Improved ballistocardiac artifact removal from the electroencephalogram recorded in fmri

Transcription

1 Journal of Neuroscience Methods 135 (2004) Improved ballistocardiac artifact removal from the electroencephalogram recorded in fmri Kyung Hwan Kim a,, Hyo Woon Yoon b, Hyun Wook Park b,1 a Department of Biomedical Engineering, Yonsei University, 234 Maeji-ri, Heungup-myun, Wonju, Kangwon-do , South Korea b fmri Laboratory, Department of Electrical Engineering and Brain Science Research Center, Korea Advanced Institute of Science and Technology, Guseong-dong, Yuseong-gu, Daejeon , South Korea Received 19 August 2003; received in revised form 22 December 2003; accepted 22 December 2003 Abstract The simultaneous recording of electroencephalogram (EEG) and functional magnetic resonance image (fmri) is a promising tool that is capable of providing high spatiotemporal brain mapping, with each modality supplying complementary information. One of the major barriers to obtain high-quality simultaneous EEG/fMRI data is that pulsatile activity due to the heartbeat induces significant artifacts in the EEG. The purpose of this study was to develop a novel algorithm for removing heartbeat artifact, thus overcoming problems associated with previous methods. Our method consists of a mean artifact wave form subtraction, the selective removal of wavelet coefficients, and a recursive least-square adaptive filtering. The recursive least-square adaptive filtering operates without dedicated sensor for the reference signal, and only when the mean subtraction and wavelet-based noise removal is not satisfactory. The performance of our system has been assessed using simulated data based on experimental data of various spectral characteristics, and actual experimental data of alpha-wave-dominant normal EEG and epileptic EEG Elsevier B.V. All rights reserved. Keywords: EEG; fmri; Artifact removal; Adaptive filtering; Wavelet transform; Teager energy operator 1. Introduction The simultaneous recording of the electroencephalogram (EEG) and functional magnetic resonance imaging (fmri) is a promising tool for functional brain mapping, which provides brain imaging with both high spatial and temporal resolution. Any neuroscientific studies that utilize event-related potential and neuroimaging such as positron emission tomography (PET) or fmri can benefit from this combined technique (Kruggel et al., 2001; Wang et al., 1999). Another advantage of simultaneous EEG/fMRI is that information from the two methods is complimentary. EEG-based fmri studies of epilepsy (Benar et al., 2002; Krakow et al., 2001) and studies on brain activation as a function of sleep stages (Czisch et al., 2002; Lovbald et al., 1999) are two typical examples. Corresponding author. Tel.: ; fax: addresses: khkim@ieee.org (K.H. Kim), hwpark@athena.kaist.ac.kr (H.W. Park). 1 Tel.: ; fax: Limitations in the combined use of these techniques come from artifacts that are induced in the EEG wave forms by the subject s movements in a high magnetic field and magnetic field switching for MR imaging. The switching of gradient magnetic fields induces severe artifacts whose amplitudes are 10 to 100 times larger than that of the EEG. This makes the monitoring of EEG wave forms difficult when MR imaging is being simultaneously performed. The second serious artifact arises from slight movement of a subject s scalp and electrode leads in the high static magnetic field. This is called ballistocardiac pulse artifact or ballistocardiogram (BCG). The amplitude of a ballistocardiac pulse can also be much higher than that of EEG (Allen et al., 1998; Bonmassar et al., 2002; Kruggel et al., 2002; Sijbers et al., 1999), even if the EEG acquisition system is specially designed for the simultaneous EEG/fMRI. Compared to the first problem of magnetic field switching, only a few studies on ballistocardiac artifacts are available in the literature. Bonmassar et al. (2002) utilized adaptive filtering derived from Kalman filter. Their technique necessitates a motion sensor to provide the reference signal for the adaptive filtering. The computational burden may /$ see front matter 2004 Elsevier B.V. All rights reserved. doi: /j.jneumeth

