Neurophysiological connectomic signatures of consciousness during propofol-induced general anesthesia

Summary

General anesthesia induces reversible changes in consciousness through cortical activity and connectivity alterations, yet the functional connectome dynamics underlying propofol-induced unconsciousness remains unclear. We analyze high-density 128-channel electroencephalogram (EEG) from 31 surgical patients using source localization to identify neurobiological connectome signatures of propofol anesthesia. Propofol anesthesia increases delta and theta functional connectivity and decreases alpha, beta, and gamma connectivity. A classification model and dynamic analysis of consciousness loss reveals that alpha-band connectivity between parietal, occipital, and subcortical regions is critical for sustaining consciousness, with its disruption marking a key transition to unconsciousness. EEG from 46 additional patients under mild sedation with low-dose propofol confirms that decreased parietal-related alpha connectivity serves as a stable marker of reduced consciousness, insensitive to subtle fluctuations but sensitive to the transition from consciousness to unconsciousness. These findings suggest that parietal, occipital, and subcortical alpha connectivity serves as a reliable neural correlate of propofol-induced unconsciousness.

Graphical abstract

Highlights

• •We identify functional connectome dynamics of propofol-induced unconsciousness
• •Propofol increases δ and θ connectivity and decreases α, β, and γ connectivity
• •Disrupted parietal-related α connectivity marks a transition to unconsciousness
• •Reduced α connectivity serves as a marker of diminished consciousness

Li et al. explore functional connectome dynamics underlying propofol-induced unconsciousness using high-density electroencephalogram. Propofol anesthesia disrupts alpha-band connectivity between parietal, occipital, and subcortical regions, marking the key transition from consciousness to unconsciousness. These findings highlight the importance of functional networks for understanding the physiology underlying anesthesia-induced changes in consciousness.

Introduction

Understanding the mechanisms underlying transitions between consciousness and unconsciousness has been a substantial challenge for modern neuroscience. While behavioral responses remain the primary reference for assessing consciousness, neurophysiological markers derived from regional electroencephalogram (EEG) have also been widely used.¹ Consequently, EEG has long been used to estimate hypnotic levels under general anesthesia,^2,3 mostly based on time- and frequency-domain analyses.^4,5,6

Four cortical areas—the frontal, parietal, temporal, and occipital regions—are considered critical zones for maintaining consciousness.^7,8,9 Integration between the parietal and subcortical areas (striatum and thalamus) is thought to be critical for consciousness and may contribute more than the frontal areas.⁹ Meanwhile, changes in thalamic activity and thalamocortical interactions appear to broadly influence arousal levels.^10,11 The insula, cingulate gyrus, hippocampus, and claustrum are additional areas supporting cognitive processes such as sensation, attention, and memory.^12,13,14,15 However, it remains unclear how changes in various brain nuclei and their connections to other nuclei result in the reversible unconsciousness that defines general anesthesia.

Network connectivity analysis improves understanding of brain processes and allows the construction of consciousness-related theoretical frameworks.¹⁶ For example, anesthesia-induced unconsciousness is strongly correlated with functional connectivity disruption, reduced network efficiency, and a constrained repertoire of functional states^17,18—although the opposite has also been reported.^8,19 In particular, network characteristics within the α band reportedly directly reflect various states of consciousness induced by anesthesia.^20,21,22 Identifying neural correlates, therefore, requires careful delineation among changes in functional networks across various frequency bands.

Analyses based on EEG source reconstruction identified dynamic changes in network connectivity and processing of human consciousness in the cortical or subcortical regions across various frequency bands.²³ Identification of brain regions and time-varying processes that control consciousness may improve the safety and effectiveness of general anesthesia.²⁴ Propofol is a widely used general anesthetic because of its rapid onset, minimal side effects, and smooth, rapid awakening.²⁵ A growing body of literature indicates that propofol-induced loss of consciousness is characterized by distinct spectral and network-level alterations. Spectrally, propofol increases low-frequency EEG power (<1 Hz), suppresses occipital alpha oscillations (8–12 Hz), and enhances coherent frontal alpha activity,^5,26 reflecting altered thalamocortical dynamics. At the network level, propofol disrupts inter-regional communication, particularly thalamocortical and frontoparietal connectivity, while largely sparing primary sensory pathways.^26,27,28,29 Graph theoretical analyses further demonstrate reduced long-range integration, increased local clustering, and topological reconfiguration of brain networks, indicating a breakdown of global brain integration underlying the loss of consciousness.^30,31,32 Nonetheless, most prior work has emphasized static measures and cortical-level connectivity; how propofol modulates the dynamic connectivity between cortical and subcortical regions during transitions in consciousness remains poorly understood.

To address these questions, we used high-density 128-channel EEG from 31 surgical patients to compute source-localized phase-locked time-varying functional connectivity and identify brain functional connectivity patterns in various states of consciousness during propofol general anesthesia. Specifically, we selected nine brain regions related to consciousness—frontal cortex (Fro), temporal cortex (Tem), parietal cortex (Par), occipital cortex (Occ), cingulate gyrus (CG), insula (INS), thalamus (Tha), hippocampus (Hipp), and claustrum (Cls),^{33,34,35,36,37} with a total of 242 nodes—and compared differences in functional connectivity networks among these areas at the different frequency bands while subjects were awake and anesthetized. Second, we identified key neural correlates of propofol-induced unconsciousness based on results from a classification model of awake and unconscious states. Third, the correlation between dynamic connectivity changes in cortical and subcortical areas and loss of consciousness was analyzed, and these findings were further verified in 46 patients who were subsequently enrolled during mild sedation induced by low-dose propofol. Our goal was to identify and localize the temporal progression of cortical connectivity during anesthesia-induced alterations of consciousness.

