Anesthesia-induced Suppression of Human Dorsal Anterior Insula Responsivity at Loss of Volitional Behavioral Response

UNDERSTANDING how the brain generates and maintains consciousness is a major challenge in clinical neuroscience. Delivery of anesthesia during functional neuroimaging, therefore, offers a powerful tool to investigate the neurobiology underlying consciousness. As anesthetic drug dose is increased, individuals transfer through a continuum of states from alert wakefulness to deep unconsciousness. A key feature in this continuum is loss of behavioral responsiveness (LOBR) to stimulation, where individuals are no longer willing or able to interact with their external environment. Loss of responsiveness is often used to define anesthesia-induced loss of consciousness in both clinical and experimental settings,¹  primarily due to the fact that this behavioral endpoint is often the only outwardly observable metric.

The advent of functional neuroimaging of anesthesia, in addition to the vast historical supporting evidence, has highlighted the fact that the loss of responsiveness and loss of consciousness are not necessarily equivalent states (see Ref. 2 for a review). Neuroimaging of the anesthetized resting brain has produced conflicting evidence with regards to the sequence and degree of network connectivity disruption in the large-scale thalamocortical and corticocortical networks observed around LOBR and deeper levels of anesthesia.^3–8  Recent work from our laboratory exploring stimulus-evoked rather than resting changes using simultaneous electroencephalography and functional magnetic resonance imaging (FMRI) has demonstrated that thalamocortical processing of multisensory stimuli persists past the point of LOBR up until slow-wave activity saturation.⁹ 

It has been postulated previously that a small cortical region could underpin the LOBR during anesthesia.^1,2  However, no study has yet explicitly shown what brain regions govern this earlier and clinically measurable loss of behavioral response. Identifying the brain regions responsible for LOBR is crucial to understanding the subsequent changes in brain function observed at higher doses of general anesthesia and has important implications for treating patients with disorders of consciousness.

Herein, we present three analyses targeted to identify changes in neural processing specifically associated with the LOBR transition during propofol anesthesia. First, we present a whole-brain FMRI analysis to identify stimulus-evoked changes to laser pain, auditory tones, and word stimuli around loss of responsiveness. We hypothesize that any brain regions demonstrating reduced activity across the LOBR transition to all of the multisensory external stimuli presented will represent key brain regions underlying the loss of volitional behavioral responsiveness during general anesthesia.

Having identified the dorsal anterior insula cortex (dAIC) a posteriori as a potential cortical gate that underlies LOBR, we hypothesize that due to its extensive structural and functional connectivity,^10,11  suppression of dAIC activity could act as the seed for the subsequent breakdown of frontoparietal feedback^6,12  and loss of information integration^13,14  observed at deeper levels of anesthesia. Therefore, we present a whole-brain FMRI analysis to identify changes in the functional connectivity of the dAIC around LOBR. Finally, we demonstrate that these reductions in the functional connectivity of the dAIC can be observed as reductions of electroencephalography synchrony between frontoparietal electrodes, corresponding to the dorsolateral prefrontal cortex (dLPFC) and inferior parietal lobule (IPL) regions, identified by the FMRI connectivity analysis.

Healthy volunteers participated in two neuroimaging experiments; the first experiment was a laboratory-based electroencephalography study, and the second was a simultaneous electroencephalography and FMRI study. Both studies were approved by the Local Research Ethics Committee (Oxford Research Ethics Committee B, Oxford, UK) and performed in the same subjects between October and December 2009. The same experimental paradigm was used for both studies and consisted of data acquisition in four phases (fig. 1A). The volunteers experienced a resting period with eyes closed and no drug administration for 10 min (phase 1), followed by an ultraslow induction to loss of consciousness using propofol sedation (phase 2). A target-controlled intravenous infusion of propofol was used with step increases of 0.2 μg/ml to achieve a maximum effect-site concentration (ESC) of 4 μg/ml over 48 min. After resting at the peak propofol dose for 10 min (phase 3), the propofol sedation was switched off, and subjects were allowed to emerge naturally from unconsciousness over 48 min (phase 4).

