Abstract
Variability of neural activity is regarded as a crucial feature of healthy brain function, and several neuroimaging approaches have been employed to assess it noninvasively. Studies on the variability of both evoked brain response and spontaneous brain signals have shown remarkable changes with aging but it is unclear if the different measures of brain signal variability – identified with either hemodynamic or electrophysiological methods – reflect the same underlying physiology. In this study, we aimed to explore age differences of spontaneous brain signal variability with two different imaging modalities (EEG, fMRI) in healthy younger (25±3 years, N=135) and older (67±4 years, N=54) adults. Consistent with the previous studies, we found lower blood oxygenation level dependent (BOLD) variability in the older subjects as well as less signal variability in the amplitude of low-frequency oscillations (1–12 Hz), measured in source space. These age-related reductions were mostly observed in the areas that overlap with the default mode network. Moreover, age-related increases of variability in the amplitude of beta-band frequency EEG oscillations (15–25 Hz) were seen predominantly in fronto-temporal and sensorimotor brain regions. There were significant sex differences in BOLD and EEG signal variability in various brain regions, but no significant interactions between age and sex were observed. Further, both univariate and multivariate correlation analyses revealed no significant associations between these two variability measures. In summary, we show that both BOLD and EEG signal variability reflect aging-related processes but are likely to be dominated by different physiological origins, which relate differentially to age and sex.
1. Introduction
Functional neuroimaging methods such as fMRI, PET, fNIRS, EEG, or MEG have allowed the non-invasive assessment of functional changes in the aging human brain (Cabeza, 2001; Cabeza et al., 2018). Most previous functional neuroimaging studies on aging have employed a task-based design (Grady, 2012) and in their data analysis the central tendency has typically been assumed to be the most representative value in a distribution (e.g., mean) (Speelman and McGann, 2013) or the “signal” within distributional “noise”. In recent years, also the variability of brain activation in task-dependent and task-independent measurements (as spontaneous variations of background activity) has been shown to provide relevant information about the brain’s functional state (Garrett et al., 2013b; Grady and Garrett, 2018; Nomi et al., 2017). These studies primarily measured the blood oxygen level dependent (BOLD) signal using fMRI. For example, it has been demonstrated that the variance of the task-evoked BOLD response was differentially related to aging as well as cognitive performance (Armbruster-Genc et al., 2016; Garrett et al., 2013a). Similarly, spontaneous signal variability in resting state fMRI (rsfMRI) has been associated with age (Grady and Garrett, 2018; Nomi et al., 2017), emotional state (state anxiety; Labrenz et al., 2018), and mental or neural disorders such as stroke (Kielar et al., 2016), Attention Deficit Hyperactivity Disorder (Nomi et al., 2018) or 22q11.2 deletion syndrome (Zöller et al., 2017). From these studies, it was concluded that reductions in BOLD signal variability might serve as an index for deficits in neural processing and cognitive flexibility (Grady and Garrett, 2014).
These conclusions of aforementioned studies imply that BOLD signal variability is mainly determined by neuronal variability. To a large extent, this is based on the premise that BOLD is related to neuronal activity: The evoked BOLD signal in task-based fMRI reflects the decrease of the deoxyhemoglobin concentration to changes in local brain activity, which is determined by vascular (blood velocity and volume: “neurovascular coupling”) and metabolic (oxygen consumption: “neurometabolic coupling”) factors (Logothetis and Wandell, 2004; Villringer and Dirnagl, 1995). The BOLD signal is therefore only an indirect measure of neural activity (Logothetis, 2008). For the variability of task-evoked BOLD signal and for spontaneous variations of the BOLD signal, in principle, the same considerations apply regarding their relationship to underlying neural processes (Murayama et al., 2010). However, since in rsfMRI there is no explicit external trigger for evoked brain activity to which time-locked averaging could be applied, the time course of rsfMRI signals is potentially more susceptible to contributions of “physiological noise”, such as cardiac and respiratory signals (Birn et al., 2008; Chang et al., 2009), but also spontaneous fluctuations of vascular tone, which is found even in isolated arterial vessels (Failla et al., 1999; Hudetz et al., 1998; Wang et al., 2006). In the same vein, the variability of task-evoked fMRI is not necessarily reflecting only the variability of evoked neuronal activity, as it may also – at least partly – reflect the variability of the spontaneous background signal on which a constant evoked response is superimposed.
In aging, non-neuronal signal fluctuations may also introduce spurious common variance across the rsfMRI time series (Caballero-Gaudes and Reynolds, 2017), thus confounding estimates of “neural” brain signal variability. Previous evidence suggests that the relationship between neuronal activity and the vascular response is attenuated with age – and so is, as a consequence, the BOLD signal (for review see D’Esposito et al., 2003). For instance, aging has been associated with altered cerebrovascular ultrastructure, reduced elasticity of vessels, and atherosclerosis (Farkas and Luiten, 2001) but also with a decrease in resting cerebral blood flow (CBF) (Ances et al., 2009; Martin et al., 1991), cerebral metabolic rate of oxygen consumption (CMRO2) (Aanerud et al., 2012), and cerebrovascular reactivity (CVR) (Liu et al., 2013). Taken together, age-related changes in BOLD signal or BOLD signal variability are related to a mixture of alterations in non-neural spontaneous fluctuations of vascular signals, neural activity, neurovascular coupling, and/or neurometabolic coupling (D’Esposito et al., 2003; Geerligs et al., 2017; Tsvetanov et al., 2015).