Results

This prospective cohort observational study investigated changes in functional connectivity among brain nuclei during the induction of general anesthesia. Figure 1 illustrates the corresponding analytical protocols for exploring the network mechanisms of general anesthesia, encompassing EEG preprocessing, frequency filtering, source reconstruction, region of interest (ROI) selection, functional network construction among ROIs, network differences analysis between awake and unconscious states, classification between awake and unconscious states, dynamic observation of the state of consciousness, and low-dose propofol anesthesia verification. The corresponding results are presented as follows.

Figure 1.

Framework of data analysis

EEG data were collected using a 128 high-density EEG cap. Following EEG preprocessing and frequency-specific filtering, source reconstruction was conducted to obtain source space projected time sequence. Time-varying functional connectivity was then calculated at the ROI level. Differences in functional connectivity networks between awake and unconscious states were examined, and a support vector machine was applied to identify wakefulness and unconsciousness. Finally, dynamic analyses of key connections were conducted to characterize transitions between states of consciousness during general anesthesia.

Differences in brain network connectivity and network properties between awake and unconscious states

Given that consciousness emerges from complex interactions among spatially distributed brain functional regions, EEG functional connectivity networks were analyzed across 242 pairs of ROIs within five carrier frequency bands (delta, theta, alpha, beta, and gamma) and two states (wakefulness and unconsciousness), derived from cortical current source density in source space. Before constructing the functional networks, we first examined the cortical distribution of power spectral density (PSD) across frequency bands during wakefulness and unconsciousness to verify that the source-reconstructed signals exhibited the expected spectral modulations under propofol-induced general anesthesia. As shown in Figure S1, we observed canonical spatial patterns, including increased frontal delta power and an anterior shift of alpha power during propofol-induced unconsciousness, consistent with previous findings.^5,38,39 These observations confirm that source-level spectral activity was appropriately modulated and validate the accuracy of our source reconstruction.

Furthermore, there were significant differences in EEG functional connectivity networks between awake and unconscious states across all regions and all frequency bands. Specifically, as displayed in Figure 2A, delta band connectivity across brain regions was significantly stronger when patients were unconscious than when they were awake (p < 0.001, Bonferroni corrected). Unconsciousness similarly significantly strengthened Fro-CG, Fro-Tha, and Fro-Cls connections of the theta band (Figure 2B, p < 0.001, Bonferroni corrected). There was also a small amount of increased Tem-Hipp connectivity at the beta and gamma bands during unconsciousness.

Figure 2.

Phase-locking value connectivity matrices and differential network topologies between awake and unconscious states across frequency bands

(A–E) δ, θ, α, β, and γ bands, respectively (n = 31 patients). Red solid lines indicate connections with stronger connectivity during wakefulness compared with unconsciousness, whereas blue solid lines indicate the opposite. WAK, wakefulness; UCS, unconsciousness; FCN, functional connectivity network.

In contrast, connections were markedly diminished during unconsciousness in the alpha (i.e., Par-Tha, Par-CG, and Par-Occ), beta (i.e., Fro-Occ, Par-Occ, and Tem-Par), and gamma (i.e., Fro-Tem and Fro-Hipp) bands (p < 0.001, Bonferroni corrected, Figures 2C–2E). Taken together, propofol-induced general anesthesia was associated with an abnormal increase of network connectivity in the delta and theta frequencies, along with a profound loss of network phase coupling in the alpha, beta, and gamma frequencies.

To evaluate these functional networks further quantitatively, we computed two network properties: clustering coefficients and characteristic path length. Next, the network properties of the awake and unconscious states were compared to explore the differences in brain efficiency. As illustrated in Figure 3, compared with the awake state, the unconscious state showed decreased characteristic path length (t = −5.20, p < 0.001, Cohen’s d = −0.93, 95% confidence interval [CI] = [−1.63 −0.48]) and increased clustering coefficients (t = 4.11, p < 0.001, Cohen’s d = 0.74, 95% CI = [0.36 1.28]) in the delta band. The unconscious state showed a larger characteristic path length (alpha: t = 4.33, p < 0.001, Cohen’s d = 0.78, 95% CI = [0.47 1.18]; beta: t = 1.80, p = 0.08, Cohen’s d = 0.32, 95% CI = [−0.02 0.78], gamma: t = 2.27, p = 0.03, Cohen’s d = 0.41, 95% CI = [0.04 1.07]) and shorter clustering coefficients (alpha: t = −5.41, p < 0.001, Cohen’s d = −0.97, 95% CI = [−1.40 −0.67]; beta: t = −4.28, p < 0.001, Cohen’s d = −0.77, 95% CI = [−1.35 −0.38]; gamma: t = −4.70, p < 0.001, Cohen’s d = −0.85, 95% CI = [−1.70 −0.41]) in the alpha, beta, and gamma bands compared to wakefulness. In contrast, no significant differences were observed in theta-band network properties between the unconscious and wakeful states.

Figure 3.

Comparison of network properties between awake and unconscious states across frequency bands (n = 31 patients)

(A) Clustering coefficients of different carrier bands.

(B) Characteristic path length of different carrier bands.Box plots show the median and interquartile range (IQR), with whiskers extending to the minimum and maximum values within 1.5 × IQR, and violin plots depict the full data distribution. ∗p < 0.05; ns, p > 0.05; ∗∗∗p < 0.001.

EEG connectivity characteristics that discriminate unconsciousness from wakefulness

Given the extensive connectivity differences between the two states, we investigated whether functional connectivity distinguishes wakefulness from unconsciousness. Specifically, 45 inter- and intra-regional connectivity features were derived from the weighted network. Inter-regional connectivity was defined as the average connectivity between all ROIs across different regions (36 features), while intra-regional connectivity represented the average connectivity within multiple ROIs in a single region (9 features).