Noxious laser stimuli, computer-generated tones, and auditory word tasks were presented to the subjects during the induction and emergence phases (i.e., phases 2 and 4). Loss of appropriate motor response to the cognitive auditory word task was used to assess loss of behavioral response (LOBR) in the healthy volunteers. No behavioral response was sought after the laser or tone stimuli. The data presented here correspond to the induction to loss of consciousness (phase 2). This is a targeted analysis of the imaging data previously presented.⁹ 

Three blocks of sensory stimulation were delivered using Presentation software (www.neurobs.com) during the induction to loss of consciousness phase. Each block of 16-min duration contained interleaved stimuli presented in a pseudorandomized order with approximately one word task, one computer-generated tone, and two noxious laser stimuli per minute. The mean interstimulus interval (ISI) and SD for all stimuli was 15 ± 2 s (range, 12 to 19 s).

All auditory stimuli (tones and word tasks) were presented binaurally using magnetic resonance compatible electrostatic headphones (MRC Institute of Hearing Research, Nottingham, UK). Auditory volume adequacy was checked before each experiment and while the scanner was running by playing sample stimuli.

The word discrimination task was based on the task used by Gaab et al.¹⁵  Subjects were presented with pairs of words and then asked to perform a simple decision-making task by responding, using a two-option button box, whether the words were the SAME or DIFFERENT. The word pairs were presented with mean ISI of 56.8 s (range, 16–104 s). Each stimulation block contained eight SAME and eight DIFFERENT word pairs. The words were selected at random from a list of 200 single syllable words from the MRC Psycholinguistics Database (Machine Usable Dictionary v2.0, Informatics Division, Science and Engineering Research Council, United Kingdom). These words had a familiarity of 488 ± 99 (mean ± SD) and concreteness of 438 ± 120 (mean ± SD) and were recorded by a male actor using Audacity software (http://audacity.sourceforge.net). Once a word stimulus had been presented, it was not used again.

Computer-generated tones (1 kHz, 60 ms) were presented with a mean ISI of 60 s (range, 16 to 92 s).

Brief laser stimuli (5-ms duration) were applied to provide an acute noxious stimulus that selectively activated the A-delta and C nerve fibers. Nd:YAP laser stimuli (STIMUL 1340, Elen Engineering, Italy) of wavelength 1.34 μm were applied to the right lower leg with a mean ISI of 30 s (range, 13 to 48 s) at a subjective intensity rating of 5 of 10. The energy required to achieve this subjective rating (anchored by 0 = no sensation, 1 = just painful, and 10 = most intense sensation) was established in a thresholding session before scanning. To avoid skin damage and nociceptor sensitization or habituation, the site of stimulation was moved after each laser stimulus within a marked 6-cm² area. For both experiments, the subject and investigators wore protective goggles. Numerical ratings for pain intensity were sought after the laser stimulation during the laboratory electroencephalography session only.

A 1% propofol solution (Diprivan, AstraZeneca UK Limited, UK) was administered intravenously in the left forearm using a pump (Alaris Asena PK, Cardinal Health, UK) and target-controlled infusion module (Diprifusor, AstraZeneca UK Limited). This incorporates the Marsh pharmacokinetic model and a pharmacodynamic delay derived from behavioral and electroencephalography parameters. During the propofol-induced loss of consciousness in phase 2, the target ESC was increased by 0.2 μg/ml every 2 min for 38 min and then continued for a further 10 min while the ESC attained the maximum target dose of 4.0 μg/ml. The target ESC of 4.0 μg/ml produced deep sedation in all volunteers, even accounting for the interindividual variation in these pharmacodynamic and pharmacokinetic parameters.

An anesthesiologist was responsible for the administration of propofol and subject physiological monitoring. The anesthesiologist did not have any other study responsibilities and stopped the experiment if they had any concerns over the safety of the volunteer. Oxygen saturation, electrocardiography, respiratory rate, and respiratory depth were recorded electronically (MP150, BIOPAC, USA). Subjects wore nasal cannulae that allowed simultaneous delivery of oxygen at 2 l/min, and capnographic measurement of expired carbon dioxide was performed (GasAnalyser module, BIOPAC). Noninvasive blood pressure was recorded manually.

Healthy volunteers between the ages of 18 and 50 yr were recruited by local advertising. All participants gave written informed consent and received financial compensation for study participation. Subjects were medically screened during an initial visit to ensure they met the American Society of Anesthesiologists physical status grade I or II and were not susceptible to airway obstruction. Potential participants were excluded for history of tobacco or illicit drug use, high alcohol intake (more than 14 units/wk), allergy to anesthetic drugs, difficult venous access, contraindications to magnetic resonance imaging (MRI), and psychiatric, neurological, or psychological pathology.

