Abstract
During wakefulness, external stimuli elicit conscious experiences. In contrast, dreams and drug-induced dissociated states are characterized by vivid internally generated conscious experiences and reduced ability to perceive external stimuli. Understanding the physiological distinctions between normal wakefulness and dissociated states may therefore disambiguate signatures of responsiveness to external stimuli from those that underlie conscious experience. The hypothesis that conscious experiences are associated with brain criticality has received considerable theoretical and experimental support. Consistent with this hypothesis, statistical signatures of criticality are similar in normal wakefulness and dissociative states but are abolished in dreamless sleep and under anesthesia. Thus, while statistical measures of criticality are associated with the ability to have conscious experience, they do not readily distinguish between perception of the external world from internally generated percepts. Here, we investigate distinct, dynamical, signatures of criticality during escalating ketamine doses in high-density EEG in human male volunteers. We show that during normal wakefulness, EEG is found at a critical point between damped and exploding oscillations. With increasing doses of ketamine, as dissociative symptoms intensify, activity is progressively stabilized—most prominently at higher frequencies. We also show that stabilization is a more reliable marker of the effects of ketamine than conventional measures such as power spectra. These findings suggest that stabilization of cortical dynamics correlates with decreased ability to respond to and perceive external stimuli rather than the ability to have conscious experiences per se. Altogether, these results suggest that combining statistical and dynamical criticality measures may distinguish wakefulness, dissociation, and unconsciousness.
Significance Statement
During wakefulness, external stimuli elicit sensory perceptions while during unconsciousness, perception is absent. Dissociated states of consciousness, including those induced by ketamine, feature internally generated experiences and, concomitantly, reduced responsiveness to stimuli. Both normal wakefulness and dissociated states have been linked to statistical criticality, a regime in which the brain operates at the transition between order and disorder. Here, we study a distinct notion of criticality—transition between stable and unstable oscillations—and show that ketamine induces dose-dependent stabilization of normally critical brain dynamics. Thus, departure from dynamical criticality is associated with states of reduced responsiveness rather than unconsciousness. Combining statistical and dynamical criticality measures may better distinguish connected and dissociated states of consciousness.
Introduction
From the dawn of human neurophysiology over a century ago, scientists have been trying to understand the relationship between brain activity and consciousness. This has been done by disrupting consciousness using anesthetics (Gibbs, 1937) and correlating brain activity with naturally occurring variations in the level of consciousness during sleep (Loomis et al., 1937). “Paradoxical” or rapid eye movement (REM) sleep (Aserinsky and Kleitman, 1953) and dissociative anesthesia induced by drugs such as ketamine (Bowdle et al., 1998; Domino and Warner, 2010) have emerged as clear outliers. Unlike slow wave sleep and conventional anesthetics which are associated with slowing of brain activity and decrease in cerebral metabolism (Alkire et al., 1995; Bazhenov et al., 2002; Nofzinger et al., 2002; Murphy et al., 2011), REM sleep and ketamine anesthesia are characterized by active EEG and preserved cerebral metabolism (Jones, 1991; Maquet et al., 1996; Akeju et al., 2016; Laaksonen et al., 2018). Perhaps the most fascinating aspect of REM sleep, ketamine anesthesia, and other similarly dissociated states is that they are characterized by vivid sensory experiences unrelated to physical stimuli (Bowdle et al., 1998; Hobson, 2009; Domino and Warner, 2010) even though responsiveness to sensory stimuli is greatly attenuated (Llinás and Ribary, 1994; Hobson, 2009; Andrillon et al., 2016). Thus, there appear to be two fundamentally distinct types of states in which consciousness is possible. During normal wakefulness, sensory experiences are triggered by and correlated with external stimuli, whereas in dissociated and dream-like states, sensory experiences arise spontaneously while responses to external stimuli are suppressed. It is therefore imperative that we identify neurophysiological correlates that track the capacity to have sensory experience and distinguish them from those that track responsiveness to external stimuli.