While BOLD fMRI signal and specifically variance measures based on fMRI are only partially and indirectly related to neural activity (Liu, 2013; Logothetis, 2008), electrophysiological methods such as EEG can provide a more direct assessment of neural activity with a higher temporal but poorer spatial resolution (Cohen, 2017). EEG measures neuronal currents resulting from the synchronization of dendritic postsynaptic potentials across the neural population; the cerebral EEG rhythms thereby reflect the underlying brain neural network activity (Steriade, 2006). Resting state (rs)EEG is characterized by spontaneous oscillations (“brain rhythms”) at different frequencies. Previously, the mean amplitude of low-frequency bands (e.g., delta and/or theta, 1-7 Hz) has been shown to correlate negatively with age (Vlahou et al., 2015), while higher-frequency bands (e.g., beta, 15-25 Hz) show the reverse pattern (Rossiter et al., 2014). However, less is known about the within-subject variability of EEG measures and their association with aging. Several studies have addressed the variability in the spectral amplitudes of different frequency bands using variance (Hawkes and Prescott, 1973; Oken and Chiappa, 1988), coefficient of variation (Burgess and Gruzelier, 1993; Maltez et al., 2004), and complexity (Fernández et al., 2012; Sleimen-Malkoun et al., 2015). For instance, reductions of the complexity in rsEEG signal have been found not only in healthy aging (Yang et al., 2013; Zappasodi et al., 2015) but also in age-related pathologies such as mild cognitive impairment (McBride et al., 2014) and Alzheimer’s disease (Smits et al., 2016). Accordingly, it has been suggested that irregular (e.g., variable) systems indicate a normal and healthy state (more integrated information) while highly regular systems often mark dysfunction or disease (Lipsitz and Goldberger, 1992; Vaillancourt and Newell, 2002).
The different methodological approaches, fMRI based “vascular” approaches on the one hand and electrophysiological methods such as EEG and MEG, on the other hand, indicate alterations of brain signal variability with aging. However, it remains unclear whether these different measures of brain variability at rest reflect the same underlying physiological changes. Evidently, there are some correlations between the two signal sources (for a review see, Jorge et al., 2014; Ritter and Villringer, 2006). For instance, in task-based EEG-fMRI simultaneous recordings, a relationship between BOLD responses and amplitude of evoked potentials has been demonstrated (e.g., Ritter et al., 2009; Seaquist et al., 2007), while in resting state EEG-fMRI studies, a negative association between spontaneous modulations of alpha rhythm and BOLD signal has also been established (e.g., Chang et al., 2013; Goldman et al., 2002; Gonçalves et al., 2006; Moosmann et al., 2003). Further, differential correlation patterns have been noted for the various rhythms of different frequencies in EEG/MEG and the fMRI signal, such that low-frequency oscillations show a negative (Deligianni et al., 2014; Mantini et al., 2007; Meyer et al., 2013), while higher frequencies oscillations demonstrate a positive correlation with the BOLD signal (Niessing et al., 2005; Scheeringa et al., 2011).
Regarding the known age-related changes in BOLD and EEG signal variability, respectively, the question arises whether these alterations are dominated by joint signal sources of fMRI and EEG or by – potentially different – signal contributions that relate to each of these two methods. Given the – potentially large – non-neuronal signal contribution, this issue is particularly relevant for rsfMRI studies. Here, we addressed this question analyzing rsfMRI and EEG measures of variability in healthy younger and older subjects. To our knowledge, the only study that compared variability in a “vascular” imaging method (rsfMRI) and an electrophysiological method (rsMEG at the sensor space) concluded that the effects of aging on BOLD signal variability were mainly driven by vascular factors (e.g., heart rate variability) and not well-explained by the changes in neural variability (Tsvetanov et al., 2015). The main aims of the present study were to explore i) age differences of brain signal variability measures, as well as to investigate ii) how neural variability derived from rsEEG related to the analogous parameters of BOLD signal variability derived from rsfMRI. We used rsfMRI and rsEEG from the “Leipzig Study for Mind-Body-Emotion Interactions” (Babayan et al., 2019). As an explanatory analysis, we further investigated sex-related differences of brain signal variability measures. To measure the brain signal variability, we calculated the standard deviation (SD) of both the BOLD signal and of the amplitude envelope of the filtered rsEEG time series for a number of standard frequency bands at the source space. We hypothesized that brain signal variability would generally decrease with aging. In addition, based on the premise that BOLD fMRI signal variability reflects neural variability as measured by rsEEG, we expected that the corresponding changes in both signal modalities would demonstrate moderate to strong similarity in their spatial distribution. Given the confounding effects of vascular factors during aging on the fMRI signal (D’Esposito et al., 2003; Liu, 2013; Thompson, 2018), we further expected to find the relationship between BOLD and EEG signal variability to be stronger in younger than older adults.
2. Method
2.1. Participants
The data of the “Leipzig Study for Mind-Body-Emotion Interactions” (LEMON; Babayan et al., 2019) comprised 227 subjects in two age groups (younger: 20-35, older: 59-77). Only participants who did not report any neurological disorders, head injury, alcohol or other substance abuse, hypertension, pregnancy, claustrophobia, chemotherapy and malignant diseases, current and/or previous psychiatric disease or any medication affecting the cardiovascular and/or central nervous system in a telephone pre-screening were invited to the laboratory. The study protocol conformed to the Declaration of Helsinki and was approved by the ethics committee at the medical faculty of the University of Leipzig (reference number 154/13-ff).
