Functional geometry of the cortex encodes dimensions of consciousness

The study addresses a fundamental question in neuroscience: how the multiple dimensions of consciousness (traditionally parsed into awareness/content and wakefulness/level, with a recently proposed behavioral dimension) are represented in neural terms. Traditional region- or network-centric localization approaches struggle to map specific functions or conscious states to isolated areas due to functional pluripotency and overlap. The authors propose that changes in conscious state reflect changes along one or more neurofunctional dimensions encoded by the brain’s intrinsic functional geometry. They posit that macroscale cortical gradients—continuous axes from unimodal sensory/motor regions to transmodal association cortices—offer a principled representation of these neurofunctional dimensions, and that understanding the coupling between this spatial geometry and temporal brain dynamics can illuminate the neural basis of consciousness.

Prior work has conceptualized consciousness along multiple theoretical dimensions (awareness and wakefulness, and behavior), but neural instantiation of these dimensions remained unclear. Traditional searches for neural correlates of consciousness often examined discrete regions (e.g., prefrontal or posterior cortex) or canonical networks, which do not map cleanly to single functions. Advances in macroscale gradient mapping reveal dominant functional axes spanning unimodal to transmodal cortices and differentiating sensory modalities, with transient cortex-wide fMRI co-activations that propagate as waves aligned to these gradients. These findings suggest that spatial functional geometry and temporal dynamics are intertwined across scales, motivating the present gradient-based account of consciousness.

Design: Comparative analysis across five fMRI datasets from three sites, encompassing pharmacological perturbations (propofol deep sedation, PDS; propofol general anesthesia, PGA; ketamine anesthesia, KA), neuropathology (unresponsive wakefulness syndrome, UWS), and psychiatric diagnoses (schizophrenia, SCHZ; bipolar disorder, BD; attention deficit hyperactivity disorder, ADHD), with corresponding healthy control or baseline conditions. Behavioral responsiveness was assessed via standardized tasks/scales (e.g., Ramsay scale, command-following). Participants/Datasets: Dataset-1 PDS (n=26; within-subject baseline and PDS), Dataset-2 PGA (n=23 after exclusions; within-subject baseline and PGA), Dataset-3 KA (n=12; within-subject baseline and KA), Dataset-4 UWS (n=7) vs healthy controls (HC; n=16), Dataset-5 psychiatric cohort from OpenNeuro (HC n=116; SCHZ n=44; BD n=49; ADHD n=39). Detailed anesthesia protocols for propofol (TCI Marsh model; effect-site 2.4 µg/ml for deep sedation; 4.0 µg/ml for general anesthesia with adjuncts for intubation) and ketamine (infusions up to 0.1 mg/kg/min) were implemented under clinical monitoring. Behavioral unresponsiveness defined PDS, PGA, KA epochs. Imaging acquisition: Multi-site 3T scanners with standard EPI parameters; runs included resting-state fMRI during baseline and perturbed states. Datasets differed in TR/TE, slice count, and resolution; anatomical T1 images acquired for all. Preprocessing: AFNI-based pipeline: discard initial frames; slice timing correction; rigid motion correction; frame censoring with FD>0.4 mm (plus previous frame); coregistration to T1; spatial normalization (Talairach); band-pass filtering 0.01–0.1 Hz with regression of motion (and derivatives), drift, WM/CSF; spatial smoothing (6 mm FWHM); z-normalization per voxel; global signal regression (main analysis), with control analyses without GSR. Cortical gradient analysis: Time series extracted from 400 cortical parcels (Schaefer atlas). For each participant/condition, 400×400 Pearson correlation functional connectivity matrices computed, z-transformed, and thresholded at 90% sparsity (retain top 10% weights per row). A normalized cosine angle affinity matrix was computed. Diffusion map embedding (BrainSpace toolbox) with parameters α=0.5, t=0 was applied to derive gradients. Group-level gradients aligned via Procrustes rotation to a reference HCP subsample to address eigenvector sign/order ambiguities; individual-level gradients computed and aligned similarly. The first three gradients were analyzed, explaining ~37% variance (G1: 15.0±1.7%; G2: 11.9±1.1%; G3: 9.8±0.9%). Metrics in gradient space: For each participant in the 3D gradient space, computed (1) gradient numerical range (max–min eigenvector values) per gradient (indexing segregation of extremes); (2) global dispersion (sum squared Euclidean distance of all parcels to the global centroid) indexing overall differentiation; (3) network eccentricity (squared Euclidean distance between a network centroid and the global centroid) for seven canonical networks (VIS, SMN, DAN, VAN, LIM, FPN, DMN); (4) between-network distances (squared Euclidean distance between network centroids). Temporal dynamics: Co-activation pattern (CAP) analysis matched each fMRI volume to one of eight CAP centroids (DMN, DAN, FPN, SMN, VIS, VAN, GN+ and GN−) based on maximal spatial similarity; occurrence rates computed per state for VAN+, DAN+, DMN+. Statistics: Bayesian Paired Samples T-tests (two-tailed) for within-subject comparisons (PDS, PGA, KA) vs baseline; Bayesian Independent Samples T-tests for between-group comparisons (UWS, SCHZ, BD, ADHD) vs healthy controls. Bayes factors (BF10) with default Cauchy prior (scale 0.707); emphasis on strong (BF10>10) evidence; classical t-tests with FDR correction (α=0.05) also reported. Bayesian Kendall correlations between selected between-network distances (DMN–VAN, DMN–DAN, VAN–DAN) and CAP occurrence rates (VAN+, DAN+, DMN+) across all scans (n=393). Robustness checks varied sparsity (0–90%) and α (0–1).