Participating volunteers were given written instructions that followed the day case anesthesia guidelines from the Association of Anaesthetists of Great Britain and Ireland specifying fasting, travel, and supervision standards required to ensure their safe participation. Experiments were carried out in the morning to minimize the discomfort of fasting.

As there is limited literature on stimulus evoked blood oxygen level dependent (BOLD) imaging during anesthesia, at this stage, the number of participants was selected based on our publications and previous experience, alongside this literature.^6,16–19  We chose a subject number equal to or in excess of these reported studies to provide a sufficient sample size to test our hypothesis. Consequently, we recruited 16 volunteers into the study initially.

Whole-brain BOLD FMRI images (resolution 3 × 3 × 3.5 mm, relaxation time = 3 s) were obtained using an echo planar imaging sequence on a 3-T MRI system with a 1 m bore magnet (Oxford Magnet Technology Ltd, United Kingdom), a birdcage radio frequency coil and a reduced bore head gradient coil for pulse transmission (Magnex SGRAD MKIII, Magnex Scientific Ltd, United Kingdom), and a four-channel helmet head coil for signal detection (PulseTeq, United Kingdom). The system was controlled by a Varian Unity Inova console (Varian Medical Systems, USA) using Siemens' gradients (Siemens Medical Ltd, United Kingdom).⁹  Each induction scan consisted of 964 volumes (including four dummy volumes). High-resolution 1-mm³ T1-weighted anatomical scans were also acquired for coregistering the individual volunteer scans to standard stereotactic space. During MRI data acquisition, cushions were used to support the subjects’ head and neck, maintaining a hands-free “jaw-thrust” maneuver to ensure airway patency when in the supine position and also minimizing head movement.

Large artifacts are induced in the electroencephalography data by the FMRI gradients and ballistocardiographic effects during data acquisition inside the scanner.^20,21  As we were interested in subtle changes in spatial correlations based on the fluctuations in the power spectrum, we analyzed the artifact-free electroencephalography data that had been collected in the first laboratory session in the same subjects rather than the electroencephalography simultaneously acquired during MRI data acquisition. The experimental protocol for both of these sessions was identical with the exception of the acquisition of numerical ratings of pain intensity of the laser stimuli.

Electroencephalography data were acquired using a 32-channel electroencephalography cap (BrainCap MR, Easycap GmbH, Germany) and MR compatible amplifier system (MRplus; Brain Products GmbH, Germany) at 5-KHz sampling rate using FCz as a reference electrode. An abrasive electrolyte conducting gel was used between the electrodes and scalp to ensure electrode impedances were kept below 5 kΩ. Filtering (high-pass filter = 0.5 Hz, low-pass filter = 70 Hz, notch = 50Hz) was applied by the acquisition software (BrainVision Recorder, version 1.10; Brain Products GmbH), which also recorded the timings of the stimuli presentation and subjects’ button presses. Electrocardiographic signals and vertical/horizontal electrooculograms (VEOG/HEOG) were simultaneously recorded through an auxiliary device (BrainAmp ExG MR, Brain Products GmbH, Germany) for offline removal of the blink artifacts.

Consistent loss of appropriate motor response to the cognitive word task was used to define each individual’s LOBR and allow the associated changes in the neural processing of multisensory stimuli to be identified. A consistent loss of motor response was defined as two consecutive missed responses to the cognitive word task in both the FMRI and the electroencephalography analyses.

FMRI–BOLD data were analyzed using Functional Magnetic Resonance Imaging of the Brain’s (FMRIB) Expert Analysis Tool (FEAT) version 6.0, part of the FMRIB Software Library version 6.0 (http://www.fmrib.ox.ac.uk/fsl; Oxford Centre for Functional MRI of the Brain, United Kingdom).²²  Automated removal of nonbrain tissue was performed using Brain Extraction Tool (FEAT v5.98; Oxford Centre for Functional MRI of the Brain), with further manual correction (if required) to remove any artifacts introduced by the presence of the electroencephalography electrodes. FMRI–BOLD data were truncated around each individual’s LOBR to include a total of 24 laser pain stimuli in the time window of interest; i.e., 12 stimuli before and 12 stimuli after each individual’s first consistent missed response to the word task (see fig. 1B). Data were trunctated to start one whole-brain acquisition (or volume) before the time of delivery of the first laser stimulus in the time window of interest and finish one volume before the onset of the stimulus after the last (i.e., 24th) laser stimulus. This equated to an FMRI–BOLD data run of approximately 12 min duration for each subject.