RsEEG recordings were available for 216 subjects who completed the full study protocol. We excluded data from subjects that had missing event information (N=1), different sampling rate (N=3), mismatching header files or insufficient data quality (N=9). Based on the rsfMRI quality assessment, we further excluded data from subjects with faulty preprocessing (N=7), ghost artefacts (N=2), incomplete data (N=1), or excessive head motion (N=3) (criterion: mean framewise displacement (FD) ≤ 0.5 mm; Power et al., 2012) (Supplementary Figure 1). The final sample included 135 younger (M = 25.10 ± 3.70 years, 42 females) and 54 older subjects (M = 67.15 ± 4.52 years, 27 females).
Flowchart of selecting participants from the Mind-Brain-Body study.
2.1. fMRI Acquisition
Brain imaging was performed on a 3T Siemens Magnetom Verio MR scanner (Siemens Medical Systems, Erlangen, Germany) with a standard 32-channel head coil. The participants were instructed to keep their eyes open and not fall asleep.
The structural image was recorded using an MP2RAGE sequence (Marques et al., 2010) with the following parameters: TI 1 = 700 ms, TI 2 = 2500 ms, TR = 5000 ms, TE = 2.92 ms, FA 1 = 4°, FA 2 = 5°, band width = 240 Hz/pixel, FOV = 256 × 240 × 176 mm3, voxel size = 1 x 1 x 1 mm3. The functional images were acquired using a T2*-weighted multiband EPI sequence with the following parameters: TR = 1400 ms, TE = 30 ms, FA= 69°, FOV = 202 mm, voxel size = 2.3 x 2.3 x 2.3 mm3, slice thickness = 2.3 mm, slice gap = 0.67 mm, 657 volumes, multiband acceleration factor = 4, duration = 15 min 30 s. A gradient echo field map with the sample geometry was used for distortion correction (TR = 680 ms, TE 1 = 5.19 ms, TE 2 = 7.65 ms).
2.2. fMRI Preprocessing
Preprocessing was implemented in Nipype (Gorgolewski et al., 2011), incorporating tools from FreeSurfer (Fischl, 2012), FSL (Jenkinson et al., 2012), AFNI (Cox, 1996), ANTs (Avants et al., 2011), CBS Tools (Bazin et al., 2014), and Nitime (Rokem et al., 2009). The pipeline comprised the following steps: (I) discarding the first five EPI volumes to allow for signal equilibration and steady state, (II) 3D motion correction (FSL mcflirt), (III) distortion correction (FSL fugue), (IV) rigid body coregistration of functional scans to the individual T1-weighted image (Freesurfer bbregister), (V) denoising including removal of 24 motion parameters (CPAC, Friston et al., 1996), motion, signal intensity spikes (Nipype rapidart), physiological noise in white matter and cerebrospinal fluid (CSF) (CompCor; Behzadi et al., 2007), together with linear and quadratic signal trends, (VI) band-pass filtering between 0.01-0.1 Hz (Nilearn), (VII) spatial normalization to MNI152 (Montreal Neurological Institute) standard space (2 mm isotropic) via transformation parameters derived during structural preprocessing (ANTS). (VIII) The data were then spatially smoothed with a 6-mm full-width half-maximum (FWHM) Gaussian kernel.
BOLD Signal Variability (SDBOLD)
Standard deviation (SD) quantifies the amount of variation or dispersion in a set of values (Garrett et al., 2015; Grady and Garrett, 2018). Higher SD in rsfMRI signal indicates greater intensity of signal fluctuation or an increased level of activation in a given area (Garrett et al., 2011). We first calculated SDBOLD across the whole time series for each voxel and then within 96 boundaries of preselected atlas-based regions of interests (ROIs) based on the Harvard-Oxford cortical atlas (Desikan et al., 2006). The main steps of deriving brain signal variability (SDBOLD) from the preprocessed fMRI signal are shown in Figure 1.
Main steps of deriving brain signal variability from the preprocessed resting state fMRI and EEG signal. We calculated the standard deviation of the blood oxygen level dependent (BOLD) signal and of the coarse-grained amplitude envelope of the rsEEG time series for a number of standard frequency bands at the source space.
The reproducible workflows containing fMRI preprocessing details can be found here: https://github.com/NeuroanatomyAndConnectivity/pipelines/releases/tag/v2.0.
2.3. EEG Recordings
Sixteen minutes of rsEEG were acquired on a separate day with BrainAmp MR-plus amplifiers using 61 ActiCAP electrodes (both Brain Products, Germany) attached according to the international standard 10-20 localization system (Jurcak et al., 2007) with FCz as a reference. The ground electrode was located at the sternum. Electrode impedance was kept below 5 kΩ. Continuous EEG activity was digitized at a sampling rate of 2500 Hz and band– pass filtered online between 0.015 Hz and 1 kHz.
