Architectural reorganization of functional connectivity after propofol anesthesia exceeds within-state measurement variability and is only marginally reversed during early recovery

Abstract

Propofol anesthesia disrupts functional connectivity, but the magnitude of whole-brain architectural reorganization has not been calibrated against the variability of repeated measurement within a single state, and it is therefore unclear how much apparent reorganization reflects state change rather than estimation noise. We applied Weighted Jaccard (WJ) similarity to fMRI from healthy adults scanned during wakefulness, propofol-induced unresponsiveness, and early recovery (OpenNeuro ds006623), computing full pairwise Spearman correlation matrices among 456 regions. To provide an interpretable reference we computed a within-condition floor: WJ between two non-overlapping, equal-length segments of the same run. Every matrix in the primary analysis was estimated from an identical number of timepoints, because run durations differ across conditions.

Between-state similarity fell significantly below the within-condition floor for all comparisons (floor WJ = 0.544; wakefulness versus recovery 0.486, wakefulness versus unresponsiveness 0.468, unresponsiveness versus recovery 0.458; paired Cohen's d 2.23 to 2.79, all p = 2.4e-07). Expressed against this reference, the state-attributable share of architectural divergence is approximately 11 to 14 percent, substantially smaller than the raw between-state dissimilarity implies. The conclusion was unchanged under four sign treatments: absolute, positive edges only, negative edges only, and a signed variant (all twelve tests p = 2.4e-07). Direct stratified sign-flip rates exceeded the within-condition floor at every magnitude stratum, by 5-fold for the task-matched contrast and 17-fold for unresponsiveness at min|r| ≥ 0.3, indicating genuine sign reorganization among strong edges rather than magnitude change alone.

Reorganization was strongly redistributive (net-versus-gross uniformity R = 0.775 to 0.871), meaning edge strengthenings and weakenings largely cancel, so scalar summaries of total connectivity are insensitive to it by construction. Consistent with this, the change in mean absolute correlation did not distinguish wakefulness from recovery (p = 0.096) while the full-matrix comparison did. Subject-level recovery was small but reliably non-zero (3.2 percent, 95 percent CI 1.8 to 5.7). Group-averaged matrices yielded a recovery fraction an order of magnitude larger; group-level WJ exceeded that of every individual subject in all three comparisons, and this inflation persisted when subject identity was randomly permuted, establishing it as a property of matrix averaging rather than of cross-subject correspondence.

These results support a modest, redistributive, and largely unreversed architectural reorganization after propofol, and indicate that group-averaged connectivity comparisons systematically overstate recovery.

Highlights

• •Similarity is calibrated against a within-condition floor of 0.544, not against 1.0
• •State-attributable architectural divergence is 11 to 14 percent, not roughly half.
• •Robust to four sign treatments; sign reorganization among strong edges is real.
• •Reorganization is redistributive, so scalar connectivity summaries cannot see it.
• •Group averaging inflates recovery and survives permutation of subject identity.

1. Introduction

General anesthesia with propofol produces reversible loss of consciousness, yet the extent to which the underlying functional connectivity architecture reorganizes, and the completeness of that recovery, remains an open question in neuroimaging research. Prior work has examined propofol effects on functional connectivity using seed-based analyses, independent component analysis, and graph-theoretic metrics (Boveroux et al., 2010; Schrouff et al., 2011; Monti et al., 2013; Chennu et al., 2016). However, these approaches typically examine selected connections or summary statistics rather than quantifying the similarity of entire correlation architectures between states (see Fig. 1, Fig. 3, Fig. 4, Fig. 5, Fig. 6, Fig. 7, Fig. 8, Fig. 9Figure R1, Figure R2, Figure R3).

Fig. 1.

Subject-level Weighted Jaccard values for pairwise condition comparisons. Values are computed from full-length runs for all 24 subjects, as in the original submission, and are therefore higher than the length-matched 300-vol segment values reported in Section 3.1 and Figure R1 (see the note opening Section 3.6). Violin plots showing individual WJ values for all 24 subjects across three comparisons: awake vs. unconscious (mean = 0.499), awake vs. recovery (mean = 0.522), and unconscious vs. recovery (mean = 0.488). Each point represents one subject. Dashed line at WJ = 1.0 indicates identical architecture.

Fig. 3.

Network-level Weighted Jaccard values for awake vs. unconscious comparison (group-averaged matrices). Horizontal bar plot with 8 canonical networks ordered by WJ value. Asterisks indicate significance after FDR correction of network-level permutation tests. ROI count indicated for each network.

Fig. 4.

Group-level Weighted Jaccard values for pairwise condition comparisons (supplementary). Bar plot showing WJ between awake and unconscious (0.646), awake and recovery (0.759), and unconscious and recovery (0.653). Error bars represent jackknife standard errors. Dashed line at WJ = 1.0 indicates identical architecture.

Fig. 5.

Subject-level network recovery fractions. Horizontal bar chart showing per-subject mean recovery fraction for each canonical network. Only the Default Mode network shows significant recovery (7.7%, p = 0.021). All other networks show recovery fractions indistinguishable from zero. Error bars represent standard deviation across subjects.

Fig. 6.

Split-half reliability. Scatter plots of WJ values from random half-splits (n = 100 splits) for each comparison.

Fig. 7.

Leave-one-out jackknife influence. Bar plots showing each subject's influence on group WJ for each comparison.

Fig. 8.

Pearson vs. Spearman sensitivity. Grouped bar chart comparing WJ values computed with each correlation method.

Fig. 9.

Timeseries length matching control. Grouped bar chart comparing WJ values with and without length truncation.

Figure R1.

Between-state similarity falls below the within-condition floor. (a) Weighted Jaccard similarity for the within-condition floor (two non-overlapping 300-vol segments of the same run) and for the three between-state comparisons. Boxes show the interquartile range and median; points are individual subjects (n = 23). The dashed line marks the floor median (0.544). (b) State-attributable divergence, computed as the paired difference between each subject's floor and their between-state value. Error bars are bootstrap 95 percent confidence intervals of the median (2000 resamples). All three comparisons p = 2.4e-07, Wilcoxon signed-rank, paired Cohen's d 2.23 to 2.79.

Figure R2.

The result is robust to sign treatment, and sign reorganization is genuine. (a) Weighted Jaccard computed four ways on identical matrices: absolute correlations, edges positive in both conditions, edges negative in both, and a sign-sensitive variant. Grey bars are the within-condition floor, blue bars the primary wakefulness versus recovery comparison. Every variant separates from its own floor at p = 2.4e-07. (b) Proportion of edges reversing correlation sign, stratified by minimum edge magnitude, on a logarithmic axis. Dashed grey is the within-condition floor computed identically. Between-state rates exceed the floor at every stratum for every comparison (all p = 2.4e-07).

Figure R3.

Group averaging inflates apparent similarity. (a) Individual subject weighted Jaccard values (grey) against the value obtained from Fisher z averaged group matrices (orange diamonds) for each comparison. The group value exceeds that of every individual subject in all three cases. (b) Subject-level median, observed group-averaged value, and the group-averaged value after randomly permuting subject identity within each condition before averaging (200 permutations, error bar shows standard deviation). Permutation leaves the group value unchanged, establishing the inflation as a property of matrix averaging rather than of cross-subject correspondence.

