ArticlePDF Available

Discovery of key whole-brain transitions and dynamics during human wakefulness and non-REM sleep


Abstract and Figures

The modern understanding of sleep is based on the classification of sleep into stages defined by their electroencephalography (EEG) signatures, but the underlying brain dynamics remain unclear. Here we aimed to move significantly beyond the current state-of-the-art description of sleep, and in particular to characterise the spatiotemporal complexity of whole-brain networks and state transitions during sleep. In order to obtain the most unbiased estimate of how whole-brain network states evolve through the human sleep cycle, we used a Markovian data-driven analysis of continuous neuroimaging data from 57 healthy participants falling asleep during simultaneous functional magnetic resonance imaging (fMRI) and EEG. This Hidden Markov Model (HMM) facilitated discovery of the dynamic choreography between different whole-brain networks across the wake-non-REM sleep cycle. Notably, our results reveal key trajectories to switch within and between EEG-based sleep stages, while highlighting the heterogeneities of stage N1 sleep and wakefulness before and after sleep.
Sensitivity and specificity of HMM states and dynamics within polysomnography stages. a Fractional occupancies of each of the 19 HMM states computed within the four PSG stages corresponded to the PSG-sensitivity of the whole-brain network states. The coloured bars and error bars show the average and standard error, respectively, across the 18 participants that included all four PSG stages. b PSG-specificity of the HMM states for each of the four PSG stages. Specificity corresponds to the probability of an HMM state occurring within a PSG stage. The bars represent the group average and the error bars the standard error (n = 18). In a and b horizontal lines show significant differences within HMM states, with p values < 0.01 as evaluated through paired t tests and permutation testing. c The mean life times of the 19 HMM states are shown by the bars, representing values averaged across the 18 participants. Error-bars represent the standard error across participants. Each HMM state is coloured according to the probability of finding it within each of the four PSG stages, i.e., their PSG specificity. Note how HMM states with high specificity for N3—and to a lesser extent N2—exhibit longer mean life times. d The dynamics of the HMM transitions were calculated within each of the four PSG stages, in terms of switching frequency (‘Switching’), and e the number of different HMM states visited per time (‘Range of HMM states’). These measures significantly separate the four PSG stages suggestive of a higher dynamical repertoire during wakefulness and N1. In d and e error bars represent standard error across participants and significant differences between PSG stages are denoted by stars: one star: p < 0.05, two stars: p < 0.01 and three stars p < 0.001; all evaluated using paired t tests and permutations. W: wakefulness, N1: N1 sleep, N2: N2 sleep, N3: N3 sleep
This content is subject to copyright. Terms and conditions apply.
Discovery of key whole-brain transitions and
dynamics during human wakefulness and non-REM
A.B.A. Stevner 1,2,3, D. Vidaurre 4, J. Cabral 1,5, K. Rapuano6, S.F.V. Nielsen7, E. Tagliazucchi8,9,10,
H. Laufs 9,10, P. Vuust3, G. Deco11,12,13,14, M.W. Woolrich4, E. Van Someren8,15 & M.L. Kringelbach1,2,3,5
The modern understanding of sleep is based on the classication of sleep into stages dened
by their electroencephalography (EEG) signatures, but the underlying brain dynamics remain
unclear. Here we aimed to move signicantly beyond the current state-of-the-art description
of sleep, and in particular to characterise the spatiotemporal complexity of whole-brain
networks and state transitions during sleep. In order to obtain the most unbiased estimate of
how whole-brain network states evolve through the human sleep cycle, we used a Markovian
data-driven analysis of continuous neuroimaging data from 57 healthy participants falling
asleep during simultaneous functional magnetic resonance imaging (fMRI) and EEG. This
Hidden Markov Model (HMM) facilitated discovery of the dynamic choreography between
different whole-brain networks across the wake-non-REM sleep cycle. Notably, our results
reveal key trajectories to switch within and between EEG-based sleep stages, while high-
lighting the heterogeneities of stage N1 sleep and wakefulness before and after sleep. OPEN
1Department of Psychiatry, University of Oxford, Warneford Hospital, OX3 7JX Oxford, UK. 2Center of Functionally Integrative Neuroscience (CFIN), Aarhus
University, 8000 Aarhus, Denmark. 3Center for Music in the Brain (MIB), Department of Clinical Medicine, Aarhus University, 8000 Aarhus, Denmark.
4Wellcome Centre for Integrative Neuroimaging, Oxford Centre for Human Brain Activity (OHBA), University of Oxford, Warneford Hospital, OX3 7JX
Oxford, UK. 5Life and Health Sciences Research Institute (ICVS), School of Health Sciences, University of Minho, 4710-057 Braga, Portugal. 6Department of
Psychological and Brain Sciences, Dartmouth College, 03755 Hanover, NH, USA. 7Department of Applied Mathematics and Computer Science, Technical
University of Denmark, 2800 Kgs. Lyngby, Denmark. 8Netherlands Institute for Neuroscience, 1105 BA Amsterdam, The Netherlands. 9Department of
Neurology, University Hospital Schleswig Holstein, Christian-Alrbrechts-Universität, 24105 Kiel, Germany. 10 Department of Neurology and Brain Imaging
Center, Goethe University, 60528 Frankfurt am Main, Germany. 11 Center for Brain and Cognition, Computational Neuroscience Group, Department of
Information and Communication Technologies, Universitat Pompeu Fabra, Roc Boronat 138, Barcelona 08018, Spain. 12Institució Catalana de la Recerca i
Estudis Avançats (ICREA), Passeig Lluís Companys 23, Barcelona 08010, Spain. 13 Department of Neuropsychology, Max Planck Institute for Human
Cognitive and Brain Sciences, 04103 Leipzig, Germany. 14 School of Psychological Sciences, Monash University, Melbourne, Clayton, VIC 3800, Australia.
15 Departments of Integrative Neurophysiology and Psychiatry GGZ-InGeest, Amsterdam Neuroscience, VU University and Medical Center, 1081 HV
Amsterdam, The Netherlands. These authors jointly supervised this work: D. Vidaurre, M. L. Kringelbach. Correspondence and requests for materials should
be addressed to A.B.A.S. (email:
NATURE COMMUNICATIONS | (2019) 10:1035 | | 1
Content courtesy of Springer Nature, terms of use apply. Rights reserved
The primary behavioural observation of sleep is a lack of
interaction with, and responsiveness to, the external world,
i.e., a decreased level of arousal1. The lack of commu-
nication with sleeping subjects implies that we rely on physiolo-
gical recordings to scientically describe and categorise sleep. The
advent of modern neuroimaging techniques and network analyses
has been explored to map and characterise spontaneous large-
scale brain activity during wakefulness with high-spatiotemporal
precision. Yet, our understanding of brain activity during sleep
remains dictated by observations in a few channels of electro-
encephalographic (EEG) recordings.
Today, the dominant description of normal human sleep is
represented by polysomnography (PSG), which relies mainly on
EEG but also electromyography (EMG), electrooculography
(EOG) and electrocardiography (ECG), as well as measures of
respiration2. On-going brain activity is recorded from a low
number of EEG electrodes and typically categorised into wake-
fulness, rapid-eye movement (REM) sleep andaccording to the
most recent set of guidelinesthree stages of non-REM (NREM)
sleep (N1N3)2. Staging is based on the visual detection of
spectral EEG qualities (e.g., alpha- and delta-frequency power)
and sleep graphoelements (sleep spindles and K-complexes),
many of which have been known since the 1930s3.
PSG has been essential in the development of modern sleep
research, and remains undoubtedly the quickest and easiest way
to establish arousal levels in individuals. Indeed, PSG-dened
sleep stages were originally devised from EEG as surrogate
markers of arousal thresholds, yet over time many have come to
see them as a more or less exhaustive set of intrinsic canonical
states that cover the full repertoire of brain activity during sleep.
However, the use of (1) xed scoring windows of 30 s and (2)
only a few EEG electrodes means that PSG involves considerable
averaging of brain activity in both time and space4arguably
leading to an incomplete representation of brain activity.
Furthermore, PSG corresponds relatively poorly to the sub-
jective perception of sleep. Participants may experience being
awake during periods with EEG signals otherwise fullling PSG
criteria of NREM sleep5,6. The relative lack of correspondence
between PSG and subjective experience becomes important in
populations with sleep complaints, where PSG is not indicated in
the clinical evaluation of insomnia, the most common of all sleep
Recent developments in whole-brain neuroimaging and ana-
lyses support the examination of more sophisticated features of
brain networks through functional connectivity (FC) and struc-
tural connectivity analyses, the detection of task-related and
resting-state functional networks9,10, and the development of
mechanistic computational models11,12. Yet, studies that have
applied these promising tools to investigate large-scale brain
activity of sleep have commonly relied upon PSG in a strict sense,
thus regressing PSG stages onto functional brain data. This
approach has yielded whole-brain correlates of PSG stages and
sleep graphoelements, in terms of activation maps13,14, FC pat-
terns1519, graph-theoretical measures20,21 and EEG-
microstates22. However, this top-down constraint by the low-
resolution PSG scoring comes at the cost of exploring only a small
fraction of the information available in the high-resolution neu-
roimaging data.
Rather than constraining analyses by traditional denitions of
sleep stages, we propose to use novel data-driven analysis
methods to elucidate whole-brain networks that can complement
and potentially expand the classical understanding of sleep. This
requires a sufciently sensitive decomposition of whole-brain
network activity in time. Building on a recent study showing that
individual PSG stages can be extracted from functional magnetic
resonance imaging (fMRI) recordings in a data-driven way23,we
here leveraged the full spatiotemporal resolution of blood-
oxygen-level dependent (BOLD) signals to nd large-scale net-
works in sleep, applying a Hidden Markov Model (HMM)24 on
fMRI recordings of 57 healthy participants, whoaccording to
simultaneously acquired EEGcycled through PSG-dened
stages of wakefulness and NREM sleep. Crucially, the HMM
decomposition was not constrained by PSG stages, but rather
allowed us to discover directly from the data, at a time-scale of
seconds, the relevant brain network transitions explored by the
human brain during the wake-NREM sleep cycle. Compared to
other methods for extracting dynamic FC25, the HMM frame-
work explicitly models the transition probabilities between its
inferred states. We show that this information can be used to
discover new whole-brain aspects of sleep, complementing the
traditional segmentation of brain activity offered by PSG.
Whole-brain network states identied by HMM. In order to
extract the large-scale networks inherent to whole-brain record-
ings of the wakefulness-NREM sleep cycle, we estimated an
HMM on fMRI data from 57 healthy participants (age 23.5 ±
3.3 years, 39 females). Participants were instructed to lie still in
the scanner with their eyes closed. Each recording had a duration
of 52 min, and was accompanied by acquisitions of EEG, EMG,
ECG and EOG, based on which PSG staging was performed by an
expert, according to the AASM criteria2(see Supplementary
Table 1). Following preprocessing, the voxel-wise BOLD time-
courses were temporally averaged over 90 region-of-interest
(ROI) timecourses, using the cortical and subcortical regions of
the automated anatomical labelling (AAL) atlas26. ROI time-
courses were demeaned and variance-normalised for each parti-
cipant, and subsequently concatenated across participants along
the temporal dimension.
The estimated HMM contained a set of whole-brain network
states, each dened as a multivariate Gaussian distribution,
including: (i) a mean activation distribution, representing the
mean level of activity in each ROI when a state is active; and (ii)
an FC matrix, summarising the pairwise temporal co-variations
occurring between the ROIs during that state. The HMM also
contained a transition probability matrix with the probabilities of
transitioning between each pair of states. Each state also had an
associated state timecourse describing the points in time (dened
by the fMRI sampling, TR =2.08 s) where the state was
active24,27. The HMM was endowed with 19 states, and, crucially,
was given no information about the PSG staging for its
estimation. An illustration of the analysis workow is given in
Fig. 1(see Methods for details).
