We examined fluctuations in band-limited power (BLP) of local field potential (LFP) signals recorded from multiple electrodes in visual cortex of the monkey during different behavioral states. We asked whether such signals demonstrated coherent fluctuations over time-scales of seconds and minutes, and would thus serve as good candidates for direct comparison with data obtained from functional magnetic resonance imaging (fMRI). We obtained the following results. (i) The BLP of the local field displayed fluctuations at many time-scales, with particularly large amplitude at very low frequencies (<0.1 Hz). (ii) These fluctuations exhibited high coherence between electrode pairs, particularly for BLP signals derived from the gamma (g) frequency range. (iii) Coherence in the BLP, unlike that in the raw LFP, did not fall off sharply as a function of cortical distance. (iv) The structure and coherence of BLP changes were highly similar under distinctly different behavioral states. These results demonstrate the existence of widespread coherent activity fluctuations in the brain of the awake monkey over very long time-scales. We propose that such signals may make a significant contribution to the high variability observed in the time course of physiological signals, including those measured with functional imaging techniques. The results are discussed in the context of combined fMRI/electrophysiological recordings.
Accurate interpretation of neurophysiological signals requires that they be taken in the context of existing activity patterns in the brain. While this tenet is clearly important for the scientist attempting to decipher highly variable neural measurements, it is also of fundamental importance to brain function — the brain must be able to understand its own variability. A number of recent studies have underscored the practical importance of this theoretical consideration for the processing of sensory patterns, demonstrating that the responses of neurons are predetermined by existing brain-states rather than simply by an externally applied stimulus (Arieli et al., 1995, 1996; Azouz and Gray, 1999; Steriade, 2001a). To reliably interact with their environment, brains must somehow disentangle stimulus-related information from that reflecting changes in its internal milieu; and they evidently do so with great success. How this is done remains poorly understood.
While it is clear that this process must draw upon a careful balance between local (intra-area) and global (inter-area) processing, it is gradually becoming apparent that sensory information does not flow unidirectionally through a chain of discrete processing stages. Response latency measurements, for example, suggest that visual stimuli are processed by a highly complex, parallel and recurrent cortical network (Bullier and Nowak, 1995; Schmolesky et al., 1998; Schroeder et al., 1998; Lamme and Roelfsema, 2000). Other studies have argued that the flow of information through this network is inherently multi-directional (Hupe et al., 1998; Lamme et al., 1998). At the same time, the basic role played by the thalamus in mediating interactions within this cortical network — when not ignored — is strongly debated (Ramcharan et al., 2000; Sherman and Guillery, 2000; Steriade, 2000, 2001b). Using traditional neurophysiological techniques to address the brain’s global scheme for processing its input is challenging, as it involves the targeted placement of multiple electrodes in diverse brain areas. While such approaches have been used successfully (Nicolelis et al., 1995; von Stein et al., 1997; Kara et al., 2000), the spatial coverage afforded by such methods is inherently limited by the positioning of the electrodes, which is, in turn, constrained by the priors and hypotheses of the investigator. In contrast, a great strength of functional brain imaging, in particular functional magnetic resonance imaging (fMRI), is that it can provide a global view of activity throughout the brain, allowing one to simultaneously monitor the functional changes in distant cortical and subcortical structures. For this reason, as well as its noninvasive nature, it has in the last decade become the dominant tool for studying human brain function (Frackowiak et al., 1997). Recently fMRI has been introduced as a much-needed addition to the repertoire of tools used to study the brain of nonhuman primates (Logothetis et al., 1999).