Propofol administration has been shown to disrupt functional connectivity within and between canonical resting-state networks, particularly the default mode and frontoparietal networks (Boveroux et al., 2010; Schrouff et al., 2011). Rapid fragmentation of neuronal networks at loss of consciousness has been demonstrated using electrophysiological measures (Lewis et al., 2012), and thalamocortical disconnection has been proposed as a key mechanism underlying propofol-induced loss of consciousness (Alkire et al., 2008; Mashour, 2024). Graph-theoretic analyses have revealed increased characteristic path length and reduced information processing efficiency during propofol-induced loss of consciousness (Monti et al., 2013; Schrouff et al., 2011), while consciousness-specific dynamic interactions have been identified using propofol fMRI paradigms (Luppi et al., 2019). However, the recovery dynamics of functional connectivity following return of responsiveness have received considerably less attention than induction effects, and most existing studies examine selected connections or summary graph metrics rather than quantifying the similarity of entire correlation architectures between consciousness states.

The Weighted Jaccard (WJ) index provides an alternative approach: rather than testing individual connections or computing summary graph metrics, WJ quantifies the overall similarity between two complete correlation matrices by treating them as weighted edge sets and computing their overlap. WJ = 1 indicates identical correlation architectures; WJ = 0 indicates complete topological independence. Unlike matrix distance metrics such as the Frobenius norm or cosine similarity, which measure aggregate magnitude differences or angular distance in matrix space, WJ quantifies the proportion of shared weighted edge structure, a set-theoretic overlap that is directly interpretable as the fraction of correlation architecture preserved between conditions. This approach is domain-agnostic by construction: it operates on any pair of correlation matrices regardless of the system that generated them, providing a general framework for quantifying structural reorganization in correlation networks.

The Jaccard index was originally introduced as a measure of similarity between sets (Jaccard, 1912) and has been extended to weighted formulations for comparing continuous distributions (Ioffe, 2010). The WJ index between two correlation matrices treats each matrix as a weighted edge set and computes the ratio of shared to total edge weight, providing a single scalar measure of architectural similarity that operates on the full correlation matrix without requiring binarization, threshold selection, or hypothesis-driven connection filtering. This domain-agnostic formulation enables direct methodological comparison across fundamentally different complex systems.

That propofol reduces functional connectivity, that frontoparietal coupling is particularly affected, and that connectivity does not fully normalize immediately after responsiveness returns are established findings. A study that re-described those effects with a different similarity index would add little. The contribution here is not the direction of the effect but its calibration and its structure, and we state the three specific gaps we address.

First, whole-brain similarity indices are routinely interpreted against a perfect-match ceiling. A WJ of 0.5 between two states is read as though half the architecture were preserved. That interpretation is only available if repeated measurement of a single state returns a value near 1.0, which it does not. Without a within-condition reference, the reported magnitude of reorganization is uninterpretable, and we show below that most of it is attributable to estimation variability rather than to state.

Second, whether anesthetic reorganization is uniform or redistributive has consequences for what can measure it. If connectivity changes largely in one direction, a scalar summary of total connectivity captures it and a full-matrix method is unnecessary. If strengthenings and weakenings cancel, scalar summaries are insensitive by construction. This is an empirical question that has not been posed directly, and its answer determines whether a matrix method is required or merely available.

Third, connectivity comparisons are commonly performed on group-averaged matrices. Averaging correlation matrices across individuals is not a neutral operation, and we quantify how it distorts the recovery estimate specifically.

2. Methods

2.1. Dataset and participants

Data were obtained from the Michigan Human Anesthesia fMRI Dataset-1 (OpenNeuro accession ds006623; Huang et al., 2026), originally collected and described in Huang et al. (2018),Huang et al., 2021a andHuang et al., 2021b. Twenty-six healthy adults underwent graded propofol sedation during functional MRI scanning. Two subjects were excluded due to incomplete data across conditions, yielding a final sample of N = 24 (13 male, 11 female; mean age 25.2±4.1 years, range 20-34; Table 1).

Table 1.

Subject-Level Within-Network WJ Similarity (Awake vs. Unconscious).

Network	N ROIs	Mean WJ	SD	JK SE	Cohen's d	W	p (FDR)	N/N Subjects
Control	52	0.505	0.045	0.216	10.98	0.0	<0.001	24/24
Visual	61	0.534	0.061	0.293	7.64	0.0	<0.001	24/24
Somatomotor	77	0.546	0.042	0.200	10.90	0.0	<0.001	24/24
Subcortical	56	0.564	0.080	0.386	5.42	0.0	<0.001	24/24
Sal/Vent Attn	47	0.571	0.057	0.272	7.56	0.0	<0.001	24/24
Dorsal Attention	46	0.573	0.063	0.302	6.77	0.0	<0.001	24/24
Default Mode	91	0.579	0.069	0.329	6.14	0.0	<0.001	24/24
Limbic	26	0.617	0.064	0.305	6.02	0.0	<0.001	24/24

Table 1. Subject-level within-network Weighted Jaccard similarity for the awake vs. unconscious comparison. WJ was computed for each subject individually by restricting per-subject correlation matrices to ROIs within each network. Statistics are derived from the distribution of N = 24 subject-level WJ values. JK SE = jackknife standard error. W = Wilcoxon signed-rank test statistic (vs. WJ = 1.0). All p-values FDR-corrected (Benjamini-Hochberg) across 8 networks. N/N Subjects = subjects showing WJ < 1.0/total subjects.

The sedation protocol involved stepwise increases in propofol effect-site concentration (ESC) in 0.4 μg/mL increments (starting from 0.4 or 1.0 μg/mL depending on protocol version) until loss of responsiveness (LOR), defined as failure to respond to verbal commands. LOR ESC ranged from 0.8 to 2.6 μg/mL across subjects (mean 1.71±0.50 μg/mL). Following LOR, propofol was maintained at a fixed concentration above LOR threshold. Scanning occurred during three conditions: (1) pre-sedation wakefulness (rest_run-1), (2) propofol-induced unconsciousness following LOR (imagery_run-2, during which subjects had been instructed to perform mental imagery but were unresponsive), and (3) early recovery after return of responsiveness (rest_run-2). This within-subject design controls for inter-individual variability in baseline connectivity.

Wakefulness was task-rest run 1, unresponsiveness was the post loss-of-responsiveness portion of task-imagery run 2, and recovery was task-rest run 2. The three conditions are therefore not matched on task: wakefulness and recovery are both resting-state acquisitions, while the unresponsive condition derives from a run in which imagery instructions were delivered. They are also not matched on duration: after trimming to post loss-of-responsiveness timepoints, median run lengths were 770, 1110 and 770 vol respectively.

Because wakefulness and recovery are matched on both task and duration, and because the incomplete-recovery question is the primary question of this study, the wakefulness versus recovery contrast is designated the primary comparison. Comparisons involving the unresponsive condition are reported as supporting analyses and are interpreted as state-as-a-whole effects rather than as isolated pharmacological effects.

