Contribution
The paper gives an information-theoretic operationalisation of a global workspace and of its input and output sides, and applies it to fMRI. It proposes the SAPHIRE architecture: specialised modules interact redundantly, “gateways” gather their information synergistically, the workspace integrates it, and “broadcasters” send copies back out redundantly. The workspace and its two sides are mapped onto known networks. Then, across two causes of unconsciousness, it finds that the consistent loss of pairwise integrated information is on the gateway side. The measure used is a corrected Φ (ΦR); the uncorrected Φ does not show the effect.
Key insight
Synergy and redundancy separate two things functional connectivity blurs: combining distinct inputs, and copying the same content to many places. A region that is synergistic with many modules looks like an entry point. A region that is redundant with many modules looks like a broadcaster. On that reading, losing consciousness in this data is losing integration at the entry points, mainly between anterior and posterior default-mode regions, not losing the broadcast.
Assumptions
- Linear-Gaussian BOLD. Regional BOLD signals are deconvolved for the haemodynamic response and treated as jointly Gaussian. Then the minimum-mutual-information PID applies, which is justified because “linear-Gaussian models are sufficiently good descriptors of functional MRI timeseries”.
- Pairwise, undirected measures. Every quantity is computed for pairs of regions (X, Y) from the past and future of each, and then treated as an undirected network. The paper says this “does not amount to measuring the integrated information of the brain as a whole”.
- The measures (Methods, Eqs. 4–7).
- Red(X,Y) = min{I(X_{t−τ}; X_t), I(X_{t−τ}; Y_t), I(Y_{t−τ}; X_t), I(Y_{t−τ}; Y_t)}.
- Syn(X,Y) = I(X_{t−τ}, Y_{t−τ}; X_t, Y_t) − max{I(X_{t−τ}; X_t, Y_t), I(Y_{t−τ}; X_t, Y_t)}.
- Φ = I(X_{t−τ}, Y_{t−τ}; X_t, Y_t) − I(X_{t−τ}; X_t) − I(Y_{t−τ}; Y_t), the whole-minus-sum measure of IIT 2.0 in Barrett and Seth’s empirical form.
- ΦR = Φ + Red, shown non-negative in Mediano et al. (2021).
- The analysis uses persistent synergy and persistent redundancy, “the information that is always carried synergistically (respectively, redundantly)”.
- Workspace membership. Taken from earlier work (Luppi et al. 2022b). A region’s synergy rank minus its redundancy rank, by nodal strength, is computed over the group-average HCP matrices. Regions favouring synergy form the workspace. The gradient is relative: a positive value “does not necessarily mean that the region’s synergy is greater than its redundancy”.
- Gateways and broadcasters, by sign. Within the workspace, the rank of each region’s participation coefficient in the synergy network minus its rank in the redundancy network is computed, with modules taken as the seven Yeo networks plus a subcortical one. Positive means gateway; negative means broadcaster. Every workspace region is therefore one or the other by construction.
- Unconsciousness. Propofol at Ramsay 5 (no response to verbal command), and chronic disorders of consciousness, which include both unresponsive wakefulness and minimally conscious patients. Both are treated as loss of consciousness.
Key results
- Data.
- The HCP sample is N = 100, with 454 ROIs (Schaefer-400 plus Tian-54) and a replication at 232 ROIs.
- Propofol volunteers are “N=15” in the Introduction, but the Methods report n = 19 recruited, 3 excluded and “a total n=16 for analysis”. The paper does not reconcile the two.
- The DOC sample is 22 adults (from 71 recruited), aged 17–70, 13 months post-injury on average, with UWS or MCS diagnoses.
- Where gateways and broadcasters are (Fig. 2). “Gateways reside primarily in the brain’s default mode network”, and “broadcasters are mainly located in the executive control network, especially lateral prefrontal cortex”. The participation coefficients correlate between a priori modules and Louvain modules (spin test), and the classification replicates at the coarser parcellation.
