Brain songs framework used for discovering the relevant timescale of the human brain

Abstract

A key unresolved problem in neuroscience is to determine the relevant timescale for understanding spatiotemporal dynamics across the whole brain. While resting state fMRI reveals networks at an ultraslow timescale (below 0.1 Hz), other neuroimaging modalities such as MEG and EEG suggest that much faster timescales may be equally or more relevant for discovering spatiotemporal structure. Here, we introduce a novel way to generate whole-brain neural dynamical activity at the millisecond scale from fMRI signals. This method allows us to study the different timescales through binning the output of the model. These timescales can then be investigated using a method (poetically named brain songs) to extract the spacetime motifs at a given timescale. Using independent measures of entropy and hierarchy to characterize the richness of the dynamical repertoire, we show that both methods find a similar optimum at a timescale of around 200 ms in resting state and in task data.

Fig. 1

General whole-brain modelling scheme for generating milliseconds time series from BOLD data. a Extracting the BOLD time series from fMRI with a typical time scale of 2 s. b These BOLD time series are generated by neural activity on the scale of milliseconds. This scale is fundamental to be able to investigate the relevant time scale of brain activity. c The transformation from slow to fast time scale is accomplished by whole-brain modelling, where we use the slow data to fit a balanced dynamic mean field model with realistic synaptic dynamics and shaped by the underlying anatomical skeleton. d The optimal working point of the model is found using the optimal global synchronisation level (red line shows the quadratic error of the difference between the empirical and simulated Kuramoto order parameters, see Methods) and fitting to the static functional connectivity (FC, black line shows the correlation between the empirical and simulated static FC matrices). At the optimal working point (corresponding to the minimum of the synchronisation fit, shown in orange box), the model generates the milliseconds time series which is used to find the relevant time scale

Fig. 2

Extraction of spacetime motifs. a In order to study the relevant time scale, we create different bin sizes of the milliseconds neural time series. The middle panel shows the data with 10, 200, 1000 ms bin sizes. These time binned time series are binarised using a point-process algorithm (shown on the right). b In order to extract the number of significant spacetime motifs, we compute the eigenvalues above the maximum of the eigenvalues of the null hypothesis distribution based on random matrix theory, following the Marčenko–Pastur distribution. c We then extract the spacetime motifs using independent component analysis (ICA) and estimate the corresponding activity, where co-activation patterns are found and used to track the activity over time. d The richness of the dynamical repertoire at different timescales can be computed from the spacetime motifs and corresponding probabilities allow using measures of entropy and hierarchy of functional brain organisation (see Methods). We show the four possible different scenarios of how this may vary with timescale whether flat, monotonic decrease or increase or having an optimum

Fig. 3

Discovering the relevant timescale of the brain. a Using the methods outlined in the two other figures, we here show the results of using these on normal human resting state activity. In particular, we show the results of using three different measures (entropy, red line; hierarchy, blue line; and mean functional connectivity, FC, orange line) on the data in different bin sizes from 10 to 3000 ms. As can be seen very clearly from the peaks in entropy and hierarchy (red and blue lines), the richness of the dynamical repertoire is found in the region of around 200 ms (light orange box). Please note that the mean FC is monotonically increasing, suggesting that this static measure is not ideal for finding the relevant time scale of the dynamic rich repertoire of brain states. We show the spacetime motifs for four timescales [very fast (10 ms), optimal (200 ms), slow (1000 ms) and very slow (2000 ms)] in terms of the b Transition Probability Matrix, c probability state space and d the patterns. At the timescale of 200 ms, it is remarkable how uniformly distributed the individual states are in terms of their probability of occurrence, which is in contrast to the other timescales of 10, 1000 and 2000 ms, where there are also fewer spacetime motifs

Fig. 4

Individual spacetime motifs at 200 ms rendered on the standard brain. Some of these brain networks resemble known resting state networks, e.g. networks 6 and 15, which correspond to the frontal part of the default mode network and the visual network, respectively. The other networks resemble sub-components and lateralised versions of the classical resting state networks, namely: 1. medial cingulate, 2. left orbitofrontal, 3. left prefrontal, 4. right prefrontal, 5. left higher order visual areas, 6. frontal DMN, 7. right parietal, 8. left parietal, 9. left auditory and insula, 10. right STG, auditory and insula, 11. right orbitofrontal, 12. left hippocampus, 13. right higher order visual areas, 14. right hippocampus, 15. visual network

Fig. 5

Comparison of resting state networks at optimal (left, 200 ms) and slow BOLD (right, 2000 ms) timescales. The figure shows the similarity of the spatial characteristics of the spacetime motifs networks found at both timescales. As an example, the visual network (bottom) and frontal part of the default mode network (top) are shown. Thus, classical methodologies used for extracting resting state networks are valid for finding the spatial components but of course less suitable for extracting the underlying temporal dynamics given the inherent temporal constraints of BOLD signals

Fig. 6

Timescale in HCP task data. We analysed HCP task data (see Methods) in order to compare the optimal timescale for resting state and neuroimaging during a social cognition task, where participants were presented with short video clips (20 s) of objects (squares, circles, triangles) that either interacted in some way, or moved randomly on the screen. We fitted the whole-brain model to the BOLD signal and plot the results of measuring entropy (red line) and hierarchy (blue line) in different bin sizes from 10 to 3000 ms. We found peaks in entropy and hierarchy of around 200 ms, which is very similar to the peak found in resting state fMRI. This means that the timescale for resting state and task condition is optimal at the same timepoint, suggesting that the 200 ms timescale is an intrinsic property of brain dynamics

Fig. 7

Timescale in MEG data. Figure shows the results of using hierarchy (left) and entropy (right) on the different delta (1–4 Hz), theta (4–8 Hz), alpha (8–12 Hz) and beta (12–30 Hz) bands of the MEG data in different bin sizes from 10 to 1600 ms, i.e. directly on the empirical data—and thus not using the whole-brain modelling part of the process used for fMRI data. The peaks in entropy and hierarchy for all bands show a peak in the region of around 200 ms, which is similar to the peak found in fMRI and thus suggest a similar richness of the dynamical repertoire