2.2. Preprocessing

Functional MRI data were preprocessed using the XCP-D pipeline (Mehta et al., 2024) with the 4S456Parcels atlas, yielding timeseries for 456 cortical and subcortical regions of interest. Preprocessing included motion correction, bandpass filtering, and confound regression. Global signal regression was not applied. ROIs were assigned to 8 canonical resting-state networks based on the Schaefer parcellation scheme: Control (Cont; 52 ROIs), Default (91 ROIs), Dorsal Attention (DorsAttn; 46 ROIs), Limbic (26 ROIs), Salience/Ventral Attention (SalVentAttn; 47 ROIs), Somatomotor (SomMot; 77 ROIs), Subcortical (56 ROIs), and Visual (Vis; 61 ROIs).

2.3. Correlation matrix construction

For each subject and each condition, full pairwise Spearman rank correlation matrices were computed among all 456 ROIs, yielding 103,740 unique pairwise correlations per subject per condition. Spearman correlation was chosen for robustness to outliers and non-linear monotonic relationships; a sensitivity analysis using Pearson correlation confirmed minimal impact on results (Section 2.7.3). Subject-level correlation matrices were used for the primary analysis. Additionally, group-averaged timeseries matrices were constructed by averaging ROI timeseries across subjects within each condition, and group-level correlation matrices were computed from these averaged timeseries for supplementary analyses (see Section 3.6.5).

2.4. Weighted Jaccard (WJ) similarity analysis

The Weighted Jaccard index between two correlation matrices was computed as:

WJ(A, B) = Σ min(wA, wB) / Σ max(wA, wB)

where w_A and w_B are the absolute correlation values for each edge in conditions A and B, respectively. Absolute values were used, treating positive and negative correlations of equal magnitude as equivalent contributions to connectivity architecture; sign-sensitive formulations exist but are beyond the current scope. WJ was computed on raw (non-Fisher-z-transformed) correlations because the min/max formulation operates directly on the bounded correlation scale; Fisher-z transformation would disproportionately upweight extreme correlations and alter relative edge contributions to the index. WJ ranges from 0 (completely dissimilar architectures) to 1 (identical architectures). This formulation treats the correlation matrix as a weighted graph and quantifies the overlap of edge weight distributions without requiring binarization or threshold selection.

For each subject, WJ was computed for three pairwise comparisons: awake vs. unconscious, awake vs. recovery, and unconscious vs. recovery, using that subject's own per-condition correlation matrices.

2.5. Statistical inference

Subject-level WJ values were tested against the null hypothesis of no change (WJ = 1.0) using Wilcoxon signed-rank tests. Recovery was assessed by testing whether awake-recovery WJ exceeded awake-unconscious WJ (paired Wilcoxon signed-rank test). Per-subject recovery fractions were computed as (WJ_awake-recovery - WJ_awake-unconscious)/(1.0 - WJ_awake-unconscious) and tested against zero.

Uncertainty in group-level WJ statistics was quantified using leave-one-out jackknife standard errors. The jackknife SE was computed as sqrt((n-1)/n * sum((x_i - x_bar)^2)) where x_i are the leave-one-out WJ values (Efron, 1982).

Additionally, permutation testing (10,000 permutations) was conducted on group-averaged correlation matrices to assess significance of the group-level WJ. For each permutation, condition labels were randomly swapped within subjects (each subject independently assigned to original or swapped conditions with 50% probability), group-averaged correlation matrices were recomputed from the permuted assignments, and WJ was computed on these permuted group matrices. P-values were computed as the proportion of permuted WJ values less than or equal to the observed WJ and corrected across the three comparisons using FDR (Benjamini-Hochberg).

Effect sizes were computed as (observed WJ - null mean) / null SD.

Amendment. The test of subject-level WJ against a null of 1.0, described above and used in the original submission, is withdrawn. As Reviewer 4 correctly noted, repeated measurement of a single state does not return WJ = 1.0, so that null is uninformative. All primary inference in this revision is against the empirically estimated within-condition floor described in Section 2.8. The jackknife, permutation and effect-size procedures described above are unchanged and are retained for the analyses in Sections 3.6 onward.

2.6. Network-level analysis

For each subject, within-network WJ was computed for each of the 8 canonical resting-state networks by restricting that subject's per-condition correlation matrices to the ROIs assigned to each network and computing WJ on the resulting submatrices. This yielded 24 subject-level within-network WJ values per network per comparison. Significance of network-level connectivity change was assessed using Wilcoxon signed-rank tests across subjects (testing WJ < 1.0), with p-values FDR-corrected across the 8 networks within each comparison. Per-subject network recovery fractions were computed identically to the global recovery fraction (Section 2.5) but restricted to within-network edges.

As a supplementary analysis, within-network WJ was also computed on group-averaged correlation matrices using network-level permutation testing (1000 permutations per network). For each permutation, condition labels were randomly swapped within subjects (identical to the global permutation procedure), group-averaged correlation matrices were recomputed, and within-network WJ was computed from the network-restricted submatrices. P-values were FDR-corrected across all 24 tests (8 networks x 3 comparisons). This supplementary analysis is reported in Section 3.6.5 for comparison with existing literature.

2.7. Robustness analyses

2.7.1. Split-half reliability

The 24 subjects were randomly split into two groups of 12, 100 times. For each split, group-level correlation matrices and WJ values were computed independently for each half. Stability was quantified using the coefficient of variation (CV) across splits.

2.7.2. Leave-one-out jackknife

Each subject was sequentially excluded and the group WJ was recomputed from the remaining 23 subjects. Jackknife standard errors and influence values were computed to identify disproportionately influential subjects.

2.7.3. Correlation method sensitivity

The entire analysis was repeated using Pearson correlation instead of Spearman. The absolute difference in WJ values between methods was computed for each comparison.

2.7.4. Timeseries length matching

To control for potential confounds from unequal scan durations across conditions, timeseries were truncated to the minimum length within each subject-condition pair before recomputing correlation matrices and WJ values.

2.7.5. Threshold sensitivity

Binary Jaccard indices were computed across 7 correlation thresholds (0.1 to 0.7 in steps of 0.1) to assess whether the rank ordering of comparisons was preserved under binarization. This analysis motivated the use of the weighted (continuous) Jaccard as the primary metric.

2.7.6. Network-level permutation testing

Described in Section 2.6.

2.8. Within-condition floor and length matching

Two non-overlapping segments of 300 vol were drawn from each run of each subject. A correlation matrix was computed for each segment. The within-condition floor is the WJ between the two segments of the same run, and the between-state value is the WJ between the first segment of one run and the first segment of another. Every matrix entering any comparison is therefore estimated from exactly 300 vol, so no comparison confounds state with estimation precision. Segment placement was randomized and the procedure repeated five times per subject, with the per-subject median taken forward. Twenty-three subjects had sufficient data in all three conditions after trimming.

We note explicitly that an interleaved (odd and even volume) split is not an appropriate within-condition reference for fMRI. Adjacent volumes share the same hemodynamic response and arousal state, and such a split yields a floor near 0.99 that reflects temporal autocorrelation rather than measurement reproducibility.