- ΦR across states (Fig. 4). The NBS F-threshold was 9. ΦR is
reorganised for DOC against awake (p < 0.001), for propofol against
pre-induction (p < 0.001), and for propofol against recovery
(p < 0.001), with both increases and decreases.
- A composite test asks for edges that change in all three contrasts. Its statistic is the minimum of the three F-scores, with one dataset reshuffled at a time. It is a least-favourable-configuration test whose “details will be described in a future publication”.
- Common decreases: bilateral precuneus and medial prefrontal cortex, bilateral inferior parietal cortex, and left temporal cortices.
- Common increases: frontal cortex, “especially lateral prefrontal cortex”.
- The net balance favours reduction: the sum of F-scores over the significant edges is −25.37.
- The original Φ (Fig. 4—supplement 2). It “did not identify any reductions in integrated information that were common across anaesthesia and disorders of consciousness, instead only identifying common increases”.
- Gateways, not broadcasters (Fig. 5). The regions with consistent ΦR reductions “play the role of gateway nodes … rather than broadcaster nodes”. The caption says “all regions are gateways (violet)”. The reductions read as an “anterior-posterior disconnection … between DMN nodes that act as gateways”.
- Robustness. Decreases in medial prefrontal and posterior cingulate or precuneus survive a coarser parcellation, an extent-based NBS threshold and F > 12. The increases do not: “no region consistently showing increases”.
- Propofol against wake alone. Reductions also appear in thalamic, auditory and somatomotor regions. The authors read this as thalamocortical isolation, with no EEG to confirm it.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Within the synergy-defined workspace, regions with synergy-dominant intermodular connections are mainly DMN, and regions with redundancy-dominant ones mainly fronto-parietal | moderate | HCP N = 100, group-average matrices, replicated at a second parcellation and module definition (Fig. 2) |
| C2 | Pairwise ΦR between DMN gateway regions falls under both propofol and DOC, and recovers after propofol | moderate | NBS with a composite shared-effect test across three contrasts; robust to parcellation and threshold (Figs. 4–5) |
| C3 | The original whole-minus-sum Φ shows no such common reduction, so ΦR is the empirically valid measure | moderate for the contrast; weak for “validity” | Fig. 4—supplement 2 |
| C4 | Loss of consciousness is a breakdown of the entry points to the synergistic core | weak | interpretation of C2; resting state, correlational |
| C5 | Different subsets of the workspace are differentially involved in supporting consciousness | weak | the non-robust increases near broadcasters, against the robust gateway decreases |
Method
- Preprocess resting-state fMRI (CONN, aCompCor, no global signal regression, 0.008–0.09 Hz) and deconvolve the haemodynamic response.
- Compute Gaussian MMI-ΦID quantities for every pair of the 454 regions with JIDT.
- Build networks of synergy, redundancy and ΦR.
- Rank by nodal strength to find the workspace (from the earlier paper), and by participation coefficient to split it into gateways and broadcasters.
- Test state differences in ΦR edges with the network-based statistic and the composite shared-effect test.
Concepts
- Synergistic workspace. As defined here: regions whose synergy rank exceeds their redundancy rank by nodal strength, relative to the rest of the brain.
- Gateway. A workspace region whose participation coefficient ranks higher in the synergy network than in the redundancy network. It is “synergistic with many modules”.
- Broadcaster. A workspace region with the opposite rank order. It “possess[es] information that is duplicated across multiple modules”. Broadcasting is read as redundancy: “Sameness of information corresponds to redundancy”.
- ΦR (revised Φ). The whole-minus-sum Φ plus the double redundancy, which removes the negative values of Φ in redundancy-dominated systems. It counts synergistic and transferred information.
