizures in scalp EEG" Clinical Neurophysiology, In

Press, Corrected Proof, Available online 18 September 2004). In the field of seizure prediction, muscle artifacts obscure the interpretation of the nonlinear measures when looking for a pre-ictal state in scalp EEG recordings (W. De Clercq, P. Lemmerling, S. Van Huffel, and W. Van Paesschen, "Correspondence : Anticipation of epileptic seizures from standard EEG recordings," The Lancet, vol. 361 , pp. 970-971 , 2003). Hence, there is a need in the art to remove these artifacts from the EEG.

Low-pass filters are commonly used to remove muscle artifact. However, as the frequency spectrum of the muscle artifacts overlaps with that of interesting brain signals, frequency filters not only suppress muscle artifacts but also valuable information (I. Goncharova, D. McFarland, T. Vaughan, and J. Wolpaw, "EMG contamination of EEG: Spectral and topographical characteristics," Clinical Neurophysiology, vol. 114, pp. 1580-1593, 2003.).

Most EEG artifact removal techniques only consider ocular artifacts, while, little effort has been made on the removal of muscle artifacts. Regression methods, investigated for eye blink and eye movement artifact correction (R. Croft and R. Barry, "EOG correction: Which regression should we use?" Psychophysiology, vol. 37, pp. 123-125, 2000; G. Wallstrom, R. Kass, A. Miller, J. Cohn, and N. Fox, "Automatic correction of ocular artifacts in the EEG: A comparison of regression-based and component-based methods," International Journal of Psychophysiology, vol. 53, pp. 105-119, 2004.), are impractical when muscle artifacts are considered, because no reference channel is available. A more recently explored approach is Independent Component Analysis (ICA) which separates the EEG into statistical independent components. This method was already successfully applied to ocular artifact removal (T. Jung, C. Humphries, T. Lee, S. Makeig, M. McKeown, V. Iragui, and T. Sejnowski, "Extended ICA removes artifacts from ElectroEncephaloGraphic recordings." Advances in Neural Information Processing Systems, vol. 10, pp. 894-900, 1998; H. Nam, T.-G. Yim, S. Han, J.-B. Oh, and S. Lee, "Independent component analysis of ictal EEG in medial temporal lobe epilepsy," Epilepsia, vol. 43, no. 2, pp. 160-164, 2002.). However, cross-talk can be observed when the separation of brain and muscle activity is considered (H. Nam, T.-G. Yim, S.

Han, J.-B. Oh, and S. Lee, "Independent component analysis of ictal EEG in medial temporal lobe epilepsy," Epilepsia, vol. 43, no. 2, pp. 160-164, 2002., E. Urrestarazu, J. lriarte, M. Alegre, M. Valencia, C. Viteri, and J. Artieda, "Independent component analysis removing artifacts in ictal recordings," Epilepsia, vol. 45, no. 9, pp. 1071-1078, 2004, S. Mozaffar and D. W. Petr, "Artifact Extraction from EEG Data Using Independent Component Analysis," Information Telecommunication and Technology Center, University of Kansas, Lawrence, KS, Tech. Rep. ITTC-FY2003-TR-03050-02, Dec. 2002.). Moreover, when applying ICA, the identification of the components containing artifacts in general, and muscle activity in particular, is not obvious, and requires additional user attention. ICA based techniques who try to overcome this problem, like constrained ICA (clCA) (C. J. James and J. Gibson, "Temporally constrained ICA: An application to artifact rejection in electromagnetic brain signal analysis," IEEE Trans. Biomed. Eng., vol. 50, no. 9, pp. 1108-1116, Sept. 2003), cannot be applied for muscle artifact removal. This method extracts only that independent component that is most similar to a specific reference signal (W. Lu and J. Rajapakse, "ICA with reference," in Proceedings 3rd In. Conf. Independent Component Analysis and Blind Signal Separation: ICA 2003, 2003, pp. 120-125). However, for muscle artifacts the construction of a good reference signal is impractical.

Present invention comprises a method and system for muscle artifact correction in EEG based on the statistical Canonical Correlation Analysis (CCA) (H. Hotelling, "Relations between two sets of variates." Biometrika, vol. 28, pp. 321-377, 1936.) applied as a Blind Source Separation (BSS) technique (O. Friman, M. Borga, P. Lundberg, and H. Knutsson, "Exploratory fMRI analysis by autocorrelation maximization." Neurolmage, vol. 16, no. 2, pp. 454^-64, 2002), which will be further referred to as BSS-CCA. The BSS-CCA method, as ICA, thus belongs to the group of data-driven techniques solving the BSS problem. The blind source separation can be defined as the determination of the original or underlying sources which generate multisensor signals, while knowing very little, if anything, of how these sources project on the sensors and making little assumptions on these source signals. The BSS-CCA method requires the sources to be

uncorrelated and to be maximally correlated with a given function, which is in our case a time shifted version of the original signal.