A note on the reported p-values. All floor comparisons are tested by Wilcoxon signed-rank across subjects. With n = 23 the smallest attainable two-sided p-value is 2/2^23 = 2.4e-07, which is obtained when every subject moves in the same direction. Several comparisons reported below reach exactly this value. That indicates a saturated test rather than a more extreme result, and it is why paired effect sizes are reported alongside every p-value rather than left implicit.

2.9. Sign treatments

Four similarity variants were computed on identical matrices: WJ on absolute correlations; WJ restricted to edges positive in both conditions; WJ restricted to edges negative in both; and a sign-sensitive variant computed after shifting correlations to non-negative values, in which inversions reduce similarity. Sign reorganization was additionally quantified directly as the proportion of edges whose correlation reverses sign, stratified by min(|r_A|, |r_B|) at thresholds of 0, 0.1, 0.2 and 0.3, and tested against the within-condition floor computed identically. Global signal regression was not applied, and 28.5 percent of edges were negative, so the negative-edge analysis is adequately powered.

2.10. Uniformity and scalar comparison

For each pair of conditions we computed the per-edge magnitude change d, and the uniformity index R = 1 - |sum d|/sum |d|. Values near zero indicate that edges move predominantly in one direction, in which case a scalar index of total connectivity tracks the change. Values near one indicate that strengthenings and weakenings cancel, in which case scalar indices are insensitive. We separately tested whether the change in mean absolute correlation distinguished the conditions.

2.11. Group-averaging control

Group matrices were formed exactly as in the original submission, by Fisher z transformation, averaging across subjects, and inverse transformation. Group-level WJ was compared against the distribution of subject-level WJ. As a control, subject identity was randomly permuted within each condition before averaging, 200 times, to test whether elevated group similarity depends on genuine cross-subject correspondence.

3. Results

3.1. Between-state similarity falls below the within-condition floor

The within-condition floor was WJ = 0.544, confirming that repeated measurement of a single state does not approach 1.0 and that comparison against a perfect-match ceiling is not informative. Against this reference, all three between-state comparisons were significantly lower: wakefulness versus recovery 0.486 (gap 0.053, d = 2.23), wakefulness versus unresponsiveness 0.468 (gap 0.068, d = 2.32), unresponsiveness versus recovery 0.458 (gap 0.071, d = 2.79), all p = 2.4e-07, Wilcoxon signed-rank across subjects.

Expressed as a proportion of observed architectural divergence, the state-attributable share is 11.4 percent for the primary contrast and 14.2 percent for wakefulness versus unresponsiveness. We therefore revise the characterization in the original submission. The reorganization is statistically robust and of large paired effect size, but its magnitude relative to within-state variability is modest, and the earlier description of approximately half the architecture being altered is not supported once the appropriate reference is applied.

3.2. The result is unchanged under four sign treatments

Section 2.6

Metric	Between	Floor	Gap	d	p
Absolute WJ	0.486	0.544	+0.053	2.23	2.4e-07
Positive edges only	0.555	0.608	+0.052	2.16	2.4e-07
Negative edges only	0.444	0.497	+0.038	1.76	2.4e-07
Signed variant	0.835	0.859	+0.024	1.99	2.4e-07

Primary contrast shown; all twelve tests across four metrics and three comparisons were significant at p = 2.4e-07.

3.3. Sign reorganization among strong edges is genuine

Sign-flip rates exceeded the within-condition floor at every stratum. Among strong edges (min|r| ≥ 0.3) the primary contrast showed 2.5 percent versus a floor of 0.5 percent, a fivefold excess; wakefulness versus unresponsiveness showed 8.2 percent, a 16.6-fold excess; unresponsiveness versus recovery 8.5 percent, 17.2-fold (all p = 2.4e-07). Inversions were directionally symmetric (46 percent positive to negative, 54 percent the reverse) and were concentrated in subcortical pairings, most prominently control-subcortical, limbic-subcortical and default-subcortical.

Reorganization is therefore predominantly magnitude-driven, since only a small minority of strong edges invert, but sign change is a real and localized component rather than an artifact of weak correlations crossing zero.

3.4. Reorganization is redistributive, and scalar summaries are insensitive

Uniformity was high in all comparisons: R = 0.775 for the primary contrast, 0.844 and 0.871 for the unresponsiveness comparisons. Strengthenings and weakenings therefore largely cancel. Consistent with this, the change in mean absolute correlation did not distinguish wakefulness from recovery (p = 0.096), while the full-matrix comparison did (p = 2.4e-07). This is a direct demonstration that a scalar connectivity summary is insensitive to the reorganization reported here, and it is the specific justification for a matrix-based method in this setting.

3.5. Recovery is small, reliably non-zero, and overstated by group averaging

Subject-level recovery was 3.2 percent (95 percent CI 1.8 to 5.7, n = 23). The confidence interval excludes zero, so partial recovery is statistically detectable, but the magnitude is small.

Group-averaged matrices produced a substantially larger recovery estimate. Group-level WJ exceeded that of every individual subject in all three comparisons (group 0.752, 0.758, 0.753 against subject maxima of 0.642, 0.582, 0.635). Randomly permuting subject identity before averaging left the group value unchanged (0.7578 permuted versus 0.7578 observed, SD < 0.0001), demonstrating that the elevation does not depend on genuine cross-subject correspondence and is a property of averaging correlation matrices. Group-averaged comparisons therefore overstate recovery, and the discrepancy between group and subject-level recovery fractions noted in the original submission is explained by this mechanism.

3.6. Findings retained from the original submission

The analyses in this section are unchanged from the original submission and are reported as previously, now with subject-level and group-level results explicitly labelled throughout.

Two estimation lengths are reported in this manuscript, and the values differ accordingly. The analyses in Sections 3.1 to 3.5 use two non-overlapping 300-vol segments per run, so that every matrix entering any comparison is estimated from an identical number of timepoints and no comparison confounds state with estimation precision; twenty-three subjects had sufficient data in all three conditions after trimming. The analyses in this section use full-length runs and all twenty-four subjects, as originally submitted. Because correlation matrices estimated from more data are less noisy, the full-run values are systematically higher: wakefulness versus recovery 0.522 against 0.486, wakefulness versus unresponsiveness 0.499 against 0.468, and a recovery fraction of 3.8 percent against 3.2 percent. The rank ordering of the three comparisons is identical under both, and the substantive conclusion of minimal early recovery is unchanged. The segment-based values are the primary ones, because only they can be compared against a within-condition floor estimated at the same data length. Values in this section should be read as the full-run estimates they are, and not compared directly with those in Sections 3.1 to 3.5.