Previous research on the neurophysiological signatures associated with conscious experiences has focused on experimental manipulation of stimuli (Dehaene and Changeux, 2011), such as using stimuli close to a psychophysics threshold (van Vugt et al., 2018), masking of stimuli to render them imperceptible (Tsuchiya and Koch, 2005; Kouider and Dehaene, 2007), modulation of attention (Simons and Ambinder, 2005), and ambiguous stimuli that elicit spontaneously fluctuating perceptions (Leopold and Logothetis, 1996; Dwarakanath et al., 2023). One of the major conclusions from these studies is that conscious perception of a stimulus is associated with an “ignition”-like event whereby much of the cortical activity is recruited in a coordinated fashion (Mashour et al., 2020). While there are some notable disagreements in the current theories of consciousness on the details of how and why this nonlinear ignition-like event occurs and the terminology used to describe it, the fact that distributed coordinated activation of the cortex accompanies conscious experience is supported by numerous experimental observations (Seth and Bayne, 2022). This naturally leads to a question: Under what circumstances can an initially localized activation by a sensory stimulus lead to the large-scale coordinated neuronal activation associated with conscious experiences?
One theory that has been gaining experimental support (Beggs, 2007; Hesse and Gross, 2014; Cocchi et al., 2017) is that, to allow for the formation of coordinated and informative neuronal firing distributed broadly across the cortex, the brain must be operating close to criticality. Criticality generally refers to behaviors exhibited by a system whose parameters are found near the boundary between two qualitatively distinct behaviors (Mora and Bialek, 2011). One challenge in experimentally testing this hypothesis is that different studies use somewhat different notions of criticality. It is presently not clear whether different signatures of criticality reflect similar underlying processes.
Following the original work by Beggs and Plenz (2003), many studies focused on statistics of neuronal “avalanches.” In physics, systems poised near a phase transition produce power-law distributed avalanches (Bak, 1996). Observations that neuronal activity avalanches are also well described by a power law (Beggs and Plenz, 2003; Mora and Bialek, 2011) have been interpreted as experimental evidence that, during wakefulness, the brain is found near a second-order phase transition (Shew et al., 2009). Nonetheless, some have proposed caution in interpreting the observation of power-law distributions as indications of criticality in the brain (Touboul and Destexhe, 2017). Because this notion of criticality is concerned with statistics of brain activity, we will refer to it as “statistical criticality.”
Changes in the level of consciousness are correlated with decrease in statistical criticality. Under anesthesia, avalanche distributions in human electrophysiological recordings (He et al., 2010; Shriki et al., 2013) deviate from a power law (Maschke et al., 2024). Interestingly, in the ketamine-induced dissociative state, distributions of avalanches are largely wake-like (Maschke et al., 2024). This result was corroborated by using a similar analysis of ECoG signals in primates (Varley et al., 2020). A somewhat related approach that focused on the power-law relationship between spectral power and frequency reached a similar conclusion (Colombo et al., 2019). These results taken together suggest that both normal waking consciousness and dream-like dissociated states are associated with a statistically critical regime while states in which perception is diminished are associated with departure from statistical criticality.
While preservation of statistical criticality may explain why ketamine, unlike conventional anesthetics, preserves the ability to have vivid sensory experienced, it does not explain why ketamine administration diminishes responsiveness to and perception of external stimuli. To investigate the origins of this decreased responsiveness here, we used a distinct, dynamical, notion of criticality (Magnasco et al., 2009; Solovey et al., 2012, 2015; Alonso et al., 2014). From a dynamical systems (Strogatz, 2015) perspective, the response of a system to an external stimulus is governed by its stability. Perturbations in purely stable systems decay and the system quickly returns to its original state. For instance, in the case of EEG oscillations, a stable system will produce damped oscillations in response to an input stimulus. Conversely, in unstable systems, the effect of the perturbation grows without bounds—response to a small input stimulus will be ever growing oscillations. Between these two regimes there is a critical point at which the system can respond to an input stimulus by producing a sustained oscillatory response.