By consequence, in our case, the last requirement opposes the sources to be maximally autocorrelated, inducing an order on the sources based on the autocorrelation index. The BSS-CCA method is computationally much faster than ICA (O. Friman, M. Borga, P. Lundberg, and H. Knutsson, "Exploratory fMRI analysis by autocorrelation maximization." Neurolmage, vol. 16, no. 2, pp. 454-464, 2002.).

SUMMARY OF THE INVENTION

The present invention solves the problems of the related art by a new method based on canonical correlation analysis as a blind source separation technique for muscle artifact removal (Myogram artifacts) from an EncephaloGram. In particular it solves the problem for muscle artifact removal (ElectroMyoGram artifacts) from an ElectroEncephaloGram, prefereably a digital ElectroEncephaloGram or for muscle artifact removal from a MagnetoEncephaloGram.

The performance of present invention was tested on both synthetic data and ictal EEG (see Examples).

One embodiment of present invention involves a system for recording source signals of a subject comprising a detecting means configured to receive said source signals characterised in that the system further comprises a processing means for removal of muscle activity signals with frequency spectrum overlap from central nervous system activity signals based on the different autocorrelation structure of the timecourses of the sources. This system can be characterised in that it comprises a canonical correlation analysis (CCA) function and a blind source separation configured for CCA decomposition and removal or separation of the muscle activity signals based on their lower level of autocorrelation than that of the central nervous system activity signals. The system can be further characterised in that it involves the steps of CCA

decomposition to order the CCA components according to their level of autocorrelation and removal of fraction of the lower autocorrelated components.

This source separation system of present invention is suitable for real-time online performance and for continuous EncephaloGraphic monitoring as for instance in an Epilepsy Monitoring Unit or an Intensive Care Unit.

The system can be used for source, separation in EncephaloGraphic operations, more particularly the system can be an electroencephalograph configured to remove electrophysiological potentials associated with muscle contraction from mixed electrophysiological potentials related to brain activity and to muscle activity or it can be a MagnetoEncephalograph configured to remove magnetic signals emanating from the muscles from mixed magnetic signals emanating from brain activity and from the muscle activity

The source separation system of present invention can be used for automatical or semi- automatical removal of muscle artifacts from mixed sources for instance for removal of artifacts associated with muscle contraction from brain signals without altering said brain signal. Such system is suitable to separate or to clean ElectroEncephaloGraphic recordings from ElectroMyoGram artifacts. This enhances the visual interpretability of the EncephaloGrams and can also be used in a pre-processing step to improve the further signal processing of the ElectroEncephaloGraphic signals.

The system of present invention is also suitable to separate or to clean MagnetoEncephaloGraphic recordings from muscle artifacts. This enhances the visual interpretability of the MEG and can also be used in a pre-processing step to improve the further signal processing of the MagnetoEncephaloGraphic signals.

The method of present invention outperformed the low-pass filter and an ICA based technique for muscle artifact removal. Moreover, the fast computational time and the

ordering of the CCA components make the proposed method applicable in medical practice and particularly clinical practice.

Further scope of applicability of the present invention will become apparent from the detailed description given hereinafter. However, it should be understood that the detailed description and specific examples, while indicating preferred embodiments of the invention, are given by way of illustration only, since various changes and modifications within the spirit and scope of the invention will become apparent to those skilled in the art from this detailed description. It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention, as claimed.

ILLUSTRATIVE EMBODIMENTS OF THE INVENTION

Disclosure of the invention

A description of at least one way of carrying out the present invention has described in the examples 2 and 3 and 4 is discussed hereunder.

Present invention concerns a new method and system for muscle artifact correction in EEG. The system is preferably based on CCA and belongs to the group of blind source separation techniques. The performance of the method was tested on simulated data with the aim to remove muscle artifacts without altering the known underlying brain signal. In the simulation study, the performance of the proposed method was compared to the performance of other commonly used methods. A low pass filter with different cutoff frequencies and ICA (JADE) (J. -F. Cardoso, "Higher-order contrast for independent component analysis", Neural Comput, Vol. II. Pp 157 - 192, 1999) were used as contenders. The BSS-CCA method succeeded in removing the superimposed muscle artifact from the brain signal and performed this task better than its competitors. The frequency spectrum of the muscle artifacts overlaps with that of the brain activity signal, leading to the fact that the low pass filters were not capable of removing the whole

artifact without altering the underlying brain signal; as shown in Fig. 2 (d-g). Due to mixing of the muscle artifact in ICA components accounting for brain activity, BSS-CCA outperformed ICA for all considered signal to noise ratios. Especially for low SNRs it significantly outperformed the other methods, as shown in Fig. 5.