3.6.1. Network-level specificity

Within-network WJ was computed for each subject individually by restricting per-subject correlation matrices to ROIs within each canonical network. All 8 networks showed significant similarity reduction during unconsciousness across subjects (Wilcoxon signed-rank test vs. WJ = 1.0, all p < 0.001, FDR-corrected; Table 1). The largest effects were observed in:

Control (mean WJ = 0.505, SD = 0.045, d = 10.98, 52 ROIs)

Visual (mean WJ = 0.534, SD = 0.061, d = 7.64, 61 ROIs)

Somatomotor (mean WJ = 0.546, SD = 0.042, d = 10.90, 77 ROIs)

Subcortical (mean WJ = 0.564, SD = 0.080, d = 5.42, 56 ROIs)

Salience/Ventral Attention (mean WJ = 0.571, SD = 0.057, d = 7.56, 47 ROIs)

Dorsal Attention (mean WJ = 0.573, SD = 0.063, d = 6.77) and Default Mode (mean WJ = 0.579, SD = 0.069, d = 6.14) showed moderate similarity reduction, while Limbic showed the smallest effect (mean WJ = 0.617, SD = 0.064, d = 6.02) but remained highly significant (p < 0.001). Notably, all 24 subjects showed WJ < 1.0 in every network, confirming that connectivity change is universal across subjects and not driven by individual outliers.

The subject-level network ranking differs substantially from the group-averaged network ranking (see Section 3.6.5), with Control showing the greatest similarity reduction at subject level but the least at group level. This divergence reflects differential averaging inflation across networks: group averaging inflated Control WJ by +0.316 but Limbic WJ by only +0.099, indicating that inter-subject variability in correlation patterns is highest in the Control network and lowest in Limbic structures. The subject-level ranking, which preserves individual variability, is reported as the primary result.

3.6.2. Network-level recovery

At the subject level, network-specific recovery was minimal and largely non-significant. Per-subject within-network recovery fractions, computed as (WJ_awake-recovery - WJ_awake-unconscious)/(1.0 - WJ_awake-unconscious), were near zero for most networks (Table 2). Only the Default Mode network showed statistically significant recovery (mean RF = 7.7%, SD = 20.0%; Wilcoxon p = 0.021). Subcortical regions trended positive (mean RF = 7.1%, p = 0.099). All other networks showed recovery fractions indistinguishable from zero (all p > 0.25), including Visual (RF = −0.2%), Somatomotor (RF = −0.6%), Salience/Ventral Attention (RF = −1.6%), Control (RF = −0.2%), Dorsal Attention (RF = −1.8%), and Limbic (RF = −0.4%).

Table 2.

Subject-level network recovery fractions.

Network	Mean RF	SD	Median RF	W	p	N Positive	N Negative
Default Mode	+7.7%	20.0%	+10.7%	221.0	0.021*	16	8
Subcortical	+7.1%	26.6%	+7.9%	196.0	0.099	16	8
Control	−0.2%	11.5%	+2.0%	163.0	0.363	12	12
Limbic	−0.4%	16.7%	+0.2%	142.0	0.594	12	12
Visual	−0.2%	18.2%	+2.5%	159.0	0.406	15	9
Somatomotor	−0.6%	13.6%	+1.4%	173.0	0.264	14	10
Sal/Vent Attn	−1.6%	14.3%	+2.1%	168.0	0.312	14	10
Dorsal Attention	−1.8%	18.1%	+3.3%	160.0	0.395	14	10

Table 2. Subject-level network recovery fractions. Recovery fraction = (WJ_awake-recovery - WJ_awake-unconscious)/(1.0 - WJ_awake-unconscious), computed per subject and then summarized. RF = recovery fraction. W = Wilcoxon signed-rank test statistic (vs. RF = 0). N Positive/Negative = subjects showing positive/negative recovery. *p < 0.05. Only the Default Mode network shows statistically significant recovery.

This pattern contrasts sharply with group-averaged recovery fractions (Section 3.7), which suggested substantial recovery in several networks (e.g., Salience/Ventral Attention 52.4%, Somatomotor 44.4%). The subject-level analysis reveals that these apparent recoveries were artifacts of timeseries averaging. At the individual level, recovery during the early post-anesthetic period is essentially absent across all networks except Default Mode (see Fig. 10).

Fig. 10.

Network-level permutation significance. Horizontal bar plot showing observed WJ, null distribution mean, and significance indicators for each network x comparison.

3.6.3. Dose-response relationship

Individual WJ dissimilarity was significantly correlated with propofol effect-site concentration at LOR. Higher LOR ESC was associated with greater dissimilarity (lower WJ) during unconsciousness (Spearman rho = −0.591, p = 0.002, N = 24; Fig. 11). Conversely, higher LOR ESC was associated with better recovery (higher WJ in the awake-recovery comparison; rho = 0.499, p = 0.013). No significant relationship was observed between LOR ESC and the unconscious-recovery comparison (rho = −0.176, p = 0.410) (see Fig. 2).

Fig. 11.

Dose-response relationship between propofol effect-site concentration at LOR and subject-level WJ dissimilarity. Scatter plots with regression lines for awake vs. unconscious (rho = −0.591, p = 0.002), awake vs. recovery (rho = 0.499, p = 0.013), and unconscious vs. recovery (rho = −0.176, p = 0.410).

Fig. 2.

Heatmap of pairwise correlation differences between awake and unconscious conditions (group-averaged matrices). Red indicates increased correlation under unconsciousness; blue indicates decreased correlation. ROIs ordered by network assignment.

These findings suggest that individual pharmacological sensitivity, as indexed by the propofol concentration required to achieve LOR, is associated with both the magnitude of acute architectural reorganization and the degree of early recovery. Subjects requiring higher propofol concentrations to lose responsiveness may experience more extensive disruption of correlation architecture, yet also exhibit greater architectural recovery during the early post-anesthetic period.

3.6.4. Edge-level mapping

To identify the specific connections most disrupted during unconsciousness, group-averaged correlation matrices were used to rank edges by absolute change in correlation (|delta_r|) between awake and unconscious conditions. Among the 9612 edges showing changes, the top 50 most disrupted edges (|delta_r| range: 0.342-0.411) were overwhelmingly concentrated at the somatomotor-visual interface: 45 of 50 edges connected somatomotor and visual ROIs, with visual network ROIs involved in 48 and somatomotor ROIs in 47 of the 50 edges (Table S6). All 50 edges represented correlation losses (weakening during unconsciousness).

This concentration was consistent across multiple thresholds: restricting to the top 25, top 50, and top 100 most disrupted edges yielded 92%, 90%, and 87% somatomotor-visual edges respectively, confirming that the pattern is not an artifact of the arbitrary top-50 cutoff. The concentration suggests that the largest individual connection disruptions during propofol unconsciousness occur at the interface between sensory and motor systems, consistent with the interruption of sensorimotor integration during loss of consciousness.

Between-network edges showed significantly larger disruptions than within-network edges overall (between mean |delta_r| = 0.157 vs. within mean |delta_r| = 0.126; Mann-Whitney p < 0.001), indicating that inter-network connectivity is more vulnerable to propofol-induced reorganization than intra-network connectivity.