2 194 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) become prohibitively high since the filter coefficients must be adapted for each of the time samples. Allen et al. (1998) described a technique based on the subtraction of the average heartbeat wave form within a predetermined interval, where the inaccuracy in heartbeat detection and the variability of heartbeat wave form may induce serious errors. In this paper, we present a novel signal processing technique that combines the advantages of the two previous techniques, i.e. mean subtraction and adaptive filtering. In addition, wavelet postprocessing is applied to remove residual pulse artifacts after the mean subtraction process. A recursive least-square (RLS) adaptive filtering is enabled only when the heartbeat pulse removal from the mean subtraction followed by wavelet postprocessing is not satisfactory. This selective exploitation of adaptive filtering is particularly advantageous for long-term EEG recording in sleep studies. In addition, no dedicated sensor is necessary for the acquisition of the reference signal in the proposed adaptive filtering. A more efficient yet still fast heartbeat detector using a slight modification of Teager energy operator (TEO) is employed. Performance assessment of the proposed method and comparison with previous methods are presented using simulated data based on experimental data and experimental data. 2. Experimental methods Six normal subjects (three females and three males, age: 21.5 ± 1.38) volunteered to participate in EEG recording within MR scanner. The original purpose of the experiment was to investigate brain activation according to the sleep stage. Before the main session, the subjects were asked to close their eyes and not to move their eyeballs so that alpha-dominant waves contaminated by heartbeat pulse could be recorded. Six subjects were healthy volunteers and one had temporal lobe epilepsy (female, age: 55). All the subjects signed a written consent form before the study began. Normal subjects were recruited by advertisement and selected if they had no history of medical, neurological or psychiatric illness. The temporal lobe epilepsy patient was selected from the outpatient clinic of the neurology department of the Asan medical center, University of Ulsan School of Medicine, Seoul, South Korea. Experiments on the epilepsy patient were performed under the supervision of a neurologist. The EEG recording was performed inside a 3 Tesla MR scanner at the fmri lab in Korea Advanced Institute of Science and Technology. The MR scanner, equipped with 3 Tesla magnet, was manufactured by ISOL technology, Kwangju, South Korea. For the acquisition of EEG within the MR scanner, an fmri-compatible EEG recording system, BrainAmp-MR (BrainProducts GmbH, Munich, Germany) was used along with a specially-designed electrode cap (BrainCap-MR). The electrode cap contains 32 EEG channels and three additional channels dedicated to electrocardiogram (ECG) or electrooculogram (EOG) acquisition. The reference electrode is located at the center between Cz and Fz. All the electrodes are ring-type sintered nonmagnetic Ag/AgCl electrodes. The impedance of each electrode site was maintained to 5k by injecting an electrode paste (ABRALYT 2000, FMS, Herrsching-Breitbrunn, Germany). The lead wires from each EEG electrode were fixed firmly to the cap surface for the immobilization. The EEG signals from the electrode cap were transferred to the amplifier via a nonferrous ribbon cable. The amplifier was designed to be placed inside the magnet bore of the scanner and was connected to the host computer outside the MR room via a fiber optic cable. The resolution and dynamic range of the amplifier were 100 nv and ±3.2 mv, respectively. The EEG and EOG wave forms were recorded with a sampling rate of 500 samples/s. Bandpass filtering from 0.5 to 80 Hz was applied along with 60 Hz notch filtering. 3. Ballistocardiac artifact removal algorithm Fig. 1 shows the block diagram of the proposed heartbeat artifact removal system. Basically, the system consists of two branches; one of which is the mean pulse wave form subtraction followed by the residual noise reduction using the selective elimination of wavelet coefficients, and the other is based on RLS adaptive filtering. Our algorithm operates on 10 s segments of the input signal. The mean subtraction branch that performs the role of heartbeat artifact removal is similar to the method described by Allen et al. (1998), with the important addition of a wavelet-based de-noising. Since we found that two EOG channels of our system provide a BCG-dominant signal that can be used as a reference signal for the heartbeat detection and the adaptive filtering, an additional motion sensor dedicated to BCG acquisition was not used. Each heartbeat is detected from the derived BCG signal (BCG ) obtained by subtracting