Joseph R Whittaker1, Molly G Bright1,2, Ian D Driver1, and Kevin Murphy1
1CUBRIC, School of Psychology, Cardiff University, Cardiff, United Kingdom, 2Sir Peter Mansfield Imaging Centre, University of Nottingham, Nottingham, United Kingdom
Synopsis
A pilot study of fMRI time-frequency dynamics,
characterized using a maximal overlap discrete wavelet transform, demonstrates
matched frequency correlations with beat-to-beat mean arterial blood pressure fluctuations.
Voxel-wise correlations between fMRI and blood pressure wavelet coefficients,
on a frequency scale centred at 0.1Hz, reveal distributed and structured
spatial variance across the brain. We demonstrate that functional connectivity
methods that include time-frequency representations of fMRI data are likely
very sensitive to these blood pressure fluctuations. Purpose
Researchers
are increasingly interested in the dynamics of functional connectivity (FC) brain
networks [1, 2], and wavelet
decompositions are popular and analogous to sliding-window Fourier methods for
inferring transient FC within given frequency ranges. FC methods are
particularly sensitive to physiological noise [3], and dynamic FC measures
are likely to be even more so. Although many sources are well understood and
can be accounted for, the impact of dynamic cerebral autoregulation and
vasomotion remains elusive. Given that spectral analysis reveals autonomic
blood pressure oscillations at the 0.1Hz frequency [4], we expect this to impact
time-frequency representations of fMRI time-series, and thus FC measures that
rely on them.
Methods
Two twenty-minute resting-state fMRI runs were
acquired for 4 subjects (1 female) on a GE 3T HDx scanner with a gradient-echo
EPI readout (TR/TE = 2000/35 ms;
33 slices; resolution = 3.5 × 3.5 × 4.0 mm3).
Concurrent beat-to-beat fluctuations in mean arterial pressure (MAP)
were obtained via a non-invasive MR compatible device (Caretaker, BIOPAC). Data
were motion corrected, and physiological noise fluctuations corresponding to PETCO2,
cardiac and respiratory phase (RETROICOR), heart rate, and respiration volume
(RVT) were regressed out along with 6 motion parameters.
De-trended resampled MAP traces were decomposed into 6
frequency scales using the maximal overlap discrete wavelet transform (MODWT).
We expect scale 2, which corresponds with the frequency range 0.0625-0.125Hz to
be of most interest. Average grey matter (GM) time-series were also decomposed
into the same frequency scales, and correlated with frequency matched MAP
wavelet coefficients to identify which scales explains the most variance in
fMRI data. For a voxel-wise analysis, de-trended, de-spiked, smoothed (5mm)
time-series were decomposed, and voxel-wise correlations between fMRI and MAP
wavelet coefficients of matching scales was performed. Spatial correlations
between correlation maps across sessions were calculated for each scale to
assess the reproducibility.
For a basic FC analysis, unsmoothed fMRI
time-series from 90 anatomically derived ROIs in the AAL template [5] were decomposed,
and the correlation matrix for each scale computed. For each scale, node-strength
(mean connectivity per node) was correlated with a vector of correlations
between MAP and each node. For comparison, correlation matrices, node strength
and node-MAP correlation vectors were also calculated from un-decomponsed fMRI
time-series and MAP traces.
Results
Fig.
1A,B shows a subject MAP trace and the corresponding wavelet coefficients for
the 6 frequency scales. Fig. 1C shows the group mean correlation between mean
GM fMRI signals and MAP. Scale 2 shows significant (p<0.01, Bonferroni corrected) negative
correlations. Fig. 1D shows the group mean spatial correlation between scans 1
& 2 MAP correlation maps. Scale 2 shows significant spatial correlation
between scans 1 and 2 (p<0.05, Bonferroni
corrected). Fig. 2 shows voxel-wise z-scores for the group mean correlation
coefficients between fMRI and MAP wavelet coefficients for scale 2.
Fig. 3Ai
shows the group mean correlation matrix between 90 ROIs for un-decomposed data,
alongside a vector of correlation coefficients between each ROI and MAP trace.
Fig. 3Aii shows the session mean node-MAP correlation vectors vs node-strengths
for each subject. Fig. 3Bi,ii are the same as Fig. 3A, but for scale 2 wavelet
coefficients. It is obvious from Figs 3A and 3B that far more variance in node
strength is explained by MAP when data are wavelet decomposed compared with
when they are not (group means ~48% and ~9% respectively).
Discussion
Here
we show that fluctuations in wavelet coefficients of a decomposed MAP
time-series are highly correlated in a spatially structured manner with fMRI
wavelet coefficients of matching frequency centred on 0.1Hz. We used node
strength as a simple network metric, and have demonstrated that in scale 2
wavelet coefficients, node-strength is significantly and highly inversely
correlated with node-MAP correlations. Although wavelet and dynamic FC
techniques provide more information than simple single correlation FC measures,
these data illustrate the potential confounds that can arise due to increased
sensitivity to dynamic physiological processes.
Further research on the
relationship between MAP fluctuations and fMRI time-series is required, i.e.
temporal lag, cross frequency effects etc. Furthermore, validation of the
Caretaker signal as an accurate representation of blood pressure fluctuations
is needed. However, these preliminary results are cause for concern regarding
the interpretability of dynamic FC studies and the effect of latent
physiological variables on fMRI time-frequency dynamics.
Acknowledgements
The Wellcome Trust funded this
work [WT090199]References
1. Chang,
C. and G.H. Glover, Time-frequency
dynamics of resting-state brain connectivity measured with fMRI.
Neuroimage, 2010. 50(1): p. 81-98.
2. Hutchison, R.M., et al., Dynamic functional connectivity: promise,
issues, and interpretations. Neuroimage, 2013. 80: p. 360-78.
3. Murphy, K., R.M. Birn, and P.A.
Bandettini, Resting-state fMRI confounds
and cleanup. Neuroimage, 2013. 80:
p. 349-59.
4. Julien, C., The enigma of Mayer waves: Facts and models. Cardiovasc Res, 2006. 70(1): p. 12-21.
5. Tzourio-Mazoyer, N., et al., Automated anatomical labeling of activations
in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject
brain. Neuroimage, 2002. 15(1):
p. 273-89.