The method was applied to a real ictal EEG recording. The muscle artifact was removed with success without altering the ictal EEG. Since the BSS-CCA method is designed for finding maximally autocorrelated signals, the method succeeded in extracting the brain activity without cross-talk of the muscle artifacts, which have very different autocorrelation structures. The muscle artifact removal based on ICA (JADE) was more difficult and again less complete than the BSS-CCA results. The muscle artifact appeared on nearly all ICA components and even accompanied the ictal component, as shown in Fig. 6 (c) and Fig. 7 (c). As a consequence, the selection of the ICA components accounting for the muscle artifact that should be excluded in the reconstruction became difficult or even problematic. These results correspond to the findings of (H. Nam, T.-G. Yim, S. Han, J.-B. Oh, and S. Lee, "Independent component analysis of ictal EEG in medial temporal lobe epilepsy," Epilepsia, vol. 43, no. 2, pp. 160-164, 2002 and E. Urrestarazu, J. Iriarte, M. Alegre, M. Valencia, C. Viteri, and J. Artieda, "Independent component analysis removing artifacts in ictal recordings," Epilepsia, vol. 45, no. 9, pp. 1071-1078, 2004). Both studies reported that the removal of muscle artifacts with ICA was more difficult and less complete than that of eye artifacts because of mixture of the muscle artifact with EEG components.

Compared to ICA, the BSS-CCA method as a blind source separation technique has three major advantages. First, the CCA components are ordered according to their autocorrelation, whereas the ICA components have no order. Second, the BSS-CCA algorithm has a much smaller computational time, which makes the algorithm applicable for real-time/on-line performance (O. Friman, M. Borga, P. Lundberg, and H. Knutsson,

"Exploratory fMRl analysis by autocorrelation maximization." Neurolmage, vol. 16, no. 2, pp. 454-464, 2002; M. Borga and H. Knutsson, "A canonical correlation approach to blind source separation." Dept. of Biomedical Engineering, Linkδping University,

Sweden, Tech. Rep. LJU-IMT-EX-0062, 2001). In addition, the outcome of the BSS-CCA method is unique, while the ICA output can be different for the same input, depending on the algorithm (O. Friman, M. Borga, P. Lundberg, and H. Knutsson, "Exploratory f M Rl analysis by autocorrelation maximization." Neurolmage, vol. 16, no. 2, pp. 454-464, 2002).

It was surprisingly found that the muscle activity is present in the lowest components in the CCA decomposition, which indicates that they are a more noise like signal (small autocorrelation) than the brain activity. A possible explanation for these observations could be found in the origin of the muscle artifact signal.

A muscle is composed out of different motor units, each compromising a motor neuron and a group of innervated muscle fibres. The recruitment of several motor units when contracting a muscle, leads therefore to a complex electrical activity. This complex signal is a summation of asynchronized potentials, caused by the muscle fibres of the different activated motor units, and results in a low autocorrelated signal. This property is not apparent in other EEG activity and consequently, their time course has a higher autocorrelation compared to that of muscle activity, resulting in a good separation between both signals.

These observations can be exploited to automatically remove the muscle artifact from the EEG. Removing all CCA components below a threshold value of the autocorrelation index is useful for automatical removal of muscle artifacts in a large subject group.

By present invention an automatic removal of muscle artifacts or a semi-automatic removal of muscle artifacts has been proposed. In this way, the BSS-CCA method becomes practical for people without experience in the field of BSS and avoids manual selection of a fairly large set of components as in ICA. ICA based techniques who try to avoid this need for a selection of artifactual components, like constrained ICA (C. J.

James and J. Gibson, "Temporally constrained ICA: An application to artifact rejection in electromagnetic brain signal analysis," IEEE Trans. Biomed. Eng., vol. 50, no. 9, pp.

1108-1116, Sept. 2003), can not be applied for muscle artifact removal. The clCA

needs a good reference signal which is lacking for the muscle artifact. An important factor in the clinical applicability of the present invention is that the user can easily determine the optimal number of components to be removed by observing the EEG of interest while removing components (see example 4), hereby avoiding the removal of important information. The proposed semi-automatic removal of muscle artifact is of great importance in the visual evaluation of ictal EEGs.