3.6.5. Group-level supplementary analysis

Group-averaged correlation matrices (computed from timeseries averaged across all 24 subjects) yielded higher WJ values than subject-level analyses, consistent with noise reduction through averaging: awake-unconscious WJ = 0.646 (jackknife SE = 0.023, permutation p < 0.001, d = 7.86), awake-recovery WJ = 0.759 (jackknife SE = 0.017, permutation p = 0.012, d = 3.46), and unconscious-recovery WJ = 0.653 (jackknife SE = 0.019, permutation p < 0.001, d = 6.86). The group-level recovery fraction (31.9%) was substantially larger than the subject-level estimate (3.8%), reflecting the smoothing effect of timeseries averaging, which reduces noise and artificially increases correlation matrix similarity.

Group-level network WJ values (Table S1) showed a different ranking from subject-level values, with Somatomotor (0.751), Salience/Ventral Attention (0.783), and Default Mode (0.800) appearing most dissimilar, while Control (0.822) appeared least affected. The Limbic network did not reach significance in group-level permutation testing (p = 0.242) despite being highly significant at subject level (p < 0.001), illustrating how group-level permutation tests on small networks (26 ROIs) can lack power even when the underlying effect is robust across individuals. Group-level recovery fractions (Salience/Ventral Attention 52.4%, Somatomotor 44.4%, Control 37.9%) were substantially inflated relative to subject-level estimates (all near zero except Default Mode), confirming that the averaging inflation is not uniform across networks and produces misleading recovery estimates. The group-level analysis is reported for comparison with existing literature, which predominantly uses group-averaged matrices.

3.6.6. Robustness

3.6.6.1. Split-half reliability

Group WJ estimates were stable across 100 random half-splits, with coefficients of variation of 3.1% (awake vs. unconscious), 2.2% (awake vs. recovery), and 2.3% (unconscious vs. recovery).

3.6.6.2. Leave-one-out jackknife

No individual subject had a disproportionate influence on group WJ estimates. The maximum deviation from any single subject's exclusion was 0.012 (subject sub-18 in the awake-unconscious and unconscious-recovery comparisons).

3.6.6.3. Correlation method sensitivity

Replacing Spearman with Pearson correlation changed group-level WJ values by less than 0.7% across all comparisons (maximum absolute difference = 0.004), indicating that the findings are not dependent on the choice of correlation method.

3.6.6.4. Timeseries length matching

Truncating timeseries to equal length within each subject-condition pair changed WJ values by less than 0.9% (maximum absolute difference = 0.006), ruling out unequal scan duration as a confound.

3.6.6.5. Threshold sensitivity

The rank ordering of comparisons varied across binary Jaccard thresholds, supporting the use of the continuous Weighted Jaccard as the primary metric, which avoids the arbitrary threshold selection inherent in binary graph analyses.

4. Discussion

4.1. Relation to current accounts of anesthetic connectivity change

Contemporary work has moved from static connectivity reduction toward dynamic and topological descriptions of anesthetic unconsciousness. Anesthesia is now widely characterized as a disturbance of the balance between functional integration and segregation rather than as a uniform loss of coupling (Jang et al., 2024; Lord et al., 2023), and the trajectory of that balance across loss and recovery of responsiveness has been described directly under sevoflurane (Dong et al., 2023), indicating that the pattern is not specific to a single agent. Dynamic connectivity analyses show that anesthetics alter the repertoire and dwell times of recurring brain states rather than simply lowering mean connectivity (Miao et al., 2023). At the circuit level, corticothalamocortical interaction remains the leading candidate mechanism (Zhou et al., 2025; Bao et al., 2023; Moody et al., 2021), and cellular work indicates that anesthesia can decouple compartments within cortical pyramidal neurons, providing a mechanism by which coupling can be reorganized rather than merely reduced (Suzuki and Larkum, 2020).

Our results sit naturally within that shift. A redistributive reorganization, in which strengthenings and weakenings largely cancel, is what an integration and segregation rebalancing looks like when it is measured as a whole-matrix comparison rather than as a summary statistic. That the most disrupted systems here were limbic and subcortical is consistent with corticothalamocortical accounts, and it is not the pattern a motor imagery confound would produce.

The uniformity index makes that correspondence quantitative rather than descriptive. An integration and segregation rebalancing predicts that some couplings strengthen while others weaken, so that the total changes little while its distribution changes substantially. That is a high R signature. A uniform loss of coupling predicts the opposite, with R near zero, because all edges move in the same direction. The observed values of 0.775 to 0.871 therefore favor rebalancing over global decoupling, and they do so on a quantity that is defined independently of any network partition. Corticothalamocortical accounts predict the same signature for a different reason: reorganization of specific loops changes which regions are coupled without requiring a net reduction in coupling overall, and the decoupling of pyramidal neuron compartments described by Suzuki and Larkum (2020) supplies a cellular mechanism by which coupling can be redirected rather than lost.

This yields a prediction that does not depend on the present dataset. If anesthetic unconsciousness reflects rebalancing rather than suppression, R should remain high across agents and doses that preserve a dynamic state repertoire, and should fall toward zero under conditions producing global suppression, such as burst suppression at deeper levels. Because R requires only the two correlation matrices that such studies already compute, the prediction can be evaluated retrospectively in existing datasets.

The persistence of altered architecture after responsiveness returns is likewise consistent with reports that recovery of consciousness is not accompanied by immediate normalization of network organization (Dong et al., 2023; Wu et al., 2022).

4.2. Similarity indices require a within-state reference

The principal methodological point is that similarity indices require calibration against repeated measurement of a single state. Reported against a perfect-match ceiling, the reorganization here would appear to involve roughly half the architecture. Against a within-condition floor, the state-attributable share is approximately an eighth. Both numbers derive from the same data and the same analysis. Only the second is interpretable.

This is not a peculiarity of the weighted Jaccard index. Functional connectivity estimates are themselves only moderately reliable across repeated sessions, as established by systematic review and meta-analysis of a decade of test-retest studies (Noble et al., 2019), with reliability depending strongly on scan duration and on the connectivity feature examined (Wang et al., 2021; Camp et al., 2024). Any index computed on such estimates inherits that variability. A between-condition value therefore cannot be read as a proportion of architecture altered unless the within-condition value is reported alongside it, and we recommend that whole-brain similarity comparisons do so routinely.

A related observation concerns individual differences. In this dataset, two different subjects scanned in the same state were approximately as dissimilar as one subject scanned across states. This is consistent with the connectome fingerprinting literature, in which individual connectivity architecture is sufficiently distinctive to identify people across sessions and conditions (Ciarrusta et al., 2022; Cipriano et al., 2023). It reinforces that within-subject, paired designs are required for questions of this kind, and that absolute similarity values are not comparable across studies without a stated reference.

4.3. Redistribution and the limits of aggregate indices

The redistributive character of the reorganization helps explain heterogeneity in the literature. Where anesthetic effects are indexed by aggregate connectivity strength, results can appear inconsistent across studies and agents. If reorganization conserves total connectivity while rewiring its distribution, aggregate indices are expected to be unstable while matrix comparisons are not. This also bears on the broader effort to identify robust measures of anesthetic depth and complexity, where the choice of summary statistic materially determines what is detected (Li et al., 2022).