the left EOG from the right EOG, using a slight modification of Teager energy operator (TEO) (Choi and Kim, 2002; Mukhopadhyay and Ray, 1998). This heartbeat detector, which is referred to as k-teo, is defined as follows: ψ k {x(n)} =x 2 (n) x(n k)x(n + k) The usual TEO corresponds to k-teo with k = 1. When x(n)iscosωn, the output of the usual TEO (k = 1) becomes approximately (Aω) 2 and the output of the k-teo becomes close to (A sin ωk) 2. Because the output of the TEO is proportional to the product of the instantaneous frequency and amplitude, it is useful for detecting the pulsatile wave form in a BCG signal. Using k-teo, the value of k can be selected according to the main frequency of the input signal, ω, thus permitting the output to be maximized at k = π/(2ω). Therefore, k-teo is more advantageous than TEO, in that the frequency sensitivity can be selected. The mean wave form of the heartbeat artifact is obtained by averaging the segmented wave forms centered at the

3 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Fig. 1. (a) Block diagram of the proposed ballistocardiac artifact removal system. (b) Heartbeat detector using k-teo. detected heartbeat time points, as described by Allen et al. (1998). Assuming that the pulse wave form varies only slightly within the time interval under consideration and that the positions of the heartbeats can be precisely detected, the subtraction of this mean wave form from the contaminated EEG can satisfactorily remove heartbeat artifacts. However, this assumption proved to be incorrect and thus, the mean subtraction gave unacceptable results for many experimental EEG recordings within our 3 Tesla MR scanner, and significant parts of the artifacts still remained in the output wave form. We used two approaches, wavelet-based de-noising and the selective use of adaptive filtering, to solve this problem. Wavelet coefficients are selectively eliminated from the output of the mean subtraction in order to reduce residual artifacts. Because the power of the heartbeat pulses (and their residuals) is concentrated on some scales and on particular time points, the residual artifacts can be reduced considerably by the selective elimination of wavelet coefficients at specific scales and time points. The mean-subtracted signal is decomposed into eight dyadic scales by discrete wavelet transform (WT) using the Coiflet-5 basis function (Daubechies, 1992). The wavelet coefficients of scales d1 and d2 are then thresholded by the method of Donoho (1992), to reduce high frequency random noise. Here, dj denotes the detail signal at the j th level of wavelet decomposition, which is given as the output of the successive application of the quadrature mirror filter and dilation by a factor of two (Daubechies, 1992). The wavelet coefficients of the scales d6 and d7 are subsequently eliminated if they are within a 1 s interval centered at the heartbeat peak or if their absolute values are larger than 1/10 of the heartbeat peak amplitude. The denoized wave form is then obtained from the inverse wavelet transform (IWT), as shown in the block diagram of Fig. 1a. Because the wavelet coefficients at a particular scale roughly correspond to the bandpass filtered signal of the original wave form, attention must be paid as to whether the input wave form does not contain important frequency components at the subbands corresponding to those scales. The elimination of these two scales, d6 and d7, reduces the frequency components located in the range of 2 7 Hz. This corresponds to the delta and theta waves of EEG (0.5 8 Hz). In some special cases, for example, slow wave sleep stages (stages 3 and 4) when the delta wave would be expected to be dominant, it is better not to exclude the scale d7 in order not to reduce the underlying delta wave (0.5 4 Hz). Other cases where the exclusion of d6 and d7 is not recommended include the case of theta-wave-like ictal epileptic EEG (3 7 Hz), and experiments aimed at recording event-related potentials (ERPs). In these cases, instead of the mean subtraction with wavelet de-noising, the adaptive filter is turned on and applied for the artifact removal. The switching-on of the adaptive filter is also recommended when the elimination of artifacts by (mean subtraction + wavelet de-noising) is not satisfactory, as explained below. After the selective removal of the wavelet coefficients, the signal is analyzed to determine whether the adaptive filter branch should be applied. This is determined by calculating the power spectrum and the ratio between the power in the Hz subband and the total power. The power spectrum is computed using Welch s method with a 1024-point fast Fourier transform and 256-point Hanning window (Hayes, 1996). The adaptive filter branch is switched on if this ratio is larger than 0.6. In addition, when the mean subtraction branch is insufficient even with the wavelet-based residual noise elimination, the adaptive filtering branch comes into