As described above, the proposed EMG removal technique can be used to clean the EEG resulting in a better visual interpretation. Next to this visual interpretation tool, the removal of muscle artifact can be used as a pre-processing step to facilitate the signal processing of scalp EEG signals. As the muscle artifacts give rise to false positives in current seizure detection methods, pre-processing the EEG with CCA could improve the sensitivity and specificity of the detectors. In the field of seizure prediction, muscle artifacts obscure the interpretation of nonlinear measures when looking for a pre-ictal state in scalp EEG recordings (W. De Clercq, P. Lemmerling, S. Van Huffel, and W. Van Paesschen, "Correspondence: Anticipation of epileptic seizures from standard EEG recordings," The Lancet, vol. 361, pp. 970-971, 2003). Removing the artifacts could again improve the performance of the nonlinear measures. Furthermore, ictal dipole source localization would benefit considerably when muscle artifacts were removed. In the domain of the event related potentials, heavily EMG contaminated epochs will no longer need to be rejected. This could lead to a better interpretability of the ERP, especially in cases where a small number of epochs is available, as for example in complex memory tasks.

The method could also be applied to MagnetoEncephaloGraphic (MEG) recordings. In MEG recordings the number of channels is usually much larger than in EEG. As a consequence, the number of components derived by ICA is also much larger and the selection of the artifactual ICA components on visual inspection becomes even more demanding than in EEG. For this reason, the ordered CCA components could be of major use to MEG signals too.

Examples

Example 1: Blind source separation by canonical correlation analysis for artifact removal

BSS by CCA is outlined, and its application as artifact corrector of EEG data is described

A. Blind Source Separation by Canonical Correlation Analysis

In the blind source separation problem, the observed timecourse x(f) = [xi(0/ *2(0/ ■•■ _> xκ(t)] ^τ ; with if = 1...N , with N the number of samples and K the number of sensors, is the result of an unknown linear mixture of a set of unknown source signals s(0 = [s- _\ (t); s ₂ (t);...; s _K {t)] ^r r-

x(f) = As(f), (1)

where A is the unknown mixing matrix. The goal is to estimate the mixing matrix and recover the original source signals s(t). This is carried out by introducing the de-mixing matrix W such that

approximates the unknown source signals in s(0, by a scaling factor. Unless there are extra constraints imposed, it is in general impossible to solve this problem. Canonical correlation analysis solves the problem by forcing the sources to be mutually uncorrelated and maximal correlated with a predefined function (M. Borga and H. Knutsson, "A canonical correlation approach to blind source separation." Dept. of Biomedical Engineering, Linkδping University, Sweden, Tech. Rep. LiU-IMT-EX-0062, 2001).

Let χ(f) be the observed data matrix with K mixtures and N samples, then we define the predefined function y(t) as a temporally delayed version of the original data matrix, to enforce the sources to be maximally autocorrelated (M. Borga and H. Knutsson, "A canonical correlation approach to blind source separation." Dept. of Biomedical Engineering,

Linkδping University, Sweden, Tech. Rep. LiU-IMT-EX-0062, 2001):

y(f) = x(M) (3)

BSS-CCA refers to this autocorrelation version.

When the mean of each row from the data matrices x(f) and y(if) is removed, canonical correlation analysis obtains two sets of basis vectors, one for x and the other for y, such that the correlations between the projections of the variables onto these basis vectors are mutually maximized. Consider the linear combinations of the components in x and y:

X = Wx ₁ X ₁ + ... + W _xK XK = W _x ⁷ X y = WyIy ₁ + ... + w _yK yκ = w _y ^r y; (4)

CCA finds the vectors W _x = [w _x1 ; ...;w _x κ] ^τ and w _y = [w _y1 ; ...;w _y κ] ^τ that maximize the correlation p between x and y by solving the following maximization problem:

W _χ ^r C _χy Wy

(5) y/{ W _x ^1' Cx _x W _x ) ( W _y ^X Cyy Wy ) ^'

with Cxx and C _yy the within-set covariance matrices of x and y, respectively, and C _xy the between-sets covariance matrix. The maximalization problem can be solved by setting

the derivatives of Eq. (5) with respect to W _x and w _y to zero, which after manipulation results in the following two eigenvalue problems:

^i _EE CWCyy Cy _x W _x = p W _x ,

These equations results in K solutions sorted by autocorrelation p _/ , from which the K estimates ztf) of the sources s,(f) can be derived (M. Borga and H. Knutsson, "A canonical correlation approach to blind source separation." Dept. of Biomedical Engineering, Linkόping University, Sweden, Tech. Rep. LiU-IMT-EX-0062, 2001).