- Gradient ranges: Distinct patterns of degradation across states. - Gradient-1 (unimodal→transmodal) decreased in PDS (t(25)=4.331, p=0.002, BF10=136.507), PGA (t(22)=4.000, p=0.002, BF10=55.220), and UWS (t(21)=5.660, p<0.001, BF10=1098.829). Not reduced in KA or psychiatric diagnoses. - Gradient-2 (visual→somatomotor) decreased in KA (t(11)=5.170, p=0.001, BF10=109.209) and SCHZ (t(158)=3.772, p=0.006, BF10=102.473). - Gradient-3 (visual/DMN→multiple-demand) decreased in PGA (t(22)=5.494, p<0.001, BF10=1416.786). - Global dispersion in 3D gradient space decreased, indicating globally reduced differentiation, in PGA (t(22)=4.736, p<0.001, BF10=272.220), UWS (t(21)=4.182, p=0.001, BF10=58.409), KA (t(11)=3.781, p=0.006, BF10=15.735), and SCHZ (t(158)=3.076, p=0.013, BF10=12.948). - Network eccentricity (distance of network centroid from global centroid): - PDS: Reduced VAN (t(25)=8.677, p<0.001, BF10≈2.39e6). - PGA: Reduced DAN (t(22)=5.009, p<0.001, BF10=493.899), VAN (t(22)=6.237, p<0.001, BF10=6940.439), DMN (t(22)=4.957, p<0.001, BF10=440.966; confirmed by non-parametric tests). - KA: Reduced VIS (t(11)=5.224, p=0.001, BF10=117.198), VAN (t(11)=4.387, p=0.003, BF10=37.348; non-parametric corroborated), FPN (t(11)=4.281, p=0.003, BF10=32.192), DMN (t(11)=4.984, p=0.001, BF10=85.025). - UWS: Reduced SMN (t(21)=4.871, p<0.001, BF10=226.944), DMN (t(21)=4.768, p<0.001, BF10=185.148), DAN (t(21)=3.960, p=0.002, BF10=38.120), FPN (t(21)=3.989, p=0.002, BF10=40.277), VAN (t(21)=3.716, p=0.002, BF10=24.030; non-parametric corroborated). - SCHZ: Reduced SMN (t(158)=3.254, p=0.009, BF10=21.188; non-parametric supportive). No strong evidence for BD or ADHD. - Between-network distances: - Shortened VAN–DMN distance was a common feature of depressed states (PDS, PGA, KA; moderate in UWS by non-parametric test), aligning with behavioral unresponsiveness. - PGA, KA, UWS (but not PDS): shortened DAN–VIS, DMN–SMN (PGA via non-parametric), DMN–DAN, FPN–VAN, DMN–FPN. - KA and SCHZ shared shortened SMN–VIS, but KA showed VIS-dominant broad shortening (e.g., SMN–VIS, DAN–VIS, VAN–VIS, FPN–VIS, DMN–VIS), whereas SCHZ showed SMN-dominant shortening (SMN–VIS, VAN–SMN, FPN–SMN). - BD showed shortened DMN–VIS; ADHD showed no strong effects. - Spatial–temporal coupling: Across participants/conditions (n=393), reduced distances among DMN, DAN, and VAN were associated with shifts in CAP occurrence rates: DMN+ and DAN+ occurrences decreased, VAN+ occurrences increased during depressed states, linking compressed VAN–DMN/DAN geometry with altered temporal dynamics of co-activations. - Overall, the first three cortical gradients accounted for ~37% of connectome variance; patterns of gradient degradation and network reconfiguration were state-specific and reproducible across datasets.

