By Topic

IEEE Quick Preview
  • Abstract



INTEREST in high-frequency electroencephalographic (EEG) rhythms (above 20 Hz) has intensified since the recognition of the involvement of high-frequency beta (15–25 Hz) and gamma (>25 Hz) rhythms in cognitive processing [1], [2], [3]. It is well known that electrical activity from muscles [electromyogram (EMG)] of the cranium and neck contaminates scalp recordings, and the spectrum of frequencies in the EMG overlaps the spectrum of the EEG particularly above 20 Hz [4], [5], and we have quantified the extent of the contamination [6], [7], [8].

Different referencing methods do not significantly attenuate but only marginally affect the distribution of EMG due to active muscle contraction [4]. Attempts have been made to identify EMG-free EEG components using blind source separation techniques, none of which have yet been demonstrably reliable [9], [10], [11].

Surface Laplacian transforms determine scalp current density using data from all active scalp electrodes [12]. As such, the method is sensitive to local sources and sinks located near to recording sites but insensitive to distant sources. There are no muscles underlying the central scalp; therefore, raw recordings from central channels should only be contaminated by potentials from distant muscles. These distant effects should be significantly attenuated in the surface Laplacian transforms. Here we report that, with certain restrictions, scalp surface Laplacian transforms substantially reduce muscle contamination in central scalp electrical recordings.



We used power spectra of electrical scalp recordings obtained from volunteers before and after the elimination of EMG activity by neuromuscular paralysis. We examined a variety of preprocessing methods to quantitate the amount of EMG contamination. In particular, we searched for any method that might sufficiently reduce EMG contamination in preparalysis recordings such that their power spectra matched the power spectra of recordings after paralysis (EMG-free).

EMG-free scalp recordings were obtained from volunteers who were pharmacologically paralyzed and artificially respired while fully conscious. The procedure has been described in detail previously [6], [7] and was approved by the Clinical Research Ethics Committee of the Flinders University and Flinders Medical Centre.

Five males (aged 28, 40, 48, 62, and 73 years) and one female (aged 47 years) participated in the experiment. They were studied in a Faraday cage, seated in a reclined position, and required no change in posture after paralysis. They were recorded, first without paralysis, while undertaking a range of behavioral tasks. Cisatracurium (a curare-like agent) was then administered to achieve full neuromuscular paralysis and intermittent positive pressure ventilation was initiated. When a complete neuromuscular blockade was achieved, recordings continued and the subjects repeated the tasks.

A. EEG Recording

EEG (115 channels) was recorded continuously (left ear reference, 5000 samples/s, 16-bit analog-to-digital conversion, 1250 Hz low-pass filter) using a NeuroScan system (Compumedics, Vic., Australia). Electrode impedances were less than 5 k Formula$\Omega$. A 128-channel electrode cap with Ag–AgCl electrodes (Easy Cap Falk Minnow, Germany) was used to provide uniform scalp coverage.

B. Behavioral Tasks

All tasks were performed in both the unparalyzed and paralyzed states. Tasks included resting with the eyes closed (eyes closed), left eye held open by a swab on a stick (left eye open), submaximal jaw clenching (bite), frowning (frown), and strobe light stimulation to the held-open left eye at 40 Hz (strobe), among several other tasks not reported in this paper. Resting, baseline tasks were recorded for 10 s with both eyes closed and then for 10 s with the left eye held open and were recorded on three occasions in the unparalyzed state (twice at the beginning and once at the end) and on two occasions in the paralyzed state (at the beginning and the end). Bite, frown, and strobe were undertaken for 10 s.

C. EEG Analysis

Recordings were edited by removing data from technically defective electrodes. All EEG data were retained except for those contaminated by electrooculographic (EOG) activity. Analysis was undertaken using in-house programs written in MATLAB (The Mathworks, Natick, MA). Data using five preprocessing methods were compared namely, left ear reference (L-ear—the original recording), common average reference (CAR), and surface Laplacian transforms [12], [13]. Three forms of surface Laplacian transforms were tested: a spline interpolation [14], Hjorth [15], and finite element [13]. Code for all estimators was obtained from the FieldTrip toolbox [16]. Power spectra were estimated using the Welchs modified periodogram method, with blocks of 5000 samples (1 s). Contamination by 50 Hz (mains frequency signal) was removed by excision of 49–51 Hz power and its harmonics.

