Dynamic coordination, consciousness, and structural connectivity

Citation: Demertzi, A. et al. Human consciousness is supported by dynamic complex patterns of brain signal coordination. Science Advances 5(2), eaat7603 (2019). DOI; PMC; institutional full PDF.

Reading record

Read on 29 September 2026: the complete published 11-page article, including introduction, results, discussion, all Materials and Methods, all three figure captions, references, declarations, and data availability. All three main figures were also inspected visually from rendered PDF pages 3–5. The local extraction was read without truncation in contiguous chunks covering lines 1–633.

Supplementary reading is complete. Initial PMC/publisher/Europe PMC attempts failed, but a later official Europe PMC supplementary-archive download succeeded. All 26 supplement pages were read, including supplementary methods 1–2, figures S1–S11, all rows of tables S1–S2 and their notes. All eleven supplementary figures and both tables were visually inspected. No raw fMRI data or code were obtained or rerun; reading the published clinical table is not a reanalysis. Provenance, earlier access failures, and hashes are in the manifest.

Question and design

Can recurring patterns of changing interregional brain-signal coordination distinguish preserved from impaired consciousness, and do these patterns generalize to covert command-following and propofol anesthesia? How closely do those functional patterns resemble a structural connectivity network?

This is principally a cross-sectional, multisite observational fMRI study, with a separate anesthetized patient cohort. It is not an experiment that reconstructs or manipulates a connectome.

Of 169 initially included people, 10 patients were excluded because signals from selected regions were partially missing. The analyzed sample was 159 people: 47 healthy controls and 112 patients. The three datasets were:

Clinical classifications used Coma Recovery Scale–Revised assessments. Table S1 records one to 21 assessments across patients with counts provided; one New York case was classified by clinical consensus without a CRS-R count. Thus repeated assessment was not equally extensive for every patient. Failure to show an overt response is not itself proof of absent experience; the covert-command group makes that limitation explicit.

Measurements and analysis

The functional measurement was BOLD fMRI, with acquisition settings varying across sites, rather than direct synaptic or electrical recording. Analyses used 42 spherical regions belonging to six networks. Preprocessing included normalization, 6-mm spatial smoothing, motion nuisance regressors, white-matter/CSF components and a 0.008–0.09 Hz temporal filter. Outlier frames were modeled rather than deleted to preserve time-series order. Reported motion-outlier counts did not differ significantly across the three main groups.

Instantaneous phases were derived using the Hilbert transform, yielding 861 interregional pair features per volume. K-means with Manhattan distance, repeated 500 times, identified four recurring coordination patterns in dataset 1. Four was the main choice; supplementary Fig. S1 supports robustness of the extreme patterns for three through seven clusters. New volumes in datasets 2 and 3 were assigned to the nearest already fitted centroid. Outcomes included occurrence probability, time spent consecutively in a pattern, transition probabilities, probability-distribution entropy and Markov entropy rate.

The structural measurement was an external, group-level diffusion spectrum imaging network from five healthy men, not the patients' own connectomes. A consensus binary network of 998 cortical regions linked regions connected in at least one of those five people. It excluded subcortical regions. The network was reduced to the study's 42-region scheme by averaging connections within spheres. Correlations between each functional centroid and this structural matrix supplied four structure–function similarity values. For each participant, a slope related their four pattern frequencies to those four similarities.

This slope is a descriptive statistic built from four centroid comparisons. It is not an estimate of a synaptic coupling constant or the fraction of consciousness explained by anatomy.

Main findings

1. The most complex pattern was more common with higher clinical levels of consciousness. Pattern 1 contained positive and negative long-range coordination; its frequency tracked the ordered groups UWS < MCS < healthy controls (Spearman rho = 0.7, P < 10^-16). The MCS–UWS frequency difference in Fig. 1B was P = 0.001. Pattern 4 had low overall interregional coordination and the reverse group ordering (rho = −0.6, P < 10^-11; MCS–UWS P = 0.007).

2. Higher functional resemblance to the structural atlas was associated with the less conscious groups. The centroid–atlas correlations displayed in Fig. 1C were approximately 0.24, 0.34, 0.42 and 0.67 for patterns 1–4. The frequency-versus-similarity slopes were 0.34 ± 0.31 in controls, 0.94 ± 0.44 in MCS and 1.5 ± 0.31 in UWS (median ± median absolute deviation). The implication is a restricted repertoire more concentrated on the atlas-like pattern, not that anatomy is absent from conscious dynamics.

3. Temporal organization also differed. Conditional transition probabilities involving pattern 1, and its contiguous duration, differed across groups; UWS patients had a higher pattern-4 self-transition probability than MCS patients. Pattern 1's contiguous duration differed between MCS and UWS (Fig. 2B, P = 0.01). The pattern-4 duration comparison between those two groups was P = 0.1 despite the significant overall three-group effect. Markov entropy rate increased across the ordered groups (Fig. S8, rho = 0.4, P = 4.7 × 10^-6), but its MCS–UWS comparison was not significant (P = 0.08). The shuffled-sequence correction in Fig. S7 retains greater pattern-1 self-transition in MCS than UWS; it changes the interpretation of several other transitions once occupancy is controlled.