Here, we sought to determine whether the ketamine-induced state, characterized by a preservation of internal experience, coupled with a decrease in responsiveness to external stimuli, is paralleled by departure from dynamical criticality. Importantly, since ketamine's effects are largely dose dependent, we sought to examine a dose–response relationship between ketamine and criticality. To address this question, we performed linear stability analysis of human EEG during exposure to escalating doses of ketamine. Consistent with previous work performed with human and nonhuman primate ECoG (Solovey et al., 2012, 2015; Alonso et al., 2014), we show that during wakefulness human EEG is dynamically critical. In contrast to statistical measures of criticality and spectral features of the EEG signals which are not reliably affected by ketamine, brain dynamics exhibited reliable and dose-dependent stabilization during ketamine administration. We also demonstrate that stability of the EEG can be used to reliably decode the ketamine dose. In contrast, classification based on the spectral features of the EEG or stability analysis performed on EEG surrogates, which preserve statistics of each signal and pairwise correlations, is significantly less reliable. Thus, in combination with previous work on statistical criticality, these findings suggest that stabilization of brain activity may be a feature of states that exhibit a reduced responsiveness to sensory stimuli, whereas statistical criticality may be more closely associated with the capacity to have sensory experience per se.
Materials and Methods
Experimental design and statistical analysis
This study was approved by the University of Pennsylvania School of Medicine's Institutional Review Board (IRB, 829800; clinicaltrials.gov NCT03498391). All study procedures were conducted at the Hospital of the University of Pennsylvania. Written informed consent was acquired from all study participants prior to study procedures. Seven healthy human volunteers (7 males, ages 25–28, weight 150–212 lbs.) were recruited from the Philadelphia community. On the day of the experiment, once the nil per os (NPO) status was verified, the subjects were fitted with a 128-channel EEG system (HydrocCel Geodesic Sensor Net), and a 20 gauge IV was placed. Throughout the experiment, EKG, blood pressure, pulse oximetry, and expired CO2 were monitored. Supplemental oxygen at 2–4 L/min was administered via a nasal cannula. All experiments were performed under direct supervision of two anesthesiologists. All experiments were carried out in a dedicated room located in the Penn Center for Phenomic Science (https://www.med.upenn.edu/chps/) located in the Hospital of the University of Pennsylvania. During the experiment, subjects were resting in a hospital bed. A battery of psychophysics experiments was conducted on the day of the study. The results presented in this manuscript only concern spontaneous EEG collected during a 2 min interval when subjects were instructed to keep their eyes closed.
The primary outcome variable in the present work is distributions of criticality indices (see below, Stability analysis, for a full description of how these distributions are generated). Significance testing was done in a within-subject fashion, comparing distributions of criticality indices between all conditions (0, 0.2, 0.4, and 1 mg/ml ketamine), utilizing the Mann–Whitney U test as implemented in MATLAB's ranksum function. An additional comparison was made between two distributions constructed from the control condition by splitting data therein into two equally sized samples. p values were corrected for multiple comparisons via the Benjamini–Hochberg procedure (Benjamini and Hochberg, 1995) as implemented in the MATLAB function fdr_bh (https://www.mathworks.com/matlabcentral/fileexchange/27418-fdr_bh). Spectral properties of the EEG were also measured and compared between ketamine doses (see below, Spectral estimation, for a full description of how these measures were generated). Statistical testing was done via a bootstrapping analysis further detailed below, Spectral bootstrapping analysis. Classifiers trained on stability and spectra were generated, and their performances were compared (see below, Classifier training/testing and PCA of features, for a full description of how these classifiers are implemented).
Inclusion and exclusion criteria
Inclusion criteria were as follows: American Society of Anesthesiologists (ASA) patient status 1 or 2, negative urine drug screen, no allergies to ketamine.
Exclusion criteria were as follows: hypertension, history of neuropsychiatric disease, history of seizures and/or cerebral aneurysms.
Ketamine infusion
Target-controlled infusion (TCI) utilizes continuously adjusted drug infusion rate to ensure stable predicted plasma concentration (Fig. 1). Ketamine infusion was controlled using STANPUMP (http://opentci.org/code/stanpump) software which implemented Domino's weight-adjusted pharmacokinetic (PK) model (Domino et al., 1984). The infusion was delivered using a Harvard 22 Basic Syringe Pump (Harvard Apparatus). After collecting baseline (no drug) recordings, ketamine was administered in two subanesthetic concentrations (0.2 and 0.4 µg/ml) and a single anesthetic dose (1.0 µg/ml). Resting EEG was selected from a period where the specified blood concentrations were achieved and stable.