The experimental session was divided into 16 blocks, each lasting 60 s, with two conditions interleaved, eyes closed (EC) and eyes open (EO), starting with the EC condition. Changes between blocks were announced with the software Presentation (v16.5, Neurobehavioral Systems Inc., USA). Participants were asked to sit comfortably in a chair in a dimly illuminated, sound-shielded Faraday recording room. During the EO periods, participants were instructed to stay awake while fixating a black cross presented on a white background. To maximize comparability, only EEG data from the EO condition were analyzed, since rsfMRI data were collected only in the EO condition.
2.4. EEG Data Analysis
EEG processing and analyses were performed with custom Matlab (The MathWorks, Inc, Natick, Massachusetts, USA) scripts using functions from the EEGLAB environment (version 14.1.1b; Delorme and Makeig, 2004). The continuous EEG data were down-sampled to 250 Hz, band-pass filtered within 1–45 Hz (4th order back and forth Butterworth filter) and split into EO and EC conditions. Segments contaminated by large artefacts due to facial muscle tensions and gross movements were removed following visual inspection; rare occasions of artifactual channels were excluded from the analysis. The dimensionality of the data was reduced using principal component analysis (PCA) by selecting at least 30 principle components explaining 95% of the total variance. Next, using independent component analysis (Infomax; Bell and Sejnowski, 1995), the confounding sources e.g. eye-movements, eye-blinks, muscle activity, and residual ballistocardiographic artefacts were rejected from the data.
2.5. EEG Source Reconstruction
Before conducting source reconstruction, preprocessed EEG signals were re-referenced to a common average. We incorporated a standard highly detailed finite element method (FEM) volume conduction model as described by Huang et al. (2016). The geometry of the FEM model was based on an extended MNI/ICBM152 (International Consortium for Brain Mapping) standard anatomy, where the source space constrained to cortical surface and parceled to 96 ROIs based on the Harvard-Oxford atlas (Desikan et al., 2006). The forward model was also expressed in MNI coordinates and determined using boundary element models as implemented in the M/EEG Toolbox of Hamburg (METH; Haufe and Ewald, 2016; Huang et al., 2016). The leadfield matrix was calculated between 1804 points located on the cortical surface to the 61 scalp electrodes. Source activity was estimated using exact low-resolution tomography (eLORETA; Haufe and Ewald, 2016; Pascual-Marqui, 2007). Following the singular value decomposition (SVD) of each voxel’s three-dimensional time course, the dominant orientation of the source signal was identified by preserving the first SVD component. We filtered into several frequency bands, associated with brain oscillations: delta (1–3 Hz), theta (4–8 Hz), alpha (8–12 Hz), and beta (15–25 Hz). The amplitude envelope of filtered oscillations was extracted using the Hilbert transform. Next, we applied temporal coarse graining by averaging data points in non-overlapping windows of length 0.5 s (Figure 1).
EEG Variability (SDEEG)
We calculated the SD of amplitude envelope of band-pass filtered oscillations on the coarse-grained signal. RsEEG signal variability (SDEEG) was obtained for different frequency bands (SDDELTA, SDTHETA, SDALPHA, SDBETA) in each of 96 ROIs. Main steps toward deriving brain signal variability from the preprocessed EEG signal are shown in Figure 1. The raw and preprocessed fMRI and EEG data samples can be found at https://ftp.gwdg.de/pub/misc/MPI-Leipzig_Mind-Brain-Body-LEMON/
2.6. Statistical Analyses
Mean SDBOLD and SDEEG
For the topographic information (based on ROIs), the mean BOLD and EEG variability were calculated by I) log-transforming the SD values, II) averaging separately for younger and older subjects, and III) back-transforming then the values (McDonald, 2014).
Age and Sex Effects
A series of 2×2 (age group vs. sex) analyses of variance (ANOVAs) were applied to each ROI of brain signal variability values separately for SDBOLD and SDEEG, using controlling false discovery rate (FDR) according to Benjamini and Hochberg (1995) as correction for multiple comparisons. Significant group differences were further examined by Tukey HSD post-hoc comparisons.
The signal variability values were log-transformed to normalize SDBOLD and SDEEG before further analyses (assessed by Lilliefors test at a significance level of 0.05). Analyses were performed using R (R core team, 2018).
SDBOLD – SDEEG Correlation
To investigate the association between each ROI of SDBOLD and SDEEG, we used pairwise Spearman’s rank correlation separately for younger and older subjects, corrected for FDR (96 ROIs). We further applied sparse canonical correlation analysis (CCA) to show that the relationship between SDBOLD and SDEEG is not missed when only mass bivariate correlations are used. CCA is a multivariate method to find the independent linear combinations of variables such that the correlation between variables is maximized (Witten et al., 2009). The sparse CCA criterion is obtained by adding a Lasso Penalty function (l1), which performs continuous shrinkage and automatic variable selection and can solve statistical problems such as multicollinearity and overfitting (Tibshirani, 2011). We used l1 penalty as the regularization function to obtain sparse coefficients, that is, the canonical vectors (i.e., translating from full variables to a data matrix’s low-rank components of variation) will contain exactly zero elements. Sparse CCA was performed using the R package PMA (Penalized Multivariate Analysis; Witten et al., 2009; http://cran.r-project.org/web/packages/PMA/). In our analyses, the significance of the correlation was estimated using the permutation approach (N=1000) as implemented in the CCA.permute function in R (pperm<0.05).