Unfortunately, the spatial and temporal resolution of the fMRI signal are currently poor compared to neurophysiological methods. Many of its limitations stem from the fact that it measures a surrogate signal rather than neuronal activity directly. Specifically, most contrast mechanisms, such as the popular BOLD (blood oxygen level dependent) technique, are driven by blood flow and volume changes that are secondary to metabolic changes caused by neural activity. Recent studies have demonstrated that while such signals are tightly coupled to neural activity, in particular to the LFP (Logothetis et al., 2001), they are sluggish in their temporal response, often requiring seconds to register a change. Compared with electrophysiological recordings, which register neural events at arbitrarily short time-scales, BOLD imaging is at least three orders of magnitude worse at tracking dynamic aspects of neural activity relevant for cognition and behavior, but see Ogawa et al. (Ogawa et al., 2000). Given that the strengths and weaknesses of electrophysiological and fMRI techniques are in some sense complementary, a number of previous human studies have sought to combine the two techniques, with the aim of gaining spatial localization from fMRI and temporal information from electroencephalographic (EEG) measurements (Heinze et al., 1994; Bonmassar et al., 2001). Of particular interest to the current study is the recent emergence of combined techniques in the nonhuman primate, where it is now possible to monitor neural activity at one or more intracranial sites during high resolution fMRI scanning (Logothetis et al., 2001). This combination affords the direct comparison of cortical signals (spiking or local field) with the BOLD fMRI signal and promises new insights into brain function that neither signal alone could provide.
The present study represents a first step toward exploring the possibility that combined electrophysiology and fMRI methods can be used to investigate endogenous activity changes in the brain, including those related to the signal variability mentioned above. The supposition is that a continuously monitored neural signal can serve as a useful reference for functional imaging, allowing one to focus on spontaneous changes in the BOLD signal that are of neural origin. Here we concentrate on one aspect of this newly starting research line: what kinds of electrophysiological activity fluctuations might best be correlated with the time course of individual voxels obtained from functional imaging? Specifically, we seek temporal patterns in the neural signal that occur very slowly and can thus be directly compared with changes in the BOLD signal. Fluctuations in the raw local field voltage itself are an improbable candidate since neurophysiological measurements are generally filtered to eliminate changes on time-scales longer than a second. While it would be possible to minimize such filtering in order to track very slow potential changes (Bauer and Rebert, 1990), we focus instead on changes in the local field power, since previous results found LFP power to be well correlated with BOLD responses to sensory stimuli (Logothetis et al., 2001). We begin by extracting a family of band-limited power (BLP) signals from a single LFP trace and demonstrate that these extracted signals possess large-amplitude fluctuations over very long time-scales (>10 s) during different behavioral states. We then show that these fluctuations are highly coherent between electrodes whose effective cortical separation exceeds even 25 mm. We propose that these very slow, highly coherent fluctuations in neural activity are perfectly suited for measurement with fMRI and are therefore likely to have a large impact on the BOLD signal. We discuss the merits of combined electrophysiological/ fMRI recordings in the context of (i) studying brain networks that contribute to this variability and (ii) removing ‘physiological noise’ that is unrelated to a fixed stimulus paradigm.
Materials and Methods
Two adult rhesus monkeys, K97 and R97, were used in the experiments, and participated in 21 and two experimental sessions, respectively. During each session, data was recorded while the animal either executed a behavioral task (K97 only, task described below) or sat quietly in a dimly lit room (both animals). Results from the two animals were similar, and are thus considered together. All experiments were performed in compliance with the guidelines of the local authorities (Regierungs-praesidium) as well as the European Community (EUVD 86/609/EEC) for the care and use of laboratory animals.
Details of the surgical procedures can be found elsewhere (Logothetis et al., 1999; Leopold et al., 2002). Briefly, the animal was implanted under sterile conditions with a custom-made, single piece titanium head holder. In a second surgery, an occipital recording chamber was implanted and a large (23 mm) craniotomy was made, permitting direct access to the brain through the dura mater. A cylindrical chamber constructed of plastic (Tecapeek, GF 30; Ensinger GmbH, Germany) was attached to the skull with ceramic screws. Following surgery, the animal spent 10 days in a recovery chair, which allowed him to stand and move freely, but did not permit touching of the fresh implants. Analgesic (Finadyne, 1.0 mg/kg) was given during the first two days, as well as antibiotic (Veracin Composite, 0.25 ml/kg) for 8 days after the surgery.
