Changes in Whole Brain Dynamics and Connectivity Patterns during Sevoflurane- and Propofol-induced Unconsciousness Identified by Functional Magnetic Resonance Imaging

The wakeful human is able to flexibly modulate his or her brain activity in order to respond to external stimuli or to execute tasks. These capabilities of the human brain can be eliminated through general anesthesia. As a consequence, anesthetic agents form an experimental framework to investigate the underlying brain physiology.

In the past decade, relevant progress has been made to elucidate the mechanisms of anesthesia-induced unresponsiveness, giving rise to the hypothesis that specific brain regions and the communication between them are necessary to be wakeful. Specifically, changes of connectivity in the default mode network, attentional network, and/or the medial thalamus have been associated with loss of responsiveness during general anesthesia. Electroencephalography analyses have repeatedly shown a significant decrease of frontoparietal information transfer during general anesthesia when compared to wakefulness.^1–3  The underlying neurophysiologic mechanism of frontoparietal connectivity was further substantiated by functional magnetic resonance imaging experiments that demonstrated a suppression of frontal activation by propofol and subsequently impaired frontoparietal connectivity.^3–5  These observations have been replicated in functional magnetic resonance imaging experiments during sevoflurane-^6,7  and ketamine-induced loss of responsiveness.^8,9  Similarly, dexmedetomidine-induced unresponsiveness was associated with reduced connectivity between the thalamus and the default mode network.¹⁰  In summary, frontoparietal information exchange in the default mode network, the attentional networks, and thalamocortical connectivity to both are thought to be necessary for consciousness. However, there is little knowledge about how changes in few networks lead to a breakdown of between-network communication that we hypothesized to be a neurophysiologic correlate of anesthesia-induced unconsciousness.

One possible explanation is that only a few brain networks are crucial for generating global, dynamic cortical interactions. Indeed, functional magnetic resonance imaging data in humans could demonstrate that “hub” regions exist, characterized by a dominance of long-range connectivity. It is therefore presumed that these regions are capable to dynamically integrate cross-modal information.¹¹  Additionally, these regions overlap anatomically with the networks affected during general anesthesia. Computational models predicted that lesioning of hub regions is capable of changing long-range connectivity patterns in the human brain.^12–15  Experimentally, it could be demonstrated that anesthesia-induced loss of responsiveness leads to a fragmentation of cortical information processing¹⁶  and to a decreased dynamic repertoire of brain states.^17–19  Taking these results into account, we hypothesized that the altered connectivity of hub regions should result in two observations: (1) decrease of the dynamic repertoire of connectivity patterns under loss of consciousness (irrespective of the anesthetic agent), and (2) the existence of specific dynamic connectivity patterns that can only be observed during consciousness.

To connect prior analyses of within-network activity changes^3,7  with between-network connectivity, we chose a spatial independent component analysis–based approach. We regarded the mean z-score of spatially independent maps or networks as a measure of within-network activity.⁷  The correlation matrix of the associated time courses was regarded as a measure of between-network connectivity. Thus, we aimed to model both the modular and intermodular connectivity structure of the human brain. The dynamics of between-network connectivity were modeled with sliding window method calculating the correlation between networks’ time courses over 1 min and k-means clustering²⁰  to identify dominant traits of between-network connectivity patterns (fig. 1).

The study was conducted in accordance with the Declaration of Helsinki, and the study protocol was approved by the ethics committee of the medical school of the Technical University Munich (Munich, Germany) and the University of Liege (Liege, Belgium), respectively. Applicants to the study were given detailed information about the procedures and potential risks before a written informed consent. All participants were reimbursed for taking part in the study. A medical history was taken from every applicant, especially inquiring about contraindications for general anesthesia using propofol or sevoflurane or examination in a magnetic resonance scanner.

The Munich data set included in this analysis has been previously published in Jordan et al.³  and Ranft et al.⁷  Further details on the study protocol can be found there. Briefly, 15 subjects, aged 21 to 32 yr, were enrolled in the propofol study. For the resting state measurement, participants lay supine in the scanner and were told not to fall asleep while keeping the eyes closed. In the rest of the article, this condition is referred to as “awake.” The measurement during propofol-induced loss of responsiveness was carried out using a target-controlled infusion pump (Open TCI; Space infusion system; Braun Medical, Germany). Propofol concentration was increased in 0.4 μg/ml steps beginning at 1.2 μg/ml until volunteers stopped responding to the verbal command “squeeze my hand” (equivalent to a Ramsay sedation scale score of 5 to 6). The concentration was then kept stable for the remaining functional magnetic resonance imaging measurement. This point was reached at plasma concentrations of 2.97 ± 0.47 µg/ml (mean ± SD). Ten minutes of equilibration time passed before the actual measurement took place. Throughout this article, this state as is referred to as propofol-induced loss of responsiveness.