Ketamine delivery protocol. Traces detailing the predicted plasma concentration (orange trace, right y-axis) in tandem with the ketamine infusion rate (teal trace, left y-axis) in a representative experiment. “Infusion On” and “Infusion Off” labels, respectively, indicate when the drug infusion started and ended. Blue bars labeled Recording Period indicate times when predicted drug concentration was stable and EEG recording was used for resting state analysis. Infusion rate is continuously adjusted to ensure a stable predicted plasma concentration of ketamine.
EEG acquisition
EEG recordings were acquired using a high-density 128-channel EGI HydrocCel Geodesic Sensor Net, at a sampling rate of 1,000 Hz. Besides the standard antialiasing hardware filter, all additional filtering and processing of the EEG was performed post hoc in MATLAB (MathWorks).
EEG preprocessing
EEG was processed using a combination of EEGLAB (Delorme and Makeig, 2004) functions and in-house code. First, raw EEG was imported into MATLAB via the EEGLAB package (https://eeglab.org/). Channels not located on the scalp, such as those placed over the mastoids, were then excluded, leaving 82 channels in the final dataset. One continuous minute of artifact-free data at each dose was identified by visual inspection and selected for further preprocessing and analysis. A bandpass fourth-order Butterworth filter with cutoff frequencies at 0.5 and 200 Hz was applied to the data via zero-phase setup as implemented in MATLAB's filtfilt function to avoid phase distortions. Bad channels were automatically identified as those whose spectra exceeded 2 standard deviations from the mean of all channels using EEGLAB's pop_rejchanspec function and interpolated using EEGLAB's eeg_interp function via spherical spline technique. Line noise centered at ∼60 Hz and its harmonics were removed using CleanLine (https://github.com/sccn/cleanline) plugin for EEGLAB. The remaining artifacts were removed by applying independent component analysis (ICA) to the data, and using EEGLAB's ICLabel (Pion-Tonachini et al., 2019; https://github.com/sccn/ICLabel) plugin to identify components associated with nonbrain sources. Finally, average rereferencing was applied to the data. One subject was excluded from analysis due to persistent electrical EEG artifact, leaving a total of six subjects.
Spectral estimation
Once data were processed as above, spectral power was estimated utilizing Welch's method (Welch, 1967) as implemented in EEGLAB's spectopo function (which itself calls upon MATLAB's pwelch function from the Signal Processing Toolbox to perform spectral estimation). To produce data in Figure 3B, the distribution of power at each location and frequency band was estimated in 2 s windows with no overlap for each subject during the baseline (no drug) period. Mean and standard deviation of power was then computed across windows. Then, the same windows were used to estimate the spectra under all ketamine concentrations and the power was expressed in units of z-score relative to the same subject in the baseline state. The average of these z-scores across time and subjects is shown in Figure 3B.
When constructing training/testing data for classifiers, further discussed below, spectra were estimated in 0.5 s time segments with no overlap.
Spectral bootstrapping analysis
To test spectral changes associated with ketamine administration, we conducted a bootstrap analysis wherein for 100 iterations, a random 5 s segment of the data was selected at each drug concentration, and the difference between the spectral power during baseline and that under each dose of ketamine was calculated. A group of six channels was chosen from each of the following regions: frontal, occipital, temporal, and parietal. Mean deviation from baseline and 95% confidence intervals (averaged across the six channels in each region) were computed for each ketamine dose and each region.
Stability analysis
Our basic objective here is to estimate how the EEG signal will change after a small perturbation. A well-established method for doing this is to linearize the dynamics—approximate the system by a linear differential equation (in a short time window). For EEG this corresponds to fitting a linear autoregressive model.