3. Results
Mean SDBOLD and SDEEG
The topographic distribution of SDBOLD in younger adults revealed the largest brain signal variability values in fronto-temporal regions while in older adults it was in the frontal and occipital areas. Further, we found strongest variability across younger subjects in occipito-temporal regions for SDDELTA, SDTHETA, SDALPHA, and in medial frontal brain regions for SDBETA, while older adults showed strongest brain signal variability in the fronto-central brain regions for SDDELTA, in parietal-central brain regions for SDTHETA, SDALPHA, and in medial frontal brain regions for SDBETA. The details of topographic distribution of SDBOLD and SDEEG across age groups are available at Neurovault (https://neurovault.org/collections/WWOKVUDV/).
Age and Sex Effects
The 2×2 ANOVA analyses with SDBOLD as dependent variable demonstrated that there was a significant main effect of age group in 49 ROIs in frontal, temporal, and occipital brain regions (F-values: 13.01–63.25; Figure 2). We also found a significant main effect of sex on SDBOLD in 17 ROIs, mainly in frontal and occipital brain regions (F-values: 13.19–32.04; Figure 3) but no significant interaction between age group and sex (all pFDR>0.05). Tukey HSD post-hoc analyses showed that older subjects had decreased SDBOLD compared to younger adults which were presented in both sexes (nROI=31, Supplementary Figure 3A). We further found that male subjects had higher SDBOLD, that was restricted to the groups of younger adults in most of the significant ROIs (nROI=12, Supplementary Figure 3B).
Scatter plots showing the distribution of F-values and p-values for the main effect of age group (upper row) and sex (lower row), derived from 2×2 ANOVAs on the brain signal variability values in 96 regions-of-interest (Harvard-Oxford anatomical atlas; Desikan et al., 2006). While x-axes show the F-values for the main effect of age-group, y-axes show the corresponding p-values, corrected for multiple comparisons by false discovery rates(FDR; Benjamini and Hochberg, 1995).
Number of brain regions in which brain signal variability as measured by rsfMRI or rsEEG (delta: 1–3 Hz, theta: 4–8 Hz, alpha: 8–12 Hz, beta:15–25 Hz) differed significantly depending on age and sex: 96 region of interests (ROIs), 2×2 ANOVA, corrected for multiple comparisons by false discovery rate (FDR; Benjamini and Hochberg, 1995). Group differences were further examined by Tukey HSD post hoc comparisons. The boxplots show Tukey’s post-hoc test results for the differences in brain signal variability measures between age groups (A), and sex (B) respectively.
Spatial maps of significant age group differences in SDBOLD and SDEEG. We calculated the standard deviation (SD) of the blood oxygen level dependent (BOLD) signal and of the coarse-grained amplitude envelope of the rsEEG time series for the delta (1– 3 Hz), theta (4–8 Hz), alpha (8–12 Hz), and beta (15–25 Hz) frequency bands at the source space. Statistical significance was determined using 2×2 ANOVAs corrected for multiple comparisons by false discovery rates (FDR; Benjamini and Hochberg, 1995). Blue colored areas indicate where brain signal variability was lower in older than in younger adults, while red color indicates the opposite.
Spatial maps of significant sex differences in SDBOLD and SDEEG. We calculated the standard deviation (SD) of the blood oxygen level dependent (BOLD) signal and of the coarse-grained amplitude envelope of the rsEEG time series for the delta (1– 3 Hz), theta (4–8 Hz), alpha (8–12 Hz), and beta (15–25 Hz) frequency bands at the source space. Statistical significance was determined using 2×2 ANOVAs corrected for multiple comparisons by false discovery rates (FDR; Benjamini and Hochberg, 1995). Light blue indicates that the brain signal variability was higher in male subjects as compared to female subjects, and yellow indicates the opposite.
The 2×2 ANOVAs with SDEEG as dependent variable showed significant main effects of age group in all frequency bands: SDDELTA in 17 ROIs in occipital and frontal lobes (F-values: 12.77–25.64), SDTHETA in 13 ROIs in frontal (F-values: 14.18–36.62), SDALPHA in 13 ROIs in occipital (F-values: 13.35–18.83), and SDBETA in 63 ROIs in central, fronto-temporal, and sensorimotor brain regions (F-values: 12.71–38.71), as shown in Figure 2. There were also significant main effects of sex in all frequency bands: SDDELTA in 9 ROIs in occipital (F-values: 12.57–19.17), SDTHETA in 46 ROIs in occipital and temporal (F-values: 12.51–25.95), SDALPHA in 4 ROIs in frontal (F-values: 12.69–17.40), and SDBETA in 61 ROIs in temporal, occipital, and frontal (F-values: 12.73–46.91), as shown in Figure 3. No significant interaction effects between age group and sex on SDEEG were observed in any frequency band (pFDR>0.05). Tukey HSD post-hoc analyses on SDEEG showed that older subjects had less brain signal variability, which was present in both sexes for SDDELTA (nROI=11), SDTHETA (nROI=10), and SDALPHA (nROI=11). Additionally, older adults showed higher SDBETA, driven by female subjects (nROI=39) (Supplementary Figure 3A). With regard to sex differences, post-hoc analyses showed that females had higher SDDELTA, SDTHETA, SDALPHA, and SDBETA than males. The sex differences in SDDELTA (nROI=8) were mostly pronounced in younger adults, while the effect of sex in SDTHETA (nROI=29) and SDBETA (nROI=44) were mainly presented in both age groups (p<0.05) (Supplementary Figure 3B). The graphical distribution of the F-values for the main effects of age group or sex for each ROI are shown in Supplementary Figure 2. Additional tables and boxplots showing SDBOLD and SDEEG for each frequency band and in each the 96 ROIs, split up by age group and sex, are presented in the Supplementary Tables 1–5.