4 196 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) action. The insufficiency is determined by judging whether the correlation coefficient between the k-teo outputs of the processed wave form and the original input wave form is larger than 0.5. For ERP recording, the adaptive filter should be applied when the overlap of the critical time range of the ERP wave form and that of the heartbeat wave form occurs within the 10 s segment of the signal under analy- sis. The RLS adaptive filter is employed because of its relative advantages of fast convergence compared to the usual least-mean square algorithm (Haykin, 2001), at the expense of computational complexity. The computational burden can be tolerated since the adaptive filter branch is only switched on occasionally. The length of the finite impulse response adaptive filter tap was 80. Fig. 2. EEG wave forms recorded (a) outside and (b) inside the MR scanner. The contaminated wave form in Fig. 2b was recorded just after the recording shown in Fig. 2a. The underlying EEG in Fig. 2b shows characteristics similar to that of Fig. 2a.

5 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Results The wave forms of the contaminated (inside fmri) and clean (outside fmri) EEGs are plotted in Fig. 2. The contaminated wave form in Fig. 2b was recorded within the MR scanner just after the recording shown in Fig. 2a outside the MR room (clean EEG). It is possible to see that the underlying EEG in Fig. 2b shows characteristics that are similar to that of Fig. 2a in both the time and frequency domains. The mean amplitudes and standard deviation (S.D.) of 20 ballistocardiac pulses were calculated for the frontal, parietal, and occipital channels and were found to be ± V for F3/F4, 58.5 ± 12.2 V for P3/P4, and 74.1 ± 16.0 V for O1/O2, respectively. The ratios between the pulse amplitude and background EEG amplitude were 7.04 for F3/F4, 4.45 for P3/P4, and 3.01 for O1/O2, respectively. Fig. 3 shows the effectiveness of the proposed heartbeat pulse detection (Fig. 1b), for the derived BCG signal (BCG = LEOG-REOG) in Fig. 2b. The peak-enhanced wave form of the third panel of Fig. 3 demonstrates the efficacy of the k-teo for the heartbeat detection from BCG. Compared to the output of the correlation detector (second panel), the determination of the heartbeat peaks by thresholding became much easier. The size of the smoothing window was 40, and the value of k was determined empirically to be 10, at a sampling rate of 500 samples/s. We first applied the proposed pulse removal algorithm to the simulated wave form based on actual experimental recordings in order to assess its ability to retrieve the underlying EEG wave form. For the actual experimental data, no ground truth is available, and thus, a comparison of the underlying EEG and the processed wave form is not possible. The simulated signal was generated by adding heartbeat wave forms obtained from the recordings within the MR scanner to alpha-dominant EEG wave forms recorded outside the scanner. The heartbeat wave forms were selected from 100 heartbeat peaks, which were experimentally measured and detected from experimental recordings, by visual inspection with assistance of the k-teo heartbeat detector. Before adding the heartbeat wave forms to the background EEG, a principal component analysis was performed to eliminate high frequency fluctuations of the measured heartbeat wave form. By retaining only the first five principal components, 99.6% of the total variance could be preserved, and high frequency components that appeared to be out of the frequency range of the BCG could be removed. Fig. 4 shows an example of the alpha-dominant simulated wave form, and the output wave forms obtained from the conventional method (mean subtraction) and the proposed method. Fig. 4a is one example where the exact locations of heartbeats are informed to the pulse removal algorithms. A better result was obtained using our algorithm (second and Fig. 3. Effectiveness of the heartbeat pulse detection block shown in Fig. 1b, tested for the EEG wave form in the same recording of Fig. 2b.