Intracortical recordings were conducted with the Eckhorn multielectrode array (Eckhorn and Thomas, 1993), permitting the simultaneous monitoring of up to 15 sites (one channel was reserved for synchronization with the computer). Electrodes were Pt90W10 wire (diameter = 20 μm) with a glass coating (external diameter = 80 mm), and were guided into the brain each day through the overlying dura mater. The recording chamber was situated over the lunate sulcus (Fig 1a), and afforded a large field for data collection. During the recordings, a custom-made adaptor was used to distribute the electrodes against the dura in a 4 × 4 square array, with an inter-electrode spacing of 2.5 mm. Thus, the closest electrode pairs were separated by 2.5 mm, while the farthest (opposite corners) were had a physical separation of 10.6 mm.
Two fundamentally different electrode-positioning strategies were applied in this study. In the first configuration (Fig. 1b), the entire electrode array was placed posterior to the lunate sulcus. This positioning ensured that the effective cortical separation between each electrode pair corresponded roughly to their horizontal distance. In the second configuration (Fig. 10), the electrode array straddled the lunate sulcus, with a portion of the electrodes on the prelunate gyrus in area V4. In this setup, the effective cortical distance was significantly larger between those electrodes on opposite sides of the sulcus (>25 mm), even though the horizontal distance between the electrodes was the same. Each day, all electrodes were positioned in the cortex, and receptive field sizes and positions were first carefully mapped (for K97 only), as shown for the example in Fig 1c. In general, the recordings were conducted in the superficial layers of cortex, shortly following penetration of the dura.
Behavioral Conditions and Data Collection
Neural activity was monitored with the electrodes in a given position during two distinctly different behavioral conditions. In the rest condition, activity was monitored from the same electrodes in the same positions, but while the monkey had no behavioral requirements. During this period, which typically lasted 30–60 min, the monkey sat alone in a dimly lit room. Eye movements during this condition revealed some periods of normal scanning movements, and others of slow drift, consistent with the animal entering light sleep. The eyes would sometimes close, although we did not have the impression that the monkey entered deep sleep since they would only stay closed for a few seconds or minutes at a time.
In addition, we tested monkey K97 under conditions in which he was actively engaged in a binocular rivalry task (task condition). Task details of this paradigm are provided elsewhere (Leopold and Logothetis, 1996). Briefly, the monkey would acquire and hold fixation for periods lasting between 10 and 20 s, during which time a variety of stimuli were shown. The stimuli were in general effective at driving neurons at the sites we recorded, although they were tailored to perceptual criteria rather than to the preferences of the neurons. The monkey’s task was to observe the stimulus and respond to changes in the perceptual dominance of two competing visual patterns. At the end of a successful observation period, the monkey received ∼0.5 ml of apple juice as a reward. Neural signals were also recorded during the 3–5 s intertrial interval, during which time the animal rested. Throughout the task, data was collected ‘pseudocontinuously’, with brief interruptions in recording for a few hundred milliseconds between trials.
During the rest and task conditions, data collection was controlled by an industrial PC (Advantech) running under the QNX operating system (QNX Software Systems, Kanata, Ontario, Canada). The local field data from each channel were amplified by a factor of 8000 and band-pass filtered between 1 Hz (two pole Butterworth filter) and 1 kHz (four pole Butterworth filter) (Alpha Omega Engineering, Nazareth, Israel). The signals were then individually digitized at a rate of 4.5 kHz on a 12-bit analog to digital board (Win30; United Electronic Industries, Watertown, MA) and stored on a PC for further analysis using custom software written in MATLAB (The Mathworks Inc., Natick, MA).
Band Limited Power Calculation
For each LFP signal, the BLP was calculated as follows. For the task condition, LFP data from individual observation periods were concatenated to provide a nearly continuous signal (see above). In the rest condition, data collection was uninterrupted. Each condition therefore yielded continuous signals corresponding to recording periods lasting between 30 (rest) and 90 (task) min.