B. Applying BSS-CCA to EEG for artifact correction

BSS-CCA is effective in performing source separation if the following assumptions are valid: (1) the mixing medium is linear and propagation delays are negligible, (2) the timecourses of the sources are uncorrelated, (3) the timecourses of the sources have different autocorrelation structures, and (4) the number of sources is equal as or less than the number of sensors. In the case of EEG signals, assumption (1) is satisfied, because volume conduction is thought to be linear and instantaneous (R. Plonsey and D. Heppner, "Considerations of quasistationarity in electrophysiological systems." Bulletin of Mathematical Biophysics, vol. 29, no. 4, pp. 657-664, 1967.). The sources in case of EEG, are related to brain activity and artifacts, which are anatomically and physiologically separate processes, meaning that those processes are uncorrelated, hence assumption (2) is fulfilled. It will be demonstrated by the present invention that brain activity produces structured signals having a high autocorrelation, whereas muscle activity is less structured and encompass more properties related to temporally white noise. Therefore different sources seem to have different autocorrelation structures and hence assumption (3) seems to be satisfied. It is questionable whether assumption (4) is

fulfilled or has to be fulfilled because the effective number of uncorrelated sources contributing to the scalp EEG is not known.

Once BSS-CCA is applied to the EEG and the sources, or components, contributing to the EEG are derived, then the manual or automatical identification (see example 4) of the components containing the artifacts is required. Next, the EEG is reconstructed without these components, by projecting the selected nonartifactual components back onto the scalp. The enhanced EEG x _c iean(f) is then:

X _c teaπrø = AcfeβπZ(f); (7)

with z(f) the sources obtained by BSS-CCA ₁ and A _c i _e an the mixing matrix with the columns representing activations of the artifactual sources, set to zero.

Example 2: Material and Methods

The synthetic data and the real life EEG recording used to test the performance of the method are described. For both the synthetic data and the EEG recording, the performance was compared with the performance of a low-pass-filter and an ICA based technique.

A. Construction of the synthetic data

The aim of the simulation study was to evaluate the performance of the proposed method in removing muscle artifacts. Several simulations with different signal to noise ratios were performed.

1) Brain activity: Three EEG epochs of 10 s containing respectively mainly delta (<4Hz), alpha (8-13Hz) and beta (>13Hz) activity, were selected as underlying brain signals. In the selected EEGs no muscle artifact was present according to the visual inspection of an experienced neurophysiologist. The data was collected from 21 scalp electrodes

placed according to the International 10-20 System (M. R. Nuwer, G. Comi, R. Emmerson, et al., "IFCN standards for digital recording of clinical EEG," Electroencephalogr am Neurophysiol, vol. 106, pp. 259-261, 1998) with additional electrodes 71 and 72 on the temporal region. The sampling frequency was 250 Hz and an average reference was used. The three EEG epochs containing delta, alpha and beta activity were stored in three 21-by-2500 dimensional matrices B-i, B ₂ and B ₃ , respectively.

2) Muscle activity: To obtain only muscle activity, necessary for the simulation study, it is not sufficient to select muscle artifacts in the EEG as these events will not only contain muscle activity but also activity from the brain. To remove the EEG activity we used ICA (SOBI) (A. Belouchrani, K. Abed-Meraim, J.-F. Cardoso, and E. Moulines, "A blind source separation technique using second order statistics," IEEE Trans. Signal Processing, vol. 45, pp. 434-444, 1997) to decompose these events. Note that for a large number of events, which we visually inspected, no clear separation was established. For those events where a clear separation between EMG and brain activity was obtained, we selected a component that accounted for muscle artifact together with the corresponding field distribution. This procedure was repeated for three different 10 s EEG epochs recordings from three different subjects. The selected components and their corresponding field distributions (topographical maps) are shown in Fig. 1 (a-c). The individual components were reconstructed into a conventional EEG format. The resulting average referenced signal M containing all three independent muscle activities is shown in Fig. 1 (d).