6 198 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Fig. 4. Performance of the proposed method for an alpha-dominant synthetic signal. (a) When the exact pulse peak position is available. (b) When the given positions of pulse peaks have considerable error. fourth panel of Fig. 4a), compared to the conventional mean subtraction algorithm (first and third panel of Fig. 4a), in that the underlying alpha rhythm became clearly visible when the proposed algorithm was used. We also tested the performance of our system in case where substantial errors exist in heartbeat peak locations (Fig. 4b). The peak positions were shifted by adding random numbers uniformly distributed within the range [0,(3 S.D. of inter-beat intervals)]. In

7 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Power spectral density (µv/ Hz) Solid line: true EEG Triangle: by mean subtraction Circle: by proposed method Frequency (Hz) Fig. 5. Frequency domain comparison of the signals shown in Fig. 4b, before (solid line) and after the application of the proposed (circle) and conventional (triangle) artifact removal algorithm. reality, an inaccurate detection of the heartbeat location can frequently occur, although we suggest an efficient detection method based on k-teo. In this case, the proposed pulse removal algorithm with additional wavelet de-noising shows a much better performance, compared to conventional mean subtraction. It can be seen that the underlying alpha wave is almost perfectly restored in the fourth panel of Fig. 4b where the pulse removal is performed by our method (second and fourth panel of Fig. 4b), while the previously described mean subtraction algorithm is insufficient (first and third panel of Fig. 4b). The reduction in amplitude in the 200 ms window centered on the position of the peak ranged Table 1 Correlation coefficient between the underlying EEG and the raw signal or output signals from pulse removal algorithms EEG band Raw signal Processed by mean subtraction Alpha (Fig. 4) Beta (Fig. 7a) Theta (Fig. 7b) Delta Processed by proposed method from V to 5.25 V (average of 20 pulses). Fig. 5 shows a frequency domain comparison of the signal before and after application of the artifact removal algorithm. The correlation coefficient (CC) with the underlying EEG increased from to using our algorithm, whereas, when the algorithm of Allen et al. (1998) was used, the correlation coefficient was The changes in CC according to the pulse removal algorithms are summarized in Table 1 for alpha, beta, theta, and delta bands. We next applied our method to EEGs measured from the epilepsy patient within the MR scanner. Here the focus is on facilitating the visual identification of an epileptic spike or spike-and-wave complex. The second panel of Fig. 6 shows the artifact-corrected wave form using our algorithm. Compared to the first panel of Fig. 6 where the mean subtraction was used for the pulse removal, the wave form corrected with our method is more suitable for the identification of the interictal spike and the spike-and-wave complex due to the superior performance in pulse removal. In the first panel of Fig. 6, it is very difficult to identify the epileptic spike-and-wave properly. It is also likely that a number of erroneous identifications of BCG pulses can be incorrectly identified as epileptic spikes Mean subtraction method 0 Amplitude (µv) Black: proposed method Time (samples, sampling rate: 500 samples/s) Fig. 6. Performance of the proposed method for an epileptic EEG measured within the MR scanner.

8 200 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Fig. 7. Performance of the proposed method on (a) beta-dominant, and (b) theta-dominant synthetic wave forms.

9 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Fig. 8. (a) Time domain and (b) frequency domain comparison of the proposed algorithm for actual experimental alpha-dominant EEG recorded inside the MR scanner. Since it is not easy to intentionally generate beta-, theta-, and delta-dominant waves, we created EEGs simulating these waves by combining the bandpass-filtered of white Gaussian signal, as described by Eugene (2000), and heartbeat wave forms as described above. The bandpass filters were eighth order Butterworth filters. Fig. 7a shows the performance of our method on beta-dominant wave and demonstrates its superiority compared to previous methods. The correlation coefficient between the underlying EEG and the output wave form of our algorithm is 0.65, while it is 0.49 for the conventional algorithm. The reduction in the frequency component outside the beta band is also superior. In cases of theta-dominant and delta-dominant waves, the wavelet de-noising adopted in our method has some disadvantages, since it reduces a large amount of the frequency component of the signal under study. However, in practice, a theta or delta-dominant wave does not occur except for an EEG under slow-wave sleep stages or in some epileptic EEGs. Moreover, this problem can be successfully resolved by the criteria of our algorithm based on the power spectrum calculation and using RLS adaptive filter instead. As shown in Fig. 7b, the output wave form of our algorithm was much better than the output of the mean subtraction in the case of a theta-dominant wave. This better performance could also be quantified by the increase in correlation coefficient between the processed wave form and the underlying EEG ( ).

