Introduction

Recording electric or magnetic fields from the scalp using electro- or magneto-encephalography (EEG/MEG) is one of the primary ways in which human brain activity can be characterised. However, the signals measured at the scalp are a mixture of the contributions from multiple generators or sources, added to background activity and system noise, meaning that it is often difficult to identify and study the dynamic activity of generators of interest at the level of the electrode/sensor space. Although the most common method of overcoming this limitation is time-domain averaging with or without source localisation (^{Key et al., 2005, Michel et al., 2004}), source separation algorithms are becoming increasingly widely accepted as a way of extracting and investigating the different neuronal sources that contribute to the measured scalp signal. The advantage of blind or semi-blind source separation techniques compared to more restricted analyses such as time-domain averaging lies in their ability to explore the data and extract dependencies that were not expected a priori. They also provide a means of investigating non-phase locked oscillatory processes and improving the signal-to-noise ratio (SNR) of the sources of interest by allowing the extraction of a specific part of the signal.

Blind Source Separation (BSS) algorithms make some assumptions about the statistical properties of the sources contributing to the measured signal, but within the constraints of those assumptions are able to unmix or decompose the scalp signal into a number of underlying sources. The most common source separation technique in terms of EEG and MEG is Independent Component Analysis (ICA, ^{Comon, 1994; Hyvärinen, 1999, Hyvärinen et al., 2001}), which contains no prior information regarding the signals. Recently, a new semi Blind Source Separation (s-BSS) method has been developed and has shown some promising results. Functional Source Separation (FSS, ^{Barbati et al., 2006; Tecchio et al., 2007}) is an extension of ICA that incorporates prior information into the decomposition, resulting in a single component which maximises a temporal or spectral constraint of interest.

The advantages of source separation techniques are most evident when dealing with trial-by-trial variations of electrophysiological signals (^{Makeig et al., 2002, 2004a,b; Porcaro et al., 2010}), or with other low amplitude and noisy aspects of the scalp signals such as oscillations in the gamma band (~ 30–90 Hz, ^{Hadjipapas et al., 2007; Barbati et al., 2008, Muthukumaraswamy et al., 2010}). Gamma Band Activity (GBA) is increasingly widely studied because of its links with processes as disparate as the binding of cortical regions and stimulus features, attention, change detection, maintenance of working memory and the processing of sensory stimuli (^{Gray et al., 1989; Fries et al., 2001; Pesaran et al., 2002; Womelsdorf et al., 2006}; see ^{Tallon-Baudry, 2003, Ward, 2003 and Fries et al., 2007} for reviews). Activity around the gamma range has also been linked with early visual processing (^{Tzelepi et al., 2000; Sannita et al., 2001; Hadjipapas et al., 2007; Perfetti et al., 2007}), although the relationship between this activity and the low frequency Visual Evoked Potential (VEP) remains unclear, with most studies concentrating on one aspect or the other (although see ^{Muthukumaraswamy et al., 2010}). Given the prevalence of these different responses to visual stimuli, and the amount of work that has been done to characterise their properties, it is perhaps surprising that more is not known about how the two are related. For example, are they generated by the same neuronal pools, or do they represent the stimulus responses of independent but spatially co-localised populations? Is there a link between the VEP/P100 and the GBA?

These questions have a wider significance when considering the combination of EEG with functional magnetic resonance imaging (fMRI). Although a number of lines of evidence from animal and human studies suggest that the tightest coupling between EEG and BOLD occurs in the gamma band (see, for example, ^{Logothetis et al., 2001, Niessing et al., 2005, Muthukumaraswamy and Singh, 2008, Zaehle et al., 2009}), most combined EEG-fMRI studies in humans are interested in and restricted to the lower frequencies of evoked potentials (see ^{Debener et al., 2006} for a review). In general, this has also proved fruitful, although even when restricting the analysis to low frequencies the goal of investigating co-dependencies between EEG and fMRI is complicated since many potential features can be investigated (^{Ostwald et al., 2010}). For example, ^{Porcaro et al. (2010)} observed a significant correlation between the mean area of the VEP and that of the BOLD response, but not when considering their respective amplitudes. What extra neurophysiological significance was carried by the area of the VEP was not clear. However, it is evident that if EEG-fMRI is to achieve its potential as a neuroimaging methodology, the nature of the link between the two signals must be more easily interpretable at all frequencies.

