A multivariate, spatiotemporal analysis of electromagnetic time-frequency data of recognition memory

Abstract

Electromagnetic indices of “fast” (above 12 Hz) oscillating brain activity are much more likely to be considerably attenuated by time-averaging across multiple trials than “slow” (below 12 Hz) oscillating brain activity. To the extent that both types of oscillations represent the activity of temporally and topographically separable neural populations, time averaging can cause a loss of brain activity information that is important both conceptually and for multimodal integration with hemodynamic techniques. To address this issue for recognition memory, simultaneous electroencephalography (EEG) and whole-head magnetoencephalography (MEG) recordings of explicit word recognition from 11 healthy subjects were analyzed in two different ways. First, the time course of neural oscillations ranging from theta (4.5 Hz) to gamma (42 Hz) frequencies were identified using single-trial continuous wavelet transforms. Second, traditional analyses of amplitude variations of time-averaged EEG and MEG signals, event-related potentials (ERPs), and fields (ERFs) were performed and submitted to distributed source analyses. To identify data patterns that covaried with the difference between correctly recognized studied (old) words and correctly rejected nonstudied (new) words, a multivariate statistical tool, partial least squares (PLS), was applied to both types of analyses. The results show that ERPs and ERFs are mainly displaying those neural indices of recognition memory that oscillate in the theta (4.5–7.5 Hz), alpha (8–11.5), and to some extent in the beta1 (12–19.5 Hz) frequency range. The sources of the ERPs/ERFs were in good agreement with the topography of theta/alpha/beta 1 oscillations in being confined to the anterior temporal lobe at 400 ms and being distributed across temporal, parietal, and occipital areas between 500 and 700 ms. Gamma oscillations covaried either positively or negatively with theta/alpha/beta1 oscillations. A positive covariance, for instance, was detected over left anterior temporal sensors as early as 200–350 ms and is compatible with studies in rodents showing that gamma and theta oscillations emerge together out of the interaction of the hippocampus and the entorhinal and perirhinal cortices. Fast beta oscillations (20–29.5 Hz), on the other hand, did not strongly covary with slow oscillations and were likely to arise from neural populations not adequately represented in ERPs/ERFs. In summary, by providing a more comprehensive description of electromagnetic signals, time-frequency data are of potential benefit for integrating electrophysiological and hemodynamic indices of brain activity and also for integrating human and animal electrophysiology.

Introduction

The local synchronized behavior of neural assemblies leads to fluctuations in local field potentials (LFPs) that can be measured using electroencephalography (EEG) and magnetoencephalography (MEG) (Lopes da Silva, 1991). Much of our knowledge about the electrophysiology of recognition memory is derived from the interpretation of EEG and MEG signals that were averaged across multiple trials resulting in the so-called event-related potentials (ERPs) and fields (ERFs). The interpretation of amplitude fluctuations of ERPs and ERFs has provided important insights into the neural mechanisms underlying recognition memory Dale et al 2000, Duzel et al 1999, Rugg et al 1998, Tendolkar et al 2000. They also continue to be the main source of information for integrating electrophysiological signals and hemodynamic measurements in multimodal imaging. Such multimodal integration is an important goal to achieve in the attempts to understand the functional neuroanatomy of cognitive events.

The utility of ERPs and ERFs for a comprehensive understanding of the electrophysiology of cognitive events and for multimodal integration could be hampered by the limited information they carry about LFP oscillations. Indeed, it has been known for a long time that time-averaged electromagnetic signals provide a limited signature of the electrophysiology of cognitive events. The relationship between ERPs/ERFs and LFP oscillations is determined by the relationship between “evoked” (Galambos et al., 1981) and “induced” (Tallon-Baudry and Bertrand, 1999) neural activity (Basar-Eroglu and Demiralp, 2001). Both types of activity critically differ in how various frequency bands [delta (0.1–3.5 Hz), theta (4–7.5 Hz), alpha (8–11.5 Hz), beta1 (12–19.5 Hz), beta2 (20–29.5 Hz), and gamma (30–80 Hz and higher)] are affected by averaging across trials. Stimulus-evoked neural activity is time- and phase-locked to the onset of a stimulus and, given adequate filtering, is apparent in the averaged EEG and MEG, and hence in ERPs and ERFs. Stimulus-induced neural activity, on the other hand, has a loose temporal relationship to stimulus onset. It consists of oscillatory bursts whose latency jitters from trial to trial. At a given time point after stimulus onset, this type of oscillation will therefore have a different phase on different trials (Tallon-Baudry and Bertrand, 1999). This phase variability will cause attenuation during averaging particularly of beta and gamma oscillations which may therefore become invisible in ERPs and ERFs Basar-Eroglu and and Demiralp 2001, Vanderwolf 1969, Varela et al 2001. Stimulus-induced delta, theta, and alpha frequencies, on the other hand, might also contribute to ERPs and ERFs if their temporal onset does not vary much from trial to trial. Therefore, to the extent that “fast” (above 12 Hz) and “slow” (below 12 Hz) oscillations represent the activity of temporally and topographically separable neural populations, time averaging will cause brain regions with fast oscillations to be underrepresented. Logothetis et al. (2001) recently showed that fast oscillations are very likely to be correlated with the functional magnetic resonance imaging (fMRI) signal. Thus, time averaging of electromagnetic data might lead to discrepancies with hemodynamic measurements to the extent that high-frequency oscillations are not retained in electromagnetic data and do not covary temporally and topographically with slow oscillations that are retained. This problem will be referred to as the problem of high-frequency retention.