Table showing the F-values for the main effect of age group (left) and sex (right) for BOLD signal variability (SDBOLD). Statistical significance was determined using 2×2 ANOVAs corrected for multiple comparisons by false discovery rates (FDR; Benjamini and Hochberg, 1995).
Table showing the F-values for the main effect of age group (left) and sex (right) for EEG signal variability (SDDELTA). Statistical significance was determined using 2×2 ANOVAs corrected for multiple comparisons by false discovery rates (FDR; Benjamini and Hochberg, 1995).
Table showing the F-values for the main effect of age group (left) and sex (right) for EEG signal variability (SDTHETA). Statistical significance was determined using 2×2 ANOVAs corrected for multiple comparisons by false discovery rates (FDR; Benjamini and Hochberg, 1995).
Table showing the F-values for the main effect of age group (left) and sex (right) for EEG signal variability (SDALPHA). Statistical significance was determined using 2×2 ANOVAs corrected for multiple comparisons by false discovery rates (FDR; Benjamini and Hochberg, 1995).
Table showing the F-values for the main effect of age group (left) and sex (right) for EEG signal variability (SDBETA). Statistical significance was determined using 2×2 ANOVAs corrected for multiple comparisons by false discovery rates (FDR; Benjamini and Hochberg, 1995).
SDBOLD – SDEEG Correlation
The correlation coefficient of pairwise associations for 96 ROIs of SDBOLD with SDDELTA, SDTHETA, SDALPHA, and SDBETA ranged in younger adults from rho=−0.200 to rho=0.223 (Figure 4A, Supplementary Table 6) and in older adults from rho=0.386 to rho=0.349 (Figure 4B, Supplementary Table 7). None of the pairwise associations between SDBOLD and SDEEG remained significant after the correction for multiple comparison corrections. Confirmatory multivariate sparse CCA further showed that correlations between SDBOLD and SDEEG across all subjects were rather low, highly sparse, and non-significant (SDDELTA; r=0.145, pperm =0.750, l1=0.367; SDTHETA; r=0.143, pperm=0.713 l1=0.7; SDALPHA; r=0.153, pperm=0.528, l1=0.1; SDBETA; r=0. 232, pperm=0.096, l1=0.633).
Spearman correlation of SDBOLD with SDEEG for each frequency band in young subjects (N=135). None of the pairwise correlations between SDBOLD and SDEEG were statistically significant.
Spearman correlation of SDBOLD with SDEEG for each frequency band in old subjects (N=54). None of the pairwise correlations between SDBOLD and SDEEG were statistically significant.
The distribution of correlation coefficients (rho) for the association between SDBOLD and SDEEG in A) younger (N=135) and B) older (N=54) adults for different frequency bands. The correlations between SDBOLD and SDEEG were tested using pairwise Spearman’s rank correlation corrected for multiple comparison by false discovery rates (FDR; Benjamini and Hochberg, 1995).
4. Discussion
Comparing healthy younger and older adults, we found widespread variability reductions in BOLD signal as well as in the amplitude envelope of delta, theta, and alpha frequency of rsEEG, whereas increased variability with aging was observed in the beta-band frequency. As a complementary analysis, we also explored sex differences and found that male subjects exhibited higher BOLD signal variability, while the sex differences in the rsEEG variability showed the opposite pattern. There were no significant correlations between hemodynamic (SDBOLD) and electrophysiological (SDEEG) measures of brain signal variability, neither in the younger nor in the older adults. Our results suggest that variability measures of rsfMRI and rsEEG – while both related to aging – are dominated by different physiological origins and relate differently to age and sex.
4.1. BOLD Signal Variability
The first aim of our study was to investigate the effect of age on BOLD signal variability, as measured by SD of spontaneous fluctuations during rsfMRI. Consistent with recent rsfMRI studies demonstrating that BOLD signal variability decreases with age in large-scale networks (Grady and Garrett, 2018; Nomi et al., 2017), we found that older subjects had reduced SDBOLD in temporal and occipital brain regions but also in cortical midline structures like the precuneus, anterior and posterior cingulate cortices, as well as orbitofrontal cortex compared to younger adults. These age-related reductions in BOLD signal variability were thus especially apparent in regions of the Default Mode (DMN) and the Fronto-Parietal Network (FPN). The DMN is an intrinsically correlated network of brain regions, that is particularly active during rest or fixation blocks (Biswal et al., 2010). It reflects the systematic integration of information across the cortex (Margulies et al., 2016) and has been frequently associated with psychological functions like self-referential thought or mind-wandering, and also memory retrieval (Andrews-Hanna et al., 2014; Raichle, 2015). The FPN is involved in cognitive control processes (Spreng et al., 2013), and closely interacts with the DMN, for example during mind-wandering state (Golchert et al., 2017). Previous studies in healthy aging noted that older subjects showed lower connectivity as well as reduced network modularity and functional segregation in DMN and FPN regions (Damoiseaux, 2017; Damoiseaux et al., 2008; Meunier et al., 2009; Petersen et al., 2014). Similarly, an altered functional connectivity in the DMN has been found in different pathologies, for example, in Alzheimer’s disease (Greicius et al., 2004) or mild cognitive impairment (Das et al., 2015). One can speculate that decreased BOLD signal variability in the DMN and the FPN, particularly in the overlapping frontal brain regions, may be related to reduced modularity and integrity in the aging brain, that may reflect functional alterations involved in cognitive processes (Campbell et al., 2012). Therefore, characterizing BOLD signal variability and its transitions in age-related health conditions promises to further our understanding of basic neurocognitive functioning in aging.