The F-score feature selection method was employed to select the top 25% promising features, which were then used to construct a classification model using the support vector machine. To ensure the reliability of the classification, a leave-one-out cross-validation strategy was adopted to quantitatively measure the performance and accuracy, as well as report on the sensitivity and specificity. As depicted in Figure 4A, classification in the delta band achieved an accuracy of 82.26% (sensitivity = 77.42%, specificity = 87.10%) and that in the alpha band achieved an accuracy of 87.10% (specificity = 87.10%, sensitivity = 87.10%) when differentiating the awake state from an unconscious state. Both bands, therefore, accurately distinguished awake and unconscious states. In contrast, when using connectivity as a discriminative feature, accuracies of 72.58%, 77.42%, and 75.81% were achieved for the theta, beta, and gamma bands, which were lower than those for the delta and alpha bands. The alpha model demonstrated the best classification performance across all frequency bands, achieving the highest classification accuracy, which was higher than that of the other models (all with classification accuracies below 85%). Furthermore, the alpha model also exhibited the highest area under the receiver operating characteristic (ROC) curve (AUC: 0.924, Figure 4B) compared to other frequency bands. Permutation tests (1,000 random label shuffles) confirmed that all classification accuracies were significantly higher than chance (p_perm < 0.001), indicating that the reported performance is robust and not due to random labeling. The permutation test results are shown in Figure S2.

Figure 4.

Identification of states of consciousness (i.e., awake and unconscious states) using EEG functional connectivity (n = 31 patients; classification performed at the subject level)

(A) Classification performance (i.e., accuracy, sensitivity, and specificity) across frequency bands.

(B) Receiver operating characteristic (ROC) curves of classification performance. AUC, area under curve.

(C) Feature weight matrix for the alpha-band classification model.

To determine the contribution of different connections to the identification of conscious states, we evaluated and visualized the weights of the alpha-band classification model, as shown in Figure 4C. Based on the feature weights, the parietal-related α-connectivity has a high contribution weight in recognizing awareness, such as the Par-Occ, Par-Tha, Par-CG, Par-Cls, and Par-Tem connections (all feature weights >0.7), which also differed significantly during the transition from wakefulness to propofol-induced general anesthesia (Figure 2C). These results underscore that parietal-subcortical, parietal-occipital, and parietal-temporal α-connectivity are key features for discerning the state of consciousness and play a critical role in maintaining consciousness.

Dynamics of functional connectivity during loss of consciousness

We then evaluated time-dependent changes in phase-locking connectivity during the transition from awake to unconscious states, focusing specifically on parietal-related connectivity within the alpha band (Par-Occ, Par-Tha, Par-CG, Par-Cls, and Par-Tem). These connections, which were identified as having the most significant contributions during anesthesia induction by conducting connectivity-level feature importance analysis, were central to the time-varying analysis. Additionally, we examined the dynamics of less-influential connections (Hipp-Hipp, Cls-Cls, Fro-Fro, Fro-Cls, and Fro-CG) in the classification of the state of consciousness, aiming to highlight the pivotal role of important network connections in modulating state transitions of consciousness. Time-varying connectivity for each patient was estimated from weighted time-varying matrices that were not binarized to any threshold.

The analysis indicates that there was considerable individual variability in the time course of information integration from the moment of awakening to unconsciousness (Figures S3–S5). To facilitate statistical analysis and visualization, we focused on the connectivity dynamics within a 20-s window before and after the loss of consciousness. As shown in Figure 5A (right), the time course of the least important connections exhibited little fluctuation during this period across all patients. In contrast, the most pivotal connections—particularly those related to parietal networks—displayed a sharp decline immediately before and after the loss of consciousness (Figure 5A, left). Figure 5B presents the results of a paired sample t test with false discovery rate correction, which revealed that parietal-related connectivity in the α band was significantly diminished following loss of consciousness compared to the pre-loss state (Par-Occ: t = 5.20, p < 0.001, Cohen’s d = 0.95, 95% CI = [0.58 1.47]; Par-Tha: t = 5.17, p < 0.001, Cohen’s d = 0.94, 95% CI = [0.58 1.55]; Par-CG: t = 4.05, p < 0.001, Cohen’s d = 0.74, 95% CI = [0.38 1.25]; Par-Cls: t = 4.09, p < 0.001, Cohen’s d = 0.75, 95% CI = [0.36 1.33]; Par-Tem: t = 4.19, p < 0.001, Cohen’s d = 0.77, 95% CI = [0.38 1.41]). However, no significant differences were observed in the less-influential network connections (p > 0.05). These findings emphasize the critical role of parietal-subcortical and parietal-occipital connectivity of the α band in reflecting the dynamic shifts associated with loss of consciousness.

Figure 5.

Time-varying alpha-band functional connectivity during the transition from wakefulness to unconsciousness (n = 31 patients)

(A) Time-varying connectivity during the 20 s before and after the loss-of-consciousness (LOC) time point. Data are shown as mean ± SE across patients.

(B) Averaged functional connectivity within the same time windows. ns, p > 0.05; ∗∗∗p < 0.001.

Low-dose propofol anesthesia verification

To validate the crucial role of alpha-band functional connectivity in characterizing fluctuations in consciousness levels, we conducted the same analysis on EEG data from patients (both awake and sedative states) during low-dose propofol anesthesia. Figure 6A presented that the sedative state exhibited a significantly reduced functional connectivity network within the alpha band compared to the awake state (p < 0.05, Bonferroni corrected), with the Fro-Occ connection showing the most pronounced decrease. As for the network properties, paired sample t tests revealed that the sedative state had a larger characteristic path length (t = −4.47, p < 0.001, Cohen’s d = −0.67, 95% CI = [−1.14 −0.34]) and a smaller clustering coefficient (t = 4.62, p < 0.001, Cohen’s d = 0.69, 95% CI = [0.34 1.19]) compared to the awake state (Figure 6B). These results further demonstrate that a decrease in consciousness level is accompanied by a reduction in alpha-band functional connectivity.