To allow for unbiased within-participant testing, when
comparing the HMM output to the PSG scoring we considered
the subset of the HMM output that corresponded to the data
from the 18 participants that reached all four PSG stages
(wakefulness, N1, N2 and N3, see Supplementary Table 1).
Whole-brain network states underlie PSG sleep stages. The 19
whole-brain network states, inferred solely from the fMRI by the
HMM, contained most of the temporal information given by
the PSG stages that were scored from the EEG independently of
the fMRI. The HMM state timecourses and the PSG scoring are
plotted together in Fig. 2a for the 18 participants that included all
four PSG stages, illustrating how the activity of the different
HMM states co-varied with specic PSG stages. We have high-
lighted the HMM state timecourses of two participants to ease
visual inspection, however, the temporal relationships between
HMM states and PSG stages were consistent across the group. It
can be observed, for example, that HMM state 8 occurred most
2NATURE COMMUNICATIONS | (2019) 10:1035 | |
Content courtesy of Springer Nature, terms of use apply. Rights reserved
often during wakefulness, HMM state 3 occurred during N2 sleep
and HMM state 16 occurred during N3 sleep.
We quantied the temporal association between the PSG stages
and the HMM state timecourses using multivariate analysis of
variance (MANOVA). This allowed us to ask if the 19 HMM
states were signicantly grouped in time by the four PSG stages
(for the 18 participants that included all four PSG stages).
Through non-parametric testing (see Methods) we conrmed this
temporal relationship (p<0.05, permutation testing, see Supple-
mentary Figure 1b). The MANOVA placed the PSG stages in the
02515 20105
ROI time courses
Group level
Nsubjects x Nsamples
Nsubjects x Nsamples
Principal component #
Variance explained (%)
100 90%
HMM states
time courses
Single participant
k M T
k MT
Back projection
–0.3 0 0.3
–0.8 0.80
HMM state 1
mean activity,
Back projection
HMM state 1
Voxel time courses
BOLD data
Hidden Markov model
Large-scale dynamic states
Fig. 1 Dynamic whole-brain networks from fMRI sleep recordings using a Hidden Markov Model. aROI timecourses were extracted by averaging BOLD
signals across voxels within each of the 90 cortical and subcortical AAL areas for each participant. Each ROI timecourse was demeaned and normalised by
its standard deviation. bThe data were concatenated across participants, and the dimensionality was reduced using PCA (principal component analysis),
such that ~90% of the variance of the ROI timecourses was retained. The HMM was run on the PCA timecourses, resulting in Knumber of states with
associated timecourses, each describing the points in time each state is active and inactive. cEach HMM state was characterised by a multivariate
Gaussian distribution comprising a covariance matrix, Σ
, and a mean distribution, μ
The state-specic mean distributions and covariance matrices were
back-projected to the MNI space of the AAL by using the mixing matrix, MTfrom the PCA decomposition, yielding a mean activation map and an FC matrix
for each HMM state
NATURE COMMUNICATIONS | (2019) 10:1035 | | 3
Content courtesy of Springer Nature, terms of use apply. Rights reserved
space of the HMM state timecourses, resulting in the clustering
dendrogram of Fig. 2b, with wakefulness and N1 sleep sig-
nicantly separated from N2 sleep, which in turn was further
separated from N3 sleep.
Whole-brain network states track different PSG stages. Next,
we examined the contribution of the individual whole-brain
network states to the multivariate relationship, established above,
between the HMM and the PSG scoring. We quantied the
temporal sensitivity and specicity of the HMM states for each of
the PSG stages. For each of the 18 participants that included all
four PSG stages, we dened the sensitivity of each HMM state as
the proportion of total time spent in a PSG stage, in which this
HMM state was active. Specicity was dened as the likelihood of
nding each HMM state active during a given PSG stage. We
compared the sensitivity and specicity for each PSG stage within
each of the HMM states, using paired ttests and a randomisation
scheme of the PSG scoring (see Methods). The results are pre-
sented in Fig. 3a, b. HMM state 8 occupied a large proportion of
PSG-scored wakefulness, i.e., it exhibited high sensitivity for
wakefulness (see Fig. 3a). Since this whole-brain network state
was signicantly more sensitive to wakefulness than to any of the
other PSG stages, i.e., it rarely occurred outside of wakefulness, its
specicity for wakefulness was also high (see Fig. 3b). This
combined sensitivity and specicity for wakefulness was also
found for HMM states 10 and 18.
5 7 15 16 17 21 23 27 34 36 38 39 41 42 43 47 52 56
HMM state timecourses
PSG sleep
3120 s
HMM states and PSG sleep stages
N1 W N2 N3
MANOVA between
HMM and PSG scoring
a b
Fig. 2 State timecourses of whole-brain network states and their association to polysomnography. aThe gure shows the 19 HMM state timecourses
describing each states probability of being active at each sample point of the fMRI sessions in the 18 participants that reached all four PSG stages. Below
the HMM state timecourses are shown the independently obtained PSG sleep scoring (based on the simultaneously acquired EEG). The coloured overlay
shows periods scored as wakefulness (red), N1 (white), N2 (blue) and N3 (green). The two dashed boxes highlight the HMM state timecourses and PSG
scoring of two representative participants. Note, how the majority of HMM timecourses varied with the PSG stages, in highly consistent ways across
participants. A few sporadicHMM states, occurring mainly in a few participants, are also visible (e.g., states 11 and 12). bQuantifying the multivariate
relationship between the HMM states and the PSG scoring, through the use of MANOVA, revealed a hierarchical grouping of the HMM states, in which
wakefulness and N1 sleep were separated from N2 sleep, which in turn was separated from N3
4NATURE COMMUNICATIONS | (2019) 10:1035 | |
Content courtesy of Springer Nature, terms of use apply. Rights reserved
10 11 12 13 14 15 16 17 18 19123456789
Sensitivity for PSG stages
Fractional occupancy (%)
HMM states
Specificity for PSG stages
10 11 12 13 14 15 16 17 18 19123456789
HMM states
*** ***
Wake N1 N2 N3
Switching frequency (Hz)
Switching Range of HMM states
*** ***
Unique states per time (Hz)
Wake N1 N2 N3
Mean life time of HMM states
across PSG stages
HMM states
Fig. 3 Sensitivity and specicity of HMM states and dynamics within polysomnography stages. aFractional occupancies of each of the 19 HMM states
computed within the four PSG stages corresponded to the PSG-sensitivity of the whole-brain network states. The coloured bars and error bars show the
average and standard error, respectively, across the 18 participants that included all four PSG stages. bPSG-specicity of the HMM states for each of
the four PSG stages. Specicity corresponds to the probability of an HMM state occurring within a PSG stage. The bars represent the group average and
the error bars the standard error (n=18). In aand bhorizontal lines show signicant differences within HMM states, with pvalues < 0.01 as evaluated
through paired ttests and permutation testing. cThe mean life times of the 19 HMM states are shown by the bars, representing values averaged across the
18 participants. Error-bars represent the standard error across participants. Each HMM state is coloured according to the probability of nding it within
each of the four PSG stages, i.e., their PSG specicity. Note how HMM states with high specicity for N3and to a lesser extent N2exhibit longer mean
life times. dThe dynamics of the HMM transitions were calculated within each of the four PSG stages, in terms of switching frequency (Switching), and e
the number of different HMM states visited per time (Range of HMM states). These measures signicantly separate the four PSG stages suggestive of a
higher dynamical repertoire during wakefulness and N1. In dand eerror bars represent standard error across participants and signicant differences
between PSG stages are denoted by stars: one star: p< 0.05, two stars: p< 0.01 and three stars p< 0.001; all evaluated using paired ttests and
permutations. W: wakefulness, N1: N1 sleep, N2: N2 sleep, N3: N3 sleep
NATURE COMMUNICATIONS | (2019) 10:1035 | | 5
Content courtesy of Springer Nature, terms of use apply. Rights reserved
Select whole-brain network states displayed similarly exclusive
sensitivity and specicity proles for N2 (HMM states 3 and 6)
and N3 sleep (HMM states 16). Notably, this was not the case for
N1 sleep. The whole-brain network states occupying most of
N1 sleep, such as HMM states 1, 4 and 15, were not found specic
for this PSG stage. Instead these states would also occur with
considerable likelihood outside of N1 sleep, although rarely
during N3 sleep.
In summary, wakefulness was found to correspond to a
collection of whole-brain networks states, while N2 and N3 were
characterised by less state-diversity, and dominated by two and
one whole-brain states, respectively. In contrast, no single whole-
brain states were found specic for N1 sleep, which instead was
modelled by a collection of HMM states with mixed PSG proles.
Changes in whole-brain network dynamics between PSG
stages. Having the whole-brain network states temporally dened
allowed us to investigate the large-scale brain dynamics of the
traditionally dened PSG stages in the 18 participants that
reached all PSG stages during their recordings.
In Fig. 3c, the HMM states are represented by a bar plot
showing their mean lifetimes, i.e., the average duration of the
state visits. The bars have been overlaid with colours depicting the
PSG specicity averaged across the corresponding HMM states.
HMM states with high specicity for N2 and N3 (HMM states 3,
6 and 16) generally expressed longer mean lifetimes than those
related to wakefulness and N1. The mean lifetimes of the HMM
states ranged from seconds to tens of seconds.
Figure 3d, e shows two summary measures for the dynamics of
the whole-brain network states during the individual PSG stages:
(i) the amount of switching dened as the average number of
transitions between HMM states during a given PSG stage
divided by the total time a participant spent in this PSG stage and
(ii) the range of HMM states dened as the number of unique
states visited during the given PSG stage divided by the total time
a participant spent in this PSG stage. Both measures were
estimated for each PSG stage, within each of the 18 participants
that included all four PSG stages, and normalised by time.
Wakefulness and N1 sleep expressed signicantly higher values
than N2 and N3. Interestingly, the amount of switching was
particularly low for N3 sleep.
In summary, unique state visits per time were few and of long
durations during N2 and N3 relative to wakefulness and N1 sleep.
Consequently, the switching between and range of HMM states
were signicantly higher in wakefulness and N1.
Sleep stages as modules of whole-brain network transitions.So
far, we have used the traditional PSG stages to organise and
evaluate the temporally resolved whole-brain network states. Yet,
the data-driven nature of the HMM also allowed us to perform
reverse inference, and consider the temporal progression of
HMM states, taking thisrather than the PSG stagingas a
starting point. This way, we were able to ask if the high-resolu-
tion, fMRI-based, HMM suggests new aspects of the wake-NREM
sleep cycle, hidden from the EEG-based PSG. For this purpose,
we examined the transition probabilities of the HMM states,
extracting modules of HMM states that transitioned more often
between each other than to other statesas recently identied for
the waking resting state in ref. 24.
The whole-brain network states organised into a transition map
as presented in Fig. 4, where the 19 × 19 transition probability
matrix (Fig. 4a) was submitted to a modularity analysis (see
Methods). By considering the most frequent transitions between
the HMM states that were consistent across participants (see
Fig. 4b), the thresholded transition matrix organised into four
partitions or transition modules (see Fig. 4c, and Methods),
suggestive of a lower time scale (see ref. 24 and Supplementary
Discussion 1). When these most consistent transitions are
presented as a transition map, and each whole-brain network
state is represented by a circle plot indicating its specicity for
each of the four PSG stages, it can be seen that the HMM states
exhibit a strong temporal structure (Fig. 4d). In line with the
MANOVA results above, this transition map describes an overall
progression from states with high specicity for PSG-dened
wakefulness (red module) through states with more activity
during, albeit not signicant specicity for, N1. From here
transitions lead towards states specic to N2 sleep and nally to a
single whole-brain network state modelling N3 sleep. The N2- and
N3-related HMM states thus grouped together in the blue module.