The dynamics given by the autoregressive matrix A can be understood through its eigendecomposition (see below, Relationship between eigenvalues and dynamics, for intuitive explanation):
Because A is generally not symmetrical, most of its eigenvalues are complex-valued and can be written in the exponential form
Relationship between eigenvalues and dynamics
Equation 1 is a numerical approximation to the first time derivative of the EEG. A snapshot of the EEG at a particular moment is a vector of voltages (one element for each of the channels). The first-order autoregressive matrix A approximates how this vector will change after one time step. In general, A will stretch (or shrink) and rotate this vector in a complicated way making the overall behavior of the system difficult to track. It is therefore desirable to separate the overall dynamics captured by A into a set of linearly separable components. This is accomplished by eigendecomposition of A. Specifically an eigenvector ϕ is a vector that is only stretched (or shrunk) but not rotated by the matrix:
Illustration of EEG signal decomposition into eigenmodes. Five-hundred millisecond window of four channels of hypothetical EEG signal is shown (Simulated EEG). These data were generated by an autoregressive model. The power spectrum below traces shows the average across channels. For the sake of illustration, the EEG contains two frequencies (alpha, 10 Hz, and gamma, 33 Hz). The overall EEG signal can be approximately decomposed into two components (modes). Note that the first mode oscillates at the alpha band, while the second mode oscillates at gamma (power spectra shown below traces). The amplitude of the signal in each channel of the mode reflects the loadings of the corresponding eigenvector. Note that channel 1 is dominated by gamma oscillations, but channel 3 has predominantly alpha oscillations. This difference in power arises because the first eigenvector has a large component at channel 3 while the second eigenvector has a large component in channel 1. While the average amplitude of the oscillation differs from channel to channel, the dynamics of the decay of the mode are the same across all channels. The modulus of the eigenvalue for the first eigenmode is ∼1 and, consequently, the signal does not decay over time. We refer to such modes as critical. The second eigenmode, in contrast, is damped (the modulus of the eigenvalue is <1). This timescale of the decay is governed by the modulus of the eigenvalue.
Surrogate data generation
Time-shuffled surrogates were created from the raw EEG timeseries to test the dependence of the results on network-level features of the data. Surrogate timeseries were generated via the following procedure for each channel:
A time-shift amount is sampled randomly from a Gaussian distribution with a mean of 0 s and standard deviation of 1 s.
That channel’s voltage trace is shifted in time by the selected amount (in either the forward or backward direction, as dictated by the sign of the sampled amount).
This modification preserves spectral features of each channel and leaves intact the coherence between each pair of channels (with an arbitrary phase shift), while disrupting higher-order network relationships in the data.
Classifier training/testing and PCA of features
To evaluate and compare the performance of criticality and spectral features in predicting ketamine dose, two random forest classifiers were trained using features from two distinct datasets: a criticality index dataset and a power spectral density (PSD) dataset both computed from the same EEG signals. The criticality dataset contained distributions of criticality indices (|l|) at each dose computed in each 0.5 s window, while the PSD dataset consisted of spectral power between 1 and 100 Hz. Data in both datasets were normalized as the difference from the mean distribution during the no-drug control condition, in order to account for individual differences in baseline EEG.
Each random forest model was trained using the randomForest (https://rdocumentation.org/packages/randomForest/versions/4.7-1.1) package in R (https://www.r-project.org/). The dose variable was set as the dependent variable. Hyperparameters for the models were set to 500 decision trees (ntree = 500) with six randomly selected features considered at each split (mtry = 6).
Performance metrics included the out-of-bag (OOB) error rate and confusion matrices. OOB error rates were extracted directly from the models and indicate error rate on unseen testing data, providing a measure of the generalization error. Confusion matrices were used to compute the number of correct and incorrect classifications, which formed the basis for statistical comparison. To assess the statistical significance of differences in classifier performance, Fisher's exact test, with Benjamini–Hochberg FDR correction, was applied to contingency tables summarizing correct and incorrect predictions for both classifiers at each ketamine concentration. A classifier was also trained on distribution of stability parameters computed on surrogate data in exactly the same fashion as for the real EEG recordings and its OOB error rate derived for comparison.
Variable importance was calculated as mean decrease in Gini impurity. Mean decrease in Gini impurity is a measure of feature importance in random forest classifiers that quantifies how much each feature helps reduce uncertainty (or impurity) when making splits in the decision trees (Breiman et al., 1984). Gini impurity is a metric that reflects how mixed the classes are at a node, with lower values indicating more pure splits. When a feature is used to split the data at a node, it reduces the impurity by dividing the data into subsets that are more homogeneous. The reduction in Gini impurity is calculated by comparing the impurity of the parent node with the weighted impurities of the resulting child nodes, where the weights are based on the number of samples in each child.
For each feature, the total reduction in Gini impurity is summed across all splits in all trees of the random forest and then averaged. Features that consistently result in larger reductions in impurity are considered more important, as they contribute more to the model's ability to distinguish between classes.