10 202 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Finally, our method was applied to an experimental alpha-dominant EEG recorded inside the MR scanner. The dominance of alpha wave could be controlled to some degree by eye closure, and verified by inspection of the alpha-dominance in the EEG recordings outside the MR scanner. By comparing the artifact-corrected wave form of the section (Fig. 8a) where the alpha-dominance would be expected with that of the alpha-dominant EEG recorded outside the MR scanner, we confirm that the proposed method properly restored the underlying alpha-dominant EEG. This was also confirmed by a frequency domain comparison of the input and output of the proposed algorithm, as shown in Fig. 8b. 5. Discussion and conclusion In this paper, an algorithm for ballistocardiac artifact removal from the EEG recorded within an MR scanner is described. The method is based on a mean wave form subtraction along with wavelet de-noising and the selective utilization of adaptive filtering. Our system also adopted a high-performance heartbeat detector based on a slight modification of the TEO, which has been proved to be useful for many pulsatile wave form detections (Choi and Kim, 2002; Kim and Kim, 2000). The performance of our method was tested on experimental and simulated wave forms with various characteristics that would be expected for EEG signals from normal subjects, and appeared to perform better than the mean subtraction algorithm of Allen et al. (1998). It also showed a better performance in the analysis of an epileptic EEG. Compared to the method of Bonmassar et al. (2002), our method has the advantage that it does not require a separate sensor for the acquisition of the pulse-dominant reference signal. We showed that a useful reference signal for heartbeat artifact removal by means of the RLS adaptive filter can be obtained from two EOG channels. The operation of the adaptive filter is necessary only when the performance of the mean subtraction/wavelet de-noising branch is not satisfactory, so that the required computation time is much shorter than the method reported by Bonmassar et al. (2002) which updates the filter coefficients for all the time samples of the input signal. The heartbeat can vary considerably during the experiment. If it becomes extremely fast (faster than 150 beats/min), heartbeat artifact wave forms become overlapped in the time domain so that the estimation of the template wave form of heartbeat artifact by averaging becomes erroneous. If this happens, we recommend the use of the RLS adaptive filter. The ballistocardiac pulse artifact removal algorithm proposed in this paper utilizes some of the relative advantages of previous methods and overcomes some of their drawbacks, providing a faster and more accurate processing of a large amount of EEG data recorded simutaneously with fmri. Our system processes 10 s segments of input signal at a time. In most cases, the recorded EEG and fmri data are analyzed by post-processing, and thus this batch-type processing is acceptable. In previous studies on the identification of an epileptic focus, functional MR imaging was triggered by the manual identification of EEG spikes (Seeck et al., 1998; Warach et al., 1996). Thus, the near real-time processing of ballistocardiac artifact removal in the EEG was required. However, epilepsy studies can now be performed by continuous fmri image acquisition and postprocessing of the EEG, as suggested by Baudewig et al. (2001) and Benar et al. (2002). Therefore, our method can also be employed in studies of epileptic subjects. As shown by Allen et al. (1998), except for some EEGs with extraordinarily high amplitude (>200 V), heartbeat artifact removal is essential for the analysis of an