Interestingly a collection of HMM states with mixed PSG-
specicity formed a transition module of their own. This white
module was intercalated between the red module of wakefulness
in the top and the blue module of N2/N3 sleep. Even if the
included HMM states were not specic for PSG-dened N1 sleep,
the white module appears in the location of the transition map,
where one would expect to nd N1 or rather sleep onset.
The transition map suggested two sub-divisions of HMM states
with high specicity for wakefulness. The red module in close
proximity to the white module of N1-related states, and the black
module sending transitions to the blue module of consolidated
NREM sleep. This apparent separation of wakefulness and the
asymmetric relationship to the sleep-related HMM states led us to
the hypothesis that one of these could represent wakefulness after
sleep onset (WASO). Given the poor correspondence between the
HMM states and the general uncertainty associated with the
staging of PSG-dened N1 sleep (see Discussion), we chose to
dene WASO as PSG-staged wakefulness, which followed after
visits to N2 sleep28. By computing the sensitivity and specicity of
the whole-brain network states in the subset of the data
corresponding to the 31 participants who woke up after having
reached N2 sleep (see Supplementary Table 2), we were able to
conrm this hypothesis. As shown in Supplementary Figure 2,
HMM states 5, 17 and 18 were all more sensitive and specicto
WASO compared to wakefulness prior to N2 sleep. Whereas
periods of wakefulness prior to and after sleep are scored equally
in PSG, the whole-brain network states separated these into two
different transition modules.
Although this transition map suggests multiple pathways from
wakefulness (red module) to the white module of NREM sleep, it
is interesting to note that HMM state 8 has direct access to HMM
state 15, which in turn guards the transition to the blue module of
N2/N3 sleep. Similarly, waking-up relates to a transition from
HMM state 4 to HMM state 10, which in turn connects with
HMM state 18 of the black WASO module. Further it is worth
noticing the strong triangular transition structure within the blue
module between the N2-specic whole-brain network states
(HMM states 3 and 6) and the N3-modelling HMM state 16.
In summary, while agreeing with the overall sequence of PSG
stages, the organisation of the transition modules also points to
aspects of sleep-related brain activity that the PSG scoring cannot
access, including the data-driven suggestions of N1 sleep, WASO-
related whole-brain network states, and multiple transition
pathways between wakefulness and sleep.
Spatial activation and FC maps of whole-brain network states.
We present the spatial maps of the whole-brain network states in
the order suggested by the transition modules of Fig. 4d. Figure 5
and Fig. 6show the mean activation maps, while the corre-
sponding FC information is presented in Supplementary Fig-
ures 35 and 1718 (see also Supplementary Note 5).
6NATURE COMMUNICATIONS | (2019) 10:1035 | |
Content courtesy of Springer Nature, terms of use apply. Rights reserved
In Fig. 5a, which shows the red module of wakefulness, the
mean activation maps of HMM states 2 and 8 resemble resting-
state network (RSN) congurations9,10. The main increases of
HMM state 8 were thus seen in key areas of the default-mode
network (DMN)29, including the bilateral posterior cingulate
cortex, bilateral angular cortex, bilateral middle temporal cortex,
and bilateral medial prefrontal cortex. These DMN-like increases
in HMM state 8 were accompanied by decreases in the so-called
anti-correlated network (ACN), involving the supramarginal
gyrus and the dorsolateral part of the frontal cortex30. In contrast,
HMM state 2 was characterised by increases in many of these
ACN-areas, including the bilateral supramarginal gyrus, middle
cingulate cortex and dorsolateral part of the frontal cortex. These
results suggest an inverse relationship between the activity of the
DMN and the ACN, which is an established trait of these RSNs30.
Since the discovery of these RSN patterns they have been
hypothesised to reect complex cognitive processes. The DMN
has been linked to inwardly directed mentation, such as
autobiographical memory and mind wandering31,32. The ACN
overlaps with areas also referred to as the dorsal attention
network33 or the central executive network (CEN)34, and has
been proposed to be involved in more externally directed
processes, including attention35. In agreement with this, we
found these high-order RSNs to be relatively exclusive for
0.5 Participants not including HMM state
Transition matrix
Transition modules
Transition map
Removal of participant-specific states,
thresholding, and module extraction
HMM destination state
HMM departure state
Total proportion of sleep stages
HMM state
HMM destination state
HMM departure state
Transition prob.
Fig. 4 Investigating modules of transitions between whole-brain network states. aThe gure shows the 19 × 19 transition probability matrix of the HMM
states calculated for the 18 participants that included all four PSG stages in their respective scanning session. This quanties the likelihood of transitioning
from any given state to any other state, yielding each matrix entry: transition probability from departure state to destination state. bA few HMM states
were sporadicand did not occur consistently across participants. HMM states not occurring in more than 25% of the participants were excluded. cThe
strongest transitions of the consistent HMM states were partitioned through a modularity analysis, and reorganised in a matrix according to the four
resulting modules. dThe transitions shown in care presented as a transition map with each state depicted as a pie plot expressing its specicity for each of
the four PSG stages. Arrows show the direction of the transitions with thickness proportional to the transition probability. The transitions describe a
passage from HMM states with high activity during wakefulness in the top, further down through HMM states including more N1, and down to HMM
states specic to N2 and nally N3. Interestingly, wakefulness appears to be represented by two modules (red and black). Even though no individual HMM
state showed clear specicity for N1 sleep, a module (white) is evident between HMM states with specicity for wakefulness and HMM states with
specicity for N2 and N3 sleep. ePie chart showing the total proportion PSG stages within the 18 participants. Transition prob.: transition probability
NATURE COMMUNICATIONS | (2019) 10:1035 | | 7
Content courtesy of Springer Nature, terms of use apply. Rights reserved
wakefulness. However, previous investigations have suggested a
rather ubiquitous presence of both the DMN and the ACN, not
just in wakefulness but, in all stages of NREM sleep15,17,36 (see
Supplementary Discussion 1).
Figure 5b shows the mean activation maps of whole-brain
network states with higher sensitivity and specicity for WASO
(black module). The mean activation map of HMM state 18
expressed a distribution similar to that of HMM state 8 (see
Fig. 5a), but with opposite signs. Hence, HMM state 18 showed
decreases in DMN-related areas, and increases in regions over-
lapping the ACN. HMM states 5 and 17 were both characterised
by mean activation increases in the frontal cortices. Interestingly,
ndings from high-density EEG studies of participants waking up
from sleep show that the posterior parts of the cortex are
particularly slowʼat returning to levels of activity seen prior to
sleep37. Consistent with this, converging evidence from PET and
fMRI have indicated frontal cortical activity to be increased
relative to that of posterior areas upon awakening28,38,39.
The whole-brain network states of the white N1-related
module are represented in Fig. 6a. A general observation for
these spatial maps is the inverse relationship between mean
activation in subcortical areas (thalamus and parts of the basal
ganglia) and primary sensory cortical areas. Increases in
subcortical activity were accompanied by decreases in primary
sensory areas of the cortex and vice versa. This was true for HMM
states 4 and 15 (and HMM state 1 although its decreases were not
State 8
–0.2 0.1
State 10
–0.2 0.1
State 2
–0.1 0.1
State 17
State 18
–0.2 0.1
State 5
WASO-related HMM states
Wake-related HMM states
Average activation
50% strongest
50% strongest
Average activation
50% strongest
50% strongest
Fig. 5 Mean activation distributions of wakefulness-related HMM states. aThree HMM states identied as sensitive and specic to wakefulness prior to
sleep. Note HMM state 8 with DMN-like conguration increases, and concomitant decreases in DAN/CEN-like areas. bThree HMM states associated to
wakefulness after sleep. There are marked frontal increases in mean activation in HMM states 5 and 17. All maps were thresholded above the 50%
strongest positive and negative changes, respectively
8NATURE COMMUNICATIONS | (2019) 10:1035 | |
Content courtesy of Springer Nature, terms of use apply. Rights reserved
State 1
–0.2 0.3
State 4
–0.9 0.2
State 15
–0.2 1.0
State 3
State 6
–0.2 0
State 16
–0.1 0.06
N1-related HMM states
N2-related HMM states
N3-related HMM states
State 13
–1.1 0.3
Average activation
50% strongest
50% strongest
Average activation
50% strongest
50% strongest
Average activation
50% strongest
50% strongest
Fig. 6 Mean activation distributions of sleep-related HMM states. aHMM states associated with N1 sleep showed opposite signs in mean activation in
subcortical areas and primary sensory areas of the cortex. bThree HMM states related to N2 sleep. HMM states 3 and 6 in particular showed peak
increases and decreases, respectively, in areas previously identied as fMRI-correlates of sleep spindles. cHMM state 16 is dominating slow wave sleep
(N3). Interestingly, there are marked decreases in mean activation in frontal areas and insula, and very localised increases in the supplementary motor area
and paracentral lobule. All maps were thresholded above the 50% strongest positive and negative changes, respectively
NATURE COMMUNICATIONS | (2019) 10:1035 | | 9
Content courtesy of Springer Nature, terms of use apply. Rights reserved
conned to subcortical areas, but supplemented by decreases in
the anterior and middle cingulate cortex). This is consistent with
intracortical studies of the sleep onset process in rats40 and in
humans41 showing that thalamic changes in dynamics precede
those of cortical areas near the onset of NREM sleep. Previous
fMRI studies of NREM sleep have suggested decreased con-
nectivity between the thalamus and cortical regions as perhaps
the most consistent trait of FC during N1 sleep16,18,19,23.
N2 sleep was dominated by HMM states 3 and 6, and the mean
activation maps of these whole-brain network states are shown in
Fig. 6b. The supplementary motor area was involved in both of
these states; in HMM state 3 as increases in concert with the
bilateral precuneus and primary motor cortices; and in HMM
state 6 as decreases together with the bilateral thalamus, middle
cingulate, supramarginal cortex, and the rolandic operculum.
Interestingly, these congurations overlap considerably with
those previously reported in studies mapping fMRI-correlates of
sleep spindles42,43, which represent a dening EEG-feature of
N2 sleep. However, no HMM state appeared to be driven solely
by either sleep spindles or K-complexes. By identifying sleep
spindles and K-complexes from the EEG data, we assessed the
temporal relationships between these graphoelements and the
HMM states. In summary the HMM states that were dominant
during N2 sleep showed comparable sensitivity and specicity to
both types of graphoelements, and hence the HMM did not
appear to have assigned individual states for either spindles or K-
complexes (for further information please see the Supplementary
Discussion 1, Supplementary Note 4, Supplementary Table 3, and
Supplementary Figures 1921).
HMM state 16 accounted for the majority of time spent in
N3 sleep. The corresponding mean activation map is shown in
Fig. 6c. Apart from some very localised increases in the
paracentral lobule and adjacent supplementary motor area the
mean activation was characterised mainly by decreases, particu-
larly in the bilateral middle and superior temporal pole, the orbital
part and the operculum of the inferior frontal cortex, bilateral
insula as well as medial temporal areas. These frontal decreases are
consistent with previous PET ndings of decreased metabolism in
these areas during N3 sleep, which in turn are believed to reect
the high, localised concentration of slow-wave activity44.
Using a data-driven exploration of large-scale brain networks and
associated dynamics from continuous fMRI recordings, we have
explored the rich dynamical complexity in spatiotemporal pat-
terns of brain activity during the healthy wake-NREM sleep cycle.
Moving beyond the traditional PSG stages of sleep, we used a
HMM to extract 19 recurring whole-brain network states, dened
in space by patterns of mean BOLD activation and FC, and
dened in time as the probability of being active at each time
point of the fMRI sampling. Comparing the temporal evolution of
the HMM-derived whole-brain network states with the inde-
pendently obtained EEG-based PSG scoring, we have discovered a
rich repertoire of brain dynamics underpinning the traditional
PSG stages. The temporal resolution of the HMM identied state
lifetimes on the order of seconds, providing a temporally ne-
grained description of the traditional PSG stages. Crucially, a
close examination of the HMM transition map furthermore
revealed a heterogeneity of large-scale network activity that PSG
cannot fully capture.