We additionally calculated variable importance as permutation variable importance (PVI). PVI is an alternative method used to assess the importance of individual features in a random forest classifier (Breiman, 2001). It provides insight into how much each feature contributes to the model's predictive performance. PVI is calculated by measuring the change in the model's performance (accuracy) when the values of a specific feature are randomly permuted (excluded). Permuting the feature breaks its association with the target variable while leaving the values of other features unchanged. The difference in the model's accuracy before and after permutation indicates the importance of the feature. A larger drop in performance suggests that the feature is more critical for accurate predictions, while little to no change indicates that the feature is less important. PVI is particularly useful because it accounts for the interactions between features and evaluates importance in the context of the trained model.
To further investigate the structure of the criticality and spectral datasets used for classification and explain the difference in classifier performance, principal component analysis (PCA) was conducted separately on the criticality and power spectral density data. Distributions of criticality indices were estimated in 99 bins (uniform width 0–1.1). Power was estimated in 1 Hz intervals from 1 to 100 Hz. Before performing PCA on the criticality data, a Gaussian smoothing window was applied to a histogram of criticality indices estimated at each temporal window. The Gaussian window was generated using a standard equation with a window size of 10 histogram bins and a shape parameter of 2.5 (standard deviation 1.8 bins). These smoothed histograms at each time window for all ketamine doses, and all subjects were submitted to PCA. Welch's spectral estimates are naturally smoothed and no further processing was performed on spectral data prior to PCA analysis.
For both criticality and PSD datasets, PCA was implemented using the prcomp (https://www.rdocumentation.org/packages/stats/versions/3.6.2/topics/prcomp) function in R, with neither centering nor scaling applied to preserve the original feature distributions. The proportion of variance explained by each principal component was calculated as the squared standard deviation of the component divided by the total variance, expressed as a percentage. The cumulative variance explained by the principal components was compared between the criticality and PSD datasets. The relationship between principal component projections and dose levels was further explored by examining the scores of the first principal component (PC1). Median PC1 projections were calculated across time windows for each dose and subject. These median values were used to generate line plots that highlight the trends in PC1 across dose levels for each subject. Boxplots overlaid with these median trends were used to display the distribution of PC1 scores across doses for both criticality and PSD measures at each ketamine dose in each subject.
Code availability
All code used to process and analyze data and generate figures is freely available at: https://github.com/proektlab/Diego_code/tree/main/KetamineEEG_Stability.
Results
Perceptual changes induced by escalating doses of ketamine
Ketamine is a dissociative anesthetic that is clinically used in two classes of settings. In surgical settings, higher doses of ketamine are used to produce a state of anesthesia. In psychiatry, subanesthetic ketamine has been used in the treatment of psychiatric disorders such as depression (aan het Rot et al., 2010) and chronic pain (Yang et al., 2020). Correspondingly, most studies that investigate the effects of ketamine on the brain have focused on ketamine effects at either “subanesthetic” or “anesthetic” regimes (Vlisides et al., 2017, 2018; Farnes et al., 2020; Li et al., 2022; Tian et al., 2023).
Yet, there does not appear to be a clear-cut qualitative distinction between the “subanesthetic” and “anesthetic” effects of ketamine. In the high ketamine concentration range typically used for surgical anesthesia, ketamine is well known to produce vivid hallucinations that are reliably recalled by subjects upon regaining responsiveness (Bowdle et al., 1998). Similar, albeit, less pronounced hallucinations are also commonly observed in patients receiving subanesthetic ketamine for the treatment of psychiatric disorders (Mathai et al., 2020). Thus, it appears that the effects of ketamine lie along a continuum with progressively more durable perceptual disturbances occurring at higher concentrations. To identify the neurophysiological features that parallel these concentration-dependent effects, here we first explored the effects of different concentrations of ketamine starting from subanesthetic (0.2 μg/ml) to anesthetic dose (1 μg/ml) typically used in surgical settings. The 0.2 mg/ml dose is similar to that used in the treatment of depression (0.5 mg/kg delivered over 30 min; Fig. 1).