Connections
The paper builds on the ΦID framework of Mediano, Rosas and colleagues, whose emergence side the record reads in LIT-025. It extends the authors’ earlier “synergistic core” paper (Luppi et al. 2022b), which supplies the workspace definition, and their DOC and propofol datasets (Luppi et al. 2019). It positions itself between GNWT (Mashour et al. 2020, LIT-631) and IIT. It aligns with Shanahan’s “connective core”, and contrasts with Deco et al.’s task-based, directed (transfer-entropy) workspace.
Bearing on the record
- QUESTION-061 asks whether conscious access is one exclusive
broadcast to every consumer, or whether a content can reach some
consumers and not others. LIT-949 was filed as “measured evidence on
whether conscious access is one uniform broadcast or has distinct
parts”. The paper bears on that weakly and indirectly.
- No contents. It measures the state of consciousness at rest, not the access of any content. There is no stimulus, no report and no consumer-by-consumer measurement, so it cannot show a content available to one system and absent from another, which is what the question says would answer it.
- The split is built in. The division of the workspace into an input side and an output side is imposed by the method: every workspace region is classed by the sign of a rank difference. What is measured is where the two classes fall (DMN against FPN), and that the consistent loss is on the input side.
- What it gives the question. The authors read the robust gateway decreases, against the non-robust broadcaster increases, as “different subsets of the global workspace may be differentially involved”. That is evidence that the workspace is not functionally uniform with respect to consciousness. It is not evidence that access is compartmented by consumer.
- THEORY-096 (GNW). The paper is consistent with lateral prefrontal cortex as a broadcaster, a role the GNW review assigns it. But the consciousness-tracking change it finds is in the default-mode gateways. Lateral prefrontal ΦR rose, not fell, in unconsciousness, though not robustly. It tests no ignition and no exclusivity, so it neither supports nor refutes THEORY-096’s mechanism.
- THEORY-098. Under propofol alone, reductions also reach thalamic, auditory and somatomotor regions. That fits a loss of sensory input to the core, as the authors say. It is a state effect at rest and does not bear on whether recurrent sensory activity is the basis of experience.
- CLAIM-003. It does not bear. The claim is about compartmented access, and the paper measures an undirected, content-free integration between pairs.
- CLAIM-158 and the record’s ΦID material. The paper uses ΦID for pairwise integration only. It does not use the downward-causation or causal-decoupling terms that the drive-profile work draws on.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- BOLD at a 2 s TR is an indirect, slow proxy. Deconvolution mitigates this and does not remove it.
- The measures are pairwise and undirected. They are not IIT’s Φ of the whole system, and they rest on IIT 2.0, not IIT 3.0 or 4.0.
- The results are correlational. The authors disclaim that ΦR causes consciousness or is identical to it.
- The gateway and broadcaster classes are relative ranks, defined on the HCP group average and transferred to the clinical datasets. They are not re-derived in those datasets.
- The DOC group includes minimally conscious patients, so “loss of consciousness” there is graded, not absent. Propofol participants may have been “merely disconnected but still conscious”.
- Resting state “may have resulted in an overestimation of the DMN’s role in consciousness, and an under-estimation of the FPN”.
- The shared-effect test’s details are deferred to a future publication. The propofol sample size is given as both 15 and 16.
Open questions
- Does the gateway/broadcaster asymmetry hold in task data with stimulus contents and report, where access can be measured per consumer? The authors name task paradigms and higher temporal resolution as next steps.
- Do causal interventions that restore consciousness (central thalamic stimulation) restore gateway ΦR? The paper proposes the test.
- Would a directed decomposition (ΦID of transfer entropy) show broadcasting as directed redundancy from the FPN, rather than as inferred from undirected participation? (Annotated 2026-10-17: a PID of transfer entropy exists, Williams and Beer’s split into unique and synergistic parts, THEORY-712. Since transfer entropy carries no redundancy at all, directed redundancy would have to come from time-delayed mutual information or a ΦID of the joint past, not from transfer entropy itself; CLAIM-335. That inference is the record’s.)