The description of brain activity offered by PSG has for long
been acknowledged as incomplete, and attempts have been made
to harvest more information from scalp EEG in a search for
features relevant for sleep, overlooked by PSG4547. Our work
adheres to this aim, while, through fMRI, incorporating evidence
of whole-brain spatial detail. Previous studies have indicated that
fMRI can be used to identify dynamic re-congurations of large-
scale brain activity during the conventional EEG-based sleep
stages, either in form of voxel-wise changes in activity48, changes
in connection strengths in resting-state networks49 or through
long-range temporal dependencies in the BOLD signal36. Rather
than direct reections of the conventional sleep stages, what has
emerged from our HMM analysis is a probabilistic representation
of the PSG scoring in the space of whole-brain network states and
transitions. Agreements as well as disagreements between the
PSG scoring and the independent HMM decomposition became
clear in the transition map (Fig. 7). Wakefulness, N2 sleep and
N3 sleep were each represented by one or more whole-brain
network states, forming a good correspondence with PSG. In
contrast, no states were found specic for N1 sleep. Furthermore,
while treated equivalently in PSG staging, wakefulness prior to
sleep and WASO were represented in the transition map as two
different modules with different repertoires of large-scale brain
networks. Consequently, the transition map also identied spe-
cic whole-brain network transitions underlying the descent to,
and ascent from, NREM sleep.
Consistent with previous neuroimaging studies that have used
regression analyses to identify consistent differences between
traditionally dened sleep stages in terms of large-scale brain
activity13,14,50, PSG-dened wakefulness, N2 and N3 sleep each
corresponded well to specic collections of whole-brain network
states (see Supplementary Discussion 1). However, the HMM
additionally provided access to the large-scale brain dynamics of
the PSG stages, showing that the state repertoire, when estimated
as amount of switching and range of states visited, is higher in
wakefulness than in both N2 and N3 sleep. That a higher and
more complex state repertoire is important for the brain to
support wakeful consciousness follows from theoretical frame-
works5153 and has received empirical support from a series of
combined TMS and EEG studies54. From a large-scale network
perspective fMRI has been used to show how an enhanced state
repertoire is associated with an expandedʼconsciousness during
the psychedelic experience55,56. In the context of sleep, however,
the large-scale network evidence is mainly represented by static
FC studies suggesting decreased information integration during
N2 and N3 sleep using graph theory20,21, as well as a higher
exploration of the structural connectome during wakefulness57.
Here, we have provided more direct evidence of a higher state
repertoire in whole-brain dynamics during wakefulness.
The transition map identied a key trajectory from wakefulness
in the red transition module to NREM sleep in the white tran-
sition module (see Fig. 7d), represented by the transition
departing from the whole-brain network state with increased
mean activation in the DMN. The proposed association between
the DMN and inwardly directed mentation31,32 makes this
nding intriguing, in the sense that it may suggest a role for the
DMN as a gatein the process of initiating sleep. Whole-brain
network evidence of sleep initiation may improve our under-
standing of sleep disorders like insomnia where PSG criteria are
difcult to apply7,58, and hypersomnia disorders59. Related
hereto, a recent study identied switching instability to and from
N2 sleep, together with difculties reaching N3 sleep as important
traits of insomnia60. In the transition map we saw N2- and N3-
related whole-brain network states forming a strong triangular
loop of transitions (see Fig. 7e). This stable conguration of
transitions may not be present in people suffering from insomnia.
The two main incongruities between the temporal segmentation
suggested by the HMM and the PSG scoring concerned N1 sleep
and WASO. N1 sleep did not correspond to any single state or any
group of states identied by the HMM. This is likely related to
the current consensus that PSG-dened N1 does not represent a
10 NATURE COMMUNICATIONS | (2019) 10:1035 | |
Content courtesy of Springer Nature, terms of use apply. Rights reserved
clear-cut sleep stage61, but rather an ill-understood mix of wake-
fulness and sleep. Compared to N2 and N3 sleep with their well-
dened EEG spectral properties, such as K-complexes, spindles,
and slow waves, N1 remains the most vaguely dened sleep stage
within PSG. A recent report by the American Academy of Sleep
Medicine (AASM) shows that staging of N1 is associated with the
highest inter-rater scoring uncertainty of all PSG stages62.Fur-
thermore, N1 sleep has proven the most difcult PSG stage to
classify from fMRI FC information in machine-learning stu-
dies18,19. Addressing the microstructure of N1, a line of evoked
response potential-studies have demonstrated a high degree of
variability in the cortical processing of external stimuli during early
NREM sleep (for reviews, see ref. 63,64). Phenomenologically, the
sleep onset period is known to be rather complex, with varying
mental content and responsiveness to sensory stimuli64,65,and
authors have long argued against the assumed homogeneity found
in PSG denitions of N1 sleep, an opposition exemplied by Horis
proposal of nine stages of early sleep45.IfPSG-dened N1 does in
fact represent a mix of wakefulness and sleep, this would explain
why we found the highest range of whole-brain states during this
PSG stage. While this primarily serves to underline the common
notion that N1 is unlikely to be a reliable demarcation between
wakefulness and sleep, the fact that the data-driven HMM was able
to identify a transition module occurring between wakefulness and
consolidated sleep (N2 and N3)represented by whole-brain
states characterised by subcortico-cortical decoupling consistent
with intracortical evidence of brain activity during sleep onset41
suggests that an improved and principled categorisation of early
sleep could be within reach.
PSG does not differentiate between brain activity prior to and
after sleep onset. However, in line with the common subjective
experience of grogginess when waking from sleep, behavioural
experiments have shown cognitive decits in the period following
awakening. The term sleep inertia is often used to describe this
phenomenon66. Our results conrm that falling asleep and
waking up are two asymmetric processes, leading to two sepa-
rated transition modules of whole-brain network states during
wakefulness, with one more likely to occur after consolidated
sleep. Like the N1-related ndings discussed above, this too serves
as a prime example of how information-rich neuroimaging data,
when treated in a data-driven way, can be carefully evaluated in
light of established knowledge (PSG in this case) to make new
- Diverse dynamic repertoire of
- High-order resting-state networks
- DMN and ACN
- Requirements of normal cognition?
- Distinct networks of wake
after sleep onset (WASO)
- Frontal activation increases
- Signs of sleep inertia?
- Key transition to sleep
passing through DMN?
- Relevance for sleep disorders
(i.e., insomnia, hypersomnia)
- Mixed states of N1 sleep
- Data-driven definition of light
- Subcortico-cortical decoupling
- Rapid switching dynamics
- Hypnagogia, hallucinations?
- N2 and N3 sleep efficiently
modelled by a few network states
- Robust triangular transition
- Slow swithing dynamics
Fig. 7 A whole-brain network perspective on the human wake-NREM sleep cycle. The main discussion points of our results have been highlighted in the
transition map of the whole-brain network states. Boxes a to e summarise the new perspectives provided by our HMM analysis
NATURE COMMUNICATIONS | (2019) 10:1035 | | 11
Content courtesy of Springer Nature, terms of use apply. Rights reserved
discoveries from, and categorisations of, brain activity. The pre-
sented ndings point ahead to a research agenda making
hypothesis-driven assessments of how the alternative, data-dri-
ven, temporal segmentations and dynamics of whole-brain net-
works across the NREM sleep cycle relate to sleep-behaviour and
cognition, when the latter is measured independently of PSG (see
Supplementary Discussion 3). Furthermore, there is scope for
HMM explorations with higher temporal detail using electro-
physiological modalities, such as magnetoencephalography
(MEG) and high-density EEG (see Supplementary Discussion 3).
For further discussion about the reproducibility of our results
across different numbers of states, different initialisations of the
HMM, and different parcellation schemes, as well as discussion
about our choice of data inclusion, pre-processing steps, such as
spatial smoothing, and the use of the RETROICOR method to
remove physiological signals from the fMRI data, we refer to the
Supplementary Discussion 2, Supplementary Notes 13, and
Supplementary Figures 1, 616, and 2225.
In summary, the work presented here demonstrates how data-
driven, temporally sensitive analyses of large-scale fMRI brain
activity can be used to explore fundamental changes in behaviour
and cognition, in the form of the wake-NREM sleep cycle. The
results reveal a higher complexity of brain activity than what
traditional sleep scoringand neuroimaging relying strictly on
PSGcan reveal. We projected the traditional stages of wake-
fulness and NREM sleep onto a probabilistic map of transitions
between whole-brain network states. By studying these transitions
we have shown a signicant decrease in whole-brain dynamics
during consolidated stages of NREM sleep; that brain activity
prior to sleep is signicantly different from just after sleep; that
whole-brain network activity do not support traditional criteria to
dene N1 sleep; and that increased activity in the DMN might
serve a gate-function for the entry into NREM sleep. By using
fMRI data we have increased the spatiotemporal resolution of
traditional NREM sleep stages, using a framework that should be
sought expanded to include other fundamental changes in brain
activity, such as REM sleep, sleep disorders, anaesthesia and
psychedelic experiences. Finally, future work should aim to
leverage even ner temporal details through modalities such as
MEG and high-density EEG.
Acquisition and processing of fMRI and PSG data. fMRI was acquired on a 3 T
system (Siemens Trio, Erlangen, Germany) with the following settings: 1505
volumes of T2*-weighted echo planar images with a repetition time (TR) of 2.08 s,
and an echo time of 30 ms; matrix 64 × 64, voxel size 3 × 3 × 2 mm3, distance factor
50%, FOV 192 mm2.
The EPI data were realigned, normalised to MNI space, and spatially smoothed
using a Gaussian kernel of 8 mm3FWHM in SPM8 (
). Spatial downsampling was then performed to a 4 × 4 × 4 mm resolution. From the
simultaneously recorded ECG and respiration, cardiac- and respiratory-induced noise
components were estimated using the RETROICOR method67, and together with
motion parameters these were regressed out of the signals. The data were temporally
band-pass ltered in the range 0.010.1Hz using a sixth-order Butterworth lter.
Please note that the fMRI data were temporally ltered with no consideration of the
later established PSG structure of the data. Hence, our ndings of relative differences
between the various PSG stages should not be affected by these pre-processing steps.
We show that this is the case for the temporal lter in Supplementary Figure 16,
where the plots of Fig. 3from the main text have beenre-computed for an HMM with
19 states on BOLD data, which had not been temporally ltered.
Simultaneous PSG was performed through the recording of EEG, EMG, ECG,
EOG, pulse oximetry and respiration. EEG was recorded using a cap (modied
BrainCapMR, Easycap, Herrsching, Germany) with 30 channels, of which the FCz
electrode was used as reference. The sampling rate of the EEG was 5 kHz, and a
low-pass lter was applied at 250 Hz. MRI and pulse artefact correction were
applied based on the average artefact subtraction method68 in Vision Analyzer2
(Brain Products, Germany). EMG was collected with chin and tibial derivations,
and as the ECG and EOG recorded bipolarly at a sampling rate of 5 kHz with a
low-pass lter at 1 kHz. Pulse oximetry was collected using the Trio scanner, and
respiration with MR-compatible devices (BrainAmp MR+, BrainAmp ExG; Brain
Products, Gilching, Germany).
Participants were instructed to lie still in the scanner with their eyes closed and
relax. Sleep classication was performed by a sleep expert based on the EEG
recordings in accordance with the AASM criteria (2007).
Results using the same data and the same preprocessing have previously been
reported in ref. 18.
Participants. We used fMRI and PSG data from 57 participants taken from a
larger data-base18. Exclusion criteria focussed on the quality of the concomitant
acquisition of EEG, EMG, fMRI and physiological recordings. Written informed
consent was obtained, and the study was approved by the ethics committee of the
Faculty of Medicine at the Goethe University of Frankfurt, Germany.
Following the HMM decomposition, two different subsets of the solution were
used for post hoc evaluation of the HMM. The rst corresponded to the 18
participants that reached all four stages of PSG, and the second corresponded to the 31
participants that woke up after having reached consolidated sleep (the WASO group).