Figure 6.

Functional connectivity network during wakefulness and low-dose propofol-induced sedation in the alpha band (n = 46 patients; independent cohort)

(A) Differential network topology between the two states. Red solid lines indicate stronger connectivity during wakefulness compared with sedation, whereas blue solid lines indicate the opposite.

(B) Comparison of network properties between the two states.Box plots show the median and interquartile range (IQR), with whiskers extending to the minimum and maximum values within 1.5 × IQR, and violin plots depict the full data distribution.

(C) Classification performance (i.e., accuracy, sensitivity, and specificity) for distinguishing between the two states.

(D) ROC curve of classification performance.

(E and F) (E) Feature weight matrix for the alpha-band classification model. (F) Time-varying functional connectivity during sedation induced by low-dose propofol. Data are shown as mean ± SE across patients.

∗∗∗p < 0.001. SED, sedative state induced by low-dose propofol.

Nonetheless, we found that the network differences between the awake and sedative states were smaller compared to those between the awake and unconscious states. This reduction in differences led to a marked decline in the model’s ability to distinguish between the two states (accuracy = 61.96%, sensitivity = 60.87%, specificity = 63.04%, and AUC = 0.638, p_perm < 0.05), with the predictive power of the features significantly diminished (Figures 6C and 6D). Notably, the parietal-related alpha connections that were critical in identifying unconsciousness in the wake-unconsciousness model no longer showed significant effects in the sedated-unconsciousness model (Figure 6E, feature weights <0.01). Moreover, time-varying parietal-related alpha connectivity (Par-Occ, Par-Tha, Par-CG, Par-Cls, and Par-Tem) exhibited only subtle fluctuations during sedation, without abrupt changes (Figure 6F). These observations suggest that the most pronounced alterations in parietal-related alpha-band connectivity occur during the transition to unconsciousness, rather than during mild sedation, supporting their close association with loss of consciousness.

Discussion

We investigated the changes in functional cortical connectivity related to propofol-induced general anesthesia by performing EEG source location analysis during the transition from wakefulness to unconsciousness. The results are summarized as follows. First, unconsciousness enhanced low-frequency (i.e., δ and θ) and attenuated high-frequency (i.e., α, β, and γ) band connectivity. Second, α-connectivity between the parietal cortex and occipital, subcortical areas effectively distinguished the state of consciousness. Third, although cortical connectivity remained dynamic during the transition to unconsciousness and no single connectivity dimension could reliably quantify levels of consciousness across states, abrupt decreases were observed in parietal-subcortical and parietal-occipital α-connectivity after the loss of consciousness, suggesting that these connections may reflect key features of consciousness transitions rather than serving as a single definitive marker. Last, these findings in the alpha band were further corroborated and extended using EEG data from an additional 46 patients who underwent mild sedation with low-dose propofol. Complex interactions in the α band between the parietal cortex, occipital cortex, and deep brain areas, therefore, appear to crucially support consciousness.

General anesthesia is characterized by dysfunctional intercommunication among brain networks.^28,40,41 Examination of network-level connections, such as those provided by functional connectivity networks, therefore offers a comprehensive perspective for measuring consciousness.^27,41 Previous theories of consciousness posit that multiple cortical and corticocortical areas of intricate coordination are required to sustain consciousness.⁴² Our results add nuance in that connectivity in the delta and theta bands increases during unconsciousness, while connectivity in the alpha, beta, and gamma bands decreases. Integration of global neural information into the brain is important for consciousness but is not entirely lost during unconsciousness, and some even increase.⁴³ Likewise, increased characteristic path length and decreased clustering coefficients of α, β, and γ bands after anesthesia quantify inefficient global information transmission and altered local brain function.⁴⁴ The δ band exhibited an increased-efficiency brain network in the unconscious state, consistent with the changes in differential network topology. Peculiarly, increased δ-connectivity and properties during unconsciousness are a vital finding in terms of reflecting coordinated bi-stability (the transition between depolarized states and hyperpolarized states) that suppresses integrated information required for consciousness.^17,39,45 Meanwhile, a large-scale breakdown in the α-, β-, and γ-connectivity of the brain directly reflects changes in states of consciousness and interrupted communication induced by anesthesia,^46,47,48,49 especially α-connectivity. In essence, reduced functional connectivity may indicate the uncoupling of certain regions from the network, impairing information integration, while increased connectivity may be interpreted as a reduction in discriminable brain states, implying more stereotypical activity patterns; in either case, such changes may reduce the level of consciousness.⁴⁰ Our findings indicate that the functional connectivity changes induced by propofol-induced general anesthesia cannot be simply attributed to interruption or decline, but rather a heterogeneous response that may be related to the reversible transition between consciousness and unconsciousness.

The differences between the two states of consciousness identified in our functional connectivity analysis, taken together with prior research on propofol-induced general anesthesia more broadly,⁵⁰ imply that EEG functional connectivity networks at the ROI level have the potential to act as neurobiological biomarkers for identifying abnormalities and estimating the state of consciousness. Notably, when network connectivity in the alpha band was employed to distinguish unconsciousness from wakefulness, the highest classification accuracy, sensitivity, and specificity of 87.10% was achieved, with the highest AUC of 0.924, demonstrating the potential of alpha-band network connectivity in identifying anesthetic status. Previous studies have proposed that alpha dynamics may underlie propofol’s disruptions of sensory and cognitive functions.⁴⁹ Other studies have shown that distinct markers of alpha activity play a causal role in shaping conscious perception.⁵¹ Our findings both converge on and extend these findings and suggest that the functional connectivity network in the alpha band shows potential as a biomarker for decoding consciousness. Feature importance analysis at the connectivity level further found that alpha-band connectivity between the parietal cortex, occipital cortex, and subcortical areas contributed most to distinguishing conscious states. This suggests that alterations in parietal-subcortical, parietal-occipital, and parietal-temporal connectivity may mediate the shift to unconsciousness,⁵² aligning with the theories that the posterior cortex is a “hot zone” of consciousness and that the integration of parietal areas is a hallmark of conscious states.^9,48,53 Importantly, previous fMRI and intracranial EEG studies have reported that propofol-induced loss of consciousness is associated with disrupted thalamocortical communication and widespread reductions in subcortico-cortical and corticocortical connectivity,^26,30,54 particularly in posterior brain regions, as well as decreased functional connectivity within the default mode network,⁵⁵ including the parietal and cingulate cortices, thereby further supporting the plausibility of our alpha-band connectivity findings.