In order to address these issues, and to investigate the ability of different methodologies to extract the sources of interest, blind (ICA) and semi-Blind (FSS) source separation methods were applied to dense-array EEG data recorded during checkerboard stimulation. FSS was performed with both temporal (^{Porcaro et al., 2008, 2009, 2010}) and spectral (^{Porcaro et al., 2008; Barbati et al., 2008}) constraints to identify specifically the generators of the main peak of the VEP/P100 and of the GBA. Source localisation and time-frequency analyses, single trial and on average data, were then used to investigate the properties and co-dependencies of the sources identified with the different methods to build up a picture of the link between VEP/P100 and GBA.

Materials and methods

Results

Discussion

This study examined the ability of blind and semi-blind source separation techniques to extract sources specifically related to the Visual Evoked Potential (VEP/P100) and Gamma Band Activity (GBA) elicited by a reversing checkerboard. The relationship between the low frequency VEP and the high frequency (~ 30–90 Hz) GBA, both of which can be generated by simple visual stimuli such as checkerboards and gratings, remains unclear. GBA is receiving an increasing amount of attention because of its purported links with many different brain processes (^{Gray et al., 1989; Fries et al., 2001; Pesaran et al., 2002; Womelsdorf et al. 2006}; see ^{Tallon-Baudry, 2003, Ward, 2003 and Fries et al., 2007} for reviews). Of particular relevance to this study, high frequency EEG activity has been linked with early visual processing (^{Tzelepi et al., 2000; Sannita et al., 2001; Hadjipapas et al., 2007; Perfetti et al., 2007}), a domain studied more commonly using the VEP (^{Di Russo et al., 2001, 2005}). Despite the considerable amount of work characterising the behaviour of both the VEP and the GBA to manipulations of stimulus properties, very little attention has been paid to the relationship between the two. For example, it is not known whether they are different manifestations of the activity of a single neuronal pool, or whether they represent the stimulus responses of independent but spatially co-localised populations. A better understanding of this relationship is important to provide a more holistic approach to the analysis of EEG and MEG data, rather than a strict parcellation of different parts of the signal. It is also crucial for a proper interpretation of the results of combined EEG-fMRI recordings, which to date have focussed almost exclusively on low frequency EEG and evoked potentials, despite the fact that it has been suggested that the primary relationship between EEG and the Blood Oxygenation Level Dependent (BOLD) effect measured with fMRI is in the gamma band.

GBA has a very low signal-to-noise ratio (SNR), meaning that methods to improve its detectability and allow a more accurate characterisation of its properties are desirable. Three methods were examined in this study: Independent Component Analysis (ICA) and Functional Source Separation (FSS) with temporal and spectral constraints. ICA is a blind source separation technique that extracts sources based purely on statistical grounds, whilst FSS can be seen as an extension to ICA which incorporates some additional information to bias the decomposition algorithm towards solutions that satisfy physiological assumptions. In this study, a temporal constraint maximised the activity around the P100 of the VEP, and a spectral constraint maximised the difference in the GBA between the rest and the task periods. The purpose of applying these different analysis methods was threefold. Firstly, to determine which technique was able to provide the best characterisation of the GBA. Secondly, to provide a degree of validation to the comparison of the GBA and the VEP. Each of the techniques extracts a different part of the raw signal which is dependent on the assumptions underlying the decomposition. In particular, the ICA and the

EEGFSγ

sources were explicitly intended to identify activity in the gamma band, whilst the