4. The covert-cognition evidence is narrower than a broad reading of the abstract suggests. In the 5 Ut+ versus 6 Ut− patients, pattern 4 was less frequent in Ut+ (P = 0.004) and the structure–function slope differed (P = 0.004). Pattern 1's frequency was numerically higher but not significant (P = 0.1). The displayed entropy comparison was also not significant (P = 0.08). Thus this cohort supports a difference in the restricted pattern and structural-resemblance statistic; it does not independently establish every proposed marker.

5. Under propofol, the groups had similar pattern distributions. Pattern 1 was relatively uncommon and pattern 4 common across the anesthetized patients irrespective of baseline diagnosis. However, nonsignificant differences between these small groups are not a formal demonstration of statistical equivalence. Dataset 3 is presented as a separate anesthetized cohort, not a reported paired before/after intervention for every person in dataset 1.

6. Neither pattern uniquely labels experience at every instant. Healthy controls also entered pattern 4, and UWS patients sometimes entered pattern 1. There was no concurrent moment-by-moment experience sampling or polysomnography. The authors consider microsleeps and mind-blanking interpretations but cannot establish them from this design.

The text sometimes describes patterns 2 and 3 as equally probable across groups, but pattern 2 has a small reported ordinal association (rho = 0.2, P < 0.01), and its overall group comparison in Fig. 1B is significant. Supplementary Methods 1 also reports a clinical-group effect for pattern 2 (F(2,116) = 3.6, P < 0.03). The defensible formulation is that patterns 1 and 4 show the clearest contrasting group profiles; exact equality of the intermediate patterns is too strong.

What the supplement adds

• Robustness is bounded. Site-specific clustering recovers broadly similar centroids (S4), and site main effects/interactions are not detected in the reported ANOVAs. S5 reports no detected demographic/etiological interactions. These are sensitivity checks, not proof that such influences are absent under every design.
• The reported graph properties are transformations of the fitted patterns. Supplementary Methods 2 bootstraps 3,033 images per pattern 10,000 times, computes graph metrics on median matrices, and forms a participant's metric by weighting those pattern metrics by that participant's occurrence probabilities. The resulting metrics do not constitute entirely independent physiological evidence in addition to the pattern frequencies. Several calculations use absolute edge weights; they measure graph properties, not directly observed information transfer or IIT's intrinsic integrated information.
• The 42-region choice was partly pragmatic. S6 describes failures of other parcellation/extraction schemes in injured brains. Table S2 defines 10-mm-diameter functional spheres, distinct from the 30-mm spheres used to reduce the structural atlas. S11 checks structural reduction at 10, 20 and 30 mm and retains the broad group ordering.
• Transfer is more convincing than independently recovering the same clusters in small cohorts. S9 shows weaker correspondence when clustering the two small validation cohorts afresh. Better correspondence follows assigning their volumes to the original centroids, as in the main analysis. This distinction limits a claim of completely independent rediscovery.
• Some apparent convergences are nonsignificant comparisons. S10 gives the cross-cohort slope comparisons and does not convert similarity to healthy controls into statistical equivalence. In combination with S8 and the main covert-cognition panel, it supports keeping each positive and null result separate.

Interpretation and limits

The authors interpret the findings as evidence that conscious states involve a richer, less structurally stereotyped repertoire of distributed dynamics. The study offers convergent associations across injury, covert command-following and anesthesia, and it tests centroid transfer to separate cohorts.

It does not identify a universal necessary-and-sufficient test for consciousness. Behavioral diagnosis is imperfect; patients are heterogeneous; the covert-cognition cohort is very small; BOLD is indirect and slow; parcel choice and preprocessing shape the results. The four patterns are model-dependent summaries, not four discovered biological modes whose number is independent of the analysis. The authors' own discussion separates global state from the particular contents of experience.

The structural comparison is especially limited for this research program: it uses a coarse healthy reference atlas, excludes subcortex, and does not measure injury-related structural changes in each patient. A departure from the atlas correlation cannot establish freedom from anatomical causation. Recurrence, indirect pathways and changing physiological parameters can generate functional relationships different from direct anatomical edges.

No held-out clinical diagnostic threshold with sensitivity/specificity is validated here. The paper is also not a causal test that producing pattern 1 restores consciousness. That is an explicitly proposed future intervention.

Relevance to consciousness and connectomics

Supported: structural and functional connectivity are different measurements; the repertoire and temporal organization of brain activity are associated with clinical consciousness, while resemblance to a fixed structural template captures only one aspect of those dynamics. The study does not quantify an incremental predictive benefit after conditioning on that resemblance.

Research inference: an emulation proposal should specify both its structural constraints and the mechanisms/state variables producing dynamics. A single static connectivity similarity score is an inadequate endpoint for testing whether a model reproduces the state changes studied here.

Unresolved: which microscopic details must be measured to predict those dynamics; whether a sufficiently rich structural/molecular snapshot could imply them under an adequate model; whether an artificial system with matching dynamics would experience anything; and whether it would preserve the identity or subjective continuity of a particular person. This paper adjudicates none of those questions.

For Kurisutina, the useful addition is a new target alongside memory or response fidelity: can a mechanistic model reproduce changes in the repertoire of activity under controlled state manipulations? Behavioral resemblance, stored autobiographical information, a structural reconstruction and a consciousness-associated dynamic signature remain separate claims.