Twenty subjects, aged 20 to 36 yr, were enrolled in the sevoflurane study. The resting state was acquired identical to the propofol setting. Image acquisition during sevoflurane-induced loss of responsiveness was carried out after intubation with a magnetic resonance tomograph-compatible laryngeal mask and during artificial ventilation using an anesthesia machine (Fabius Tiro, Dräger, Germany). Sevoflurane was kept stable at 2 volume percent end-tidal concentration during this condition. Subjects were unresponsive to the command “squeeze my hand” during this condition, tolerated the laryngeal mask well, and showed reduced movements when compared to the wakeful state (see also Ranft et al.⁷ ). It is also noteworthy that clinically this sedation was deeper when compared to propofol-induced loss of responsiveness (corresponding to a Ramsay scale score of 6). In the following we refer to loss of responsiveness irrespective of anesthetic agent as loss of responsiveness. All participants were asked for any memories of the unresponsive state and all reported an amnesia for the procedure.

The Liege data set included in this analysis consisted of 20 subjects during propofol-induced loss of responsiveness. The data was previously published in Boveroux et al.⁴  From the complete data set (4 men and 16 women, aged 18 to 31 yr), 6 subjects with recording lengths of 300 time points could be included. For anesthesia, propofol was infused through an intravenous catheter placed into a vein of the right arm using a computer-controlled intravenous pump (Alaris TIVA; Carefusion, USA). Subjects were spontaneously breathing throughout the experiment. During the state examined here, propofol-induced loss of responsiveness, subjects were unresponsive to verbal command “squeeze my hand.” Clinically, the depth of anesthesia was classified as Ramsay 5. Propofol-induced loss of responsiveness was reached with plasma propofol concentrations of 3.2 ± 0.99 µg/ml (mean ± SD).

Subjects with head motion greater than 2 mm between two subsequent images were excluded from the final analysis. This resulted in 11 datasets from the Munich propofol cohort, 5 subjects from the Liege propofol cohort, and 14 subjects from the sevoflurane cohort.

Both Munich propofol and sevoflurane image acquisitions were carried out on the same 3-Tesla whole-body magnetic resonance tomographic scanner (Philips Achieva Quasar Dual 3.0T 16CH, The Netherlands) employing an eight-channel, phased-array head coil. Functional magnetic resonance imaging was acquired with a gradient echo planar imaging sequence (echo time = 30 ms, repetition time = 1.838 ms, flip angle = 75°, field of view = 220 × 220 mm², matrix = 72 × 72, 32 slices, slice thickness = 3 mm, and 1-mm interslice gap; 300 volumes were acquired in the propofol cohort and 350 volumes were acquired in the sevoflurane cohort; of these 350 frames, the last 50 were discarded). Anatomy was acquired before the functional scan using a single T1-weighted sequence and 1 × 1 × 1-mm voxel size per subject.

The Liege propofol data were acquired on a 3 Tesla Siemens Allegra scanner (Siemens AG, Germany) For the functional magnetic resonance imaging imaging, an Echo Planar Imaging sequence using 32 slices (repetition time = 2,460 ms, echo time = 40 ms, field of view = 220 mm, voxel size = 3.4375 × 3.4375 × 3.9 mm, and matrix size = 64 × 64 × 32) was employed. The individual T1 image was acquired at the end of the whole experiment.