3) Data and measures to test the performance: In the simulation study the average referenced muscle artifact signal M was superimposed on the average referenced signal B containing only brain activity:

X(λ) = B +λ.M; (8)

with λ representing the contribution of muscle activity. The Root Mean Squared (RMS) value of the brain signal is then equal to

with N equal to the number of time samples and K equal to the number of EEG channels. The RMS value of the muscle artifact signal is equal to

An important measure is the Signal to Noise Ratio (SNR) which is defined as follows,

MIS(B) ^' RMS(X - MV ^{{ }}

Changing the λ parameter alters the signal to noise ratio of our simulated signal. Several simulations with different signal to noise ratios were performed. The performance is expressed in terms of the Relative Root Mean Squared Error (RRMSE):

R _n R _n M _w S _? E _iP - _. RM _RM S( _S B _(B - ₎ B) , (12)

where B is the signal after muscle artifact removal.

The simulations were carried out for the three different brain signals B1 , B2 and B3, described in section A1.

B. Methods for comparison

For comparison, the proposed technique for muscle artifact removal is compared with other commonly used techniques for this purpose:

• Low pass filter: a low pass butterworth filter of order 8 was used with four different cut-off frequencies equal to 10Hz, 15Hz, 20Hz and 30Hz.

• ICA: as it is not in the scope of this paper to compare all different ICA algorithms, we limit ourselves to the Joint Approximate Diagonalization of Eigen-matrices (JADE) algorithm (J .-F. Cardoso, "Higher-order contrast for independent component analysis," Neural Comput, vol. 11 , pp. 157-192, 1999). This ICA algorithm was previously used in a study for eye and muscle artifact removal in ictal scalp EEG recordings (E. Urrestarazu, J. Iriarte, M. Alegre, M. Valencia, C. Viteri, and J. Artieda, "Independent component analysis removing artifacts in ictal recordings," Epilepsia, vol. 45, no. 9, pp. 1071-1078, 2004). The muscle artifact was removed as follows. The ICA components were calculated by using the JADE algorithm. The selection of the components accounting for the muscle artifact was based on visual inspection and the signal was reconstructed, excluding the components related to the artifact. It is worth mentioning that, for the muscle artifact removal based on ICA (JADE) to be unbiased, a different ICA algorithm (namely SOBI) as the JADE algorithm was used in the construction of the simulation study in section A.2.

C. Ictal EEG recording

Because ictal EEG is often severely contaminated with muscle artifact, complicating the determination and localization of the ictal onset, the investigation had been focussed on the artifact correction of ictal EEG. Fig. 6 (a) shows a 10 s epoch of a scalp EEG recording from a long-term Epilepsy Monitoring Unit. The acquisition settings were the same as for the brain activity in section A.1. This EEG contains ictal activity from a patient with Mesial Temporal Lobe Epilepsy (MTLE). The epileptic activity is mainly present on the 72, FQ and 74 electrodes. The seizure EEG is contaminated with muscle artifact and eye blinks. Again we applied CCA, ICA and frequency filtering to this epoch.

The results were then visually inspected. Error measures could not be used as we did not have a-priori information available of the brain (epileptic) sources.

Example 3: Results

The obtained results are shown on the synthetic data and the real life EEG recording used to test the performance of the method described above. For both the synthetic data and the EEG recording, the performance was compared with the performance of a low- pass-filter and an ICA based technique.

A. Results on the synthetic data

Fig. 2 (a) shows a simulated signal X with a signal to noise ratio equal to 0.65. For this simulated signal, the brain activity signal B ₂ , which is shown in Fig. 2 (b), is used. The (normalized) components obtained by applying the BSS-CCA algorithm on X are shown in Fig. 3 (a). The autocorrelation of the components is shown in Fig. 3 (b). The components accounting for the muscle artifact are present in the three lowest autocorrelated components. The corresponding field distributions are shown in Fig. 3 (c- e). Both the derived time courses and field distributions are very similar to the ones of the original muscle sources shown in Fig. 1 (a-c). Fig. 4 shows the RRMSE as a function of the number of lowest autocorrelated CCA components that are excluded in the EEG reconstruction. A minimum is reached when the last three components are removed. Excluding those components in the reconstruction of the EEG results in the cleaned EEG shown in Fig. 2 (c). The RRMSE of the BSS-CCA processed data was equal to 0.11.