HMM general overview. In order to resolve dynamic whole-brain networks in the
fMRI signals in a data-driven way, we applied a HMM24,27 to timecourses extracted
from 90 ROIs dened by the cortical and subcortical areas of the AAL26, however,
please see Supplementary Note 1 for a demonstration of the robustness of our
results using an alternative parcellation.
We used the FSL ( function fslmeants to
average over voxels within each ROI to get the representative timecourses. The
participant-specic sets of 90 ROI-timecourses were demeaned, divided by their
standard deviation, and concatenated across participants, yielding a data matrix of
dimensions 90 × (57 × 1500), with 1500 samples corresponding to 52 min given a
TR of 2.08 s. The HMM inference estimated a number of recurring discrete states,
each of which was characterised by a unique conguration of data statistics. We
employed a Gaussian HMM, implemented using the Matlab toolbox HMM-MAR
(, such that each state was
modelled as a multivariate normal distribution with rst (mean activity) and
second order statistics (covariance matrix). The parameters of the states were
dened at the group level, whereas the state timecourses are dened for each
subject separately. Therefore, the HMM inference identied periods of time of
quasi-stationary activity, where the 90 ROI timecourses could be described by
certain congurations of mean activity and FC. The HMM represents a tool for
decomposing multivariate data into fewer dimensions. Given the high spatial
dimensionality of fMRI, it is common to use principal component analysis (PCA)
to reduce the number of parameters to be estimated in the decomposition,
increasing the signal-to-noise ratio of the data and improving the robustness of the
results24,27. Accordingly, we submitted the demeaned, standardised and
concatenated BOLD timecourses to PCA prior to the HMM inference. Keeping
approximately 90% of the signal variance, we used the top 25 principal components
(see Fig. 1), yielding a data matrix of dimensions 25 × (57 × 1500), which were then
fed to the HMM. An overview of the analysis workow is given in Fig. 1of the
main text. For certain analyses, such as the MANOVA and the test for WASO-
specic HMM states, we used subsets of the full set of HMM states.
Choice of number of HMM states. The HMM was implemented with variational
Bayes inference, which was used to probabilistically estimate the state statistics and
transition probabilities24,27. The number of states of the HMM was a free para-
meter, which had to be chosen before further evaluation. Determining the number
of states present in recordings of spontaneous brain activity is a non-trivial task,
which may be approached in a number of ways. We ran the HMM for model
orders spanning 445, and evaluated each solution by a number of summary
statistics, the most important of which are plotted in Supplementary Figure 1.
Supplementary Figure 1a shows the minimum free energy as a function of the
HMM model order. The free energy is the statistical measure that is minimised
during the (variational inference) Bayesian optimisation process. Technically
speaking, it is an approximation of the model evidence, and includes two terms: how
well the model ts the data, and the complexity of the model (measured as how
much it departs from the prior distribution). Whereas the free energy is a reasonable
criterion for choosing the ideal number of states for the HMM, its biological validity
remains unclear in so far as the HMM does not represent a biophysical model. As
apparent from the plot in Supplementary Figure 1a, the minimum free energy was
monotonically decreasing over the large range of tested numbers of states, showing
no negative peaks. Hence, like in previous applications of the HMM24,27, the free
energy was not informative for choosing the number of states in our case.
We dened fractional occupancy as the temporal proportion of a recording, in
which an HMM state was active27. In Supplementary Figure 1c is plotted the
development of the median fractional occupancy across HMM states as a function
of model order. While the curve decreases rapidly for low values of K, meaning
that, as expected, each HMM state on average accounted for less of the total
recording time as the number of states was increased, this trend ceased from
around K=19. This stagnation for higher model orders was caused by the
occurrence of sporadicstates, which modelled very (participant-) specic subparts
of the data (see Supplementary Figure 6). This phenomenon was also reected in
the development of the average HMM state lifetime, which too stabilised around
the same value of K, as shown in Supplementary Figure 1d.
12 NATURE COMMUNICATIONS | (2019) 10:1035 | |
Content courtesy of Springer Nature, terms of use apply. Rights reserved
To test whether the fMRI-based HMM states showed a signicant relationship with
the EEG-based sleep scoring, we used a multivariate analysis of variance (MANOVA).
The built-in MATLAB function manova1 provided the summary statistic WilksΛ,
which described how well the Knumber of HMM state timecourses could be grouped
according to the sleep scoring. In order to test if the relationship was signicant, we
performed MANOVAs on 1000 permuted cases of the sleep scoring, collecting the
WilksΛfor each run. Each permutation was constructed, such that the original stage
transition points, stage counts, and periods were preserved, while the sleep-stage
labelling of each of these periods was shufed randomly within participants. For each
permutation, each participant thus retained the same sleep stages, but the temporal
orderings of these were random. Supplementary Figure 1b plots WilksΛas a function
of the HMM model order for both the original sleep scoring and the permuted cases.
For number of states above K=7, the HMM state timecourses were grouped
signicantly better by the original compared to the permuted sleep scoring. This result
suggests that only when using more than seven states, the HMM identied states with
asignicant dependency on the PSG scoring.
The results in Supplementary Figure 1bd are computed from a subset of the
HMM solutions corresponding to the participants including all four available PSG
stages (18 participants with: wakefulness, N1 sleep, N2 sleep, and N3, see right part of
Supplementary Table 1). This was done to minimise the unevenness in the
representation of PSG stages. Based on the evaluations above, we concluded that the
HMM was able to, in data-driven fashion, estimate the temporal structure given by the
EEG-based sleep scoring for model orders above 7. We chose K=19 states, because
increasing the number of states above this point mainly resulted in the addition of
HMM states of low-fractional occupancy and participant-specic occurrences.
Signicance testing. In order to evaluate the PSG-sensitivity and -specicity, we
used paired ttests to test for signicant differences within the HMM states. As
such, with the original PSG stages, we compared the sensitivity and the specicity,
respectively, for W, N1, N2 and N3 within each HMM state for the 18 participants
that included all of these four stages. This yielded six comparisons within each
HMM state (W-vs.-N1, W-vs.-N2, W-vs.-N3, N1-vs.-N2, N1-vs.-N3, N2-vs.-N3),
for both the sensitivity and specicity measures, and each of these comparisons was
associated with a t-statistic. To test if these t-statistics were larger than random, we
used permutation testing as explained above (see section Choice of number of
HMM states), where the EEG-based sleep scoring vector was permuted 1000 times
with number of PSG stages and periods kept constant, but with their temporal
order randomly shufed within each participant. We computed the PSG-sensitivity
and -specicity for each permutation and performed paired ttests for every case.
The resulting t-statistics were used to build null-distributions, and the original t-
statistics were compared against these to get pvalues for the original tests. The bar
plots of PSG-sensitivity and -specicity in Fig. 3a, b of the main text include line
crossbars indicating the cases with pvalues < 0.01 (paired ttests with permuta-
tions). Please note that the use of paired ttests ensured that the identied differ-
ences were consistent within participants, and not merely as a group effect.
We evaluated the hypothesis that certain HMM states had higher activity in
periods of wakefulness after sleep onset (WASO) by considering the subgroup of
participants that, according to PSG scoring, woke up after having reached N2 sleep.
This corresponded to 31 participants, and we dened WASO as
polysomnographically estimated periods of wakefulness that followed N2 sleep.
From here we followed the same steps outlines above for the original PSG stages,
and the results are shown in Supplementary Figure 2.
To test for differences in the measures of dynamics, amount of switching and
range of HMM states, we also employed paired ttests and the permutation scheme
explained above to establish chance levels.
Analysis and visualisation of HMM transitions. The matrix of transition
probabilities, which were explicitly modelled by the HMM, contained a clear
organisation, in which subnetworks of HMM states expressed more frequent
transitions within each other than to states outside. In other words, the transition
matrix represented a directed graph with modular organisation. We demonstrated
this by submitting the transition matrix (shown in Fig. 4a) to a modularity analysis,
using Matlab functions from the Brain Connectivity Toolbox (
com/site/bctnet/Home)69, based on Newmans spectral community detection70.
Prior to running the modularity algorithm, we excluded the transitions of the
HMM states that did not occur consistently across participants, i.e., sporadic states
(see Methods, Choice of number of states, and Fig. 4b), and thresholded the
remaining transition matrix to include the strongest elements. The choice of this
latter threshold was done for visualisation purposes (for the results shown in the
main text using 19 states we included the 21% strongest transitions), however,
different thresholds resulted in highly similar module partitions. The modular
organisation is presented in a reordered matrix (Fig. 4c) and as a map (Fig. 4d).
Visualising mean activation maps of HMM states. The mean activation maps
have been overlaid on brain surfaces in Figs. 5and 6of the main text, using the
Human Connectome Project software Connectome Workbench (https://www. The HMM was inferred in
volumetric MNI152 space and mapped to the surface of the Conte-69 template using
the Workbench function wb_com mand volume-to-surface-mapping. The presented
surface maps are shown with the 50% strongest increases and the 50% strongest
decreases in activation for each HMM state relative to baseline averaged over all
HMM states.
Data availability
The datasets generated during and/or analysed during as well as code used during the
current study are available from the corresponding author on reasonable request.
Received: 23 February 2018 Accepted: 11 February 2019
1. Laureys, S., Perrin, F. & Bredart, S. Self-consciousness in non-communicative
patients. Conscious. Cogn. 16, 722741 (2007).
2. Iber C., Ancoli-Israel S., Chesson A. & Quan, S. F. The AASM Manual for the
Scoring of Sleep and Associated Events: Rules, Terminology and Technical
Specications (American Academy of Sleep Medicine, Westchester, IL, 2007).
3. Loomis, A. L., Harvey, E. N. & Hobart, G. Potential rhythms of the cerebral
cortex during sleep. Science 81, 597598 (1935).
4. Himanen, S. L. & Hasan, J. Limitations of Rechtschaffen and Kales. Sleep. Med.
Rev. 4, 149167 (2000).
5. Ogilvie, R. D. & Wilkinson, R. T. The detection of sleep onset: behavioral and
physiological convergence. Psychophysiology 21, 510520 (1984).
6. Bonnet, M. H. & Moore, S. E. The threshold of sleep: perception of sleep as a
function of time asleep and auditory threshold. Sleep 5, 267276 (1982).
7. Manconi, M. et al. Measuring the error in sleep estimation in normal subjects
and in patients with insomnia. J. Sleep Res. 19, 478486 (2010).
8. Edinger, J. D., Ulmer, C. S. & Means, M. K. Sensitivity and specicity of
polysomnographic criteria for dening insomnia. J. Clin. Sleep. Med. 9,
481491 (2013).
9. Raichle, M. E. Two views of brain function. Trends Cogn. Sci. 14, 180190
10. Smith, S. M. et al. Correspondence of the brains functional architecture during
activation and rest. Proc. Natl Acad. Sci. USA 106, 1304013045 (2009).
11. Deco, G., Jirsa, V. K. & McIntosh, A. R. Emerging concepts for the dynamical
organization of resting-state activity in the brain. Nat. Rev. Neurosci. 12,
4356 (2011).
12. Woolrich, M. W. & Stephan, K. E. Biophysical network models and the
human connectome. Neuroimage 80, 330338 (2013).
13. Maquet, P. Functional neuroimaging of normal human sleep by positron
emission tomography. J. Sleep Res. 9, 207231 (2000).
14. Dang-Vu, T. T. et al. Functional neuroimaging insights into the physiology of
human sleep. Sleep 33, 15891603 (2010).
15. Sämann, P. G. et al. Development of the brains default mode network from
wakefulness to slow wave sleep. Cereb. Cortex 21, 20822093 (2011).
16. Spoormaker, V. I. et al. Development of a large-scale functional brain network
during human non-rapid eye movement sleep. J. Neurosci. 30, 1137911387
17. Larson-Prior, L. J. et al. Cortical network functional connectivity in the
descent to sleep. Proc. Natl Acad. Sci. USA 106, 44894494 (2009).