The LFP signals were resampled to 1 kHz, and then band-pass filtered into seven frequency ranges using a second order, bi-directional, zero-phase Chebyshev type-1 filter. The frequency ranges, as well as their rough mapping onto classically defined electroencephalographic conventions, were the following: δ (1–4 Hz), 𝛉 (5–8 Hz), α (9–14 Hz), β (15–30 Hz), gL (30–50 Hz), gH (50–100 Hz), and gVH (100–150 Hz). The resulting band-limited signals were full-wave rectified by taking their absolute value. They were then resampled to 20 Hz after low-pass filtering with an eighth order Chebyshev type-1 filter (cutoff = 8 Hz). The most relevant conceptual steps in this processing are illustrated schematically in Figure 4a. Note that a particularly important step in the LFP signal processing is the rectification stage, which provides a measure of the time-varying envelope amplitude of each band-pass filtered signal. Once rectified, the signal can contain arbitrarily low frequencies, and is no longer a measure of voltage but rather power (in actuality, the square root of the power). This rectification stage is directly analogous to that used in the demodulation of an AM radio signal, where the temporal structure of the envelope is of interest, rather than the carrier frequency itself. By filtering a given LFP signal according to each of the predefined bandwidths and then rectifying the outcome, we obtain a family of BLP signals, each of which represents the temporal structure of power fluctuations within a given frequency range (e.g. α, β, γ, etc.). From another perspective, a given BLP signal is roughly equivalent to mean of several rows of a spectrogram, or time-frequency plot. Averaging several adjacent rows of a spectrogram together similarly provides a time-varying signal corresponding to the evolving power in a given frequency range.
Coupling of Raw LFP Signals
We began by examining the power and coherence characteristics of the raw (unprocessed) local field signals during task and rest conditions. For this first phase of the analysis, these two conditions provided three distinct periods for comparison: (i) activity during active engagement in the visual task, (ii) activity in the inter-trial period of the task condition, where the monkey relaxed for 2–3 s, and (iii) activity during the rest condition, when the monkey sat uninterrupted for 20–40 min. Figure 2a depicts the average spectral power of a representative session under these three different states. Each curve represents the mean power spectrum from 15 electrodes at the positions depicted in Figure 1. Note that each of the curves shows its highest power at low frequencies, with a decline in spectral power between 3 and 30 Hz. For the task condition, there is an additional well-defined peak in the γ-range (30–80 Hz) that is absent in the other conditions. The transmission of frequencies <1 Hz was limited by a high-pass hardware filter.
To examine coupling between the signals measured at different electrodes, we computed the coherence functions between all electrode pairs, which is shown in Figure 2b for the three conditions. In general, coherence values decreased with increasing frequency between 8 and 100 Hz. For lower frequencies, the pattern was highly dependent on the behavioral state, with the task condition showing a marked absence of coherence (i.e. desynchronization) below 6 Hz compared to the other conditions. In order to evaluate the impact of cortical distance on the LFP coupling, we computed coherence functions for each of the possible inter-electrode distances. Figure 3 demonstrates that the separation between cortical electrodes is a major factor in their coherence, which is in agreement with previous observations (Frien and Eckhorn, 2000). Each plot in Figure 3a represents the coherence value in a given frequency range, as a function of inter-electrode distance, for each of the behavioral states in one session. Note that in nearly all cases, coherence values fell to <0.15 for distances >10 mm, even when their values at 2.5 mm were >0.5. In this session, the drop-off was approximately linear for most frequencies. While this was true for some sessions, others displayed significant deviations from linearity. The decline of coherence with distance varied as a function of frequency, as well as behavioral state. For example, during the task, coherence declined more sharply with distance in the 5–8 Hz range than in the 1–4 Hz range. However, this trend was not true in the inter-trial, in which case the decline was similar for the two frequency ranges. Figure 3b shows in three dimensions how coherence varied with distance and frequency for each of the behavioral conditions. Note that, in general, coherence fell sharply for both long time-scales (i.e. low frequencies) and large cortical distances.