Findings support the central hypothesis that dimensions of consciousness are encoded in the brain’s functional geometry. Three cortical gradients map onto putative neurofunctional dimensions: Gradient-1 reflects awareness (integration from unimodal to transmodal processing), Gradient-2 reflects sensory organization (differentiation among sensory modalities), and Gradient-3 reflects cortical arousability (differentiation between visual/DMN and multiple-demand/attention systems). Depressed states with loss of awareness (propofol deep sedation, general anesthesia, UWS) showed Gradient-1 degradation, whereas ketamine—associated with disconnected conscious experiences—spared Gradient-1 but degraded Gradient-2, consistent with sensory disorganization. Loss of arousability under high-dose propofol aligned with Gradient-3 degradation, suggesting a link to cortical arousal mechanisms. Network-level analyses revealed a convergent reduction in VAN–DMN separation across depressed states, correlating with behavioral unresponsiveness and implicating altered salience–default interactions in gating access to conscious contents. The spatial compression of network geometry covaried with temporal dynamics: reduced propagation of DMN+/DAN+ and increased VAN+ co-activations, indicating that altered functional geometry constrains dynamic state transitions. These results unify diverse etiologies (pharmacologic, neuropathologic, psychiatric) within a multidimensional framework, reconcile prior region-centric debates, and suggest gradient-informed markers for assessing consciousness.

Cortical gradients provide a principled, multidimensional neurofunctional framework for consciousness. The unimodal→transmodal, visual→somatomotor, and visual/DMN→multiple-demand gradients index awareness, sensory organization, and cortical arousability, respectively. Disruptions of consciousness manifest as state-specific degradations of one or more gradients and convergent reconfigurations in network geometry, notably reduced VAN–DMN separation linked to behavioral unresponsiveness, with corresponding shifts in temporal co-activation dynamics. This framework advances beyond cognitive-behavioral taxonomies to a data-driven characterization of conscious states across health and disease. Future work should: (1) determine the optimal number and hierarchy of gradients; (2) incorporate subcortical gradients and their coupling to cortical geometry (e.g., ARAS/thalamus); (3) test causal perturbations (e.g., stimulation) to probe Gradient-3 and arousability; (4) examine dose–response relationships (e.g., ketamine); and (5) develop translational, behavior-independent biomarkers and machine-learning tools for clinical assessment of disorders of consciousness.

- Gradient scope: Analyses focused on the first three cortical gradients; the optimal dimensionality to represent consciousness remains undetermined. - Subcortical omission: Subcortical gradients (e.g., thalamus, brainstem ARAS) were not analyzed due to resolution constraints; the relative contributions of thalamocortical vs corticocortical interactions to cortical gradients are unresolved. - Sample constraints: Limited UWS sample (n=7) due to cortical distortions and parcellation coverage; smaller sizes in KA and UWS reduce power. - Multi-site heterogeneity: Differences in acquisition parameters and protocols across datasets; comparisons controlled within dataset, but residual site effects may remain. - Study design: Non-randomized designs in within-subject datasets; investigators not blinded to condition during analyses. - Method parameters: Some effects (e.g., Gradient-2 range) were sensitive to sparsity thresholds (e.g., detected at 90%), warranting further methodological validation.