18. Tagliazucchi, E. & Laufs, H. Decoding wakefulness levels from typical fMRI
resting-state data reveals reliable drifts between wakefulness and sleep. Neuron
82, 695708 (2014).
19. Altmann, A. et al. Validation of non-REM sleep stage decoding from resting state
fMRI using linear support vector machines. Neuroimage 125,544555 (2016).
20. Boly, M. et al. Hierarchical clustering of brain activity during human nonrapid
eye movement sleep. Proc. Natl Acad. Sci. USA 109, 58565861 (2012).
21. Tagliazucchi, E. et al. Large-scale brain functional modularity is reected in
slow electroencephalographic rhythms across the human non-rapid eye
movement sleep cycle. Neuroimage 70, 327339 (2013).
22. Brodbeck, V. et al. EEG microstates of wakefulness and NREM sleep.
Neuroimage 62, 21292139 (2012).
23. Haimovici, A., Tagliazucchi, E., Balenzuela, P. & Laufs, H. On wakefulness
uctuations as a source of BOLD functional connectivity dynamics. Sci. Rep. 7,
5908 (2017).
24. Vidaurre, D., Smith, S. M. & Woolrich, M. W. Brain network dynamics are
hierarchically organized in time. Proc. Natl Acad. Sci. USA 114, 1282712832
25. Preti, M. G., Bolton, T. A. W. & Ville, D. V. D. The dynamic functional
connectome: state-of-the-art and perspectives. NeuroImage 160,4154
26. 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 15, 273289 (2002).
NATURE COMMUNICATIONS | (2019) 10:1035 | | 13
Content courtesy of Springer Nature, terms of use apply. Rights reserved
27. Vidaurre, D. et al. Discovering dynamic brain networks from big data in rest
and task. Neuroimage 180(Pt 4), 646656 (2018).
28. Balkin, T. J. et al. The process of awakening: a PET study of regional brain
activity patterns mediating the reestablishment of alertness and
consciousness. Brain 125, 23082319 (2002).
29. Raichle, M. E. et al. A default mode of brain function. Proc. Natl Acad. Sci.
USA 98, 676682 (2001).
30. Fox, M. D. et al. The human brain is intrinsically organized into dynamic,
anticorrelated functional networks. Proc. Natl Acad. Sci. USA 102, 96739678
31. Buckner, R. L., Andrews-Hanna, J. R. & Schacter, D. L. The brains default
network: anatomy, function, and relevance to disease. Ann. N. Y. Acad. Sci.
1124,138 (2008).
32. Spreng, R. N., Mar, R. A. & Kim, A. S. The common neural basis of
autobiographical memory, prospection, navigation, theory of mind, and the
default mode: a quantitative meta-analysis. J. Cogn. Neurosci. 21, 489510 (2009).
33. Fox, M. D., Corbetta, M., Snyder, A. Z., Vincent, J. L. & Raichle, M. E.
Spontaneous neuronal activity distinguishes human dorsal and ventral
attention systems. Proc. Natl Acad. Sci. USA 103, 1004610051 (2006).
34. Koechlin, E. & Summereld, C. An information theoretical approach to
prefrontal executive function. Trends Cogn. Sci. 11, 229235 (2007).
35. Menon, V. & Uddin, L. Q. Saliency, switching, attention and control: a
network model of insula function. Brain Struct. Funct. 214, 655667 (2010).
36. Tagliazucchi, E. et al. Breakdown of long-range temporal dependence in
default mode and attention networks during deep sleep. Proc. Natl Acad. Sci.
USA 110, 1541915424 (2013).
37. Ferrara, M. et al. The electroencephalographic substratum of the awakening.
Behav. Brain Res. 167, 237244 (2006).
38. Tsai, P. J. et al. Local awakening: regional reorganizations of brain oscillations
after sleep. Neuroimage 102(Pt 2), 894903 (2014).
39. Braun, A. R. et al. Regional cerebral blood ow throughout the sleep-wake
cycle. An H2(15)O PET study. Brain 120(Pt 7), 11731197 (1997).
40. Baker, R. et al. Altered activity in the central medial thalamus precedes
changes in the neocortex during transitions into both sleep and propofol
anesthesia. J. Neurosci. 34, 1332613335 (2014).
41. Magnin, M. et al. Thalamic deactivation at sleep onset precedes that of the
cerebral cortex in humans. Proc. Natl Acad. Sci. USA 107, 38293833 (2010).
42. Schabus, M. et al. Hemodynamic cerebral correlates of sleep spindles during
human non-rapid eye movement sleep. Proc. Natl Acad. Sci. USA 104,
1316413169 (2007).
43. Caporro, M. et al. Functional MRI of sleep spindles and K-complexes. Clin.
Neurophysiol. 123, 303309 (2012).
44. Werth, E., Achermann, P. & Borbely, A. A. Fronto-occipital EEG power
gradients in human sleep. J. Sleep Res. 6, 102112 (1997).
45. Hori, T. et al. Proposed supplements and amendments to a manual of
standardized terminology, techniques and scoring system for sleep stages of
human subjects, the Rechtschaffen & Kales (1968) standard. Psychiatry Clin.
Neurosci. 55, 305310 (2001).
46. Olbrich, E. & Achermann, P. Analysis of oscillatory patterns in the human
sleep EEG using a novel detection algorithm. J. Sleep Res. 14, 337346 (2005).
47. Abeysuriya, R. G. & Robinson, P. A. Real-time automated EEG tracking of
brain states using neural eld theory. J. Neurosci. Methods 258,2845 (2016).
48. Picchioni, D. et al. fMRI differences between early and late stage-1 sleep.
Neurosci. Lett. 441,8185 (2008).
49. Watanabe, T. et al. Network-dependent modulation of brain activity during
sleep. Neuroimage 98,110 (2014).
50. Tagliazucchi, E. & van Someren, E. J. W. The large-scale functional
connectivity correlates of consciousness and arousal during the healthy and
pathological human sleep cycle. Neuroimage 160,5572 (2017).
51. Tononi, G. Consciousness as integrated information: a provisional manifesto.
Biol. Bull. 215, 216242 (2008).
52. Tononi, G., Edelman, G. M. & Sporns, O. Complexity and coherency:
integrating information in the brain. Trends Cogn. Sci. 2, 474484 (1998).
53. Marshall, W., Gomez-Ramirez, J. & Tononi, G. Integrated information and
state differentiation. Front. Psychol. 7, 926 (2016).
54. Casali, A. G. et al. A theoretically based index of consciousness independent of
sensory processing and behavior. Sci. Transl. Med 5, 198ra105 (2013).
55. Carhart-Harris, R. L. et al. The entropic brain: a theory of conscious states
informed by neuroimaging research with psychedelic drugs. Front. Hum.
Neurosci. 8, 20 (2014).
56. Tagliazucchi, E., Carhart-Harris, R., Leech, R., Nutt, D. & Chialvo, D. R.
Enhanced repertoire of brain dynamical states during the psychedelic
experience. Hum. Brain Mapp. 35, 54425456 (2014).
57. Tagliazucchi, E., Crossley, N., Bullmore, E. T. & Laufs, H. Deep sleep divides
the cortex into opposite modes of anatomicalfunctional coupling. Brain
Struct. Funct. 221, 42214234 (2016).
58. Littner, M. et al. Practice parameters for using polysomnography to evaluate
insomnia: an update. Sleep 26, 754760 (2003).
59. Larson-Prior, L. J., Ju, Y. E. & Galvin, J. E. Corticalsubcortical interactions in
hypersomnia disorders: mechanisms underlying cognitive and behavioral
aspects of the sleepwake cycle. Front. Neurol. 5, 165 (2014).
60. Wei, Y. et al. Sleep stage transition dynamics reveal specic stage 2
vulnerability in insomnia. Sleep 40, zsx117 (2017).
61. Carskadon, M. A. & Dement, W. C. in Principles and Practice of Sleep
Medicine5th edn (eds. Kryger, M. H., Roth T. & Dement, W. C.) Ch. 2 (W.B.
Saunders, Elsevier, Philadelphia, 2011).
62. Rosenberg, R. S. & Van Hout, S. The American Academy of Sleep Medicine
inter-scorer reliability program: sleep stage scoring. J. Clin. Sleep. Med. 9,
8187 (2013).
63. Colrain, I. M. & Campbell, K. B. The use of evoked potentials in sleep
research. Sleep Med. Rev. 11, 277293 (2007).
64. Ogilvie, R. D. The process of falling asleep. Sleep Med. Rev. 5, 247270
65. Goupil, L. & Bekinschtein, T. A. Cognitive processing during the transition to
sleep. Arch. Ital. Biol. 150, 140154 (2012).
66. Tassi, P. & Muzet, A. Sleep inertia. Sleep Med. Rev. 4, 341353 (2000).
67. Glover, G. H., Li, T. Q. & Ress, D. Image-based method for retrospective
correction of physiological motion effects in fMRI: RETROICOR. Magn.
Reson. Med. 44, 162167 (2000).
68. Allen, P. J., Polizzi, G., Krakow, K., Fish, D. R. & Lemieux, L. Identication of
EEG events in the MR scanner: the problem of pulse artifact and a method for
its subtraction. Neuroimage 8, 229239 (1998).
69. Rubinov, M. & Sporns, O. Complex network measures of brain connectivity:
uses and interpretations. Neuroimage 52, 10591069 (2010).
70. Leicht, E. A. & Newman, M. E. J. Community structure in directed networks.
Phys. Rev. Lett. 100, 118703 (2008).
G.D. was supported by the Spanish Research Project PSI2016-75688-P (AEI/FEDER) and
by the European Unions Horizon 2020 research and innovation programme under
Grant agreement no. 720270 (HBP SGA1). M.L.K. was supported by the ERC Con-
solidator Grant CAREGIVING (615539) and the Center for Music in the Brain, funded
by the Danish National Research Foundation Grant DNRF117. J.C. was supported under
the project NORTE-01-0145-FEDER-000023. The authors would like to thank Martin
Dietz for his inputs regarding the hemodynamic response function convolution.
Author contributions
H.L. oversaw data acquisition. E.T., A.B.A.S. and D.V. preprocessed the data. A.B.A.S., M.L.
K., D.V., S.F.V.N. and M.W. conceptualised the data analyses. A.B.A.S., D.V., M.L.K. and
K.R. analysed the data. A.B.A.S., D.V. and M.L.K. wrote the manuscript. A.B.A.S., D.V., M.
L.K., M.W., G.D., E.V.S., P.V., J.C., H.L., E.T., S.F.V.N. and K.R. edited the manuscript.
Additional information
Supplementary Information accompanies this paper at
Competing interests: The authors declare no competing interests.
Reprints and permission information is available online at
Journal peer review information: Nature Communications thanks the anonymous
reviewers for their contribution to the peer review of this work. Peer reviewer reports are
Publishers note: Springer Nature remains neutral with regard to jurisdictional claims in
published maps and institutional afliations.
Open Access This article is licensed under a Creative Commons
Attribution 4.0 International License, which permits use, sharing,
adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative
Commons license, and indicate if changes were made. The images or other third party
material in this article are included in the articles Creative Commons license, unless
indicated otherwise in a credit line to the material. If material is not included in the
articles Creative Commons license and your intended use is not permitted by statutory
regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this license, visit
© The Author(s) 2019
14 NATURE COMMUNICATIONS | (2019) 10:1035 | |
Content courtesy of Springer Nature, terms of use apply. Rights reserved
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
use bots or other automated methods to access the content or redirect messages
override any security feature or exclusionary protocol; or
share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at
... We highlight that several of our results were consistent with the previous literature, regardless of the phenomenological nature of the Hopf bifurcation model [23][24][25] . It also worthwhile to point out that SC-FC similarity as a metric bias the results to Gaussian approximation of data cloud, pushing the model into the linear regime (which was not certified by peer review) is the author/funder. ...