In our exploratory analysis of sex differences, we found that male subjects exhibited higher BOLD signal variability in frontal, temporal, and occipital regions than female subjects. Sex-specific differences in brain structure and function have been previously shown (for a review see, Gong et al., 2011; Ruigrok et al., 2014; Sacher et al., 2013). For example, larger total brain volume has been reported in male as compared to female subjects (Gong et al., 2011), whereas higher cerebral blood flow (Gur et al., 1982; Rodriguez et al., 1988) and stronger functional connectivity in the DMN (Tomasi and Volkow, 2012) were found in females than males. Further, a cross-sectional study has demonstrated sex differences in functional connectivity in large-scale resting state networks in aging (Scheinost et al., 2015). Although the relationship between BOLD signal variability and functional connectivity has not been extensively examined, it has been suggested that they may capture similar functional properties (Grady and Garrett, 2018; Leo et al., 2012; Nomi et al., 2017). Assuming this relationship to be genuine, our results correspond to previous studies showing that males have higher functional connectivity strength in occipital, temporal and parietal regions than females (Biswal et al., 2010; Filippi et al., 2012; Ritchie et al., 2018).
4.2. Electrophysiological Signal Variability
Measures of neural variability were derived from rsEEG for several main frequency bands (delta, theta, alpha, beta) as the standard deviation of their amplitude of envelope time series data, analogously to the BOLD signal variability. Multimodal imaging studies have shown that the amplitude envelope of neural oscillatory activity across frequency bands relates to different rsfMRI networks (Brookes et al., 2011; Deligianni et al., 2014), confirming the neurophysiological origin of the resting state networks measured with BOLD fMRI. Additionally, these studies also concluded that different frequency bands can be related to the same functional network, but also differentially to distinct networks (Brookes et al., 2011; Laufs et al., 2006; Mantini et al., 2007; Meyer et al., 2013). For instance, Mantini et. al. (2007) reported that the visual network is associated with all frequency bands with the exclusion of the gamma rhythm, while the sensorimotor network is primarily associated with beta-band oscillations. We found age-dependent EEG signal variability changes within networks which were associated with more than one frequency band, thus confirming that neurons generating oscillations at different frequencies may contribute to the same network. More precisely, we found age-related reductions in SDDELTA and SDALPHA mainly in a visual network (including calcarine regions, cuneal cortex, and occipital pole), SDTHETA in posterior DMN, while an enhancement of SDBETA was mainly seen in the fronto-temporal and sensorimotor networks. These results align with previous reports of age-dependent changes of electrophysiological activity using spectral power (Dustman et al., 1993; Vlahou et al., 2015), and signal variability (Dustman et al., 1999; Tsvetanov et al., 2015). For instance, age-related decreases of alpha amplitude and alpha band variability (measured by SD of the oscillatory signal) were found in posterior and occipital brain regions (Babiloni et al., 2006; Tsvetanov et al., 2015). Alpha rhythm is a classical EEG hallmark of resting wakefulness (Laufs et al., 2003) that is modulated by thalamo-cortical and cortico-cortical interactions (Bazanova and Vernon, 2014; Goldman et al., 2002; Lopes Da Silva et al., 1997; Moosmann et al., 2003). It has been suggested that the posterior alpha-frequency plays an important role in the top-down control of cortical activation and excitability (Klimesch, 1999). Accordingly, decreased alpha variability in occipital regions might be associated with altered functioning of the cholinergic basal forebrain, affecting thalamo-cortical and cortico-cortical processing. Our finding of an higher fronto-temporal and sensorimotor SDBETA in the elderly is in line with previous findings (Rossiter et al., 2014; Tsvetanov et al., 2015). Aging has previously been associated with an increase in movement-related beta-band attenuation, suggesting an enhanced motor cortex GABAergic inhibitory activity in older individuals (Rossiter et al., 2014). Similarly, beta-band activity is thought to play a key role in signaling maintenance of the status quo of the motor system, despite the absence of movement (Engel and Fries, 2010). Therefore, greater SDBETA in sensorimotor brain regions could be interpreted as a compensatory mechanism to account for a decline of motor performance during aging (Quandt et al., 2016).
In addition to the effect of age on rsEEG signal variability, an exploratory analysis showed sex differences in distinct brain regions and EEG frequencies. More precisely, we found higher SDDELTA in occipital, SDTHETA in occipito-temporal, SDALPHA in frontal, and SDBETA in frontal as well as occipito-temporal brain regions in female compared to male subjects. Previously, higher alpha (Aurlien et al., 2003), theta (Duffy et al., 1993), and beta power (Jaušovec and Jaušovec, 2010; Matsuura et al., 1985; Veldhuizen et al., 1993) have been reported in female relative to male subjects, while the reverse pattern was found in delta power (Zappasodi et al., 2006). Notably, different EEG frequencies were related to distinct hormonal level fluctuations (Becker et al., 1982; Solis-Ortiz et al., 1994). For instance, reduced absolute theta and alpha power have been reported during the preovulatory phase while increased absolute beta power has been shown during and after the menstrual phase (Solis-Ortiz et al., 1994). Based on this evidence, one could speculate that the sex differences in rsEEG signal variability are related to hormonal variations.