After truncation of the echo planar imaging data, data were preprocessed by performing motion correction with Motion Correction FMRIB’s Linear Registration Tool, spatial smoothing using a Gaussian kernel of 5 mm, global intensity normalization, and applying a high-pass temporal filter of 50 s. The resulting functional scans were registered to each individual’s T1 high-resolution structural image using Boundary-Based Registration²³  and then to the Montreal Neurological Institute (MNI) standard brain using FMRIB’s Non-Linear Registration Tool. As the final stage of preprocessing, Multivariate Exploratory Linear Optimized Decomposition into Independent Components was used to identify cardiac, respiratory, movement, and scanner artifacts in the FMRI–BOLD data. FMRIB’s independent component analysis–based Xnoiseifier v1.06^24,25  was used to automatically classify these components, and subsequently these nonsignal components were removed from the data.

We performed an event-related FMRI analysis on a narrow time window that was centered on each individual’s LOBR (see fig. 1B). The denoised data were analyzed using a two-level general linear model (GLM) to identify the regions of brain activity associated with the laser, tone, and word task stimuli. Six regressors were included at the first level to account for each stimulus type pre- and post-LOBR. Each of the stimuli were modeled to be 1 s in duration and were convolved with a γ hemodynamic response function with mean lag = 6 ± 3 (SD) s. Further confound regressors were used to account for the timing of each individual’s motor responses and any subject motion during the induction to loss of consciousness. Temporal filtering and temporal derivatives were also applied to account for intersubject and interarea differences in the hemodynamic response function.

This is an event-related design, so the remaining unmodeled time serves as the baseline period. As the first-level model depends on the timing of each individual’s LOBR, each model is different, and there is some variability in number of events across volunteers. The variation of the percentage signal change (and SD) required in the first-level contrasts for the pre-LOBR > post-LOBR contrasts for the laser, word, and tone stimuli was 1.58 ± 0.14%, 2.52 ± 0.44%, and 2.04 ± 0.1%, respectively, as estimated in FEAT.²⁶ 

At second level, a mixed-effects analysis was performed across subjects using FMRIB’s Local Analysis of Mixed Effects to identify the main effect of each regressor and any differential processing of these stimuli across the LOBR transition. This higher level estimation method in FEAT estimates the higher level parameter estimates and mixed effects variance using a two-stage estimation technique that includes Metropolis–Hastings Markov Chain Monte Carlo sampling method. Cluster-based thresholding was used in FEAT to reveal significant group-level brain activation that was corrected for multiple comparisons across space. A Z-statistic cluster threshold of 2.3 was used to define contiguous clusters, and then each cluster’s estimated significance level (from Gaussian random field theory) was compared with a cluster probability threshold of P < 0.05. For visualization, the thresholded clusters in MNI space were applied to the very inflated surface from the Human Connectome Project Workbench tutorial dataset Beta version 0.51 (www.humanconnectome.org; WU-Minn Consortium, USA).²⁷ 

A conjunction analysis was performed to identify any brain regions that demonstrated significant reductions in stimulus-evoked activity across the LOBR transition that were common for all stimulation types. The thresholded Z-statistic maps for each of the differential laser, words, and tones stimuli contrasts (i.e., pre-LOBR > post-LOBR contrasts) were binarized and then multiplied together. This created a mask in standard space that corresponded to where activity was commonly reduced across the LOBR transition in response to all of the multisensory stimuli. Again, for visualization, the brain regions identified by the conjunction analysis in MNI space were resampled onto the very inflated surface from the Human Connectome Project Workbench tutorial dataset.

For demonstration purposes, the average stimulus-evoked percentage BOLD signal change within the region identified by the conjunction analysis was calculated for each type of stimulus pre- and post-LOBR. First, the conjunction mask in standard space was transferred back into each individual’s subject space. For each individual, the average stimulus-evoked BOLD time series within this mask were temporally resampled using cubic spline interpolation and then averaged across trials and subjects for each stimulus type.

A two-level GLM analysis was performed to identify changes in the functional connectivity of the dAIC cluster (identified in the previous conjunction analysis) across the LOBR transition. At the individual subject level, separate event-related GLMs were set up for the time periods before and after the loss of behavioral response. These models included regressors for the timing of (1) the delivery of all stimuli, (2) the motor responses where appropriate, and (3) motion regressors. These first two regressors were used to account for the original task within the functional connectivity analysis. Consequently, they were convolved with a γ hemodynamic response function with mean lag = 6 ± 3 (SD) s. Temporal filtering and temporal derivatives were applied as per the previous analysis described in the FMRI General Linear Modeling section. Six motion regressors were also included to account for any confounds of subject motion during the pre- or post-LOBR periods.