Although the averaged connectivity provides valuable information for identifying stereotypical signatures associated with unconsciousness, it does not reveal the full picture of cortical connectivity. According to the classification results, general anesthesia is characterized by disrupted α-connectivity in the parietal cortex, occipital cortex, and subcortical areas. We found large individual differences in time-varying functional connectivity, as each individual exhibited distinct fluctuation patterns and the time from wakefulness to unconsciousness (Figures S3–S5). Here, using the time alignment method based on the clinical label, we revealed the evolution of cortical connectivity at the group level and found that there were specific changes in patterns in multiple consciousness-related connectivity (Par-Occ, Par-Tha, Par-CG, Par-Cls, and Par-Tem). Specifically, these connections abruptly declined before and after the loss of consciousness time point for all patients. The decline may reflect a breakdown in the integration of information across a widespread region, with the remarkable changes in connectivity being the tipping point between consciousness and unconsciousness.⁵⁶ The results are consistent with the theory that disruption of parietal cortex-related connectivity is a final common pathway to unconsciousness.⁹ Our findings further suggest that the connections between the parietal cortex and subcortical areas may play a key integrative role and contribute more than intra-region interactions (Hipp-Hipp, Cls-Cls, and Fro-Fro) for consciousness.

The EEG data from patients undergoing low-dose propofol-induced sedation further corroborates and extends the above findings on brain network dynamics during transitions in consciousness. Sedation is a transitional state from wakefulness to general anesthesia, exhibiting significant alterations in functional connectivity.^57,58,59 We observed a significant reduction in alpha-band connectivity and network properties in the sedative state compared to wakefulness. Nonetheless, the reductions observed in the sedative state were smaller compared to those in unconscious states, resulting in a reduced ability of the classification model to distinguish between wakefulness and sedation. Accordingly, the reductions in alpha-band functional connectivity may represent a stable correlate of reduced consciousness. Crucially, subtle fluctuations in parietal-related alpha connectivity during sedation further highlight that the disruption of these connections may merely serve as a critical pathway to unconsciousness during anesthesia, rather than being directly influenced by reductions in consciousness level. However, differences in anesthetic dosage and the absence of a direct sedation-anesthesia comparison may limit the interpretation of these findings. Future studies systematically manipulating propofol dosage or incorporating intermediate states of sedation and recovery will be essential to determine whether the observed parietal alpha connectivity decreases genuinely reflect the transition to unconsciousness, rather than dose-dependent effects.

Our results suggest the importance of functional networks in understanding the physiology underlying changes in consciousness induced by anesthesia. Propofol anesthesia enhanced low-frequency (i.e., δ and θ) and attenuated high-frequency (i.e., α, β, and γ) band connectivities. A classification model and dynamic analysis of consciousness loss suggest that alpha connectivity between the parietal cortex, occipital cortex, and subcortical areas is critical for consciousness and disruption of these connections signifies a key transition to loss of consciousness. Taken together, this study integrates high-density EEG source localization, high temporal resolution, and multi-band connectivity analyses to reveal dynamic changes in parietal-subcortical α-band during propofol-induced unconsciousness, complementing previous cortex-focused or static connectivity approaches and providing insights into potential biomarkers of brain states under general anesthesia.

Limitations of the study

Several limitations of the present study should be acknowledged. First, EEG analysis was restricted to anesthetic induction and loss of consciousness because surgical stress and other drugs required for surgery would complicate analysis. Nonetheless, extending monitoring to the intraoperative period and, especially, recovery of consciousness are obvious next steps. Second, surface EEG electrodes cannot directly examine the deep cortex and the reliability of subcortical source reconstructions remains debated. Therefore, our findings should be regarded as exploratory and interpreted with caution. Incorporating stereo-EEG and intracranial electrodes in future studies would complement and extend our results and provide a more definitive understanding of neural dynamics during anesthesia. Third, although sLORETA reduces sensor-level volume conduction, residual spatial leakage inherent to linear inverse solutions may still influence phase-locking value (PLV)-based connectivity estimates. PLV was chosen for its robustness in short sliding windows and its suitability for tracking rapid anesthesia-induced state transitions, and future studies may further refine source-space connectivity estimates by integrating additional leakage-mitigation strategies (e.g., leakage-resistant phase metrics, orthogonalization or model-based geometric correction) under analysis frameworks that allow for longer temporal windows or model-based connectivity estimation. Fourth, our analysis is highly mathematical and depends on much previous work on EEG signal processing and assumptions about localization. We also made many statistical comparisons. While somewhat protected by Bonferroni corrections, some apparently significant associations may prove spurious. Finally, we evaluated a relatively small number of homogeneous patients under highly structured circumstances. Results in diverse patients and populations will surely be more variable.