Several scales have been developed to interrogate particular domains of dissociation and/or psychedelic states, such as the Clinician Administered Dissociative States Scale (CADSS; Bremner et al., 1998). However, this scale may not capture the entirety of the acute effects of ketamine (van Schalkwyk et al., 2018). Accordingly, we opted for open-ended self-report. Consistent with previous findings in psychiatric literature (Bowdle et al., 1998; van Schalkwyk et al., 2018), we observed that at 0.2 µg/ml study participants variably reported somatosensory disturbances such as the feeling that one's body is not their own and lack of coordination in movements. Escalation of the dose to 0.4 µg/ml resulted in more durable out-of-body experiences (“I was floating above my body”). Visual hallucinations at this concentration typically involved simple shapes (e.g., “blue dots”). Several subjects reported experiencing motion (“flying”). During escalation of the dose from 0.4 to 1.0 mg/ml, participants experienced intense hallucinations. While at the lower concentration most subjects endorsed hallucinations, they had little difficulty distinguishing them from real sensory stimuli. This distinction between real and hallucinatory perceptions was notably absent during the increase from subanesthetic to anesthetic dose. Once the concentration of 1.0 mg/ml was achieved, all but one subject became unresponsive to voice commands or tap on the shoulder. Upon recovery, subjects reported bizarre and vivid dream-like experiences (“I was in… I was falling along the side of an edge […] I plummeted down into a sort of a big… like in Star Wars 4, 5, 6, the first movies, when they fall into the trash pit […] and there were these big walls coming up over me and I was flying basically”). Thus, increasing doses of ketamine result in more intense and durable hallucinations. Consistent with previous reports (Domino and Warner, 2010; Sarasso et al., 2015; Vlisides et al., 2018), ketamine anesthesia is associated with intense visual dream-like experienced that are accompanied by decreased responsiveness to sensory stimuli. No discrete differences between the subanesthetic and anesthetic doses of ketamine was observed at the behavioral level. Instead, we observed a progressive dose-dependent increase in the dominance of hallucinatory percepts which at higher concentrations was accompanied by decreased responsiveness to external stimuli.
The administration of ketamine is associated with decrease of alpha power and increase in frontotemporal gamma power
Prior work investigating the effects of ketamine on EEG spectra have delineated prototypical changes (Vlisides et al., 2017). These previous observations were largely consistent with our findings (Fig. 3). While at rest with eyes closed, human EEG is typically dominated by high alpha power chiefly in the occipital region. This alpha power is reduced during administration of ketamine in a dose-dependent manner, along with a global power reduction (Fig. 3B). As noted previously (Vlisides et al., 2017), we observe that subanesthetic ketamine administration is associated with an increase in gamma power, primarily in the frontal and temporal regions (Fig. 3B).
Ketamine-induced changes to the EEG power spectrum. A, 5 s of EEG from an occipital channel in one subject at baseline and all ketamine concentrations. Note the attenuation of alpha and the increase in higher frequency oscillations that accompany ketamine administration. B, Spectral power expressed as z-scores relative to baseline (scale shown by color bar), calculated within each frequency band, and averaged across all subjects, plotted with respect to location on the scalp surface. While we observe a global power reduction, an increase in gamma power is also observed, most notably at the lower ketamine concentration.
Results of the bootstrap analysis (see Materials and Methods) of these changes in the spectra (Fig. 4) confirm reduction in spectral power observed across most frequency bands at higher doses of ketamine. The reduction is most prominent in the alpha band, particularly in the occipital regions. We observe a slight increase in frontal and temporal gamma power at the 0.2 µg/ml dose, as noted in prior work using a similar dose (Vlisides et al., 2017). Thus, we have replicated prior findings on the effects of low-dose subanesthetic ketamine on spectral and elaborated changes that emerge at higher subanesthetic concentrations. This points to a highly dose-dependent pattern of spectral features associated with subanesthetic ketamine concentrations.
Spectral changes by frequency and region. Bootstrapped estimates of the mean change in power at each frequency at each region (error bars indicate 95% confidence intervals). Bootstrapping procedure is described in methods subsection “Spectral Bootstrap Analysis.” As seen in Figure 4, the most pronounced change is the decrease in alpha oscillations.