... This also suggests that nonlinear compression via variational autoencoders could represent a novel method to infer scalar signatures of consciousness from neuroimaging data. Accordingly, other methods for dimensionality reduction have revealed consistent results when applied to neural activity measured during sleep and anaesthesia[23][24][25] .While previous computational efforts addressed the outcome of simulated perturbations in terms of the global state of the brain19,22,[26][27][28][29][30][31] , our work provides a series of distinct new insights. We demonstrated that the overall effect of stimulating the cortex of unconscious individuals is to displace the state towards conscious wakefulness, as clearly visualized by the arrows in the latent space ofFig. ...
Brain states are frequently represented using a unidimensional scale measuring the richness of subjective experience (level of consciousness). This description assumes a mapping between the high-dimensional space of whole-brain configurations and the trajectories of brain states associated with changes in consciousness, yet this mapping and its properties remain unknown. We combined whole-brain modelling, data augmentation and deep learning for dimensionality reduction to determine a mapping representing states of consciousness in a low-dimensional space, where distances parallel similarities between states. An orderly trajectory from wakefulness to brain injured patients is revealed in a latent space whose coordinates represent metrics related to functional modularity and structure-function coupling, both increasing alongside loss of consciousness. Finally, we investigated the effects of model perturbations, providing geometrical interpretation for the stability and reversibility of states. We conclude that conscious awareness depends on functional patterns encoded as a low-dimensional trajectory within the vast space of brain configurations.
... Similarly to clustering methods, the HMM method does not represent a biophysical model and hence requires that the number of states be specified a priori. The commonly used criterion for choosing an "optimal" number of states is to compare different models in terms of goodness-of-fit using the free-energy index (Stevner et al., 2019). The free-energy is the cost function that the variational Bayesian inference aims to minimize, and is typically used for model comparison and selection (i.e., the lower the better). ...
Studies of resting-state functional connectivity (FC), measured by functional magnetic resonance imaging (rsfMRI), have revealed extensive functional connections between the cerebellum and association regions in the brain, supporting an important role for the cerebellum in cognition. These findings have been based on static FC measures averaged across entire scans spanning a few minutes. However, this is a narrow view that has been recently challenged, with findings pointing to the presence of an ongoing, behaviorally relevant dynamics in resting-state FC occurring at short timescales of a few seconds, which, given the dynamic nature of the brain, is a more natural view that may encode information about complex cognitive functions. So far, however, the cerebellum has been overlooked in most, if not all, studies of dynamic FC, despite its well-recognized role in coordinating complex cognitive functions. In this thesis, we hypothesized that the dynamics of cerebro-cerebellar FC, during rest, may be behaviorally relevant, capturing aspects of cognition and behavior not accounted for by static FC and exhibiting alterations in brain disorders commonly associated with cerebro-cerebellar dysfunction, such as alcohol use disorder (AUD). We tested these hypotheses in two separate studies focusing on the dynamics of cerebro-cerebellar FC in relation to complex traits and disorders, such as impulsivity (first study) and AUD (second study). The first study has been motivated by a recent hypothesis for a role of the cerebellum in impulsivity; a complex personality trait defined as the tendency to act without foresight. We hypothesized that individual differences in normal impulsivity traits could be associated with the (static) strength and (dynamic) temporal variability of cerebro-cerebellar resting-state FC. We tested this hypothesis using rsfMRI data and self-report questionnaires of impulsivity (UPPS-P and BIS/BAS) collected from a group of healthy individuals. In particular, we employed data-driven techniques to identify cerebral and cerebellar resting-state networks, compute summary measures of static and dynamic FC, and test for associations with self-reported impulsivity. We observed evidence linking multiple forms of impulsivity to the strength and temporal variability of resting-state FC between the cerebellum and a set of highly dynamic and integrative brain networks that support top-down cognitive control and bottom-up reward/saliency processes, supporting our hypothesis that cerebro-cerebellar FC dynamics are behaviorally relevant. In the second study, we hypothesized that the dynamics of cerebro-cerebellar FC at short timescales would differ between AUD and controls, especially in the frontocerebellar circuits. To test this hypothesis, we explored the differences in the dynamic cerebro-cerebellar FC between an AUD group (N=18) and a group of unaffected controls (N=18) by comparing groups on different dynamic connectivity measures. Results revealed altered cerebro-cerebellar FC dynamics in the AUD group characterized by hypervariability of FC within fronto-parieto-cerebellar networks, reduced cerebellar flexibility, and increased cerebellar integration, compared with controls. These results suggest a possible role for the dynamics of fronto-parieto-cerebellar networks in the pathophysiology of this disorder. Taken together, the findings from this thesis highlight the utility of complementing static FC approaches with dynamic FC analysis in furthering our understanding of the functional repertoire of cerebro-cerebellar networks and the neurobiological architecture of complex behaviors and brain disorders.
... In contrast, if the firing frequency fluctuates randomly, microglia will detect it. Interestingly, a recent human study demonstrated that the complexity of neuronal activity during sleep SWA was significantly lower as compared to both wakefulness and REMS (Stevner et al., 2019). We propose that neuronal activity during the biphasic cortical SWA with its predictable and recurring activity pattern, would not stimulate microglial ramification. ...
In nocturnal animals, waking appears during the dark period while maximal non-rapid-eye-movement sleep (NREMS) with electroencephalographic slow-wave-activity (SWA) takes place at the beginning of the light period. Vigilance states associate with variable levels of neuronal activity: waking with high-frequency activity patterns while during NREMS, SWA influences neuronal activity in many brain areas. On a glial level, sleep deprivation modifies microglial morphology, but only few studies have investigated microglia through the physiological sleep-wake cycle. To quantify microglial morphology (territory, volume, ramification) throughout the 24 h light-dark cycle, we collected brain samples from inbred C57BL male mice (n = 51) every 3 h and applied a 3D-reconstruction method for microglial cells on the acquired confocal microscopy images. As microglia express regional heterogeneity and are influenced by local neuronal activity, we chose to investigate three interconnected and functionally well-characterized brain areas: the somatosensory cortex (SC), the dorsal hippocampus (HC), and the basal forebrain (BF). To temporally associate microglial morphology with vigilance stages, we performed a 24 h polysomnography in a separate group of animals (n = 6). In line with previous findings, microglia displayed de-ramification in the 12 h light- and hyper-ramification in the 12 h dark period. Notably, we found that the decrease in microglial features was most prominent within the early hours of the light period, co-occurring with maximal sleep SWA. By the end of the light period, all features reached maximum levels and remained steadily elevated throughout the dark period with minor regional differences. We propose that vigilance-stage specific neuronal activity, and SWA, could modify microglial morphology.
... Metastability refers to the brain lingering in a state before switching to another state. In functional magnetic resonance imaging (fMRI) literature, dynamic functional connectivity has revealed brain connectivity states using unsupervised machine learning methods (Cabral et al., 2017;Preti et al., 2017), and elucidated how the activity of these states varies following perturbations to the resting state, e.g., sleep (Stevner et al., 2019) or the administration of psychedelic drugs (Lord et al., 2019;Olsen et al., 2021). However, the frequency content in blood-oxygen-level-dependent (BOLD) fMRI is limited to very slow oscillations (<0.1 Hz) and thus does not allow for investigation of "real-time" brain state transitions and complicates, for instance, the analysis of evoked responses. ...
Full-text available
Metastable microstates in electro- and magnetoencephalographic (EEG and MEG) measurements are usually determined using modified k-means accounting for polarity invariant states. However, hard state assignment approaches assume that the brain traverses microstates in a discrete rather than continuous fashion. We present multimodal, multisubject directional archetypal analysis as a scale and polarity invariant extension to archetypal analysis using a loss function based on the Watson distribution. With this method, EEG/MEG microstates are modeled using subject- and modality-specific archetypes that are representative, distinct topographic maps between which the brain continuously traverses. Archetypes are specified as convex combinations of unit norm input data based on a shared generator matrix, thus assuming that the timing of neural responses to stimuli is consistent across subjects and modalities. The input data is reconstructed as convex combinations of archetypes using a subject- and modality-specific continuous archetypal mixing matrix. We showcase the model on synthetic data and an openly available face perception event-related potential data set with concurrently recorded EEG and MEG. In synthetic and unimodal experiments, we compare our model to conventional Euclidean multisubject archetypal analysis. We also contrast our model to a directional clustering model with discrete state assignments to highlight the advantages of modeling state trajectories rather than hard assignments. We find that our approach successfully models scale and polarity invariant data, such as microstates, accounting for intersubject and intermodal variability. The model is readily extendable to other modalities ensuring component correspondence while elucidating spatiotemporal signal variability.
Traditionally, in neuroimaging, model-free analyses are used to find significant differences between brain states via signal detection theory. Depending on the a priori assumptions about the underlying data, different spatio-temporal features can be analysed. Alternatively, model-based techniques infer features from the data and compare significance from model parameters. However, to assess transitions from one brain state to another remains a challenge in current paradigms. Here, we introduce a “Dynamic Sensitivity Analysis” framework that quantifies transitions between brain states in terms of stimulation ability to rebalance spatio-temporal brain activity towards a target state such as healthy brain dynamics. In practice, it means building a whole-brain model fitted to the spatio-temporal description of brain dynamics, and applying systematic stimulations in-silico to assess the optimal strategy to drive brain dynamics towards a target state. Further, we show how Dynamic Sensitivity Analysis extends to various brain stimulation paradigms, ultimately contributing to improving the efficacy of personalised clinical interventions.
Rich dynamics are the intrinsic features in brain activity, which could be characterized as sequences of multiple spatio-temporal activity events. However, how to efficiently apply brain dynamics for the recognition of brain states is still unclear and need more investigations. The spiking neural network (SNN) is a promising model with better performance in the pattern recognition of event streams. Thus, this paper proposes an algorithm framework for brain states recognition by fusing brain dynamics and SNN, where the brain dynamics are estimated as the dynamic functional connectivity (DFC) matrices. Through applying the DFC-SNN algorithm to the dataset of resting state electroencephalograph signals from healthy subjects and obsessive compulsive disorder patients, we observed that this algorithm was competent to perform the recognition of pathological brain states. It showed that the convergence of SNN model was rapid within less than 20 epochs, and the accuracy was 87.5% under optimal threshold of DFC matrices. In summary, this is the first attempting for the recognition of brain states via the aspect of brain dynamics. The algorithm framework would be beneficial for the applications of SNN model in the field of neuroscience.
Multivariate time series (MTS) classification is an emerging field with increasing demand. Existing representation learning methods for MTS classification are generally based on self-supervised learning. This results in their inability to maximize the use of labels. In addition, the inherent complexity of MTS makes it difficult to learn latent features. To this end, we introduce a new Mixed Supervised Contrastive Loss (MSCL) for MTS representation learning. To effectively leverage labels, the MSCL is calculated by mixing self-supervised, intra-class and inter-class supervised contrastive learning approaches. Then, based on MSCL, we further propose a novel MIxed supervised COntrastive learning framework for MTS classification (MICOS). It uses the spatial and temporal channels to extract the complicated spatio-temporal features of MTS. Additionally, the MSCL is applied at the timestamp level to capture the multiscale contextual information. Experiments were carried out by performing supervised and self-supervised classification tasks on 30 public datasets from the UEA MTS archives. The results show the reliability and efficiency of MICOS compared with 13 competitive baselines.