In the following, we describe the analysis steps irrespective of the anesthetic agent, since in this study we were more interested in general mechanisms of loss of responsiveness than in mechanisms of a specific anesthetic agent. The preprocessing was carried out using Statistical Parametric Mapping (SPM12, Wellcome Trust Center for Neuroimaging, University of London, United Kingdom) and Data Processing Assistant for Resting-State fMRI (DPARSF, Release V4.3_170105, State Key Laboratory of Cognitive Neuroscience and Learning, Beijing Normal University, China). We largely followed the parameters in Allen et al.²⁰  in order to make results comparable. Precisely, the first three time points were removed to avoid magnetic coil saturation, slice timing was corrected, and functional images were realigned. Functional images were coregistered to the anatomical image. Six movement parameters and their first derivatives together with five principal components of the white matter and cerebrospinal fluid signal were regressed out, and the data was band-pass filtered between 0.01 to 0.1 Hz. The data was then normalized to Montreal Neurologic Institute space, resliced to a voxel size of 2 × 2 × 2 mm with a third-degree spline interpolation, and smoothed with a 4 × 4 × 4-mm Gaussian kernel.

In a first step, the preprocessed functional magnetic resonance imaging data were analyzed using a group level spatial independent component analysis as implemented in the Gift toolbox (Version 4.0b²¹ ) and GICA3 back reconstruction. The INFOMAX algorithm was employed together with an ICASSO and 20 repetitions to decompose the data into 75 spatially independent components (fig. 1). We preferred this high number of independent components to allow a subdivision of known networks (e.g., the default mode network or its parts) and to be able to analyze both the communication between and within networks, in summary a similar approach to Allen et al.²⁰  We used the GICA3 back reconstruction to generate individual spatial maps and associated (potentially correlated) time courses for each individual map.

This resulted in two types of data: (1) spatially independent components for each subject that were standardized to be standard normally distributed, so-called z-values (we refer to these maps as S_i), and (2) one time course for each component and each subject representing the individual components contribution the overall signal over time. The time courses were also standardized to be standard normally distributed to account for different scanners and amplitudes of the raw blood oxygenation level dependent signal. We refer to these time courses as R_i.

Both types of data were further analyzed independently with the exception that spatial maps were used to separate networks from noise components. This was necessary since independent component analysis is also known to separate presumable noise from signal of neuronal origin.²²  Signal from neuronal origin was assumed if the component projected on gray matter showed no similarity to venous vessels and if it showed a characteristic frequency spectrum with clear peak less than 0.1 Hz. Noise components usually showed a flat frequency spectrum similar to Gaussian noise. Fifty-six such components with visually high similarity to components in the literature^20,23  were deemed as functional networks and included into the further analysis. This approach is widely used in the literature even though it carries a certain risk to classify neural signal as noise or vice versa.

The group level S_i was thresholded at P < 0.001, uncorrected in SPM12 to create one mask per functional network. Mean z-scores were then calculated from the single subject, single session back-reconstructed S_i. In the following, we interpret the mean z-values as a measure of activity within a given network and refer to them as “within-network activity.”

The time courses R_i were despiked and low-pass filtered less than 0.15 Hz. Correlation matrices were calculated in a sliding window manner identical to Allen et al.^20,23  and as implemented in GIFT 4.0b’s Temporal dFNC toolbox using 30 time points, equivalent to 1 min, from each measurement, resulting in 267 correlation matrices per measurement (which is 300 recorded volumes minus 3 discarded during preprocessing minus sliding window length). The size of the resulting matrices was 56 × 56, and due to symmetry of the correlation matrices, they contained 56 × 55/2 independent data points. To make correlation matrices more accessible for visual inspection, we sorted the functional network similar to Allen et al.²⁰  in groups, namely basal ganglia networks, auditory networks, somatomotor and somatosensory networks, visual networks, default mode networks, attentional networks, and cerebellar networks. The complete list of networks can be found in the Supplemental Digital Content, https://links.lww.com/ALN/B925. It is noteworthy that this sorting into functional groups has no effect on the later k-means clustering algorithm, since the algorithm only uses the sum of the absolute differences (L₁)-distance of the correlation matrices as a measure of distance and is therefore independent of the order of correlation values. In general, we interpret the correlation values of the networks’ time courses as a measure of between-network connectivity and therefore refer to them as “between-network connectivity.”

We used the k-means clustering algorithm also implemented in GIFT 4.0b with the L₁-(“Manhattan”) distance with 20 repetitions to assign the between-network connectivity matrices to between-network connectivity patterns. These patterns are not predefined and are generated by the k-means algorithm as “focal points” or clusters of the between-network connectivity matrices. Only the number of such clusters has to be defined. Since the actual number of clusters k is unknown, we varied the number between k = 2 and k = 10. The results presented in the following are from k = 7 and representative for all k = 2,…, 10 (see also Supplemental Digital Content, https://links.lww.com/ALN/B925).