We estimated relative power spectra, obtained as the ratio of power spectra for the unparalyzed condition (power due to brain activity plus power due to muscle activity plus power due to noise, B+M+N) over the power in the paralyzed condition (B+N).

To measure the quality of the output of a preprocessing method, we also estimated B:M ratios (viz. an SNR). The power due to brain activity B was estimated as the power spectrum during paralyzed recordings. The power due to muscle activity M was estimated as the difference in power spectra between unparalyzed and paralyzed recordings. When M is small, noise significantly affects this estimate, and so two steps were taken to reduce its effect. First, all calculations were averaged within frequency bands. Second, we estimated the noise in scalp recordings as the difference in power spectra between two samples of baseline paralysis recordings (measured at the beginning of paralysis and at the conclusion of the mental tasks approximately 20 min later). Neither of these samples contained EMG. Estimates of muscle M less than the 95 percentile of the noise were regarded as unreliable and replaced by the threshold value. This conservative estimate of muscle activity was then used to calculate B:M ratios for a frequency band.

D. Statistical Analysis

The statistical significance of changes in power due to paralysis was evaluated using paired Formula$t$-tests if normal, and Wilcoxon signed rank test if not normal. The Lilliefors test was use to determine normality.



A. Resting Tasks

Paralysis reduced spectral power compared to unparalyzed (due to muscle) over a broad range of frequencies during all tasks. During nonactive tasks such as eyes closed and left eye open, the lower limit of power due to muscle varied in different regions of the scalp and with different methods of preprocessing. EEG traces, power spectra, and group-mean relative spectra, unparalyzed:paralyzed for the eyes-closed task, using the five methods are illustrated in Figs.1 2 3. Relative to paralysis, EMG contamination was greatest in circumferential scalp areas (frontotemporal, temporal, and temporooccipital) and least centrally, consistent with our previous reports in smaller numbers of subjects [6], [7].

Figure 1
Fig. 1. (Left) Illustrative EEG at Cz from a single-subject for left ear (L-EAR), CAR, and nearest neighbor (Hjorth), spline, and finite-element surface Laplacian transforms. EEG was recorded during the eyes-closed task in unparalyzed (blue) and paralyzed (red) states. (Right) Power spectra for each segment.
Figure 2
Fig. 2. Illustrative mean (Formula$n$ = 6) relative power spectra of electrical scalp recordings for the five electrodes depicted in the scalp map (top right). The relative power is of eyes-closed unparalyzed:paralyzed conditions, calculated using five methods. The detail of the central scalp electrode (Cz) is illustrated in Fig. 3. Left=left; anterior=up.
Figure 3
Fig. 3. Detail of central scalp electrode (Cz) from Fig. 2. Relative power (before versus after paralysis) for CAR (purple) and surface Laplacian transforms (cyan, red, and blue). The surface Laplacian relative power centrally is close to a ratio of 1.

Centrally, where contamination was least, both left-ear reference and CAR- EEG were still significantly contaminated by EMG (Formula$P$-values all <0.05 at Cz; see Table I). In contrast, surface Laplacian transforms at central and paracentral leads revealed small and nonsignificant contamination by EMG (Formula$P$-values nearly all > 0.05 at Cz; see Table I). The spline surface Laplacian was the only method for which residual muscle contamination was not significant at any frequency range.

Table 1

Topographies of relative power for low gamma and higher frequency bands (selected according to [17]) demonstrate that there is very low central scalp muscle contamination using surface Laplacian tranforms (see Fig. 4). These contrast to the considerable EMG contamination of EEG using left ear referencing or, to a lesser extent, using CAR. B:M ratios for these frequency ranges are given in Table II.

Figure 4
Fig. 4. Average topographies (Formula$n$ = 6) of relative power for left ear, CAR, and nearest neighbor (Hjorth), spline, finite-element surface Laplacian transforms. The Laplacian methods provide an extensive central and paracentral area of signals barely contaminated by muscle.
Table 2

B. Muscle Activation

The impact of active contraction, with frowning and jaw clenching, as opposed to resting muscle tension, is illustrated in Fig. 5. Using CAR, EMG contaminated all electrodes albeit less at the central scalp. In contrast, spline surface Laplacian transforms exhibited a very small, albeit statistically significant, increase in relative power with ratios very close to 1. During bite, for Cz in the range 30–80 Hz, B:M ratios were 0.87 for CAR, and 8.1 for spline surface Laplacian. During frown, for Cz in the range 30–80 Hz, B:M ratios were 0.61 for CAR, and 9.1 for spline surface Laplacian.

