Neural mechanisms of the EEG alpha-BOLD anticorrelation

Abstract

An experimentally tested neural field theory of the corticothalamic system is used to model brain activity and resulting experimental EEG data, and to elucidate the neural mechanisms and physiological basis of alpha-BOLD anticorrelation observed in concurrent EEG and fMRI measurements. Several studies have proposed that the anticorrelation originates from a causal link between changes in the alpha power and BOLD signal. However, the results in this study reveal that fluctuations in alpha and BOLD power do not generate one another but instead respectively result from high- and low-frequency components of the same underlying cortical activity, and that they are inversely correlated via variations in the strengths of corticothalamic and intrathalamic feedback, thereby explaining their anticorrelation.

Introduction

Scalp encephalography (EEG) is widely used to noninvasively record cortical activity aggregated at spatial resolutions of few centimeters on a millisecond time scale. Measurements using EEG have shown that spontaneous brain activity includes some resonant oscillations, one of the most dominant in normal adult humans being the 7.5–12 Hz alpha rhythm (Berger, 1929), which is especially prominent during a waking resting state with the eyes closed.

Since its discovery, the alpha rhythm has been used as a marker of the level of arousal, vigilance, and attention (Strijkstra et al., 2003; Babiloni et al., 2004; Niedermeyer and Lopes da Silva, 2005; Thut et al., 2006; Rihs et al., 2007; Olbrich et al., 2009). Several works have also argued that it plays an important role in the perception of external stimuli (Babiloni et al., 2006; Thut et al., 2006; Haegens et al., 2011; Händel et al., 2011) and many cognitive abilities (Klimesch, 1999; Başar et al., 2001; Jensen et al., 2002; Sauseng et al., 2005; Zumer et al., 2014).

Because EEG is spatially coarse-grained, many recent studies have used functional magnetic resonance imaging (fMRI), based on the blood oxygen level-dependent (BOLD) signal, to probe brain dynamics at finer spatial resolution ($\sim$1 mm) but on a longer time scale (seconds) corresponding to the dynamics of spontaneous BOLD fluctuations. Notably, covariances of BOLD fluctuations are often used to define functional connectivity between discrete brain regions (Fox and Raichle, 2007). Hence, there is a strong interest in exploring the links between EEG and BOLD measurements to better understand structure–function relationships.

Even though EEG and fMRI investigate brain function in different scales, their signals are tightly coupled and originate from the same underlying neural dynamics. In particular, EEG is recorded after volume conduction effects have removed high spatial frequencies of the neural activity (Nunez et al., 1994; Nunez and Srinivasan, 2006). In contrast, fMRI measures BOLD signal resulting from slow hemodynamic responses (typically in frequencies $\mathrm{\lesssim }$ 0.2 Hz) that suppress high temporal frequencies of the neural activity (Zarahn et al., 1997; Robinson et al., 2006; He et al., 2008; Nir et al., 2008; Scheeringa et al., 2011; Aquino et al., 2012; Wang et al., 2014). Thus, they can be combined to complement one another and obtain data with high spatial and temporal resolutions [see for example the reviews of Ritter and Villringer (2006) and Laufs et al. (2008)].

The fusion of EEG and fMRI has led several studies to investigate the relationship of the alpha rhythm and the BOLD signal. These studies have shown that, in various brain structures, fluctuations in the BOLD signal are correlated or anticorrelated with fluctuations in the alpha-rhythm power (Goldman et al., 2002; Laufs et al., 2003; Moosmann et al., 2003; Gonçalves et al., 2006; Danos et al., 2001; Feige et al., 2005). However, these correlations or anticorrelations are not consistently found across different studies, which is possibly due to high inter- and intra-subject variabilities (Gonçalves et al., 2006), or methodological factors such as differences in recording paradigms and the manner in which the alpha-rhythm power was quantified from EEG recordings. Nonetheless, the most robust finding is for the occipital cortex where alpha is strongest and modulations of its power are negatively correlated with those of the BOLD signal. This concurs with several studies using other neuroimaging modalities that show similar relationships between fluctuations in alpha-rhythm power and those in blood flow and oxygenation signals (Jacquy et al., 1980; Sadato et al., 1998; Moosmann et al., 2003).

Even with the abundance of statistical evidence relating the alpha rhythm and the BOLD signal, the underlying physiological mechanisms of the observed alpha-BOLD anticorrelations have not been established. Some studies have conjectured that a decrease in occipital BOLD signal is a result of ‘idling’ of cortex and decreased neural activity (Chawla et al., 1999; Wenzel et al., 2000; Mayhew et al., 2013), or alternatively linked to other functionally coupled processes such as vigilance (Moosmann et al., 2003) and wakefulness (Horovitz et al., 2008); however, mechanistic links remain elusive and concepts such as ‘idling’ remain poorly defined.

Most studies have proposed that there is a direct causal link between fluctuations in the alpha power and the BOLD signal, as shown in Fig. 1a (Goldman et al., 2002; Laufs et al., 2003; Moosmann et al., 2003; Schirner et al., 2018). This was found by employing data processing steps, including high-pass filtering the EEG signal (usually $>$ 0.5 Hz), calculating the power spectrum from the squared amplitude of the filtered signal's Fourier transform, and convolving the time course of the alpha power with an a priori hemodynamic response function (HRF) before comparing with the BOLD signal. These processing steps have the following shortcomings. First, high-pass filtering the EEG signal inevitably throws away the important frequencies that mainly drive the slow activity of BOLD. Second, it is well known that the HRF can take into account the slow processes mediating neural activity and BOLD; hence the neural signal itself, not its power, should be convolved with the HRF to predict the BOLD signal (Friston et al., 2000). This is what convolution of the alpha power with the HRF described above aims to achieve, but we argue that it does not follow the actual neural-HRF-BOLD relationship. Moreover, even though the convolution culminates to a low-frequency signal that appears to be comparable to BOLD, it represents the windowed low-frequency envelope of alpha, not the original low-frequency neural signal that drives BOLD.

In order to overcome the abovementioned shortcomings and to not introduce confounds caused by ad hoc data processing, we instead directly extracted the low-frequency power of the actual EEG signal because it best represents the slow activity of BOLD. In this way, we can relate the powers of the low-frequency and alpha bands, which have identical dimensions. Thus, the type of analysis in this study is certainly different from what has been done previously, but it investigates the mechanisms underlying the alpha-BOLD anticorrelation in a new way that has a more compelling physical basis. In particular, we aim to show that fluctuations in the alpha power do not drive those in the BOLD power, nor vice versa. Instead, they respectively result from high- and low-frequency components of the same underlying cortical activity, which ensue from variations in the strengths of corticothalamic and intrathalamic feedback, as shown in Fig. 1b.

Here, we use an experimentally tested and verified neural field theory of the corticothalamic system (Robinson et al., 2001, 2002, 2004, 2005) to model and analyze experimental EEG data recorded for a cohort of subjects in a relaxed waking eyes-closed state. First, properties of the low-frequency (combination of the $<$ 1 Hz slow-wave and the 1–4 Hz delta-wave frequencies) and alpha bands of the EEG power spectra are analyzed to show that the anticorrelation between alpha and BOLD power (proxied by the low-frequency power) can be reproduced for a large group of subjects. Then, the model is used to fit the data and obtain time series of estimated corticothalamic gains within a single subject and show the validity of the causal relationship in Fig. 1b.