In order to investigate how information from both within-network activity and between-network connectivity is able to separate wakefulness from unresponsiveness, we formed a summary data matrix containing the within-network activity from the 56 networks together with the absolute number of between-network connectivity pattern appearances in each subject. Due to the high dimensionality (n = 63) and relatively low number of subjects (n = 30), we chose classical Fisher scoring²⁴  for a prior feature selection. Otherwise the high number of features would make the following analysis prone to overfitting and thus spurious results. Feature selection is a common strategy to reduce dimensionality without missing relevant information. In the analysis presented here, feature selection mainly favors results from within-network activity due to its high number of features (n = 56).²⁵  This particular Fisher scoring was also chosen because the data matrix contained both normally distributed data (mean z-scores) and sparse data (absolute number of appearances) and because selection of single variables from the data matrix made interpretation of the selection process simpler. After variable ranking, one to five variables were chosen for a support vector machine. For cross-validation, the 30 data sets were randomly divided into 15 training and 15 testing data sets. Single subjects were assigned to either the training or test data set. Since it was not feasible to repeat the process for all partitions of the data set, we repeated the process 1,000 times and averaged across runs. We also conducted the support vector machine with a random permutation of groups of the test data set to illustrate chance level. These steps were carried out with FSLib (Version 5.1, Roffo et al.²⁶ ). All analysis steps were carried out in Matlab (R2016a; Mathworks, USA).

Group comparisons were conducted for both the spatial and the temporal domains. Each group contained 30 subjects during awake or loss of responsiveness that were therefore regarded as paired variables.

In the spatial domain, the 75 spatially independent components’ mean z-scores were tested for a decrease using the one-sided Mann–Whitney/rank sum test. P values were corrected for multiple comparisons using the Bonferroni method (using the multiplicator n = 75; see also Supplemental Digital Content, table 1, https://links.lww.com/ALN/B925). The supplementary table and the results section give the median change of each component’s z-score and the approximated CI using a normal distribution. We chose this approximation since, according to the central limit theorem, the mean of a large number (ranging from 894 to 20,606) of voxels’ z-scores is approximately normally distributed.

The number of transitions between the different patterns of between-network connectivity was statistically tested for a difference between awake and loss of responsiveness, propofol-induced loss of responsiveness, and sevoflurane-induced loss of responsiveness using the two-sided Mann–Whitney/rank sum test. The average correlation of the average between-network connectivity matrix was also tested for differences between awake and loss of responsiveness, propofol-induced loss of responsiveness, and sevoflurane-induced loss of responsiveness using the two-sided Mann–Whitney/rank sum test. We chose this test since the actual distribution of the random variables was unknown a priori. Differences of pattern distribution between the groups awake, loss of responsiveness, propofol-induced loss of responsiveness, and sevoflurane-induced loss of responsiveness (on the nominal scale of pattern numbers) were tested using a chi-square test with Pearson’s chi-square distance.

Our analysis reproduced previous findings about within-network activity changes during loss of responsiveness. More precisely, bilateral ventral and dorsal medial prefrontal cortices (median change, 95% CI: ventral: –0.34 [–0.48; –0.11], dorsal: –0.27 [–0.40; –0.13]), dorsolateral prefrontal cortices (–0.23 [–0.35; –0.091]), the bilateral middle frontal gyruses (–0.17 [–0.31; –0.065]), posterior cingulate cortex and precuneus (–0.31 [–0.42; –0.065]), the right frontoparietal network (–0.18 [–0.34; –0.05]), the bilateral intraparietal (–0.13 [–0.25; –0.049]) cortices, and the salience network (–0.18 [–0.31; –0.044]) showed a significant reduction of within-network activity during loss of responsiveness when compared to awake at a significance level of P < 0.001 (Bonferroni-corrected; see Supplemental Digital Content, https://links.lww.com/ALN/B925, for effect sizes and CI of all components). This dominant effect on frontal networks during loss of responsiveness has been shown for both propofol-induced and sevoflurane-induced loss of responsiveness^3,7  as well as significant effects on the medial thalamus only during sevoflurane-induced loss of responsiveness.^6,7  Additionally, loss of responsiveness resulted in a significant decrease of within-network activity in wide-spread areas at a significance level of P < 0.05 (range of change from –0.36 [–0.48; –0.040] in the medial thalamus to –0.10 [–0.33; –0.020] in the bilateral frontal pole). These areas encompass premotor and motor areas, large portions of the parietal lobes and hippocampal network. Within-network activity in the occipital lobes, the temporal lobes, areas around the central sulcus and basal ganglia networks including the thalamus (but not the medial thalamus) were not significantly decreased (range of change from –0.19 [–0.44; 0.063] in the bilateral cuneus to 0.21 [–0.061; 0.42] in the left middle occipital gyrus). In summary, the results replicated previous findings showing a hierarchical and specific effect of anesthesia-induced loss of responsiveness on the brain’s networks rather than a uniform decrease of cerebral activity.