The aim of the present study was to investigate the issue of high-frequency retention by characterizing the relationship between ERPs/ERFs and oscillations in the range between 4.5–42 Hz. We identified oscillations during recognition memory and related their time course and topography to the known ERP indices of recognition memory. Comparisons between oscillations and ERFs were performed on the basis of combined ERP/ERF current density source reconstructions rather than ERF field distributions. Amplitudes of oscillations can be more directly compared to current density reconstructions than to field distributions of ERFs, because both types of analyses lack phase information. The technique of wavelet transformation provides a reliable method to measure frequency information of oscillations with high temporal resolution. Tallon-Baudry and Bertrand (1999) have pointed out that in order to identify stimulus-induced oscillations, wavelet transforms must be calculated for single trials of EEG/MEG recordings and the obtained absolute amplitude or the power must be subsequently averaged across trials. Here, we followed this approach and used a combination of EEG and MEG recordings because, due to the relatively low influence of materials with high impedance (e.g., bone) on magnetic compared to electric fields, MEG has a higher topographic specificity than EEG, but much of the existing literature on the electrophysiology of recognition memory is based on EEG recordings. Therefore, the comparison of both data sets allows us to relate MEG findings to existing data.

A major problem in the analysis of oscillations of EEG/MEG recordings is the statistical analysis of the large amount of possible comparisons between electrodes/sensors, conditions, and frequency bands. The traditional approach of performing comparisons between conditions and electrodes/sensors using univariate statistics is somewhat unsatisfactory here because a priori knowledge is required to constrain analysis to a reasonably limited number of statistical comparisons. To achieve a comprehensive analysis of the time-frequency distribution of different oscillations and their topography, and to assess covariance between fast and slow oscillations, we utilized a multivariate statistical analysis tool, partial least squares (PLS) Lobaugh et al 2001, Mcintosh et al 1996, that does not require any a priori bias with respect to the location and time course of effects. PLS was not only applied to time-frequency-amplitude information (obtained from wavelet transforms of single-trial MEG data for correctly judged old and new words) but also to ERP/ERF-based spatially normalized (Talairach and Tournoux, 1988) current density reconstructions. By relating the two types of PLS datasets to each other, we were able to determine the frequency bands in which the topography of the amplitude of oscillations was most compatible with the distribution of neural generators identified on the basis of ERPs and ERFs. Additionally, we performed PLS analysis on ERP data alone in order to be able to relate our findings to the preexisting literature on the electrophysiology of memory.

We used a task requiring conscious recognition of studied (“old”) and new words. Such recognition is associated with an increase in ERP positivity for the studied words between 400 and 800 ms after the onset of word presentation Duzel et al 1997, Rugg et al 1998. This so-called “old/new effect” is composed of an earlier (300 to 500 ms) and more frontal part (henceforth termed N400 effect) as well as a later (500–800 ms) and more parietal part, henceforth called late-positive component, or LPC effect Curran 2000, Duzel et al 1997, Paller et al 1995, Rugg et al 1998. Recent data suggest that the inferior temporal cortex is one of the possible generators of the N400 and the LPC effects (Duzel et al., 2001). Using a semantic task (a variant of a living/nonliving judgment) Dale et al. (2000) showed that under indirect task instructions, ERF repetition effects are generated in the inferior temporal and left frontal cortices in the N400 time window. These data cannot be directly related to the explicit recognition task employed here, however, because it has been shown that old/new effects differ in both the N400 and the LPC time window under direct and indirect task instructions Duzel et al 2001, Rugg et al 1998. Therefore, the neural generators underlying the ERP and ERF effects of recognition memory have so far not been sufficiently characterized.

To summarize, the goal of the present study was to investigate the extent to which poor high-frequency retention of ERPs and ERFs causes a loss to brain activity information. To that end, we assessed whether the time course and topography of fast (above 12 Hz) and slow (below 12 Hz) oscillations during recognition memory represent the activity of temporally and topographically separable neural populations. We were thus able to determine if time averaging causes brain regions with fast oscillations to be underrepresented. We hypothesized that the sources of the ERP/ERF old/new effect should be mostly compatible with the topography of amplitude differences between slow oscillations elicited by old and new words. We expected that frequency analyses should additionally identify neural events that oscillate in the beta2 to gamma frequency range but we had no a priori hypothesis as to whether the topography of these fast oscillations would differ from the slower ones and therefore would be less compatible with the location of the sources of ERPs and ERFs.