Fig. 2 (d-g) shows the results of applying the low pass filter on the simulated signal X with cut-off frequencies of 10Hz, 15Hz, 20Hz and 30Hz respectively. Because the frequency spectrum of the muscle artifacts overlaps with that of the brain signal, the low pass filters were not capable of removing the whole artifact without altering the underlying brain activity. The 10 Hz low pass filter removed most of the artifact activity,

but also altered the fast activity present in the B ₂ EEG recording. Moreover, the slow activity related to the muscle artifact between t= 0-2 s was not removed. The higher the cut-off frequency, the less complete the filter removed the muscle artifact. In Fig. 2 (h) the result of removing the artifactual JADE components is shown. As some of the muscle activity was present in ICA components containing brain activity, the removal of the muscle artifact based on the three muscle artifact components, which were selected by visual inspection, was not complete. The RRMSE of the 10HZ , 15Hz, 20Hz and 30Hz low pass filtered data was equal to 0.35, 0.33, 0.36 and 0.51 respectively. The RRMSE of the JADE processed data was equal to 0.25.

The RRMSE as a function of signal to noise ratio on the B ₂ brain signal for the different artifact removal techniques (BSS-CCA, ICA(JADE) and the low pass filter with cut-off frequency at 10Hz, 15Hz, 20Hz and 30Hz) is shown in Fig. 5. For high signal to noise ratios the 10 Hz filter performed worst because the brain signal B ₂ has mainly alpha (8- 13Hz) activity, and consequently, the 10 Hz filter altered the brain signal the most. For lower SNR the performance of the 10 Hz filter compared to the other low pass filters improved as it removed most of the muscle artifact. Compared to the low pass filters, the performance of ICA was better for all signal to noise ratios. The BSS-CCA method outperformed all methods for all SNRs. The results of the simulation study on the B ₁ and B ₃ brain signals containing delta and beta activity are similar to that of B ₂ .

B. Results on the ictal EEG recording

This section shows the results obtained by applying the method to the epoch of multichannel EEG containing ictal EEG and muscle artifacts shown in Fig. 6 (a). The (normalized) components as obtained by BSS-CCA are depicted in Fig. 7 (a); their corresponding autocorrelation is shown in Fig. 7 (b). We observed that the muscle activity is well separated from the ictal component (component 3), and is present in the lowest autocorrelated CCA components. Fig. 6 (b). shows the reconstructed EEG after excluding components (9-21).

For comparison, the low pass filter with cut-off frequency equal to 10 Hz and the JADE algorithm were also applied. Fig. 7 (c) shows the ICA components as obtained by JADE. The selection of the artifactual components by visual inspection becomes difficult because the muscle artifact is present in almost all components and even accompanies the ictal component (component 3). Fig. 6 (c) shows the reconstructed EEG based on excluding suspected muscle artifactual components 4, 7, 8, 13, 14, 16-19, 20 and 21. The muscle artifact removal based on ICA is less complete as the CCA result. Fig. 6 (d) shows the result of the low pass filter (10 Hz). As the frequency of the epileptic activity is lower than 10 Hz, the low pass filter was able to remove most of the muscle artifact without altering the epileptic activity, although a flattening of the signal is observed.

Example 4 Semi-automatic and automatic removal of muscle artifacts

In the following we explain the removal of muscle artifacts for better interpretation of the EEG. CCA components are gradually removed from bottom upwards based on the ordering of the components. An example of use is given below. Fig. 8 (a) shows a 10 s EEG epoch containing, in the opinion of an experienced neurologist, sharp epileptic activity at channel 02. This EEG epoch is severely contaminated by repeated eye blinking and muscle artifacts. The activity on channel 02 is used to inspect the performance of the semi-automatic artifact removal of muscle artifacts. Fig. 8 (b-c) shows, respectively, the reconstructed EEG with exclusion of the 7, 14 lowest autocorrelated CCA components. By gradually excluding CCA components the muscle artifact is progressively removed, while the background EEG and the sharp epileptic activity on channel 02 remain unaltered. Fig. 8 (d) shows the EEG reconstruction with exclusion of the 15 lowest autocorrelated CCA components. Here, too many components were removed as the clinical relevant activity was removed on channel 02. This means that the maximal number of CCA components which can be removed is equal to 14.

This proposed semi-automatic method for muscle artifact removal can be implemented in a Graphical User Interface (GUI) with a scrolling system where moving the scroll bar corresponds to removing/adding CCA components.

A fully automated system can be obtained by removing all CCA components below a treshold value of the autocorrelation index.

Example 5 Improved interpretation of the ictal electroencephalogram by a new muscle artifact removal subspace-based technique

Purpose: The aim of this study was to investigate the clinical relevance of a recently developed method for muscle artifact removal in the ictal electroencephalogram (EEG). Methods: One ictal EEG of 26 patients with refractory partial epilepsy and an ictal onset zone that was well-defined during a presurgical evaluation, were processed with a new subspace-based muscle artifact removal technique. The cleaned EEGs were compared with the band pass (0.3-35 Hz) filtered original EEGs by an unblinded neurologist.