The between-network connectivity was modeled with correlation matrices from the independent component’s time courses. For visualization of the correlation matrices and increased comparability with other studies, we grouped the 56 functional networks into 7 groups: basal ganglia networks, auditory networks, somatosensory networks, visual network, default mode networks, cognitive control networks, and cerebellar networks (fig. 2, A and B). A complete list of the identified functional networks can be found in the Supplemental Digital Content, https://links.lww.com/ALN/B925. The average between-network connectivity matrix during awake (fig. 2B) showed several remarkable features that have already been described in the previous literature²⁰ : high between-network connectivity within the groups of somatomotor, visual, and auditory networks and between these groups. Cognitive control networks showed a high between-network connectivity to sensory networks, while some parts of the default mode network also showed low to negative between-network connectivity to these. The average between-network connectivity matrix during loss of responsiveness showed a significant reduction of average between-network connectivity by means of average correlation (P < 0.001, median change, 95% CI: –0.031 [–0.058; –0.025], fig. 2C). A differentiated analysis of propofol-induced loss of responsiveness (fig. 2D) and sevoflurane-induced loss of responsiveness (fig. 2E) showed reduction of average between-network connectivity during sevoflurane-induced loss of responsiveness (P < 0.001, –0.050 [–0.077; –0.026]) and during propofol-induced loss of responsiveness (P < 0.05, –0.023, [–0.057; –0.010]). The between-network connectivity was not significantly smaller during sevoflurane-induced loss of responsiveness when compared to propofol-induced loss of responsiveness (P = 0.24, medians, 95% CI: propofol-induced loss of responsiveness: 0.029 [0.022; 0.055], sevoflurane-induced loss of responsiveness: 0.018 [0.016; 0.027]). The visual assessment suggested a decreased between-network connectivity between sensory systems, as well as a less pronounced negative between-network connectivity between cognitive control and default mode networks. Similar observations have successfully been used to distinguish responsive from unresponsive patients²⁷  and recognized as effects of general anesthesia.^4,8  However, due to the high number of between-network connectivity values in a single matrix and hence unfavorable correction of alpha errors, both observations are not accessible to direct statistical testing in the framework presented here. Consequently, instead of looking at individual between-network connectivity values, we suggest looking at entire between-network connectivity matrices and interpreting their content as a measure of between-network communication. It is plausible that this between-network connectivity should have characteristic traits and should be subject to dynamic fluctuations that make complex behavior possible in the first place.

We further examined the between-network connectivity matrices using a sliding window method and k-means clustering (see “Materials and Methods” for details) to further quantify the dynamics of between-network connectivity and to isolate characteristic patterns of between-network connectivity. The results are presented for k = 7 in accordance with previous literature,²⁰  but we also conducted the analyses with k ∈ [2, …, 10] (see Supplemental Digital Content, fig. 1, https://links.lww.com/ALN/B925). The presented results are representative for all examined k. We found that the between-network connectivity patterns and their dynamic change were different during awake and loss of responsiveness. We observed two main findings: (1) during awake, different patterns appeared dynamically (patterns 2, 4, 6, and 7, fig. 3A); and (2) loss of responsiveness was characterized through a dominant pattern of between-network connectivity (pattern 5, fig. 3A). Consequently, the distribution of frames across different patterns was significantly different during loss of responsiveness (relative number of pattern appearances: pattern 1: 0.047, pattern 2: 0.11, pattern 3: 0.14, pattern 4: 0.0026, pattern 5: 0.66, pattern 6: 0.0017, pattern 7: 0.031) when compared to awake (relative number of pattern appearances: pattern 1: 0, pattern 2: 0.096, pattern 3: 0.032, pattern 4: 0.25, pattern 5: 0.056, pattern 6: 0.15, pattern 7: 0.42). This was also true for awake versus propofol-induced loss of responsiveness (relative number of pattern appearances: pattern 1: 0.089, pattern 2: 0.20, pattern 3: 0.28, pattern 4: 0, pattern 5: 0.43, pattern 6: 0.0033, pattern 7: 0), awake versus sevoflurane-induced loss of responsiveness (relative number of pattern appearances: pattern 1: 0, pattern 2: 0, pattern 3: 0, pattern 4: 0.0056, pattern 5: 0.93, pattern 6: 0, pattern 7: 0.066) and propofol-induced loss of responsiveness versus sevoflurane-induced loss of responsiveness (all P < 0.001). Additionally, we found patterns that occurred exclusively or almost exclusively during loss of responsiveness (numbers 1, 3, and 5) or dominantly during awake (patterns 4, 6, and 7). Among those were one low correlation pattern (7) and two high correlation patterns (4 and 6). Especially pattern 4 showed high similarity with the average correlation matrix. Although patterns 7 and 5 were visually similar, they show a high specificity for either awake or loss of responsiveness.