Coupling of BLP Signals
Given our aim of exploring very slow neural fluctuations, we next evaluated the same data using a method that allowed us to evaluate changes in the local field at arbitrary time-scales. Specifically, we reasoned that although high pass filtering prohibited measurement of slow changes in the raw LFP signal, significant and biologically relevant fluctuations in LFP power might proceed over arbitrarily long time-scales. Furthermore, given the tight coupling between BOLD responses and LFP power (Logothetis et al., 2001), such slow fluctuations would be likely to have a significant impact on signals measured in fMRI. We therefore explored the natural time course of power signals in the visual cortex derived from different frequency ranges of the LFP. We evaluated two aspects of BLP signals (see Materials and Methods and Fig. 4) that are relevant for comparison with functional imaging data. First, we computed the power spectrum of the BLP for each frequency range during the rest condition (Fig. 5a). As one might predict from the raw spectra shown in Figure 2a, power in the lowest frequency bands (e.g. 1–4 Hz) was higher than that in upper frequency bands (e.g. 100–150 Hz). However, it is interesting to note that very slow BLP fluctuations, displaying changes on the order of tens of s (i.e. <0.1 Hz), had comparatively higher amplitude than the faster BLP fluctuations, proceeding as roughly as 1/f for frequencies <2 Hz.
Secondly, we evaluated the mean coherence in the BLP, which is shown in Figure 5b, averaged over all electrode pairs. This plot reveals several interesting features about coherence between BLP signals. For example, the coherence values are generally quite large, revealing that a portion of the BLP changes is shared between cortical sites. Also, there is a roughly monotonic trend toward higher coherence at low frequencies, demonstrating that slow changes contribute the most to the inter-electrode covariation. Finally, the frequency ranges displaying the largest coherence at long time-scales were the γ-range frequencies between 30 and 100 Hz. This observation was common in both the rest and task conditions.
The nature of BLP covariation is perhaps best observed in the time domain over periods of several minutes, as shown in Figure 6. In this figure, the δ and γL BLP signals for all electrodes are shown over a period of 40 min with the monkey in the rest state. Inspection revealed many common motifs aligned in time between the different channels (gray shaded regions). This was particularly true in the case of the γL power, where some fluctuations were present in all channels, despite electrode separation >10 mm. In accordance with the coherence plots in Figure 5b, these shared fluctuations were particularly pronounced for very low frequency changes (<0.1 Hz).
Spatial Properties of BLP Coupling
We next investigated how coherence of the BLP depended on inter-electrode separation. Figure 7 illustrates the coherence as a function of frequency and inter-electrode distance for each of the BLP ranges in the rest condition. Note that in each of the plots, the coherence values are particularly high for the lowest frequencies and that this coupling is present over long cortical distances. This is particularly true for BLP signals derived from the γ-range. Figures 8 and 9 compare the coherence-distance relationships for LFP and BLP signals during the rest and task conditions, respectively. For comparison, data from Figure 3a is replotted for all frequency bands in Figures 8a and 9a. In Figures 8b and 9b, the coherence–distance relationship is shown for the BLP for the same sessions at two time-scales. Each of the colored lines within each panel corresponds to the BLP calculated from a particular frequency band. Note that for the very slow BLP signals, the coherence remains high with increasing inter-electrode distance. This is particularly true for the low and high γ-range power fluctuations (red and cyan curves), which display very high coherence even at separations of >10 mm. The similar appearance of the BLP coherence plots in Figures 8b and 9b suggest that the basic observation is due neither to the engagement in a task, nor to hypersynchronous cortical activity patterns that might be related to sleep.
Finally, Figure 10 illustrates that strong temporal covariation exists between BLP signals over great cortical distances. Data is shown from a session in which the square electrode array was placed such that a subset of electrodes fell on either side of the lunate sulcus (bottom right). The effective cortical distance between electrodes lying on opposite sides of the sulcus exceeded 25 mm. BLP traces from five frequency bands are shown for prelunate (red) and postlunate (green) electrodes. Note the high correspondence between temporal features in the traces. While there is clearly more similarity between wave-forms on the same side of the sulcus, electrodes on opposite sides clearly share a great deal of temporal structure, particularly for BLP signals derived from the higher frequencies.