4.3. The Association between BOLD and EEG Variability
We further assessed how neural variability in source-reconstructed rsEEG related to the analogous parameters of BOLD signal variability in rsfMRI using univariate and multivariate correlation analyses. Previously, simultaneous EEG-fMRI studies have shown meaningful relationships between fluctuations in EEG power, frequency, phase, and local BOLD changes (for a review see, Jorge et al., 2014; Ritter and Villringer, 2006). Due to age-related physiological (particularly cardiovascular) alterations in the brain, we expected the relationship between BOLD and EEG signal variability to be stronger in younger than older adults. However, in the present study, both univariate and multivariate analyses showed no significant correlations between SDBOLD and SDEEG neither in the younger nor in the older adults. This finding was supported by the distinct anatomical distributions of age-related changes in BOLD and EEG signal variability, that barely showed a spatial overlap, suggesting different underlying physiological processes. The precise nature of these physiological processes, however, remains speculative, but it seems likely that they include both neuronal and vascular components. The former are likely to be dominant for EEG- and MEG-based variability measures. However, BOLD signal variability seems to reflect both vascular and neural processes (Garrett et al., 2017). The vascular factors in the elderly are, among other things, related to the known morphological changes of blood vessels and metabolic changes with aging which are reflected in CBF (Ances et al., 2009; Martin et al., 1991), CMRO2 (Aanerud et al., 2012), and CVR (Liu et al., 2013). While neuronal components in the rsfMRI BOLD signal are also likely to be relevant (Garrett et al. 2017), the main finding of our study – little correlation with measures of EEG variability – indicates that these neuronal contributions do not dominate BOLD variability measures. While it cannot be excluded that our EEG-based variability measures reflect different aspects of neuronal function than BOLD variability, given these data, it would be desirable to perform – in future studies – concurrent electrophysiological and vascular neuroimaging for a comprehensive assessment of neuronal as well as vascular factors related to aging.
5. Limitations
There are several limitations of our study: EEG and MRI scans were not recorded simultaneously. Therefore, we could not directly relate the two signals in a cross-correlation analysis. Furthermore, EEG and MRI were performed with different body postures (fMRI; supine, EEG; seated) known to affect brain function, for example, changes in the amplitude of the EEG signal have been related to different body postures presumably due to the shifts in cerebrospinal fluid layer thickness (Rice et al., 2013). Moreover, subjects were instructed to lay or sit calm during the recording and not to think of anything particular. However, the participants’ actual mental states at rest cannot be controlled or measured accurately, thereby, there might have been differences between the fMRI and EEG recordings. Yet, resting state measures of EEG (Näpflin et al., 2007) and fMRI (Shehzad et al., 2009; Zuo et al., 2010) have been shown to be reliable within-individuals across time. Thus, it is unlikely that there were systematic differences during resting state due to different timing of EEG and fMRI experiments. Finally, the computation of the source reconstructed rsEEG required the parcellation of the brain into relatively large anatomical ROIs. It could well be that the analysis with a higher spatial resolution (e.g., at the voxel-level) may provide additional insights about brain signal variability.
6. Conclusion
In this study, we report age and sex differences of brain signal variability obtained with rsfMRI and rsEEG from the same subjects. We demonstrate extensive age-related reduction of SDBOLD, SDDELTA, SDTHETA, and SDALPHA mainly in the DMN and the visual network, while a significant increase of SDBETA was seen in fronto-temporal and sensorimotor brain regions. We could not demonstrate significant associations between SDBOLD and SDEEG. Our findings indicate that measurements of BOLD and EEG signal variability, respectively, are likely to stem from different physiological origins and relate differentially to age and sex. While the two types of measurements are thus not interchangeable, it seems, however, plausible that both markers of brain variability may provide complementary information about the aging process.
7. Funding
This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.
8. Acknowledgements
We gratefully acknowledge the Mind-Body-Emotion group at the Max Planck Institute for Human Cognitive and Brain Sciences.
Footnotes
Abbreviations: BOLD – Blood Oxygenation Level Dependent; CBF – cerebral blood flow; CBV– cerebral blood volume; CCA – canonical correlation analysis; CMRO2 – cerebral metabolic rate of oxygen; CVR – cerebrovascular reactivity; DMN – Default Mode Network; EEG – electroencephalography; EC – eyes closed; EO – eyes open; FDR - false discovery rate; FEM – finite element method; fMRI – functional Magnetic Resonance Imaging; fNIRS – functional Near-Infrared Spectroscopy; FWHM – full-width half-maximum; ICBM – International Consortium for Brain Mapping; MEG – magnetoencephalography; MNI – Montreal Neurological Institute; rho – Spearman’s rank correlation coefficient; PET –Positron-emission tomography; ROI – regions of interests; rs – resting state; SD – standard deviation; SVD – Singular Value Decomposition
http://fcon_1000.projects.nitrc.org/indi/retro/MPI_LEMON.html