EEGFSP100

source employed a temporal constraint designed to maximise the VEP. If similar conclusions regarding the relationship between the VEP and GBA are drawn from examination of these different sources then some confidence can be gained that they represent a realistic interpretation of the underlying data. Finally, the use of FSS with multiple constraints allows the relationship between the constraints to be studied directly. In this case, the temporal constraint centred on the P100 of the VEP selectively identified the most probable generator of that peak. The question can then be asked as to whether there is any evidence that this source also generates GBA. Conversely, a completely orthogonal spectral constraint centred on the gamma band was used to select the generator of the GBA, and the low frequency behaviour of that source examined. The ability of FSS to selectively identify specific sources has previously been demonstrated (^{Barbati et al., 2006, 2008; Porcaro et al., 2008, 2009, 2010; Tecchio et al., 2007, 2008; Betti et al., 2009}) and, especially given the very different constraints employed in the current study, there is no reason to suspect that the two extracted sources should artefactually be related. That is, only if there is a genuine relationship between the VEP and the GBA will the activity of the two sources be similar. The results from this study provide clear evidence that the neuronal pools generating the VEP and GBA have a close spatial and statistical relationship. Localisation of the blind (ICA) and the semi blind method (FSS with both temporal and spectral constraint) was very consistent, with equivalent current dipoles (ECD) demonstrating very precise spatial co-localisation (Fig. 1 third column). This level of overlap supports the idea that the VEP and GBA are generated by spatially concordant neuronal populations. In particular, the comparison of the waveforms, localisation and time-frequency dynamics for the FSS with temporal and spatial constraint revealed a high degree of symmetry: maximising the signal extraction for the low frequency VEP lead to a source containing strong GBA, whilst maximising for the GBA resulted in a source with a clear VEP.

In order to characterise the GBA in more detail, and specifically to determine whether it was phase locked to stimulus presentation (i.e. evoked or induced), the event-related spectral perturbation (ERSP) was calculated in two ways (Fig. 3). Firstly, ERSP was calculated for each single trial, after which the spectra were averaged, and secondly the data were averaged in the time domain before performing spectral analysis. The second procedure maintains only the evoked (stimulus phase-locked) part of the GBR, whilst the first procedure will include both evoked and induced contributions. The results indicate a clear induced GBA for the semi blind methods with the spectral constraint, which is less apparent but also present in the other two data sets (Fig. 3 third column), in addition to an evoked component. This behaviour is consistent with previous work, particularly with grating stimuli (^{Hall et al., 2005; Muthukumaraswamy et al., 2010; Hadjipapas et al., 2007; Barbati et al., 2008}).

A concern when using source separation techniques is whether in selecting a few ICs or only one FS some of the signal of interest has been lost. This was addressed in Fig. 2, which summarises the discrepancy (i.e. the difference between the raw data and the FSS or ICA data). According to this metric, the residual signal not extracted by FSS or ICA is comparable to the level of the baseline, indicating that most of the signal of interest was kept. In the same figure, it is demonstrated that the discrepancy results from

EEGICk

and

EEGFSP100

are comparable, whilst the discrepancy of the

EEGFSγ

is lower than the other two methods. It is of note that although the constraint used in the

EEGFSγ

decomposition was specifically targeted towards the gamma activity, it appears from the discrepancy results that all of the peaks of the VEP were also extracted into the

EEGFSγ

method. This is further evidence that the VEP and the GBA are generated by similar neuronal pools.

These analyses point to a robust relationship between the VEP and the GBA on average. To investigate this further, and to examine other frequency bands to determine whether the VEP is solely related to the gamma band, the relationship was examined on a single trial basis. Trials were ordered based on properties of the VEP (normalised area of the VEP or normalised area of the P100 peak), and the frequency content examined. If there is a link between the VEP and the frequency content of the EEG, there should be a difference between the frequency content of the highest and lowest quartiles of the ordered data. Figs. 5, 6 and 7 show that this is indeed the case. Specifically, there is a clear increase in alpha, beta and low/high gamma power for the top quartile of trials when ordered by VEP/P100 area. However, when the average evoked response was regressed out of the data, this effect was maintained only for the data ordered according to VEP area and not for data ordered by P100 amplitude. In the latter case, the effect of ordering was drastically reduced especially for the beta and gamma bands, whilst having much less of an effect on the alpha band. Taken as a whole, these results suggest that there is a robust relationship between the induced beta and low/high gamma activity and the VEP, indicating that this effect is not a result of contamination of the time-frequency analysis by fast components of the evoked potential.