EEG signal recorded within an MR scanner, in spite of the specialized instrumentation system, even in a 1.5 Tesla scanner. As high-field scanners ( 3 Tesla) are becoming popular for fmri, the heartbeat artifact problem becomes more serious. Therefore, our development of an efficient artifact removal algorithm will become more significant for combined fmri/eeg studies. Acknowledgements This study was supported in part by grant M from the Ministry of Science and Technology, Korea. References Allen PJ, Polizzi G, Krakow K, Fish DR, Lemieux L. Identification of EEG events in the MR scanner: the problem of pulse artifact and a method for its subtraction. Neuroimage 1998;8: Benar CG, Gross DW, Wang Y, Petre V, Pike B, Dubeau F, et al. The bold response to interictal epileptic discharges. Neuroimage 2002;17: Baudewig J, Bittermann HJ, Paulus W, Frahm J. Simultaneous EEG and functional MRI of epileptic activity: a case report. Clin Neurophysiol 2001;112: Bonmassar G, Purdon PL, Jaaskelainen IP, Chiappa K, Solo V, Brown EN, et al. Motion and ballistocardiogram artifact removal for interleaved recording of EEG and Eps during MRI. Neuroimage 2002;16: Choi JH, Kim T. Neural action potential detector using multi-resolution TEO. Electron Lett 2002;38: Czisch M, Wetter TC, Kaufmann C, Pollmacher T, Holsboer F, Auer DP. Altered processing of acoustic stimuli during sleep: reduced auditory activation and visual deactivation detected by a combined fmri/eeg study. Neuroimage 2002;16: Daubechies I. Ten lectures on wavelets. Philadelphia: Society for Industrial and Applied Mathematics, Donoho DL. Denoising via soft thresholding. Technical report 409. Stanford: Department of Statistics, Stanford University, Eugene BN. Biomedical signal processing and signal modeling. New York: Wiley, Hayes MH. Statistical digital signal processing and modeling. New York: Wiley, Haykin S. Adaptive filter theory, 4th ed. Englewood Cliffs: Prentice Hall, 2001.

11 K.H. Kim et al. / Journal of Neuroscience Methods 135 (2004) Krakow K, Messina D, Lemieux L, Duncan JS, Fish DR. Functional MRI activation of individual epileptiform spikes. Neuroimage 2001;13: Kruggel F, Herrmann CS, Wiggins CJ, von Cramon DY. Hemodynamic and electroencephalographic responses to illusory figures: recording of the evoked potentials during functional MRI. Neuroimage 2001;14: Kim KH, Kim SJ. Neural spike sorting under nearly 0 db signal-to-noise ratio using nonlinear energy operator and artificial neural network classifier. IEEE Trans Biomed Eng 2000;47: Kruggel F, Wiggins CJ, Herrmann CS, von Cramon DY. Recording of the event-related potentials during functional MRI at 3.0 tesla field strength. Magn Reson Med 2002;44: Lovbald KO, Thomas R, Jakob PM, Scammell T, Bassetti C, Griswold M, et al. Silent functional magnetic resonance imaging demonstrates focal activation in rapid eye movement sleep. Neurology 1999;53: Mukhopadhyay S, Ray GC. A new interpretation of nonlinear energy operator and its efficacy in spike detection. IEEE Trans Biomed Eng 1998;45: Seeck M, Lazeyras F, Michel CM, Blanke O, Gericke CA, Ives J, et al. Non-invasive epileptic focus localization using EEG-triggered functional MRI and electroencephalographic tomography. Electroenceph Clin Neurophysiol 1998;106: Sijbers J, Michiels I, Verhoye M, van Audekerke J, van der Linden A, van Dyck D. Restoration of MR-induced artifacts in simultaneously recorded MR/EEG data. Magn Reson Imaging 1999;17: Wang J, Zhou T, Qiu M, Du A, Cai K, Wang Z, et al. Relationship between ventral stream for object vision and dorsal stream for spatial vision: An fmri + ERP study. Hum Brain Mapp 1999;8: Warach S, Ives JR, Schlaug G, Patel MR, Darby DG, Thangaraj V, et al. EEG-triggered echo-planar functional MRI in epilepsy. Neurology 1996;47:89 93.