4.4. Group averaging distorts recovery estimates

The group-averaging result has implications beyond this dataset. Because averaging smooths idiosyncratic structure, group matrices resemble one another more than any individual resembles themselves. Any group-level comparison of averaged connectivity matrices will therefore tend to understate between-condition differences and overstate recovery. That this survives random permutation of subject identity indicates that it is arithmetic rather than biology.

Because subjects reorganize idiosyncratically, per-subject analysis is the appropriate level for this question, and work on disorders of consciousness, sleep and intraoperative monitoring should be constructed accordingly. A per-subject architectural index that is calibrated against that subject's own within-state variability is, in principle, deployable in those settings, and we regard this as the most promising extension of the present approach.

There is a second and independent reason to prefer the subject level. Twenty-three paired within-subject observations support formal hypothesis testing directly, whereas a comparison of two averaged matrices yields a single value per contrast and no within-design estimate of variability. The group-averaged results are retained here because they are what a conventional analysis would report, and because quantifying their inflation is one of this study's findings, not because they are the better estimate.

4.5. Architectural reorganization, not connection-level change

The WJ approach quantifies similarity at the level of entire correlation architectures rather than individual connections. A between-state WJ of 0.486 does not mean that half of all connections break. It means that the weighted overlap across 103,740 pairwise correlation values is approximately 0.486, against a within-condition floor of 0.544 obtained when the same state is measured twice. The state-attributable component is the gap between those two values, not the distance from 1.0. That reduction could reflect changes in correlation magnitude, sign, or topology; the edge-level analysis (Section 3.6.4) confirms genuine topological change, specific edges gained and lost, rather than uniform attenuation of correlation strength, and Section 3.3 shows that a minority of strong edges additionally invert. Individual connections may strengthen or weaken; WJ captures the collective change in network topology.

This architectural perspective complements prior work examining specific connections or graph-theoretic summary statistics under anesthesia. Where seed-based analyses identify particular disrupted circuits (e.g., thalamocortical disconnection; Boveroux et al., 2010) and graph metrics summarize global properties (e.g., reduced modularity; Monti et al., 2013), WJ quantifies the degree to which the entire correlation architecture has been rearranged.

4.6. Clinical implications of minimal early recovery

The finding that approximately 3 percent of propofol-induced architectural reorganization reversed by the recovery scan has potential clinical implications. Behavioral recovery, operationalized as return of verbal responsiveness, occurred before meaningful architectural recovery. This dissociation between behavioral and architectural recovery is consistent with prior reports of cognitive impairment following emergence from anesthesia (Monk et al., 2008) and suggests that return of responsiveness is a necessary but insufficient indicator of neural recovery.

The dose-response analysis partially addresses the question of individual predictors: propofol concentration at LOR was significantly correlated with both dissimilarity magnitude (rho = −0.591, p = 0.002) and recovery degree (rho = 0.499, p = 0.013), suggesting that individual pharmacological sensitivity is associated with both the depth of architectural disruption and the degree of early recovery.

Two features of the present results constrain the prospect of individualized or real-time application directly, and both argue for caution. First, group averaging systematically inflates apparent recovery. Group-level WJ exceeded that of every individual subject, and permuting subject identity left the group value unchanged, establishing that the inflation is a property of averaging matrices rather than of genuine cross-subject correspondence. Any monitoring approach that compares an individual against a group-averaged template will therefore tend to overstate how far that individual has recovered, which is the clinically unsafe direction of error. Second, the quantity available to a monitor during emergence is small. The subject-level recovery estimate of 3.2 percent (95 percent CI 1.8 to 5.7) means that the architectural change accompanying return of responsiveness is a small fraction of the change accompanying its loss, so the dynamic range of the index across the emergence window is narrow relative to its own within-state variability.

A third constraint is temporal. The within-condition floor reported here was estimated from 300-vol segments, whereas real-time application would require considerably shorter windows, and the floor rises as segment length falls, compressing the separation between states. We therefore regard these results as supporting group-level inference and per-subject research applications rather than individual-level classification. Clinical deployment would require prospective validation that includes within-subject calibration against a pre-induction baseline and explicit characterization of how both the floor and the state separation vary with window length.

4.7. Network-level patterns

The subject-level network analysis reveals that the Control network shows the greatest similarity reduction during unconsciousness (mean WJ = 0.505, d = 10.98), followed by Visual (0.534) and Somatomotor (0.546). This ranking differs from the group-averaged analysis, where Control appeared least affected (WJ = 0.822). The reversal arises because group averaging inflates WJ differentially across networks: Control shows the largest inflation (+0.316), indicating high inter-subject variability in control network correlation patterns, while Limbic shows the smallest inflation (+0.099), suggesting more consistent cross-subject patterns. This differential inflation is itself an informative finding: networks with high inter-subject variability in baseline connectivity show greater apparent preservation under group averaging because the averaging process suppresses the subject-specific disruption patterns.

The Control network's vulnerability is consistent with preferential disruption of higher-order associative cortices under general anesthesia, as the frontoparietal control network supports executive function and flexible task-directed behavior, capacities that are abolished during unconsciousness (Mashour, 2024). The Visual network's large effect (d = 7.64) is consistent with disruption of sensory processing architecture, and the concentration of the 50 most disrupted edges at the somatomotor-visual interface (Section 3.6.4) suggests that inter-network connections bridging sensory and motor systems are particularly vulnerable to propofol-induced disconnection.

The Limbic network, which did not reach significance in the group-level permutation test (p = 0.242), is highly significant at the subject level (p < 0.001, d = 6.02), confirming that the group-level null result was a power limitation rather than a biological finding. All 24 subjects showed Limbic reorganization (WJ < 1.0), consistent with limbic structures' known high GABA-A receptor density (Zilles and Palomero-Gallagher, 2017).

At the network level, recovery is essentially absent during the early post-anesthetic period. Only the Default Mode network showed significant recovery (mean RF = 7.7%, p = 0.021). All other networks, including Visual, Somatomotor, and Control, showed recovery fractions indistinguishable from zero. This contrasts with the group-averaged analysis, which suggested substantial recovery in several networks (e.g., Salience/Ventral Attention 52.4%), but those estimates were artifacts of timeseries averaging. The Default Mode network's unique recovery may reflect the default mode's known role in maintaining baseline resting-state activity patterns (Boveroux et al., 2010), which may be preferentially restored during early behavioral recovery.

4.8. Edge-level architecture of disruption

The concentration of the 50 most disrupted edges at the somatomotor-visual interface (45/50 edges) reveals a specific architectural vulnerability that is not apparent from network-level WJ values alone. While the visual and somatomotor networks rank 4th and 2nd respectively in within-network disruption magnitude, their inter-network interface shows the largest individual edge disruptions. This suggests that propofol-induced unconsciousness preferentially disrupts the connections bridging sensory and motor systems, consistent with the interruption of sensorimotor integration as a hallmark of loss of consciousness.

The predominance of between-network over within-network disruptions (mean |delta_r| 0.157 vs. 0.126, p < 0.001) further indicates that propofol preferentially targets integrative connections linking functional modules, rather than disrupting the internal coherence of individual networks.

