BOLD fractional contribution to resting-state functional connectivity above 0.1 Hz

Abstract

Blood oxygen level dependent (BOLD) spontaneous signals from resting-state (RS) brains have typically been characterized by low-pass filtered timeseries at frequencies ≤ 0.1 Hz, and studies of these low-frequency fluctuations have contributed exceptional understanding of the baseline functions of our brain. Very recently, emerging evidence has demonstrated that spontaneous activities may persist in higher frequency bands (even up to 0.8 Hz), while presenting less variable network patterns across the scan duration. However, as an indirect measure of neuronal activity, BOLD signal results from an inherently slow hemodynamic process, which in fact might be too slow to accommodate the observed high-frequency functional connectivity (FC). To examine whether the observed high-frequency spontaneous FC originates from BOLD contrast, we collected RS data as a function of echo time (TE). Here we focus on two specific resting state networks - the default-mode network (DMN) and executive control network (ECN), and the major findings are fourfold: (1) we observed BOLD-like linear TE-dependence in the spontaneous activity at frequency bands up to 0.5 Hz (the maximum frequency that can be resolved with TR=1s), supporting neural relevance of the RSFC at a higher frequency range; (2) conventional models of hemodynamic response functions must be modified to support resting state BOLD contrast, especially at higher frequencies; (3) there are increased fractions of non-BOLD-like contributions to the RSFC above the conventional 0.1 Hz (non-BOLD/BOLD contrast at 0.4-0.5 Hz is ~4 times that at <0.1 Hz); and (4) the spatial patterns of RSFC are frequency-dependent. Possible mechanisms underlying the present findings and technical concerns regarding RSFC above 0.1 Hz are discussed.

Keywords: Above 0.1Hz; BOLD-like TE-dependence; High frequency; Resting state functional connectivity.

Fig. 1

ROI atlases of the DMN and ECN (network templates downloaded from http://findlab.stanford.edu/functional_ROIs.html).

Fig. 2

Functional connectivity signal amplitudes vs. TE. Statistics shown here are standard linear model vs. Gaussian noise tests; lack-of-fit tests were also performed and failed to reject the linear hypothesis for all frequency bands (not shown). To account for the inter-subject variability, each subject’s correlated signal amplitudes were first normalized (divided by the mean correlated amplitude across different TEs of B0 band) before group fitting.

Fig. 3

Ratios of fitted intercepts (non-BOLD contrast) over slopes (BOLD contrast), suggesting increased importance of non-BOLD contributions at higher frequencies; see Table 2 for descriptions of ‘Voxel-pairs’, ‘ROI-pairs’, ‘Reproducibility’.

Fig. 4

(A) Conventional HRFs and (B) frequency responses; (C) Simulated RS-HRFs and (D) frequency responses; (E, F) BOLD signal amplitudes (fitted slopes, BOLD-like contrast only) vs. frequency estimated from experimental data for DMN and ECN, respectively, normalized by the maximum amplitude of subject mean. As predicted, BOLD signal amplitudes dropped dramatically with frequency, but were still nonzero at the highest frequency (0.5 Hz).

Fig. 5

DMN/ECN patterns of representative subjects at <0.1 Hz (upper rows) and 0.2~0.4 Hz (bottom rows) (|r|>0.2, p<0.002, uncorrected). The power spectra of the chosen network seed signals at (<0.1 Hz) and (0.2~0.4 Hz) are highlighted in blue and red separately. Numbers in the parenthesis indicate the scan trial for subjects who participated twice.

Fig. 6

(A) The POV (percentage of overlapped voxels) similarity index between the high-frequency RSFC pattern (0.2~0.4 Hz, thresholded at r = 0.15) and the low-frequency RSFC pattern (<0.1 Hz, thresholds were varied from 0 to 0.8). Each dashed blue line represents the result of one single scan. Dark line presents the mean; (B) The number of voxels with significant positive/negative correlations with PCC (p < 0.05, |r|>0.13, uncorrected) at different frequency bands, the numbers of voxels significantly anti-correlated with PCC (blue) are multiplied by ten to contrast those associated with positive correlations (red) on the same legend scale; error bars are the standard deviations across subjects.

Fig. 7

(A) Correlated signal amplitudes (averaged across ROI signal pairs) and noise amplitudes (root mean square of the uncorrelated residuals, also estimated from ROI signal pairs) at different frequency bands; each dot represents the result from a single scan (scans with TE = 25 and 30ms are displayed). (B) SNR as a function of frequency bands.

Fig. 8

The influence of heterogeneous HRFs within a RS network on the frequency specificity of the exhibited network patterns. (A) Basis functions employed in HRF fitting: canonical HRF (blue), temporal derivative (green), dispersive derivative (red); (B) ROIs within the DMN (sub01 (2)) chosen for simulations; (C) Estimated HRFs and the normalized frequency responses; (D) Frequency specificity of signal intensity patterns (un-normalized response power of estimated HRFs in (C) integrated within the corresponding frequency bands); (E) Disparate correlation patterns of simulated signals (‘Simulated signals‘) (shown by ‘Correlation matrix’) with stimulus input (‘Events’) given at different frequency scales.