Although it is very difficult to guarantee that all of the evoked component has been removed by the regression process, and as stated previously it explicitly assumes the standard model of evoked potential generation which is a matter of considerable debate (^{Min et al., 2007; Hanslmayr et al., 2007, Becker et al., 2008, Mazaheri and Jensen, 2006}), several lines of evidence suggest that the majority of the evoked component has been taken out of the data (see Figure S1). The evoked model of VEP generation has been shown to be appropriate for early components of the VEP (< 175 ms, ^{Becker et al. 2008}), and the fact that the regressed data contains very little VEP in the single trial plots of Figs. 5 and 6 would support this. As is evident from Figs. 5 and 6, there is very little latency or shape variability in these early peaks of the VEP, which should at least give the best chance of removing the evoked component successfully (see also Figure S1).

This result may also have implications within the context of EEG-fMRI recordings, and in particular regarding a previous observation that the area of the VEP was correlated with the area of the BOLD response, but that the amplitudes were not related (^{Porcaro et al., 2010}). Given the wealth of evidence that has been accumulated suggesting that the BOLD response is most strongly correlated with gamma band EEG (^{Logothetis et al., 2001, Niessing et al., 2005, Muthukumaraswamy and Singh, 2008, Zaehle et al., 2009}), it is perhaps surprising that EEG-fMRI studies of evoked potentials have had such success. One potential explanation suggested by the results presented in this study is that the evoked potential acts as a surrogate marker for induced gamma activity. At the single trial level, variability in the area of the evoked potential was correlated with variability in the GBA in the time window 0–200 ms (Figs. 5–7), perhaps representing a mechanism whereby evoked potentials could link with BOLD. The relationship between scalp EEG and BOLD remains to be clarified, and this putative intervening link remains speculative and open to further investigation, and is no doubt complicated by the presence of a more prolonged gamma band response throughout the stimulation time period of five seconds (see Fig. 4, FSS Spectral Constraint in the low gamma band). For example, it would be important to determine whether the observations that have been made in primary visual cortex extend to other cortical regions, although the ideal scenario of examining these issues with concurrent EEG-fMRI is complicated by the current difficulty in cleanly recording high frequency activity in the MRI scanner (^{Mullinger et al., 2008}).

In this study, the improvement in data quality provided by ICA and FSS facilitated a more detailed investigation of the spatiotemporal properties of the pattern-reversal EEG response than is generally possible. GBA was not clearly seen in the Raw data (Fig. 3 first row), perhaps explaining why the gamma band response to checkerboard stimuli has not been widely investigated. This improvement in the extraction and characterisation of particular signal properties allowed the relationship between the GBA and the low-frequency VEP to be investigated, the results of which suggested common behaviour and potentially common generators of these two well-studied phenomena. Future work could investigate the effects of stimulus properties on this link, for example static grating stimuli which are more effective at generating GBA (^{Hadjipapas et al., 2007}) as well as higher order cognitive processes such as attention (^{Womelsdorf and Fries, 2007}). More generally, the methodology proposed here suggests a way in which the links between the two main methods of analysing EEG and MEG data could be explicitly investigated to provide a much more complete view of the brain's response to external stimuli.

Supplementary materials related to this article can be found online at doi:10.1016/j.neuroimage.2011.03.008.

The following are the supplementary materials related to this article.

Acknowledgments

This work was supported by grant numbers EP/F023057/1 from the UK Engineering and Physical Sciences Research Council (EPSRC) and BBE0174361 from the Biotechnology and Biological Sciences Research Council (BBSRC).

The WTCN is supported by a strategic award from the Wellcome Trust. The authors would like to thank Dr. Franca Tecchio for her valuable input regarding the interpretation of the results described in this manuscript.