Association between neurovascular coupling and neural signals in the resting state as revealed by a combined fMRI and magnetoencephalography study in humans

Abstract

The blood oxygenation level-dependent (BOLD) activation reflects hemodynamic events mediated by neurovascular coupling. During task performance, the BOLD hemodynamic response in a relevant area is mainly driven by the high levels of synaptic activity (reflected in local field potentials, LFPs) but, in contrast, during a task-free, resting state, the contribution to BOLD of such neural events is small, as expected by the comparatively (to the task state) low level of neural events. Concomitant recording of BOLD and LFP at rest in animal experiments has estimated the neural contribution to BOLD to ∼10%. Such experiments have not been performed in humans. As an approximation, we recorded (in the same subject, n = 57 healthy participants) at a task-free, resting state the BOLD signal and, in a different session, the magnetoencephalographic (MEG) signal, which reflects purely neural (synaptic) events. We then calculated the turnover of these signals by computing the successive moment-to-moment difference in the BOLD and MEG time series and retaining the median of the absolute value of the differenced series (BOLD and TMEG, respectively). The correlation between normalized turnovers of BOLD (TBOLD) and turnovers of MEG (TMEG) was r = 0.336 (r² = 0.113; P = 0.011). These results estimate that 11.3% of the variance in TBOLD can be explained by the variance in TMEG. This estimate is close to the aforementioned estimate obtained by direct recordings in animal experiments.

NEW & NOTEWORTHY Here, we report on a weak positive association between turnovers of blood oxygenation level-dependent (TBOLD) and magnetoencephalographic (TMEG) signals in 57 healthy human subjects in a resting, task-free state. More specifically, we found that the purely neural TMEG accounted for 11.1% of the TBOLD, a percentage remarkably close to that found between resting-state local field potentials (LFPs) and BOLD recorded concurrently in animal experiments.

INTRODUCTION

The widespread use of functional magnetic resonance imaging (fMRI) for noninvasive imaging of brain activity via the blood oxygen level-dependent (BOLD) signal has led to unparalleled advances in brain research; yet, critical knowledge gaps and misconceptions hinder the interpretation of the BOLD signal in a task-free state. In particular, though the BOLD signal has been shown to increase directly and monotonically with neural activity during stimulus presentation or task performance (1), the extent to which the BOLD signal reflects neural activity in a task-free state is uncertain. At face value, the BOLD signal reflects changes in blood flow, volume, and oxygenation (2). Thus, unlike magnetoencephalography (MEG), which directly measures synaptic activity (3), the BOLD signal provides an indirect measure of brain function in which the increased blood flow (i.e., functional hyperemia) is presumed to be coupled with, and in support of, the energy demands of neural activity (4). In a “resting” (i.e., task-free) state, however, synaptic activity is weak, and recent evidence suggests this “neurocentric” view in which the resting-state BOLD signal is thought to reflect neural processes may be inaccurate (5). To that end, animal studies concurrently recording electrophysiological and hemodynamic responses at rest found weak correlations between local field potential (LFP) and BOLD despite correlated time courses (6, 7). Furthermore, experimental animal studies and computational models have shown a mismatch, or decoupling, between BOLD signal and neural events under various conditions (5, 8). Here, in an extension to humans, we acquired task-free, resting-state fMRI and MEG recordings in cognitively healthy women from which the median absolute change between successive time samples, or turnover, in the BOLD (TBOLD) (9) and MEG (TMEG, this study) signals were derived and their correlation was evaluated to determine the correspondence between neurovascular and neural signals in the resting state.

MATERIALS AND METHODS

Participants

Fifty-seven cognitively healthy women (age 54.05 ± 1.51 yr, mean ± SE; range 32.17–74.58 yr) participated in this study after providing written informed consent, in adherence to the Declaration of Helsinki, and were financially compensated for their time. All study protocols were approved by the appropriate Institutional Review Boards. Their cognitive status was assessed using the Montreal Cognitive Assessment (MoCA; http://www.mocatest.org/); all women had MoCA scores >25.

fMRI

fMRI data acquisition.

During resting-state functional magnetic resonance imaging (rfMRI), acquisition participants were instructed to keep their eyes fixated, relax their minds, and avoid blinking and moving. All data were acquired using a 3 T MR scanner (Achieva, Philips Healthcare, Best, The Netherlands) with an eight-channel phased-array SENSitivity Encoding SENSE-HEAD-8 head coil for reception. For each subject, a high-resolution T1-weighted anatomical image turbo field echo (TFE) was obtained (168 sagittal slices, TR = 8.0928 ms, TE = 3.697 ms, voxel size = 0.9375 × 0.9375 × 1 mm). The fMRI was acquired using fast field echo (FFE) transversal (axial) with whole brain coverage (repetition time TR = 2,000 ms, echo time TE = 30 ms, flip angle = 90°, 40 axial slices per volume with slice scanning order ascending, no interleaving, voxel size = 2.75 × 2.75 × 3.5 mm). Altogether, 200 samples were acquired for a total acquisition time of 400 s (6.67 min).

fMRI image preprocessing.