The final regressor included in the first-level analyses was the mean BOLD time series within the dAIC cluster for the time period of interest, i.e., either pre- or post-LOBR. This regressor was orthogonalized to the stimuli regressor and the response regressor where it was present. This regressor represents the main regressor of interest as it identifies any voxels within the brain that demonstrate altered functional connectivity with the dAIC that are not already explained by activation due to the stimuli or responses.

At second level, a mixed-effects paired t test analysis was performed across subjects using FMRIB’s Local Analysis of Mixed Effects to identify any changes in the dAIC’s functional connectivity around the LOBR transition. The individual dAIC whole-brain connectivity contrasts for pre- and post-LOBR periods for the n = 15 subjects were used as input to the model, resulting in 30 inputs in total at second level. The higher level GLM was set up with one regressor to identify the change in dAIC’s functional connectivity across the LOBR transition and 15 further regressors to control for each subject’s mean effect. Cluster-based thresholding (Z = 2.3, P < 0.05) and multiple comparisons correction across space was used, as described in the FMRI General Linear Modeling section.

To corroborate our FMRI findings, we examined changes in frontoparietal electroencephalography synchrony at electrode locations corresponding to the brain regions that had been identified to have reduced functional connectivity with the dAIC around LOBR. As the BOLD signal typically detects changes in activity over time courses of perhaps 3 to 15 s, we sought to use electroencephalography methods that varied over a comparably slow time course. With the proviso of some distortion by the filtering effects of the skull, scalp, and electrode interface, the peaks of the slow electroencephalography oscillations tend to reflect cortical pyramidal cell depolarization and high neuronal activity, whereas the electroencephalography troughs indicate hyperpolarization and relative inactivity.

The main problem with direct current recording methods is that of low-frequency electrode drift. However, the slow modulation of electroencephalography activity can be detected indirectly by examining the amplitude envelope of fluctuations in the power of the traditional electroencephalography waves.²⁸  The modulation of the envelopes of the lower frequencies has been shown to be a reasonable predictor of fMRI connectomes,²⁹  and, thus, might be reasonably considered to be an index of fluctuations in cortical activity, somewhat analogous to those seen in the BOLD signal. We used the electroencephalography data that was recorded during an identical propofol administration protocol outside the scanner in the same subjects (bench electroencephalography data). The bench electroencephalography has significantly lower noise levels than the scanner electroencephalography data as is required to detect subtle changes in long-range correlations. Although most published methods use a bandpass-filtered Hilbert envelope method, we found that a short-time Fourier transform was a simpler method of quantifying the slow fluctuations in power over about 3 to 15 s.

To spatially localize the electroencephalography signal, we used a nearest neighbor modified Hjorth Laplacian derivation. This involved subtracting the mean values of the surrounding three electrodes from the electrode that overlay the regions of interest. We then compared the changes in synchrony between the electrode lying over the right dLPFC (Fp2) and the right inferior parietal regions (P8). As a control site, we also examined the synchrony between electrodes overlying regions of the cortex that were close to, but distinct from, the regions of interest; namely by comparing the synchrony of the dLPFC (Fp2) and the midline parietal cortex (Pz).

Data were exported from BrainVision Analyser Version 2.0 (Brain Products GmbH) and analyzed in MATLAB (Mathworks Inc., USA). Electroencephalography data were downsampled to 125 Hz and truncated to examine a period of 6 min before until 6 min after the time of LOBR for each subject. For each of the electroencephalography channels, we used the short-term Fourier transform “spectrogram.m” function (window = 4 s, overlap = 3 s, 0.25 Hz resolution) to obtain the mean electroencephalography power (expressed in decibels) between 0.5 and 45Hz. The synchrony between the two regions was quantified by using a simple zero-lag Pearson cross correlation coefficient (r) applied to 60-s sections of slow-wave power (0.5 to 1.5Hz). Statistical comparisons were done using paired t tests on the mean values for the 6 min before the LOBR with the mean of the 6 min after LOBR.