Resource availability

Lead contact

Further information and requests for resources should be directed to and will be fulfilled by the lead contact, Tao Xu (balor@sjtu.edu.cn).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• •All original data reported in this paper will be shared by the lead contact upon reasonable request.
• •All data analysis and visualization in this study were performed using MATLAB 2018b. Custom MATLAB code for EEG functional connectome analysis is archived at Zenodo (DOI: 10.5281/zenodo.18091237) and is mirrored on GitHub (https://github.com/yuqin19970312/EEG-Functional-Connectome-Analysis-for-Propofol-Induced-Unconsciousness/tree/main). The code is publicly available as of the date of publication.
• •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.

Acknowledgments

This study was supported by grants from the National Natural Science Foundation of China (NSFC) (82525022, 82371284, including the Young Scientists Fund, Category A), the National Key R&D Program of China (2024YFA1306902), the Research Physician Program of Shanghai Jiao Tong University School of Medicine (20240814), the Shanghai Eastern Talent Plan—Leading Talent Project (BJWS2024040), and the China Postdoctoral Science Foundation (2022T150428).

Author contributions

T.X., T.-F.Y., P.X., Y.L., and H.Z. conceived and designed the study. X.C., Q.T., and C.W. collected the data. Y.L., S.L., and F.L. analyzed the data. Y.L., H.Z., Z.X., D.S., Z.Z., P.X., T.-F.Y., and T.X. wrote the manuscript together.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Software and algorithms
MATLAB 2018b	The MathWorks	https://ww2.mathworks.cn/products/matlab.html
MATLAB package: EEGLAB 14.1.1	EEGLAB	https://sccn.ucsd.edu/eeglab/index.php
MATLAB package: BrainNet Viewer 1.63	BrainNet Viewer	https://www.nitrc.org/projects/bnv/
MATLAB package: REST_v1.2_20200818	Yao, 200160	https://www.neuro.uestc.edu.cn/name/shopwap/do/index/content/96
LORETA (v20170220)	LORETA	https://www.uzh.ch/keyinst/SubPages/loreta/loreta.html
MATLAB package: Brain Connectivity Toolbox	Brain Connectivity Toolbox	https://www.nitrc.org/projects/bct/
Custom MATLAB code: EEG Functional Connectome Analysis	This paper	https://github.com/yuqin19970312/EEG-Functional-Connectome-Analysis-for-Propofol-Induced-Unconsciousness/tree/main https://doi.org/10.5281/zenodo.18091237

Experimental model and study participant details

This prospective cohort observational study was approved by the Ethics Committee of Shanghai Sixth People’s Hospital (Approval ID: 2019-KY-038(K)). Written informed consent was obtained from all participants prior to enrollment. The study was reported in accordance with the STROBE statement. Details of the recruitment process and patient exclusion criteria are presented in Figure S6.

Participants undergoing general anesthesia

A total of 37 patients undergoing general anesthesia were initially enrolled in the present study, of whom 6 were excluded due to excessively noisy data across multiple EEG channels during preprocessing. The final sample therefore comprised 31 patients (aged 23–69 years, mean age: 44 ± 13 years), including 11 men and 20 women who required general anesthesia for surgery. Patients’ characteristics are summarized in Table S1. Inclusion criteria were: (1) American Society of Anesthesiologists physical status I–III; (2) scheduled for limb surgery (including internal fixation, removal of internal fixation, and tumor biopsy); and (3) age between 18 and 70 years. Exclusion criteria included a history of head trauma, inability to cooperate with study procedures, or diagnosed neurological or psychiatric disorders.

Apart from participating in the standard pre-anesthesia assessments, all patients were also tested for normal hearing and urine toxicology to ensure that they had not taken drugs that might interact adversely with propofol or confound the EEG or behaviors. General anesthesia was induced with propofol 2–5 mg/kg and inhaled sevoflurane, which was maintained at 2–3% end-tidal concentration in 100% oxygen without nitrous oxide. The Bispectral Index (BIS) was maintained between 40 and 50, and routine anesthetic variables were continuously monitored.

Participants undergoing low-dose propofol sedation

EEG data from an independent cohort of 57 patients recorded during both awake and mild sedation states (low-dose propofol) were used as a validation dataset to corroborate and extend the main findings derived from the general anesthesia cohort. To ensure the reliability of dynamic functional network analyses, participants exhibiting excessive artifacts across multiple EEG channels were excluded, yielding a final sample of 46 patients (aged 19 to 70 years, mean age: 40 ± 15 years; including 20 men and 26 women). The inclusion and exclusion process for this cohort is detailed in Figure S6, and patients’ personal information is summarized in Table S2.

During routine clinical procedures immediately prior to induction of general anesthesia, the propofol dosage was reduced to record EEG data under low-dose sedation conditions, without any additional experimental intervention. Sedation was induced with propofol at a dose of 0.6–1 mg/kg, maintaining the BIS value between 60 and 70. Heart rate, oxygen saturation, respiration, expired carbon dioxide, and blood pressure were continuously monitored throughout the procedure.

Method details

EEG data acquisition

EEG data during general anesthesia were recorded using a 128-channel EEG cap and an EGI amplifier (EGI, Eugene, OR). All recording parameters were predefined. Data were sampled at 500 Hz with a bandpass filter of 0.01–100 Hz. Cz served as the reference electrode, and electrode impedance was maintained below 5 kΩ throughout the recording. EEG, electrocardiography (ECG), and blood pressure were recorded while patients were awake and in a calm state for 5 min with eyes closed. Participants were then instructed to continue keeping their eyes closed while performing gripping movements with each hand. During anesthesia induction, anesthesiologists administered 5 μg of sufentanil intravenously, followed by propofol (2–5 mg/kg). Loss of consciousness was defined as the time when participants lost their eyelash reflex and were no longer able to make a fist upon request. EEG recordings continued for 5 min during anesthesia induction. Resting-state and anesthesia-induced EEG data were acquired continuously, and the onset of anesthesia induction and loss of consciousness were marked to facilitate subsequent data extraction and analysis.