A 704-core high-performance Linux computer cluster (Rocks 6.1.1, CentOS 6.5) with MATLAB R2016 (64 bit, MathWorks), Analysis of Functional NeuroImages (AFNI) program (10), Surface Mapping (SUMA) (11), and FreeSurfer (FS) (12, 13) v. 6.0.0 were used for data processing. All imaging data were analyzed using the AFNI software package. First, the skull-stripped anatomical images were registered to a 2 mm Montreal Neurological Institute (MNI) ICBM MNI 152 template image using nonlinear registration (“warping”). Next, the first two volumes of each blood oxygenation level-dependent (BOLD) time series were discarded to allow for equilibrium magnetization. All remaining volumes were slice-time corrected. For motion correction and alignment, a robust method was used to provide an alignment to a volume clean of artifacts, motion, and dropout (14). Standard space transformation, applied to the anatomical image, was applied to the functional image. Alignment of anatomical to functional image was done with cost function local Pearson’s correlation (15). The transformations were concatenated to allow a single resampling using an AFNI Python script. The anatomical image was then used to create a cortical model in FS. Each participant’s FS parcellation and segmentation were regions of interest (ROIs) used as masks to extract the voxel time series. Data from 70 ROIs of Desikan-Killiany atlas (35 in the left and 35 in the right hemisphere) were used (16): banks of the superior temporal sulcus, caudal anterior cingulate gyrus, caudal middle frontal gyrus, cuneus, entorhinal gyrus, frontal pole, fusiform gyrus, hippocampus, inferior frontal gyrus pars opercularis, inferior frontal gyrus pars orbitalis, inferior frontal gyrus pars triangularis, inferior parietal lobule, inferior temporal gyrus, insula, isthmus of cingulate gyrus, lateral occipital gyrus, lateral orbitofrontal gyrus, lingual gyrus, medial orbitofrontal gyrus, middle temporal gyrus, paracentral gyrus, parahippocampal gyrus, pericalcarine gyrus, postcentral gyrus, posterior cingulate gyrus, precentral gyrus, precuneus, rostral anterior cingulate gyrus, rostral middle frontal gyrus, superior frontal gyrus, superior parietal lobule, superior temporal gyrus, supramarginal gyrus, temporal pole, and transverse temporal gyrus. Each processing step quality was inspected. Finally, the coefficient of variation of the BOLD time series of each voxel was calculated, and voxel series with a coefficient of variation greater than 5% were eliminated from further analysis due to their likely proximity to large vessels (17).

fMRI data processing.

A robust, nonparametric estimate of BOLD turnover was obtained in three steps, as described in detail elsewhere (9). Briefly, 1) single BOLD time series for a given k^th voxel were first-order differenced; 2) the median of the absolute values of the differenced time series was computed; and 3) this median value was divided by 2 to yield BOLD turnover per second (TBOLD), as TR = 2 s:

$${\text{TBOLD}}^k=\text{median\hspace{0.17em}of\hspace{0.17em}absolute\hspace{0.17em}values\hspace{0.17em}of\hspace{0.17em}BOLD\hspace{0.17em}change}/2$$ (1)

Next, the mean TBOLD for each ROI (across M voxels in the ROI) was computed and used for further analyses:

$$\text{TBOLD}=\frac{1}{M}{{\displaystyle \sum }}_k^{k=1,M}{\text{TBOLD}}^k$$ (2)

Finally, TBOLD was normalized by computing its rankit:

$${\text{TBOLD}}^{\text{R}}=\text{rankit}(\text{TBOLD})$$ (3)

MEG

MEG data acquisition.

All participants underwent a MEG scan. As described previously (18), subjects lay supine within the electromagnetically shielded chamber and fixated their eyes on a spot 65 cm in front of them, for 60 s. MEG data were acquired during those 60 s using a 248-channel axial gradiometer system (Magnes 3600WH, 4-D Neuroimaging, San Diego, CA), band-filtered between 0.1 and 400 Hz, and sampled at 1,017.25 Hz. Data with artifacts (e.g., from excessive subject motion, eye blinks, etc.) were eliminated from further analysis.

MEG data processing.

Similar to the fMRI data processing, 1) single MEG time series for a given m^th sensor were first-order differenced and 2) the median of the absolute values of the differenced time series was computed for further analyses:

$${\text{TMEG}}^m=\text{median\hspace{0.17em}of\hspace{0.17em}the\hspace{0.17em}absolute\hspace{0.17em}values\hspace{0.17em}of\hspace{0.17em}MEG\hspace{0.17em}change}$$ (4)

Next, the mean TMEG across the 248 sensors was computed and used for further analyses:

$$\text{TMEG}=\frac{1}{248}{\displaystyle {\sum }_m^{m=1,248}{\text{TMEG}}^m}$$ (5)

Finally, TMEG was normalized by computing its rankit:

$${\text{TMEG}}^{\text{R}}=\text{rankit}(\text{TMEG})$$ (6)

Statistical analyses.