Results: In 24 of 26 cases (92%) muscle artifact contaminated the ictal EEGs significantly. In these 24 cases the ictal EEG was easier to interpret after the muscle artifact was removed. The time of ictal onset on EEG was detected earlier in 9 out of 26 cases (35%), compared to the original EEG. The onset of the seizure was better localized in 7 out of 26 cases (27%) and the ictal-pattem of the onset was located in a higher frequency range for 8 out of 26 cases (31%). Localised ictal-onset beta activity was observed only after removal of muscle artifact in 5 patients (19%). The muscle artifact removal technique did not degrade the EEG signal in any of the patients. Conclusions: Our muscle artifact removal subspace-based technique was easy to implement, fast and user-friendly in a clinical environment, and improved the interpretation of ictal EEG in a clinically significant way in around 50% of patients. We are currently conducting the same study with two blinded EEG readers and a larger sample of ictal EEGs.

Example 6 automating muscle artifact removal in ictal EEG recording

Rationale: The aim of this study was demonstrate the performance of automated muscle artifact removal from ictal electroencephalograms (EEG) based on canonical correlation analysis (CCA). Methods: CCA separated a 10s EEG epoch in a set of components (or sources) with a decreasing autocorrelation index (Al). It has been noticed that muscle artifact components have a lower Al than genuine EEG components. A neurologist removed muscle artifact from 317 epochs of ictal EEG of 40 patients by gradually removing the components starting from the one with the lowest Al. The neurologist determined subjectively the separation point where as much muscle activity was removed as possible, without affecting the brain activity significantly. This yielded a gold standard binary classification of the CCA components into muscle artifact components and EEG components. Next a fully automated classification of each CCA component was performed based on the ratio of spectral energy of the frequency band 25-50Hz to the band 10-15Hz. It is anticipated a large value of this feature will be observed when muscle activity is present in a component. The optimal feature threshold was selected by maximally remaining EEG components and maximally suppressing muscle artifact components.

Results: The developed automated method was able to remain on average 99±0.3 % of the EEG component energy, while removing oh average 71 ±33 % of the muscle artifact component energy.

Conclusion: An automated muscle artifact removal technique was presented, that could be used as pre-processing step in the early detection of ictal activity or as a filter in the visual evaluation of ictal EEG.

BRIEF DESCRIPTION OF THE DRAWINGS

The present invention will become more fully understood from the detailed description given herein below and the accompanying drawings which are given by way of illustration only, and thus are not limitative of the present invention, and wherein:

Figures 1(a), (b) and (c) demonstrates three selected independent components and their corresponding field distributions, (d) The resulting EEG epoch containing all three independent muscle activities.

Figure 2 demonstrates a (a) Simulation matrix X containing: alpha activity B ₂ and superimposed muscle activity M with SNR = 0.65; (b) Original EEG epoch B ₂ ; (c) BSS- CCA processed data; (d), (e), (f) and (g) low pass filtered data with cut-off frequency 10Hz, 15Hz ₁ 20Hz and 30Hz respectively; (h) JADE processed data

Figure 3 demonstrates . (a) CCA components (normalized) obtained by applying BSS- CCA on the 10 sec EEG epoch shown in Fig. 2 (a); (b) the autocorrelation of the CCA components; (c-e) Distribution of the three lowest autocorrelated CCA components ((c) component 21 , (d) component 20, (e) component 19).

Figure 4 displays the RRMSE as a function of the number of lowest autocorrelated CCA components excluded in the reconstruction of the EEG

Figure 5 displays the RRMSE as a function of SNR on the EEG epoch B ₂ containing alpha (8-13Hz) activity

Figure 6 demonstrates an (a) Original 10 s EEG epoch; (b) EEG reconstruction after excluding artifactual CCA components, (c) EEG reconstruction based on JADE components, (d) EEG filtered by the low pass filter with cut-off frequency equal to 10 Hz

Figure 7 demonstrates (a) the components (normalized) obtained by applying BSS-CCA on the ictal EEG recording shown in Fig. 6 (a), (b) the autocorrelation of the CCA components, (c) ICA components obtained by applying JADE on the same recording

Figure 8 demonstrates the semi-automatic muscle artifact removal (a) an original EEG with sharp epileptic activity on channel 02; (b) an EEG after exclusion of components

(15-21); (c) an EEG after exclusion of components (8-21) and (d) an EEG after exclusion of components (7-21)