EEG data during low-dose propofol sedation were also acquired with a high-density electrode array (EGI, Eugene, OR) using high-impedance amplifiers. All scalp electrodes were adjusted to maintain impedance below 5 kΩ. Data were sampled at 500 Hz and offline bandpass filtered between 0.01 and 100 Hz. In the awake state, 5 min of EEG data were recorded continuously with the eyes closed. Subsequently, EEG data were recorded for an additional 5 min during low-dose propofol-induced sedation, with propofol administered intravenously at a dose of 0.6–1 mg/kg. The onset of propofol administration was marked to facilitate data extraction for subsequent analyses.

EEG preprocessing

EEG data were analyzed with MATLAB (v2018b; MathWorks, Natick, MA) and the EEGLAB toolbox (Swartz Center for Computational Neuroscience, La Jolla, CA). Raw EEG dataset was first notch filtered to remove 50 Hz power frequency noise and bandpass filtered between 1 and 45 Hz. These time series were then re-referenced using the reference-electrode standardization technique (REST).^60,61 To support dynamic analysis of the anesthesia induction process, EEG data were visually inspected, and channels exhibiting substantial artifacts were excluded. Six patients were excluded from further analysis because more than 10 EEG channels were identified as bad. For the remaining patients, bad channels were corrected using spherical interpolation prior to subsequent analyses. Independent component analysis (ICA) was subsequently applied to remove artifacts, including blinks, eye movements, heartbeat, and myoelectricity,⁶² using the Infomax ICA algorithm implemented via the ‘runica’ function in EEGLAB. The identification and removal of artifact components were based on a visual examination of each component’s topography, time course, and spectrum, in combination with MARA⁶³ and IClabel diagnoses.⁶⁴ On average, 1 to 2 artifact components were removed per patient. Thereafter, we used bandpass filtering to isolate the five conventional EEG bands: δ (1–4 Hz), θ (4–8 Hz), α (8–13 Hz), β (13–30 Hz), and γ (30–45 Hz). Filtering was performed using a zero-phase sinc finite impulse response (FIR) filter with an order of 200 and a Hamming window, implemented via the ‘pop_eegfiltnew’ function in EEGLAB. The data were then down-sampled to 100 Hz to reduce computational load. For each patient, four-minute segments of offline-filtered EEG data across the five frequency bands were extracted, centered on the time point of anesthetic injection. The continuous EEG data were further segmented into 5-s windows with an 80% overlap, yielding a temporal resolution of 1 s. This procedure resulted in time-varying, artifact-free EEG series for subsequent analyses.

Source reconstruction

Cortical and subcortical brain activity was reconstructed from scalp EEG signals using standardized low-resolution brain electromagnetic tomography (sLORETA), which has been shown to mitigate volume conduction effects and improve spatial resolution. A three-shell realistic head model implemented in the LORETA software (v20170220) was used for source estimation. The lead field matrix was computed using a standardized boundary element method (sBEM) volume conductor model,⁶⁵ with each voxel modeled as a three-dimensional dipole. The solution space consisted of 6239 voxels distributed with a spatial resolution of 5 mm⁶⁶ and aligned to the Montreal Neurological Institute (MNI) template.⁶⁷ With the 128-channel EEG system, the resulting lead field matrix had dimensions of 128 × 3 × 6239. With respect to the inverse algorithm, sLORETA⁶⁸ was employed to accurately reconstruct the cortical sources from scalp EEG.

EEG data were segmented into 5-s windows with an 80% overlap (128 electrodes × 500 time points × windows), and voxel-wise time courses were reconstructed, resulting in source-space data with dimensions of 6239 voxels × 500 time points × windows. Time courses for each anatomical region of interest (ROI) were extracted from the reconstructed sources using the Brainnetome atlas,⁶⁹ with a spherical radius of 6 mm around each ROI center, consistent with previous EEG source-space studies.⁷⁰ In addition, previous work has demonstrated that high-density EEG combined with sLORETA allows reasonable reconstruction of subcortical neuronal activity, including the cingulate gyrus (CG), insula (INS), thalamus (Tha), hippocampus (Hipp), and claustrum (Cls).^{23,71,72,73,74} Based on these findings, nine major brain regions were selected for subsequent analyses: frontal (Fro), temporal (Tem), parietal (Par), occipital (Occ), CG, INS, Tha, Hipp, and Cls. This resulted in a total of 242 ROIs, and the final ROI time series had dimensions of 242 ROIs × 500 time points × windows. The spatial distribution of these ROIs is illustrated in Figure 1.

Following source reconstruction, power spectral density (PSD) was estimated for each ROI using Welch’s method. To reduce inter-subject variability in absolute power and emphasize relative spatial patterns, PSD values were min–max normalized across ROIs within each time window for each participant. Group-level spatial maps were then obtained by averaging normalized PSD values across subjects.

Network construction

Functional connectivity was estimated using time series extracted from 242 source-reconstructed ROIs, which helps mitigate the influence of residual volume conduction effects.⁷⁵ Connectivity strength was quantified using the phase-locking value (PLV), a phase-based metric that measures the consistency of phase differences between pairs of signals.⁷⁶ PLV was selected for its robustness in short sliding windows and its suitability for capturing rapid, non-stationary phase synchronization during anesthesia-induced state transitions⁷⁷; however, it is not inherently leakage-resistant. Instantaneous phase time series were obtained using the Hilbert transform applied to band-pass filtered signals. PLV values range from 0 to 1, with higher values indicating stronger phase synchronization. For each participant, PLV was computed for each sliding time window to generate a time-resolved 242 × 242 weighted adjacency matrix. For state-level analyses, including network property estimation and classification, segment-level matrices were further averaged within each frequency band (δ, θ, α, β, and γ) to obtain a single weighted adjacency matrix per participant and condition (awake and unconscious states).