Neural dynamics related to cognition and behavior are generally rapid, proceeding at a rate of milliseconds to seconds. In fMRI, the evaluation of such dynamics remains limited despite the development of techniques to optimize temporal information (Rosen et al., 1998). While the BOLD signal can clearly register the occurrence of brief, singular neural events, it cannot consistently track fast activity changes. For example, recent results demonstrated that neural signals could not be accurately reconstructed from the BOLD signal when changes were faster than ∼0.21 Hz (Logothetis, 2002). In the present study, we considered this limitation, and posed the following question: might neural dynamics also express themselves at much slower time-scales that are more compatible with fMRI measurements? We report here that very slow, coherent activity fluctuations do indeed encompass the visual cortex. We believe that these results indicate that large-scale networks in the brain contribute significantly to neural variability observed at a cortical site and furthermore suggest that fMRI may be an excellent tool to monitor and visualize neural activity in such networks.
The BLP signals, which were derived from the local field, displayed maximal amplitude at very low frequencies. These slow changes were largely shared between electrodes, even for the largest cortical distances tested. This global coherence may provide insight into the nature of the fluctuations themselves, as it is suggestive of interplay with subcortical structures that can simultaneously influence activity across distant cortical sites. It is interesting to note that the frequency bands whose BLP signals showed the highest temporal coherence were in the γ-range. This finding is in some respects counterintuitive, since global synchronization in the brain is often associated with lower frequencies, such as slow cortical rhythms, alpha waves, and sleep spindles (Berger, 1929; Amzica and Steriade, 1995; Contreras et al., 1997; Steriade, 2000). In contrast, γ-range coupling is thought to be more representative of local processing, as observed for the raw LFP coherence in the present study, as well as in several other studies (Steriade et al., 1996a, b; Frien and Eckhorn, 2000), but see Murthy and Fetz (Murthy and Fetz, 1996). While one might therefore predict that γ-derived BLP fluctuations would also fall off quickly with distance, we report here that it is, in contrast, the most coherent of all frequency bands, with only minimal decrease in coherence over cortical distances exceeding 10 mm. Note that the coherence described here is not cycle-to-cycle coherence in the γ voltage, but rather slow, shared fluctuations in the γ power. Interestingly, a fundamentally similar technique was previously applied in a study involving human subdural recordings (Bruns et al., 2000). In that study, the authors evaluated coherent fluctuations in the amplitude envelope of the filtered γ-range activity (equivalent to the γ-BLP) and similarly found that coupling could be detected in the modulation of this ‘carrier’ signal when it was absent in the raw signals.
The far-reaching spatial coherence of power fluctuations, contrasted with the highly localized coherence of the raw voltage signals, raises the interesting possibility that these two aspects of neural dynamics represent different dimensions of cortical, and perhaps thalamocortical, processing. Much evidence has supported the idea that coupling of the local sensory circuitry, particularly in the γ range, may important for the pattern encoding, and perhaps for aspects of higher perceptual processing (Singer and Gray, 1995; Eckhorn, 2000; Gail et al., 2000; Fries et al., 2001). In contrast, the slow, coherent power fluctuations we report here are widespread and are therefore unlikely to be involved in highly specific pattern analysis. Instead, these fluctuations might represent a more global, coordinative aspect of thalamocortical processing, where the strength or efficacy of localized processing at distant brain regions can be modulated, without disrupting specific coupling mechanisms within a region. Further experiments are needed to determine whether these slow power modulations are restricted to the visual cortex, or if they also impact other sensory and non-sensory cortical areas.
Possible Consequences and Origins of Slow Fluctuations
It is interesting to note that despite the brain’s remarkable capacity to contend with its own variability, there may indeed be behavioral consequences of the slow fluctuations of the type reported here. Evidence for this possibility comes from a comparison of our results with human behavioral studies. When humans perform a cognitive task repetitively over extended periods of time, their task performance is not stationary, but instead tends to fluctuate over periods of s and min. Careful analysis of this temporal instability revealed that the changes have a ‘1/f noise’ quality (Gilden et al., 1995; Gilden, 2001). It is intriguing to speculate that such fluctuations arise from endogenous activity changes that influence cognitive processing. It is therefore notable that the BLP spectra reported in the present study (Fig. 5a) bear a striking resemblance to those derived from human performance fluctuations, for example as shown in Figure 1a of Gilden et al. (Gilden et al., 1995). It is possible that spontaneous activity changes in the cortex over long time-scales contribute significantly to the fine-tuning of our perceptual and behavioral capabilities.