We regarded the absolute number of transitions between connectivity patterns as a measure of the ongoing dynamics. We found a significant reduction of transitions during loss of responsiveness (median during loss of responsiveness: 0, 95% CI: [0.26; 1.47]) when compared to awake (median during awake: 4, 95% CI: [3.35; 6.12], P < 0.001) as well as during propofol-induced loss of responsiveness (median during propofol-induced loss of responsiveness: 0.5, 95% CI: [0.45; 2.55], P < 0.05) and sevoflurane-induced loss of responsiveness (median during sevoflurane-induced loss of responsiveness: 0, 95% CI: [–0.17; 0.45], P < 0.001, both fig. 3B) when compared to awake. Additionally, the number of transitions was significantly reduced during sevoflurane-induced loss of responsiveness when compared to propofol-induced loss of responsiveness (P < 0.05). In total, 27 out 30 subjects showed reduction or no change of transitions during loss of responsiveness when compared to awake. All subjects who had an actual increase of transitions were in the propofol-induced loss of responsiveness group (n = 3, fig. 3B). Thirteen out of 14 subjects of the sevoflurane-induced loss of responsiveness group had 0 transitions during loss of responsiveness, probably reflecting the deeper level of anesthesia during sevoflurane-induced loss of responsiveness when compared to propofol-induced loss of responsiveness. Variation of k ∈ [2, …, 10] also led to a separation into distinct patterns for either the awake or loss of responsiveness state (see Supplemental Digital Content, fig. 1, https://links.lww.com/ALN/B925).

In order to further evaluate the information content of the acquired information about the subject being either in the awake or loss of responsiveness condition, we employed a support vector machine approach after prior feature selection via the classical Fisher score.²⁴  The top five features calculated from the complete data set were appearance of between-network connectivity pattern 5, within-network activity in the dorsal portion of the medial prefrontal cortex, appearance of between-network connectivity pattern 7, within-network activity in the dorsolateral prefrontal cortex, and within-network activity in the ventral portion of the medial prefrontal cortex (fig. 4A). We calculated the accuracy of this classification on the test data with 1, …, 5 top features based on the training data set (fig. 4B) using either within-network activity (mean classification accuracy and 95% CI: 1 feature: 0.72 [0.716; 0.725], 2 features: 0.74 [0.737; 0.745], 3 features: 0.74 [0.738; 0.746], 4 features: 0.75 [0.743; 0.751], 5 features: 0.74 [0.739; 0.747]), between-network connectivity (mean classification accuracy and 95% CI: 1 feature: 0.79 [0.788; 0.793], 2 feature: 0.80 [0.801; 0.808], 3 features: 0.86 [0.858; 0.865], 4 features: 0.90 [0.893; 0.902], 5 features: 0.89 [0.882; 0.893]), both pieces of information (mean classification accuracy and 95% CI: 1 feature: 0.75 [0.742; 0.749], 2 features: 0.82 [0.813; 0.824], 3 features: 0.84 [0.837; 0.847], 4 features: 0.86 [0.850; 0.860], 5 features: 0.87 [0.862; 0.872]), or chance level (mean classification accuracy and 95% CI: 1 feature: 0.50 [0.490; 0.501], 2 features: 0.50 [0.493; 0.504], 3 features: 0.50 [0.494; 0505], 4 features: 0.50 [0.491; 0.502], 5 features: 0.50 [0.490; 0.502]). We observed that both between-network connectivity and the combined data set performed significantly better in terms of accuracy than within-network activity at P < 0.001 when compared to a support vector machine-based on within-network activity alone. A significant difference between the combined data set and the between-network connectivity-based support vector machine could be found for 1, …, 5 selected features at P < 0.05. In summary, the most important information about discriminating awake from loss of responsiveness was contained in the between-network connectivity patterns 5 and 7 as well as in the within-network activity of the medial prefrontal cortex and the dorsolateral prefrontal cortex. Additionally, the separation between awake and loss of responsiveness using a support vector machine had a significantly higher accuracy using the combined data set or the between-network connectivity patterns when compared to within-network activity in specific networks. All approaches perform significantly better than chance level (P < 0.001).