Network properties are lower-dimensional quantitative indices of complex networks that can measure the statistical characteristics of brain networks and further assess brain efficiency.⁷⁸ Two graph-theoretical measures, clustering coefficients (CC) and characteristic path length (CPL), were calculated using the brain connectivity toolbox (BCT, http://www.nitrc.org/projects/bct/)⁷⁹ based on the weighted PLV networks. The clustering coefficient reflects functional segregation, indexing the tendency of nodes to form locally interconnected clusters, whereas characteristic path length reflects functional integration and indexes the efficiency of information transfer across distributed network nodes. The detailed definitions of these two properties are formulated as follows:

$$CC=\frac{1}{N}\sum _{i\in \theta }\frac{\sum _{j,l\in \theta }{\left({\mathrm{C}}_{ij}{\mathrm{C}}_{il}{\mathrm{C}}_{jj}\right)}^{1/3}}{\sum _{j\in w_{ij}}\left(\sum _{j\in \theta }{\mathrm{C}}_{ij}-1\right)}$$ (Equation 1)

$$CPL=\frac{1}{N}\sum _{i\in \theta }\frac{\sum _{j\in \theta ,j\ne i}{d}_{ij}}{N-1}$$ (Equation 2)

where Cij denotes the PLV value between nodes i and j, N represents the total number of nodes, and dij denotes the shortest path length between nodes i and j.

To characterize the temporal evolution of unconsciousness, we quantified time-varying connection strength from the PLV-weighted functional networks during the loss-of-consciousness period. Both intra-regional (within the same anatomical region) and inter-regional (between different regions) connectivity were examined. For each time window, connection strength was defined as the mean PLV value across all ROI pairs belonging to the corresponding regional combination (e.g., parietal–parietal, parietal–occipital, and parietal–subcortical). This procedure yielded time-resolved measures of regional and inter-regional functional coupling throughout the transition from wakefulness to unconsciousness.

Discrimination of unconscious state from awake state

We used PLV-based functional connectivity to construct features for classification. Specifically, the average weighted connectivity between brain regions was extracted and used as input features. To reduce feature dimensionality, an F-score–based feature selection method was applied to select the top 25% most discriminative features.⁸⁰ Classification between awake and unconscious states was performed using a support vector machine (SVM) with a radial basis function (RBF) kernel, combined with a leave-one-out cross-validation (LOOCV) strategy.^20,81,82 For each LOOCV iteration, one sample was held out for testing, while the remaining m–1 samples (m = 62, including both awake and unconscious states) were used for training. Feature selection was performed using only the training data within each LOOCV fold to avoid information leakage. The SVM hyperparameters, including the regularization parameter C and the kernel width γ, were optimized within the training set of each LOOCV fold using a grid-search strategy. The SVM classifier was retrained at each LOOCV iteration, and this procedure was repeated until every sample had been used once as the test set. After completing the LOOCV procedure, classification performance was evaluated using accuracy (Acc), sensitivity (Sen), and specificity (Spe). The definitions of the three indices are as follows:

$$Acc=\frac{n_{WAK}+n_{UCS}}{N_{WAK}+N_{UCS}}\times 100\%$$ (Equation 3)

$$Sen=\frac{n_{WAK}}{N_{WAK}}\times 100\%$$ (Equation 4)

$$Spe=\frac{n_{UCS}}{N_{UCS}}\times 100\%$$ (Equation 5)

where N_WAK and N_UCS denote the total numbers of samples of the awake and unconscious states, and n_WAK and n_UCS denote the correctly classified samples of the awake and unconscious states.

To further assess the model’s discriminative performance, the area under the receiver operating characteristic (ROC) curve (AUC) was computed using the ‘perfcurve’ function in MATLAB. AUC quantifies the model’s ability to distinguish between awake and unconscious states across decision thresholds,⁸³ with higher values indicating better class separability and overall classification performance.

Low-dose propofol anesthesia verification

To further corroborate and extend the findings observed during propofol-induced general anesthesia, we examined time-varying PLV-based functional networks and network properties in the alpha band during low-dose propofol anesthesia. For each participant, four-minute segments of offline-filtered alpha-band EEG data were extracted, centered on the time of low-dose propofol administration, including 2 min of wakefulness and 2 min of low-dose sedation. The EEG data were segmented using a 5-s sliding window with an 80% overlap. Source localization, ROI time-series extraction, PLV network construction, and graph-theoretical property calculations were then performed following the same procedures described for the general anesthesia cohort. Differences in functional network topology and network properties between the awake and sedated states were statistically assessed. In addition, time-varying changes in key functional connections were examined to characterize the dynamic effects of low-dose propofol on alpha-band connectivity. Classification of wakefulness and sedation was performed using a support vector machine, applying the same feature extraction, parameter optimization, and validation procedures as described previously.

Quantification and statistical analysis

All statistical analyses were performed using MATLAB (R2018b; MathWorks, Natick, MA). Data normality was assessed prior to hypothesis testing. When the data followed a normal distribution, within-subject differences in PLV-based functional connectivity and network properties between awake and unconscious states, between awake and low-dose sedation states, as well as before and after loss of consciousness were evaluated using two-tailed paired-sample t-tests (p < 0.05). For edge-wise network comparisons, multiple comparisons arise due to the large number of connections tested simultaneously. To control for family-wise error rate, Bonferroni correction was applied by adjusting the significance threshold according to the number of tested connections. This procedure reduces the likelihood of false-positive findings when identifying functional connectivity differences between conditions. For classification analyses, statistical significance was assessed using permutation testing. Empirical chance levels were derived from 1000 permutations, in which class labels were randomly shuffled prior to classification. The resulting classification accuracies formed a null distribution, against which the observed accuracy was compared to obtain a permutation-based p-value.

Published: January 29, 2026