Sixteen subjects participated in the bench electroencephalography and the simultaneous electroencephalography and FMRI sessions. For one subject, the induction to loss of consciousness was stopped as the participant developed an obstructive respiratory pattern during the simultaneous electroencephalography and FMRI session. For another subject, FMRI data acquisition was interrupted as the participant was removed from the scanner and repositioned once the safety of the individual was reestablished. Data from the first FMRI acquisition were sufficient to perform this analysis as the subject had lost responsiveness several minutes earlier. Consequently, FMRI and bench electroencephalography data are presented for the same 15 healthy volunteers (8 women, age 19 to 43 yr, mean age ± SD of 28.7 ± 7.3 yr).

The loss of an appropriate motor response to an auditory word discrimination task was used to define LOBR. In the scanner, the LOBR was abrupt in 14 of the 15 volunteers and the mean estimated propofol ESC at which it occurred was 1.27 μg/ml with 95% confidence limits of 0.92 to 1.62 μg/ml. The mean propofol ESC ± SD increase for all subjects across the narrow LOBR time window was 1.1 ± 0.09 μg/ml with a range of 1.0 to 1.2.

Similarly for the bench electroencephalography data, the mean ESC ± SD at which LOBR occurred was 1.53 ± 0.6 μg/ml with 95% confidence limits of 1.20 to 1.86 μg/ml. This was equivalent to a propofol ESC increase for all subjects across the 12-min window used for the electroencephalography analysis of 1.2 μg/ml.

The FMRI analysis on a narrow time window (of approximately 12 min in total) that was centered on each individual’s LOBR allowed us to identify specific changes in the processing of auditory tones, words, and noxious pain stimuli associated with this transition. This window length was used to provide a compromise between the number of events required for adequate signal to noise in the FMRI findings and also to limit the temporal window of investigation so that we can still observe the distinct state change that occurs at LOBR. By examining stimulus-evoked FMRI–BOLD changes, we demonstrate that propofol anesthesia did not disrupt the stereotypical perceptual activation patterns around the LOBR transition (fig. 2). For example, by examining the mean effects before and after LOBR separately in response to laser stimulation, we found that widespread thalamocortical processing persisted after LOBR. Activity was observed after LOBR in response to laser stimulation in brain regions known to be responsible for conscious pain perception, including the thalamus and the insula, anterior cingulate, and somatosensory cortices. Similarly, thalamocortical processing persisted past LOBR for both types of auditory stimuli with sensory-specific perceptual activation patterns seen for word and (to a lesser degree) tone stimuli.

In addition, a comparison of the mean stimulus-evoked activity before and after each individual’s loss of behavioral response (i.e., contrasting pre-LOBR and post-LOBR periods) showed significant localized reductions across the transition (fig. 3; table 1). In contrast, no brain regions were found to be more active to any stimuli after LOBR than before (i.e., the post-LOBR > pre-LOBR contrasts).

By performing a conjunction of these contrasts across all sensory modalities, we demonstrated that propofol suppressed activation in a specific region of the right dAIC after LOBR (fig. 4A). For demonstration purposes, the average time-locked BOLD percentage signal change for each stimulus type within this dAIC region is presented for the time periods before and after LOBR (fig. 4B) to confirm the suppression of stimulus-evoked activity.

A functional connectivity analysis was performed to identify changes in functional connectivity of the dAIC cluster across the LOBR transition. Reductions in the functional connectivity of the dAIC were observed in the dLPFC, IPL, and cerebellum after loss of responsiveness (fig. 5; table 2). No brain regions were found to increase their functional connectivity with the dAIC after the LOBR transition.

We found that the synchrony between the electrode pair that most closely overlaid the dLPFC and the IPL (Fp2 and P8) decreased significantly around the point of LOBR as predicted by the functional connectivity results (P = 0.007, t test; fig. 6). The synchrony between the Fp2 and Pz electrodes, which are separated by a similar distance, but correspond to the nearby regions of the midline parietal cortex, did not show a significant decrease (P = 0.26, t test).

In this article, we have focused on the changes in neural activity associated with LOBR. Our results suggest that the dAIC could be a key cortical region that underpins LOBR under anesthesia. We found that activity within this brain region was commonly reduced in response to both painful and two types of auditory stimuli around the point of LOBR in healthy volunteers (fig. 3). Identification of anterior insula suppression at LOBR is perhaps unsurprising given that the insula is the most frequently reported area of activation across all neuroimaging studies.³⁰  Furthermore, there is a myriad of work postulating the dorsal anterior insula division specifically to be the embodiment of the “sentient self”³¹  and a critical hub of a salience/threat detection network along with the anterior cingulate cortex.³² 

Recent meta-analyses^10,33  have also suggested the region of the dAIC identified in our conjunction analysis to be in the cognitive rather than sensorimotor domain, supporting the idea that suppression of dAIC activity is related to a loss of volition/willingness rather than the ability to produce motor responses. We should note that subjects in the study were instructed to respond only after the auditory word stimuli were presented; therefore, any loss of responsiveness does not necessarily indicate an inability to move. Indeed, examination of the motion correction data indicates that many subjects continued to move slightly after they had lost their volitional response.