Standard parametric and nonparametric statistical methods were used to analyze the data using the IBM-SPSS statistical package (v. 29). TBOLD and TMEG data were normalized by converting them to rankit scores using the RANK command in SPSS. As neural events (measured by TMEG) are expected to contribute to TBOLD (but not the opposite), a linear regression model was fit between TBOLD^R (dependent variable) and TMEG^R (independent variable). All P values reported are two-tailed.

RESULTS

There was a total of 57 pairs of TBOLD and TMEG values available (one pair per participant). We found a statistically significant dependence of TBOLD^R on TMEG^R (Fig. 1; r = 0.336, P = 0.011); the regression equation was

$${\text{TBOLD}}^{\text{R}}=0.336{\text{TMEG}}^{\text{R}}$$ (7)

Figure 1.

Normalized TBOLD is plotted against the corresponding normalized TMEG. n = 57. See text for details. TBOLD, turnover of blood oxygenation level-dependent; TMEG turnover of magnetoencephalographic.

The percent of variance in TBOLD^R explained by TMEG^R is 100 × 0.336² = 100 × 0.113 = 11.3%. As both variables were normalized, the constant term is zero and the slope is the same as the correlation. Equation 7 essentially means that ∼30% of TMEG contributes to TBOLD, accounting for 11.3% of the latter’s variance.

DISCUSSION

Different neuroimaging modalities reflect distinct underlying physiological processes. Although MEG directly measures neural activity, BOLD measures changes in blood flow and blood oxygenation coupled with neural activity. That is, BOLD indirectly reflects neural activity, yet the BOLD signal is typically interpreted as a surrogate of neural activity. Under some conditions, that interpretation likely holds; however, the extent to which resting-state BOLD reflects neural activity has increasingly been called into question (5). Here, we acquired resting-state fMRI and MEG recordings in healthy adult women and evaluated the correlation between turnover of BOLD and MEG signals. We found that neural turnover accounted for 11.3% of the BOLD turnover in this sample of cognitively healthy adult women. Notably, our findings are remarkably similar to findings from experiments in animals in which the correspondence between simultaneously recorded resting-state BOLD and LFP was also found to be equally low (6, 7). These findings in addition to evidence of uncorrelated hemodynamic and neural activity in monkeys (19), spontaneous BOLD independent of neural activity in anesthetized rodents (20), and neural activity in the absence of blood flow changes in mice during locomotion (21) point to decoupling of neural and vascular response under some conditions. Evidence of neurovascular uncoupling in addition to the relatively low correspondence between purely neural signals and neurovascular signals at rest as observed here and reported previously (6, 7) suggests that the BOLD signal partially derives from nonneural processes. Indeed, astrocyte activity (22), changes in arousal (23), and various behaviors including twitching, fidgeting, tongue movement, blinking, and respiration (24) have been shown to contribute to spontaneous hemodynamic changes.

In corroboration with previous animal studies, the present study in adult humans suggests that resting-state BOLD is modestly associated with synaptic activity. These novel findings, however, must be considered in the context of several qualifications. Unlike animal studies in which LFP and BOLD were recorded concurrently (6, 7), here MEG and fMRI were acquired separately. Consequently, and despite evidence that neurovascular coupling differs across brain regions (4, 25), the absence of temporal and spatial correspondence between brain activity for the MEG sensors and fMRI voxels recorded sequentially precluded us from evaluating regional differences in the correspondence between the MEG and BOLD signals; instead, we focused on overall TBOLD and TMEG means. Thus, the present study provides a rough overall estimate of the correspondence between moment-to-moment variation in neural and neurovascular signals in a resting state, an estimate that is almost identical to that reported in studies simultaneously recording neural and hemodynamic responses in animals. Finally, in light of changes in neurovascular coupling associated with brain development and neurodegeneration (4, 26), the degree to which the present findings in cognitively healthy adult women generalize across the lifespan and along a spectrum from brain health to disease is uncertain, and is an avenue for future investigation.

DATA AVAILABILITY

Data will be made available upon reasonable request.

GRANTS

Partial funding for this study was provided by the University of Minnesota (the Anita Kunin Chair in Women’s Healthy Brain Aging, the Brain and Genomics Fund McKnight Presidential Chair of Cognitive Neuroscience, the American Legion Brain Sciences Chair) and the US Department of Veterans Affairs.

DISCLAIMERS

The sponsors had no role in the current study design, analysis or interpretation, or in the writing of this paper. The contents do not represent the views of the US Department of Veterans Affairs or the US Government.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

A.P.G. conceived and designed research; P.C. and A.C.L. performed experiments; A.P.G. and P.C. analyzed data; A.P.G., P.C., A.C.L., and L.M.J. interpreted results of experiments; A.P.G. prepared figures; L.M.J. drafted manuscript; A.P.G., P.C., A.C.L., and L.M.J. edited and revised manuscript; A.P.G., P.C., A.C.L., and L.M.J. approved final version of manuscript.