Additional evidence linking our results with behavioral studies arises from several electrophysiological studies that correlate performance with activity in the brainstem. For example, activity in the locus coeruleus was shown to co-fluctuate with behavioral performance of monkeys in a visual discrimination task over long time-scales (Aston-Jones et al., 1994; Usher et al., 1999). Noradrenergic neurons in this structure project diffusely throughout the cortex and are thought to play a modulatory role in nonspecific aspects of behavior, such as motivation and arousal (Foote and Morrison, 1987; Berridge and Foote, 1991). Similarly, the role of the midbrain reticular formation (MRF) in cortical activation and arousal has long been recognized (Moruzzi and Magoun, 1949). Early studies revealed that electrical stimulation of the MRF enhanced the processing of sensory stimuli (Dumont and Dell, 1960) and could actually improve tachistoscopic performance in a monkey performing a discrimination task (Fuster, 1957). More recently, functional imaging in humans revealed that this area was specifically activated during tasks that required a high attentional demand (Kinomura et al., 1996). Together, these studies suggest that the slow fluctuations we observe may be under brainstem control, and may be related to behavioral fluctuations. The predominance of synchronization of γ-range BLP signals may also be of particular interest given the postulated role of γ-range activity and synchronization in cortical activation (Moruzzi and Magoun, 1949; Munk et al., 1996), sensory processing (Eckhorn et al., 1988; Singer and Gray, 1995), attention (Fries et al., 2001) and perhaps even conscious perception (Llinas and Ribary, 1993, Engel et al., 1999).
However, it is important to note that in the present study the slow, coherent fluctuations in the BLP signals were not unique to the task condition, but were equally pronounced during the rest conditions, where the animal’s external behavior was very different. Preliminary results also indicate that BLP signals, including pronounced coherence of the γ-derived signals, are similar in monkeys under light general anesthesia (D.A. Leopold, unpublished observations). Thus, while the very slow, coherent fluctuations we report may be related to active elements of attention and perception, they also appear to be present in states of diminished consciousness, suggesting that their generation is not specific to a particular cognitive or behavioral context.
In order to understand better the nature of these changes, it might be informative to learn how they relate to other aspects of brain physiology. For example, recent work from Raichle and colleagues in human fMRI and positron emission tomography (PET) has also emphasized the need to examine activity in the absence of a stimulus or a task in order to correctly interpret an evoked response (Gusnard and Raichle, 2001). These experiments have not examined neural activity directly, but have used measures of hemodynamic and metabolic signals acquired through imaging to attempt to establish a general notion of ‘activation’ for brain tissue. One computed variable, the oxygen extraction fraction (the ratio of oxygen utilized in a portion of tissue to that delivered) can, under conditions of rest, show remarkable spatial uniformity across the entire brain despite large local differences in the measured variables used to compute it. This finding led Raichle to suggest that the very notion of an ‘activity baseline’ in the brain need not be an arbitrary experimental reference point, but might instead reflect a well-defined physiological state of brain tissue. How then might the slow, spontaneous electrical fluctuations reported here relate to such a baseline? While it is still too early to answer this question, preliminary data from our laboratory suggest that the baseline activity might be significantly affected by the BLP fluctuations reported here (Leopold et al., 2002). We simultaneously measured LFP from one electrode in the visual cortex and fMRI signals over the entire brain in the anesthetized monkey whose eyes were closed. We found that the time course of the BOLD signal was often highly correlated with the spontaneous neural activity (BLP) fluctuations. Moreover, although the electrical signal came from a single electrode, the correlations often reached over large stretches of the cerebral cortex — in some cases nearly the entire brain. Given that the BOLD changes are correlated with those of the oxygen extraction fraction, this finding might suggest that even the spatially uniform ‘default state’ relevant for imaging studies is subject to spontaneous fluctuations resulting from underlying slow changes in neural activity. We are currently investigating the degree to which these slow fluctuations impact the firing rate of individual neurons. It would be, for example, particularly interesting if one type or layer of cortical neurons consistently modulated with global baseline fluctuations, while another did not.