We hypothesized that suppression of activity in a cortical gate around LOBR (such as the dAIC) could act as the seed for the subsequent breakdown of frontoparietal feedback and the loss of information integration seen at deeper levels of anesthesia. Subsequently, we have shown using FMRI that dAIC’s functional connectivity to frontoparietal regions is reduced after LOBR (fig. 5). Specifically, we found that the dLPFC and IPL had reduced functional connectivity with the dAIC. Interestingly, we also observed a reduction in dAIC’s functional connectivity with the cerebellum around LOBR. The center of gravity for this cluster was located in the Crus I region, which is known to be involved in executive functions through its functional connectivity with prefrontal regions of the brain.^34,35 

The dLPFC, identified in the functional connectivity analysis, forms part of the executive control network and is highly implicated in working memory and decision-making. The IPL cluster we identified covers the PF, PFt, and PM divisions defined by Casper et al.,³⁶  with its center of gravity within the PM division. These regions of the IPL have been shown to have the strongest interactions with dLPFC through a combination of diffusion tractography–based parcellation and human resting–state data analyses.³⁷  Interestingly, the PM region of the IPL has also been shown to be more active when subjects switch their response strategy.³⁸  This fits well in the context of our experimental data where the subjects’ loss of responsiveness was defined according to a cognitive auditory word task where subjects were being asked to choose between two options (i.e., when SAME or DIFFERENT pairs of words were presented).

Our finding that dAIC suppression at LOBR is associated with altered stimulus-evoked functional connectivity in frontoparietal regions fits well with previous findings and global theories of consciousness that postulate that loss of consciousness (or often more correctly loss of responsiveness) is based on the disruption of frontoparietal communication.^{3,6,12,39–41}  Further support for this being (either directly or indirectly) mediated by suppression of the dAIC is added through our electroencephalography power synchrony analysis (see fig. 6). Although electroencephalography techniques have poor spatial resolution compared with that of FMRI, these results, however, suggest that loss of long-range synchrony around the point of LOBR is greatest between these relatively localized regions identified by the functional connectivity analysis (i.e., the IPL and dLPFC).

Due to anatomy, it is difficult to identify any localized changes in dAIC activity around LOBR using scalp electroencephalography without using source localization. However, the fact that these electroencephalography data reproduce our dLPFC and IPL findings gives us additional confidence in the widespread applicability of our results, especially as these data were collected in the same subjects but in a different recording session. We also note that these effects could be seen using a medium-density scalp montage and did not require any source localization. It is also important to note that the shape of the curves in figure 6 is suggestive of an abrupt decrease in synchrony around LOBR. However, because we were examining correlations between power fluctuations in the slow-wave frequency band (0.5 to 1.5Hz), the time resolution is limited to around 60 s and limits our ability to assess the true shape of this curve. This resolution is comparable with that with which we are able to assess LOBR, due to the variability in timing of the presentation of the auditory word task.

It has been postulated that the disruption of synchronous activity is indicative of disruption in the interaction between regions and hence linked to a disruption in functional interregional communication.⁴²  Although our methods do not have the temporal resolution to reliably infer directionality or causality, we would suggest that the observed loss of synchrony between these frontoparietal regions potentially indicates failed information transfer between the regions caused by the increasing concentrations of propofol.

Selfhood may be defined as the experience of being a distinct entity that is capable of self-control and attention.⁴³  On the basis of the results presented and the proposed role of the dorsal anterior insula division as the site of the sentient self, we hypothesize that general anesthesia challenges an individual’s selfhood through suppression of the dAIC. We propose that this LOBR transition lies at the boundary of the transitive and intransitive definitions of consciousness,⁴⁴ i.e., where individuals may be conscious but are not necessarily consciously aware that stimuli relate to them; namely, when an individual loses selfhood. A central component of selfhood is the experience of body ownership, which is thought to arise from predictive (top–down) integration of (bottom–up) multisensory information from interoceptive and exteroceptive domains.^43,45  When this is lost, as potentially occurs at LOBR through direct (or indirect) suppression of dAIC activity, there is no need to respond to incoming sensory events because they are not perceived as happening “to me.”