C. Physiological Responses

We evaluated the application of spline surface Laplacian in identifying brain responses in unparalyzed subjects, first using a low-frequency response, alpha augmentation with eye closure, and then using a high-frequency response to sustained 40 Hz visual stimulation with strobe.

Figure 5
Fig. 5. Central (Cz) electrode mean (Formula$n$ = 6) relative power spectra of electrical scalp recordings illustrate the effect of strong muscle contraction on power in unparalyzed versus paralyzed conditions, calculated using CAR and spline surface Laplacian transforms. Topographies display relative power averaged for 30–80 Hz for both methods, and illustrate the improved performance of spline surface Laplacian transforms in minimizing muscle contamination, even in the presence of strong muscle contraction.

Examining the low-frequency (alpha) response to the eyes-closed versus the left eye-open state, the finding with CAR and with spline surface Laplacian was similar. In the four subjects with good alpha activity, spline surface Laplacian revealed the expected posterior enhancement of alpha activity [17]. Although qualitatively similar, peak alpha frequencies were different, as were distributions, so that the group average was not informative and a single representative subject is shown in Fig. 6. In addition, spline surface Laplacian did not project spurious changes to more anterior sites that occurred with CAR (see Fig. 6).

Figure 6
Fig. 6. Relative spectra of eyes closed versus left eye open (unparalysed), calculated with CAR and with spline surface Laplacian transforms, from electrodes on the left (PO7) and right hemisphere (PO8) posteriorly from one subject. Topographies display relative power for 9–10 Hz for the two methods. Both the spectra and the topographies reveal the expected increase in 9–10 Hz power due to eye-closure. The topographies also illustrate the improved localization of a source expected with spline surface Laplacian transformation.

Examining the high-frequency (gamma) response during 40 Hz strobe visual stimulation, spline surface Laplacian and CAR both revealed the expected posterior expression of 40 Hz steady-state visual evoked response, while the Laplacian did not project spurious 40 Hz changes to more anterior sites seen with CAR (see Fig. 7).

Figure 7
Fig. 7. Mean (Formula$n$ = 5) spectra from a right hemisphere posterior electrode (number 125, between POz and O2 in the extended 10–20 EEG system) illustrating the maximal relative increase in 40 Hz power due to strobe (unparalyzed) for CAR and spline surface Laplacian methods. Topographies display relative power at 40 Hz for the two methods. Both methods reveal the 40 Hz steady-state responses due to strobe. The spline surface Laplacian transforms also improve the localization of the 40 Hz steady-state response.


By comparing scalp electrical recordings from normal (unparalyzed) and paralyzed awake volunteers undertaking various mental and motor tasks, we have demonstrated that spline surface Laplacian transform of signals from central scalp positions has very high B:M ratios over all frequencies at which cerebral rhythms have been considered to be likely [18]. Particularly important is our finding that spline surface Laplacian transform is insensitive to EMG contamination centrally in the presence elsewhere of marked EMG-contaminated data (see all figures). Blink artifact was the only biological contaminant actively excluded in our study. Sustained nonforceful muscle contractions due to muscle tone in resting subjects and nonintended transient movements were not excluded. Importantly, however, we also tested powerful contractions of jaw (temporalis) and forehead (orbicularis oculi) muscles. Activity in these muscles caused very slight, albeit significant, EMG contamination in the central scalp, an observation that provides, for the first time, evidence that useful data pertaining to all frequencies can be recorded in the face of submaximal scalp muscle contraction.

Concurrently with its resistance to EMG contamination in the central scalp region, application of the method, importantly, did not eliminate changes in power of cerebral origin, around 10 Hz posteriorly due to eye closure, and the 40 Hz visual steady-state response posteriorly due to 40 Hz visual stimulation. The focal expression of alpha with spline surface Laplacian compared to the broad expression with CAR is a good example of the spatial filtering properties of the Laplacian in which broad spatial activity (low spatial frequency) is attenuated [12].

The surface Laplacian approach transforms voltage EEG into a surface current density (SCD) space. Thus, the relationships between electrodes will change in both magnitude and phase and it cannot simply be assumed that standard EEG analysis will be suitable for SCD. While there is evidence to suggest that surface Laplacian will benefit estimates of synchrony [19], [20], caution should be exercised to ensure that analyses, particularly phase analyses (e.g. phase coherence, phase-locking value, etc.), are appropriate for SCD and that the limitations of each approach are accommodated when interpreting results [20].