Our analyses clearly illustrate the coexistence of effects both in specific networks of the brain and between-network connectivity during anesthesia-induced unconsciousness. We could demonstrate in line with previous publications on propofol-, sevoflurane-, ketamine-, and dexmedetomidine-induced unconsciousness in humans^3,4,7,8,10  that general anesthesia predominantly affects the within-network activity in the frontal lobe as well as in the medial thalamus during sevoflurane-induced loss of responsiveness.²⁸  Similar effects have been reported in rodent experiments showing a dominant decrease of connectivity in higher-order brain areas when compared to primary sensory areas during anesthesia.²⁹  Nevertheless, this finding should not be interpreted as a localized effect of the agent itself. This would not be a plausible explanation since propofol mainly acts by increasing the effect of γ-aminobutyric acid on γ-aminobutyric acid type A receptors,³⁰  which are known to be widely distributed across the human brain, not only in the above-mentioned brain areas.³¹  Similarly, fluranes are thought to affect multiple types of ion channels, potentiate γ-aminobutyric acid–mediated inhibition, increase potassium permeability, and inhibit sodium channels.³²  We therefore propose that a uniform reduction of connectivity in the brain leads to a disproportionate decrease of within-network activity in specific brain networks whose functioning strongly depends on/results from the tight balancing of incoming excitatory and inhibitor input.

The prominently affected dorsal prefrontal areas have been associated with attentional mechanisms. It has repeatedly been demonstrated that activation in this network is associated with conscious perception and thus reportability of sensory stimuli.^33,34  It has also been proposed in the framework of the global work space theory that the dorsolateral prefrontal cortex might mediate a crucial function: it is supposed to serve as a “bottleneck” for sensory input competing for global distribution in the brain and thus serve both conscious perception and information integration across time and modality.³⁵  Similarly, both the medial prefrontal cortex’s and medial thalamus’s connectivity have been repeatedly associated with altered consciousness during general anesthesia.^3,7,8,10,36  It is therefore plausible that these networks form a crucial set of brain regions for intact consciousness. However, it is still conceivable that some of the affected networks are not a necessary requirement, e.g., the cerebellum. Nevertheless, we conclude that at least a subset of the affected networks is required for conscious processing (as assessed by responsiveness).

Additionally, to these hierarchical changes on the cortical level, the subcortically located medial thalamus showed prominently decreased within-network activity during sevoflurane-induced loss of responsiveness. Other studies using functional magnetic resonance imaging,⁷  positron emission tomography,³⁷  or animal models²⁸  of general anesthesia showed similar results. It is known that the medial thalamus mainly projects to the frontal lobes³⁸  and is thought to play an important role in the modulation of vigilance.³⁹  A view that is also backed by the fact that lesioning of the medial thalamus through stroke in humans results in sudden but usually not permanent onset of coma. We conclude that different anesthetic agents have a different mechanisms of action on the network level that nevertheless impair cortical information integration through a common endpoint.

Beyond effects of general anesthesia on specific brain networks, our data clearly show that anesthesia-induced unconsciousness leads to a significant decrease of cortical dynamics irrespective of the investigated agents. This effect led to a shift of the between-network connectivity patterns to only few (propofol-induced loss of responsiveness) or one single pattern (sevoflurane-induced loss of responsiveness) depending on the depth of anesthesia. Similar observations have been made in rat¹⁷  and monkey¹⁹  brains during general anesthesia. Simulation studies of the brain predicted that a general decrease of coupling strength leads to decrease of functional connectivity repertoire.⁴⁰  Additionally, it has been predicted that damaging of specific brain regions leads to an overproportionate reduction in the brain’s dynamic repertoire.⁴¹  This study clearly demonstrates that both the restriction of functional repertoire and the decrease of activity within functional networks are phenomena associated with loss of responsiveness during general anesthesia, although the causal link between these two phenomena could not be ascertained here. Together with the theoretical predictions made above, we propose that specific brain areas are more important than others to generate a highly dynamic repertoire that is an expression of the wakeful state.