In clinical practice, patients typically lose behavioral responsiveness to external auditory stimuli at concentrations of general anesthetic around 30% of those required for full general anesthesia (as measured by lack of somatic movement to surgical incision). Over the years, there has been indirect evidence of substantial perceptual processing in lightly anesthetized patients; namely, a high incidence of dreaming,⁴⁶  implicit memory formation,⁴⁷  and even explicit recall of intraoperative events.⁴⁸  Patients, and also experimental study participants,⁴⁹  often report a loss of engaged consciousness or disconnection from the environment. Importantly, our dAIC finding is supported by a recent case study in an awake epilepsy patient, which found electrical (rather than pharmacologic) disruption of the dAIC/claustrum region led to a reproducibly disrupted consciousness and LOBR.⁵⁰  In reference to disorders of consciousness patients, the ability to engage with the outside world is ultimately used to discriminate patients as minimally conscious rather than in a persistent vegetative state. Our findings linking the LOBR associated with suppression of activity within the dorsal anterior insula potentially provides the link between the internal and the external awareness networks that have been discussed in the context of these patients.⁵¹ 

Interestingly, we did not observe any change in the activity of the anterior cingulate cortex around the LOBR transition (see figs. 2 and 3; table 1), despite its proposed role alongside the anterior insula as part of the salience detection network.³²  This may be due to limited statistical power, either due to the experimental design and/or the relatively small study sample size. The signal-to-noise ratio in stimulus-evoked FMRI experiments depends on many factors, for example, the number, duration, and intensity of stimulation; the data acquisition parameters; and the analysis methods used. Given the need to remain temporally close to the LOBR transition of interest, the number of word stimuli was potentially close to minimum in our experimental design. The final sample size for our analysis was equivalent to⁶  or in excess of other contemporaneous FMRI experiments performed under anesthesia,^9,17–19  which in some way compensates for the reduced signal-to-noise ratio at an individual level. However, a secondary study with optimal statistical efficiency is required to identify whether other brain regions exist, in addition to the dAIC, which have functional relevance for LOBR but with smaller effect sizes. Further work will also be required to assess whether our current dAIC finding holds for other anesthetic agents in addition to propofol.

In conclusion, our results show for the first time the explicit neurobiology that underlies LOBR under general anesthesia. We conclude that pharmacologic lesioning of the right dAIC by propofol anesthesia is (either directly or indirectly) responsible for the measurable loss of volitional behavioral response seen in clinical practice. We have provided evidence that this suppression of dAIC activity results in reductions of functional connectivity and electroencephalography synchrony in frontoparietal brain regions around LOBR. In summary, the dAIC provides a potential cortical gate for LOBR observed under general anesthesia.

The authors thank Mark Jenkinson, D.Phil. (Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom), and Jesper Andersson, Ph.D. (Nuffield Department of Clinical Neurosciences, University of Oxford), for their advice with data analysis. They also thank the reviewers for their helpful suggestions and comments that have significantly improved the manuscript.

The study was supported by funding from the Medical Research Council of Great Britain and Northern Ireland, Swindon, United Kingdom (to Prof. Tracey, Dr. Ní Mhuircheartaigh, Dr. Jbabdi, and Oxford Centre for Functional Magnetic Resonance Imaging of the Brain [FMRIB], Oxford, United Kingdom); the Wellcome Trust, London, United Kingdom (Prof. Tracey) and the Wellcome Trust through the Scottish Translational Medicine and Therapeutics Initiative (Dr. Seretny); the National Institute for Academic Anaesthesia, London, United Kingdom (Drs. Seretny and Ní Mhuircheartaigh); the International Anesthesia Research Society, San Francisco, California (Dr. Ní Mhuircheartaigh); the Scottish Society of Anaesthetists, Scotland, United Kingdom (Dr. Seretny); and St Cross College, University of Oxford, Oxford, United Kingdom (Knoop Fellowship—Dr. Warnaby).

Patent applications were filed by Isis Innovation, the technology transfer company of the University of Oxford, Oxford, United Kingdom, on perception loss detection. All authors except Prof. Sleigh and Dr. Seretny are listed as inventors. The other authors declare no competing interests.