Our results, that spline surface Laplacian transforms provide the best preprocessing approach to avoiding muscle contamination of scalp signals, are consistent with the fact that there are no muscles underlying the central scalp so that central channels are only subject to the effects of potentials in distant muscles. As distant potentials affect closely located electrodes, similarly they will largely be eliminated in the surface Laplacian transforms. In the earlier study of Goncharova and colleagues [4], next-nearest neighbor Laplacian transforms of 15% contraction of cranial muscles versus rest performed worse than other referencing methods. Based on our findings, spline and, to a lesser extent, Hjorth Laplacians do appear to be more successful at identifying local current. Consistent with Goncharova and colleagues, however, we found high levels of EMG contamination with surface Laplacian transforms in peripheral locations (see Figs. 1 5).

The results do not challenge the clinical use of EEG in the temporal domain in which, apart from epilepsy, rhythms with frequencies below 20 Hz are usually assessed. However, in the frequency domain, using left ear or common average referencing, it is clear that EMG contamination is extensive at frequencies close to and above those normally of clinical interest [4], [5], [6]. This is particularly problematic for investigators examining higher frequency activity associated with cognitive processes and cortical binding [1], [2], [3]. The surface Laplacian transform appears to provide a window in the central scalp region for accessing high-frequency cortical activity without significant muscle contamination. Based on CAR data, EMG contamination is apparent even when careful visual editing of scalp recordings has been attempted [6], [8]. In contrast, editing to excise EMG contamination appears to be unnecessary when using spline surface Laplacian transforms for the study of central scalp recordings.

It is widely accepted that new approaches are necessary for extracting EEG from mixed EEG/EMG signals [9], [10], [11]. This view still remains relevant to channels circumferential to the central scalp. Evaluation of methods such as independent components analysis, principal components analysis, and others requires an EMG-free record to allow reliable evaluation. Until now it has not been clear that, in tasks not requiring forceful muscle contractions, central electrode data largely avoid muscle contamination when estimated with spline surface Laplacian transforms. Given this, the spline surface Laplacian approach provides an ideal source of signals for investigators to use in developing and testing algorithms to separate signals from brain and signals from muscle. An additional benefit of this observation is that a robust evaluation of possible relationships between neuropsychiatric disorders and high-frequency brain signals is feasible for the first time, at least in the central scalp window.


This work was supported in part by the Australian Research Council, in part by the National Health and Medical Research Council, in part by the Epilepsy Foundation of South Australia, in part by the Flinders Medical Centre Foundation, and in part by the Clinicians Special Purpose Fund of the Flinders Medical Centre. The work of E. W. Whitham was supported by a Pfizer Neuroscience Research Fellowship and an Epilepsy Society of Australia Fellowship from the Union Chimique Belge (UCB) Pharma. The EEG Research Unit was supported by an equipment grant from the Wellcome Trust, London, U.K. Asterisk indicates corresponding author.

S. P. Fitzgibbon is with the School of Computer Science, Engineering and Mathematics, and School of Medicine, Flinders University, Adelaide, S.A. 5001, Australia (e-mail:

T. W. Lewis, D. M. W. Powers, and K. J. Pope are with the School of Computer Science, Engineering and Mathematics, Flinders University, Adelaide, S.A. 5001, Australia (e-mail:;;

E. W. Whitham is with the Department of Neurology, Flinders University and Medical Centre, Adelaide, S.A. 5001, Australia (e-mail:

J. O. Willoughby is with the Department of Neurology and Centre for Neuroscience, Flinders University and Medical Centre, Adelaide, S.A. 5001, Australia (e-mail:


No Data Available


No Photo Available

Sean P. Fitzgibbon

Authors’ photographs and biographies not available at the time of publication.

No Photo Available

Trent W. Lewis

No Bio Available
No Photo Available

David M. W. Powers

No Bio Available
No Photo Available

Emma W. Whitham

No Bio Available
No Photo Available

John O. Willoughby

No Bio Available
No Photo Available

Kenneth J. Pope

No Bio Available

Cited By

No Data Available





No Data Available
This paper appears in:
No Data Available
Issue Date:
No Data Available
On page(s):
No Data Available
INSPEC Accession Number:
Digital Object Identifier:
Date of Current Version:
No Data Available
Date of Original Publication:
No Data Available

Text Size