We also want to highlight that specific patterns of between-network connectivity were exclusively observed during wakefulness. These patterns, characterized by high correlation between sensory areas and negative correlation between default mode and attentional networks, have been reported repeatedly.^20,23  The absence of these features has successfully been used to distinguish levels of consciousness in patients with disorders of consciousness,⁴²  and during general anesthesia.⁸  As a consequence, these specific connectivity patterns can be regarded as a modus operandi of the wakeful brain. Similarly, we also observed a pattern of between-network connectivity highly specific for loss of responsiveness. This pattern was characterized by low correlation between functional networks’ time courses. This observation has the implication that even though only specific brain networks’ within-network activity is significantly diminished, the between-network communication is substantially reduced even between networks that show no significant reduction of within-network activity. Therefore, we conclude that specific brain networks’ within-network activity is an important factor influencing between-network communication in distant brain regions.

Of course, the experimental setup and performed analyses presented here have several limitations. It is conceivable that a subset of the effects observed during general anesthesia are unspecific, i.e., changes in brain areas’ activity or changes of between-network connectivity patterns might reflect epiphenomena that are independent of the loss of consciousness itself. As an example, we also observed changes in cerebellar within-network activity that are in line with reports of a diminished metabolic rate.³⁷  However, absence of the cerebellum is not incompatible with consciousness.⁴³  These unspecific effects might be further emphasized by an increase of the anesthetic agents’ concentration beyond the point where loss of responsiveness is actually achieved. From the technical perspective, it is, however, very difficult to obtain functional magnetic resonance imaging data without relevant movement artifacts at this point, a prerequisite for the very complex analysis presented here. In addition, the inclusion of an independent data set from the University of Liege (Liege, Belgium) resulted in stable results both in terms of dynamics of the between-network connectivity and in terms of the patterns encountered. This further underlines that the findings presented here are both reproducible and independent of details of the anesthetic protocol used.

One should also consider that the abstract methodology presented here complicates the interpretation of the results. Nevertheless, there are several arguments why the mean z-score reflects “within-network activity”: (1) mathematically the independent components quantify the spatial extension of the contribution to overall signal, (2) the decrease seen during loss of responsiveness in specific networks is in line with previous experiments using different analytical techniques of functional magnetic resonance imaging data^4,6  or other modalities like positron emission tomography,³⁷  and (3) independent components classified as artifacts did not show a significant change during loss of responsiveness. Likewise, we propose that the correlation of the associated time courses is a measure of “between-network connectivity” since (1) the time courses represent a measure of the temporal modulation of a given network’s contribution to the overall signal, (2) the concept of temporal correlation is a basic principle in functional magnetic resonance imaging to identify functional connectivity, and (3) the encountered patterns have high similarity with patterns from the literature, and the patterns were reproducible in our data set across subjects from different anesthetic protocols and different scanners. In our opinion, the identification of global connectivity patterns that are relevant for consciousness and the concomitant high number of potentially participation brain regions justify this step of data reduction. However, it should also be mentioned that the approach using k-means clustering has one key flaw: the actual alpha error of this step cannot be quantified directly, and hence we encourage any effort to analyze similar data sets using the same method.

The Munich experiments were funded by the Departments of Anesthesiology, Neurology, and Neuroradiology of the Klinikum rechts der Isar of the Technical University Munich. The Liege experiments were supported by the Belgian National Funds for Scientific Research (Brussels, Belgium), the European Commission (Brussels, Belgium), the James McDonnell Foundation (Saint Louis, Missouri), the Mind Science Foundation (San Antonio, Texas), the French Speaking Community Concerted Research Action (Brussels, Belgium), the Fondation Medicale Reine Elisabeth (Brussels, Belgium), the University of Liege (Liege, Belgium), and the University Hospital of Liege (Liege, Belgium).

The authors declare no competing interests.