A Neural Reference for Functional Imaging Studies
Given the comparable time courses of the BLP and BOLD signals, as well as the preliminary results mentioned above, there are at least two reasons why combined electrophysiological and fMRI techniques may be of great value in exploring brain function. First, combined measurements might be performed to explore functional connectivity of global networks that are otherwise inaccessible. Traditionally, the use of fMRI to assess functional interactions in the brain has exploited covariation in the time course of voxels in distant brain regions, since even in the resting brain voxel intensity varies on the order of 1–2% (Biswal et al., 1997). Such covariation analysis has been applied in studies with subjects at rest (Biswal et al., 1997; Lowe et al., 1998; Xiong et al., 1999; Cordes et al., 2001), or engaged in a motor or cognitive task (Lumer and Rees, 1999; Cordes et al., 2000). Unfortunately, coherent temporal fluctuations can arise from a diversity of sources that are not directly linked to underlying neural activity. Such sources include periodic cardiac and respiratory noise (Frank et al., 2001; Lund, 2001), slow, autonomous hemodynamic regulation (Obrig et al., 2000) and magnetic field drift (Durand et al., 2001). While several methods have been developed to identify and remove artifacts caused by the periodic physiological signals (Glover et al., 2000; Chuang and Chen, 2001) these methods often rely upon high image acquisition rates in order to meet the Nyquist criterion for temporal sampling. High-speed scanning usually comes at the expense of spatial coverage and image quality, which are both particularly important for investigating functional interactions between distant brain areas. Other strategies for removing such artifacts include the use of blind source separation techniques (McKeown et al., 1998; Arfanakis et al., 2000).
The simultaneous monitoring of BOLD and LFP signals may significantly refine one’s ability to focus on temporal covariation that is related to underlying changes in neural activity. By using the BLP signal as a reference signal it may be possible to focus on BOLD fluctuations that are specifically tied to neural changes. Previous approaches using combined electrical and fast opticalimaging techniques have allowed for remarkable visualization of localized cortical networks (Arieli et al., 1996; Tsodyks et al., 1999; Seidemann et al., 2002). Given the spatial coverage of fMRI, it would be of great value to use a signal such as the BLP to isolate and visualize global brain networks that contribute to its fluctuations. Interestingly, its profound coherence between distant electrodes suggests that a single cortical electrode could serve as a reference for very large cortical region, perhaps even for the entire brain. As mentioned above, preliminary experiments from our laboratory suggest that this is indeed the case. Thus, monitoring the signal of a single electrode during imaging might facilitate the visualization of global brain networks that could not otherwise be tapped.
The second reason to combine these techniques centers on the potential contribution of the slow BLP modulations reported here to the 1–2% physiological fluctuations in the BOLD signal. As argued above, these changes may be of great important to understanding brain function, and may even underlie dynamic aspects of our perception and cognition. However, in many fMRI paradigms, such slow changes are unrelated to the presentation of a stimulus or the execution of a task and might therefore be a considered a source of neural ‘pollution’ for a particular experiment. In such cases, it might be of great value to measure and remove fluctuations in the signal related to these slow neural changes, a manipulation that might greatly improve the signal-to-noise ratio within an fMRI experiment and thereby reduce the required number of averages. In the experimental animal, this might take the form of an implanted intracortical electrode, through which the LFP is monitored during functional imaging. In humans, it is possible that a single EEG reference might be able to provide a similar function. Additional experiments are needed to determined whether the combined monitoring of electrical and fMRI signals will lead to significant improvement in the visualization of brain function, either in the context of mapping networks involved in generating spontaneous changes as reported here, or for eliminating unwanted neural contributions in a traditional fixed fMRI paradigm.
The authors would like to thank Dr A.A. Ghazanfar, A. Maier and M. Wilke for comments on an earlier version of the manuscript, Drs J. Pfeuffer and P. Mitra, and A. Oeltermann for discussion, and Dr W. S. Geisler for pointing out the connection with performance fluctuations in human psychophysics. In addition, we thank C. Moya and J. Werner for technical assistance. This work was supported by the Max Planck Society.