4.9. Methodological considerations

WJ similarity analysis offers several advantages for this application. It operates on the full correlation matrix without requiring threshold selection, binarization, or hypothesis-driven connection selection. The threshold sensitivity analysis confirmed that binary Jaccard indices yield inconsistent rank orderings across thresholds, validating the continuous weighted approach. The metric is intuitive (proportion of shared weighted architecture) and directly comparable across domains and datasets.

The robustness battery demonstrates that the findings are stable across analytical choices. Split-half reliability (CV 2-3%), leave-one-out jackknife (max influence 0.012), Pearson vs. Spearman sensitivity (<0.7%), and timeseries length matching (<0.9%) all indicate that the results are not fragile or dependent on specific methodological decisions.

4.10. Broader methodological implications

The WJ framework is domain-agnostic by construction: it operates on any pair of correlation matrices regardless of the generating system. The present application to brain connectivity data demonstrates that WJ similarity analysis can quantify functional connectivity change at the whole-brain level with a comprehensive robustness profile. Because the computation requires only paired correlation matrices, the same approach may be applicable to other state transitions in neuroscience (e.g., sleep stages, pharmacological challenges, disease states) and, more broadly, to any system where correlation network reorganization is of interest.

5. Limitations

• 1.The recovery scan was acquired shortly after return of responsiveness. The 3.2 percent global (95 percent CI 1.8 to 5.7) and near-zero network-level recovery fractions reflect very early recovery; longer follow-up periods may reveal more complete architectural normalization.
• 2.Sample size (N = 24) limits generalizability and may reduce power for detecting smaller network-level effects, as illustrated by the Limbic network result.
• 3.The conditions are not matched on task or duration. Wakefulness and recovery are resting-state acquisitions; the unresponsive condition derives from a run in which imagery instructions were delivered, and after trimming it remains longer than the other two. Any comparison involving the unresponsive condition therefore confounds pharmacological state with task instruction and with estimation precision, and is interpreted as a state-as-a-whole effect. Two considerations constrain, without eliminating, the task explanation. The primary comparison, wakefulness versus recovery, is matched on both task and duration. And the anatomical pattern of disruption does not match a motor imagery confound: limbic and subcortical systems were the most disrupted, with somatomotor third and visual fifth of eight, whereas an imagery confound predicts somatomotor and visual systems to be most affected.
• 4.Correlation-based analyses, including WJ, cannot distinguish direct from indirect functional connections, and correlation magnitude is influenced by signal-to-noise ratio and head motion.
• 5.WJ quantifies architectural similarity but does not identify the specific neurobiological mechanisms driving reorganization. The observed patterns are consistent with GABA-A-mediated pharmacological effects but do not establish this mechanism.
• 6.The 4S456Parcels atlas provides moderate spatial resolution. Limbic network parcellation (26 ROIs) is coarse relative to the anatomical complexity of limbic structures, contributing to reduced statistical power for this network.
• 7.Head motion may differ systematically across consciousness states (e.g., reduced motion during unconsciousness). Although XCP-D preprocessing includes motion correction and confound regression, residual motion-related artifacts cannot be fully excluded and may contribute to observed correlation differences between conditions.
• 8.Global signal regression was not applied; results may differ under alternative preprocessing pipelines.
• 9.The findings are based on a single dataset (Michigan Human Anesthesia Dataset-1). Replication in independent cohorts is needed to establish generalizability.
• 10.Subject-level WJ values include both true architectural reorganization and noise inherent in single-subject correlation estimates. The within-condition floor reported in Section 3.1 quantifies that noise contribution directly, and the group-averaged analysis is shown in Section 3.5 to inflate similarity rather than to bracket it.
• 11.Segment-based estimates. The within-condition floor and the comparisons calibrated against it use 300-vol segments, so all values are estimated at that data length. The floor, the between-state values and the gap between them may differ at other durations, and the gap should not be assumed invariant to acquisition length.
• 12.Residual variability is not fully attributed. The within-condition floor bounds the contribution of estimation variability but does not identify the sources of the remaining variance, which may include arousal fluctuation, motion, physiological noise and genuine moment-to-moment reconfiguration.

6. Conclusions

Calibrated against repeated measurement of a single state rather than against a perfect-match ceiling, propofol-associated architectural reorganization is statistically robust with large paired effect sizes, modest in magnitude, redistributive in character, accompanied by genuine sign reorganization among a minority of strong edges, and only marginally reversed during early recovery. Scalar summaries of total connectivity are insensitive to this reorganization by construction, and group-averaged matrices systematically overstate its reversal.

At the network level, all eight canonical resting-state networks showed significant similarity reduction at the subject level, the control network was the most affected, and the default mode network was the only one to show significant early recovery. Those results are unchanged from the original submission and are independent of the reference correction reported here.

Code availability

All analysis pipelines that reproduce every value reported in this manuscript are publicly available at https://github.com/nwharbert8-ui/brain-connectivity-wj, release v1.1.0. Analyses were performed in Python 3.11 using SciPy, NumPy, pandas and matplotlib, with a fixed random seed of 42 throughout, so that all reported permutation p-values and bootstrap confidence intervals are exactly reproducible.

Ethics statement

This work is a secondary analysis of de-identified human neuroimaging data obtained from a public repository (OpenNeuro ds006623). Ethical approval and informed consent were obtained by the investigators of the original study at the originating institution. No additional ethical approval was required for the present secondary analysis of these publicly available de-identified data.

Declaration of generative AI use

This work was produced in an author-directed, AI-assisted research workflow, and the author regards full transparency about that workflow as part of its rigor.

Author's contribution (all scientific content and responsibility): The research direction, questions, and hypotheses; the theoretical framework and methodology; the selection and design of the analyses; the interpretation of all results; and the conclusions are entirely the author's. The author takes full and sole responsibility for the accuracy and integrity of every statement, number, figure, and reference in this manuscript.

AI-assisted components (tooling used under the author's direction): Anthropic's Claude (Claude Opus 4.6 and Claude Opus 5), within the Claude Code environment, assisted with implementing and executing the analysis pipelines in Python; drafting and formatting manuscript text from the author's own dictated interpretations, notes, and specifications; literature and reference retrieval; and code and consistency review. No pipeline, draft, or citation produced with AI assistance was accepted without the author's review, verification, and revision. Every reference was independently confirmed against its primary source.

Boundaries: Generative AI did not originate the scientific ideas, formulate the hypotheses, design the methodology, or make any interpretive or scientific judgment; these are the author's. In accordance with ICMJE and COPE guidance, generative AI is not an author and is not credited as one. The author alone is accountable for this work.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following are the Supplementary data to this article.

Table S1

Complete network-level permutation results (8 networks x 3 comparisons): observed WJ, null mean, null SD, uncorrected p-value, FDR-corrected p-value, effect size, and significance indicator.

Table S2

Leave-one-out jackknife results (24 subjects x 3 comparisons): LOO WJ value and influence score for each subject.

Table S3

Split-half WJ distributions (100 splits × 2 halves x 3 comparisons).

Data availability

The neuroimaging data analysed in this study are openly available from the OpenNeuro repository under accession ds006623 at https://openneuro.org/datasets/ds006623. No data were generated by the authors.