Creativity, the ability to generate original and valuable products, has long been linked to semantic retrieval processes. The associative theory of creativity posits flexible retrieval ability as an important basis for creative idea generation. However, there is insufficient research on how flexible memory retrieval acts on creative activities. This study aimed to capture different dynamic aspects of retrieval processes and examine the behavioral and neural associations between retrieval flexibility and creativity. We developed 5 metrics to quantify retrieval flexibility based on previous studies, which confirmed the important role of creativity. Our findings showed that retrieval flexibility was positively correlated with multiple creativity-related behavior constructs and can promote distinct search patterns in different creative groups. Moreover, high flexibility was associated with the lifetime of a specific brain state during rest, characterized by interactions among large-scale cognitive brain systems. The flexible functional connectivity within and between default mode, executive control, and salience provides further evidence on brain dynamics of creativity. Retrieval flexibility mediated the links between the lifetime of the related brain state and creativity. This new approach is expected to enhance our knowledge of the role of retrieval flexibility in creativity from a dynamic perspective.
Conference Paper
Nowadays, high amounts of data can be acquired in various applications, spurring the need for interpretable data representations that provide actionable insights. Algorithms that yield such representations ideally require as little a priori knowledge about the data or corresponding annotations as possible. To this end, we here investigate the use of Kohonen's Self-Organizing Map (SOM) in combination with data-driven low-dimensional embeddings obtained through self-supervised Contrastive Predictive Coding. We compare our approach to embeddings found with an auto-encoder and, moreover, investigate three ways to deal with node selection during SOM optimization. As a challenging experiment we analyze nocturnal sleep recordings of healthy subjects, and conclude that - for this noisy real-life data - contrastive learning yields a better low-dimensional embedding for the purpose of SOM training, compared to an auto-encoder. In addition, we show that a stochastic temperature-annealed SOM-training outperforms both a deterministic and a non-temperature-annealed stochastic approach. Clinical relevance - The hypnogram has for decades been the clinical standard in sleep medicine despite the fact that it is a highly simplified representation of a polysomnography recording. We propose a sensor-agnostic algorithm that is able to reveal more intricate patterns in sleep recordings which might teach us about sleep structure and sleep disorders.
Full-text available
Human brain dynamics and functional connectivity fluctuate over a range of temporal scales in coordination with internal states and environmental demands. However, the neurobiological significance and consequences of functional connectivity dynamics during rest have not yet been established. We show that the coarse-grained clustering of whole-brain dynamic connectivity measured with magnetic resonance imaging reveals discrete patterns (dynamic connectivity states) associated with wakefulness and sleep. We validate this using EEG in healthy subjects and patients with narcolepsy and by matching our results with previous findings in a large collaborative database. We also show that drowsiness may account for previous reports of metastable connectivity states associated with different levels of functional integration. This implies that future studies of transient functional connectivity must independently monitor wakefulness. We conclude that a possible neurobiological significance of dynamic connectivity states, computed at a sufficiently coarse temporal scale, is that of fluctuations in wakefulness.
Full-text available
Brain activity is a dynamic combination of the responses to sensory inputs and its own spontaneous processing. Consequently, such brain activity is continuously changing whether or not one is focusing on an externally imposed task. Previously, we have introduced an analysis method that allows us, using Hidden Markov Models (HMM), to model task or rest brain activity as a dynamic sequence of distinct brain networks, overcoming many of the limitations posed by sliding window approaches. Here, we present an advance that enables the HMM to handle very large amounts of data, making possible the inference of very reproducible and interpretable dynamic brain networks in a range of different datasets, including task, rest, MEG and fMRI, with potentially thousands of subjects. We anticipate that the generation of large and publicly available datasets from initiatives such as the Human Connectome Project and UK Biobank, in combination with computational methods that can work at this scale, will bring a breakthrough in our understanding of brain function in both health and disease.
Full-text available
Resting-state functional magnetic resonance imaging (fMRI) has highlighted the rich structure of brain activity in absence of a task or stimulus. A great effort has been dedicated in the last two decades to investigate functional connectivity (FC), i.e. the functional interplay between different regions of the brain, which was for a long time assumed to have stationary nature. Only recently was the dynamic behaviour of FC revealed, showing that on top of correlational patterns of spontaneous fMRI signal fluctuations, connectivity between different brain regions exhibits meaningful variations within a typical resting-state fMRI experiment. As a consequence, a considerable amount of work has been directed to assessing and characterising dynamic FC (dFC), and several different approaches were explored to identify relevant FC fluctuations. At the same time, several questions were raised about the nature of dFC, which would be of interest only if brought back to a neural origin. In support of this, correlations with electroencephalography (EEG) recordings, demographic and behavioural data were established, and various clinical applications were explored, where the potential of dFC could be preliminarily demonstrated. In this review, we aim to provide a comprehensive description of the dFC approaches proposed so far, and point at the directions that we see as most promising for the future developments of the field. Advantages and pitfalls of dFC analyses are addressed, helping the readers to orient themselves through the complex web of available methodologies and tools.
Full-text available
Integrated information (Φ) is a measure of the cause-effect power of a physical system. This paper investigates the relationship between Φ as defined in Integrated Information Theory and state differentiation (D), the number of, and difference between potential system states. Here we provide theoretical justification of the relationship between Φ and D, then validate the results using a simulation study. First, we show that a physical system in a state with high Φ necessarily has many elements and specifies many causal relationships. Furthermore, if the average value of integrated information across all states is high, the system must also have high differentiation. Next, we explore the use of D as a proxy for Φ using artificial networks, evolved to have integrated structures. The results show a positive linear relationship between Φ and D for multiple network sizes and connectivity patterns. Finally we investigate the differentiation evoked by sensory inputs and show that, under certain conditions, it is possible to estimate integrated information without a direct perturbation of its internal elements. In concluding, we discuss the need for further validation on larger networks and explore the potential applications of this work to empirical studies of consciousness, especially concerning the practical estimation of integrated information from neuroimaging data.
Full-text available
The coupling of anatomical and functional connectivity at rest suggests that anatomy is essential for wake-typical activity patterns. Here, we study the development of this coupling from wakefulness to deep sleep. Globally, similarity between whole-brain anatomical and functional connectivity networks increased during deep sleep. Regionally, we found differential coupling: during sleep, functional connectivity of primary cortices resembled more the underlying anatomical connectivity, while we observed the opposite in associative cortices. Increased anatomical-functional similarity in sensory areas is consistent with their stereotypical, cross-modal response to the environment during sleep. In distinction, looser coupling-relative to wakeful rest-in higher order integrative cortices suggests that sleep actively disrupts default patterns of functional connectivity in regions essential for the conscious access of information and that anatomical connectivity acts as an anchor for the restoration of their functionality upon awakening.
Significance We address the important question of the temporal organization of large-scale brain networks, finding that the spontaneous transitions between networks of interacting brain areas are predictable. More specifically, the network activity is highly organized into a hierarchy of two distinct metastates, such that transitions are more probable within, than between, metastates. One of these metastates represents higher order cognition, and the other represents the sensorimotor systems. Furthermore, the time spent in each metastate is subject-specific, is heritable, and relates to behavior. Although evidence of non–random-state transitions has been found at the microscale, this finding at the whole-brain level, together with its relation to behavior, has wide implications regarding the cognitive role of large-scale resting-state networks.
Study Objectives Objective sleep impairments in Insomnia Disorder (ID) are insufficiently understood. The present study evaluated whether whole-night sleep stage dynamics derived from polysomnography differ between people with ID and matched controls, and whether sleep stage dynamics features discriminate them better than conventional polysomnographic parameters. Methods Eighty-eight participants aged 21–70 years, including 46 with ID and 42 age- and sex-matched controls without sleep complaints, were recruited through and completed two nights of laboratory polysomnography. Data of 100 people with ID and 100 age- and sex-matched controls from a previously reported study were used to validate the generalizability of findings. The second night was used to obtain, in addition to conventional sleep parameters, probabilities of transitions between stages and bout duration distributions of each stage. Group differences were evaluated with non-parametric tests. Results People suffering from ID showed significantly higher empirical probabilities to transition from stage N2 to the lighter sleep stage N1 or wakefulness, and a faster-decaying stage N2 bout survival function. The transition probability from stage N2 to stage N1 discriminated people with ID better than any of their deviations in conventional sleep parameters, including less total sleep time, less sleep efficiency, more stage N1, and more wake after sleep onset. Moreover, adding this transition probability significantly improved the discriminating power of a multiple logistic regression model based on conventional sleep parameters. Conclusions Quantification of sleep stage dynamics revealed a particular vulnerability of stage N2 in insomnia. The feature characterizes insomnia better than—and independently of—any conventional sleep parameter.
Advances in neuroimaging have greatly improved our understanding of human sleep from a systems neuroscience perspective. However, cognition and awareness are reduced during sleep, hindering the applicability of standard task-based paradigms. Methods recently developed to study spontaneous brain activity fluctuations have proven useful to overcome this limitation. In this review, we focus on the concept of functional connectivity (FC, i.e. statistical covariance between brain activity signals) and its application to functional magnetic resonance imaging (fMRI) data acquired during sleep. We discuss how FC analyses of endogenous brain activity during sleep have contributed towards revealing the large-scale neural networks associated with arousal and conscious awareness. We argue that the neuroimaging of deep sleep can be used to evaluate the predictions of theories of consciousness; at the same time, we highlight some apparent limitations of deep sleep as an experimental model of unconsciousness. In resting state fMRI experiments, the onset of sleep can be regarded as the object of interest but also as an undesirable confound. We discuss a series of articles contributing towards the disambiguation of wakefulness from sleep on the basis of fMRI-derived dynamic FC, and then outline a plan for the development of more general and data-driven sleep classifiers. To complement our review of studies investigating the brain systems of arousal and consciousness during healthy sleep, we then turn to pathological and abnormal sleep patterns. We review the current literature on sleep deprivation studies and sleep disorders, adopting the critical stance that lack of independent vigilance monitoring during fMRI experiments is liable for false positives related to atypical sleep propensity in clinical and sleep-deprived populations. Finally, we discuss multimodal neuroimaging as a promising future direction to achieve a better understanding of the large-scale FC of the brain during sleep and its relationship to mechanisms at the cellular level.
Introduction Maxillomandibular advancement (MMA) is an effective alternative way to treat severe obstructive sleep apnea (OSA). However, the convex facial profiles of Chinese patients significantly limit its promotion. In order to gain enough upper airway enlargement without esthetic disaster, we added counterclockwise rotation of maxillomandibular complex in the routine MMA. For the severe OSA is life threatening and the psychological status changes by aging, it is important to conclude a surgical criteria for middle aged patients. In this study, we investigated the esthetic outcomes of counterclockwise maxillomandibular advancement (CMMA) in middle aged patients, which will be important for making surgical decision. Methods 16 severe OSA patients accepted CMMA aged from 40 to 60 were enrolled in this study. None of them took the advice of orthodontic treatment around surgery. The patients were followed up for 6 to 12 months and underwent physical measurement, facial photographic assessment, cephalometry, polysomnography (PSG) and Epworth sleepiness scale (ESS). Patients were asked to score their satisfaction of facial changes by 5 points Likert scales. 30 medical students were invited to blindly evaluate the pre- and postoperative aesthetic appearance by a 10-point visual analogue scale. Cephalometric changes before and after surgery were also analyzed. Results After CMMA, the apnea-hypopnea index (AHI) decreased from 56.2 ± 14.4 to 11.2 ± 5.8 (P<0.001), minimum SpO2 (pulse oxygen saturation, %) increased from 76.9 ± 9.1 to 87.9 ± 5.5 (P<0.001), and ESS decreased from 12.8 ± 2.0 to 7.3 ± 1.7 (P<0.001). The cephalometric measurement showed that although maxilla and mandible protruded and the occlusal plan decreased after MMA, the facial convexity angle, aesthetic plan were not getting worse. The 5 points Likert scales revealed 14 patients (87.5%) were satisfied or very satisfied with their facial changes. The esthetic scores from the medical students indicated that 10 of 16 patients (62.5%) were significantly higher postoperatively, and 1 patient (6.3%) was less attractive because of maxilla protrusion. Conclusion CMMA provides a possible way to achieve a balance between OSA release and facial appearance for middle aged OSA patients in China. Support (If Any) The authors acknowledge the financial support from the Science and Technology Commission of Shanghai Municipality (16140903900).