During the administration of subanesthetic and anesthetic concentrations of ketamine, dynamics become more stable relative to wakefulness, in a dose-dependent manner
We find that during the administration of ketamine, neural dynamics become more stable relative to wakefulness. Consistent with previously reported results using ECoG in both humans (Alonso et al., 2014) and macaques (Solovey et al., 2015), we note that during normal resting wakefulness, dynamics are poised just below criticality. This is indicated by the observation that during the resting no-drug condition, criticality indices near 1 are most prevalent (Fig. 5A, 0.0 µg/ml section). As the concentration of ketamine is increased, the proportion of criticality indices near 1 decreases in a dose-dependent fashion. This stabilization is observed at both subanesthetic concentrations (0.2, 0.4 µg/ml) and a full anesthetic dose (1.0 µg/ml; Fig. 5A). It is notable that the largest stabilization typically occurs when moving from a 0.2 to 0.4 µg/ml concentration. This particular jump in dose–response is consistent with the notion that ketamine's effects are highly variable across doses in the subanesthetic range and invites more granular investigation of ketamine's unique effects at distinct concentrations. FDR-corrected p values are reported in Table 1, rounded to 4 significant digits. To determine whether the distribution of criticality indices was consistent across time during baseline, the 1 min awake period was split into two halves and compared using the same statistical procedures. The fact that no significant differences were observed suggests that the distribution of criticality indices stays approximately constant during wakefulness. Thus, upon entry into a ketamine-induced dissociated state, where awareness persists but responsiveness to external stimuli is reduced and eventually extinguished, dynamical criticality is lost.
Stabilization of dynamics during the administration of increasing concentrations of ketamine. A, Distribution of criticality indices (|l|) is shown for each 0.5 s window during baseline (0 μg/ml) and three ketamine concentrations. The distributions are averaged across subjects (x-axis shows the time during 1 min recording in each condition; y-axis denotes |l|, color encodes fraction of modes represented). As many criticality indices are close to zero and do not contribute to the dynamics, the y-axis starts at 0.7 to emphasize the modes that contribute to the overall dynamics. As ketamine dose increases, the fraction of modes near criticality decreases. B, Distribution of criticality indices computed for shuffle surrogate datasets (Materials and Methods) are plotted in the same fashion as A. Note significant disruption of overall pattern of criticality in time-shuffled data.
Comparison of criticality index distributions across doses
Surrogation analysis
It is natural to ask whether the stability analysis based on autoregressive models is merely capturing spectral changes that occur during the administration of ketamine. This can be tested via a simple surrogation procedure. The basic strategy is to shift each EEG channel's voltage trace by a randomly chosen small amount. This time shifting is performed once on the entire recording (rather than repeated in each window). In addition to the spectral features, this surrogation procedure also preserves the coherence between all pairs of channels albeit with an arbitrary phase shift but destroys all higher-order correlations among signals.
We find that these time-shifted surrogates exhibit markedly distinct distributions of criticality indices (Fig. 5B), indicating that criticality and stabilization observed in actual data do not simply recapitulate spectral changes, or features of the relationship between individual pairs of channels. Rather, they represent distinct phenomena that are a result of coordination at the level of the brain's large-scale neuronal network. While surrogation dramatically alters the overall distribution of eigenvalues, ketamine nevertheless appears to be associated with stabilization. Thus, some small component of ketamine-induced stabilization of the EEG dynamics may reflect changes in the spectra of individual channels and pairwise correlations.
Stabilization occurs over a broad range of frequencies, occurring most dramatically at higher frequencies
Until this point, we considered just the modulus of the eigenvalues of the autoregressive models fit to the EEG data. However, because the autoregressive matrices are generally not symmetric, most of the eigenvalues are complex and thus contain frequency information (see Materials and Methods). Below we further dissect the relationship between stability and the frequency of the oscillations.
We observe that across all subjects, stabilization occurs over a broad range of frequencies, although most dramatically at higher frequencies (above ∼20 Hz; Fig. 6). This effect is dose dependent, with more prominent stabilization observed as the dose increases. Thus, while ketamine produces complex changes in the spectra that depend on the dose, location, and the frequency band, signals in all frequency bands are progressively stabilized with increasing doses of ketamine. This is particularly interesting, as prior work applying stability analysis to intracranial recordings found that stabilization occurs at higher frequencies. This divergence may point to a more global effect on overall dynamics at the scale of scalp-level analysis, as well as potential attenuation of higher frequencies by scalp EEG.
Ketamine-induced stabilization is observed across most frequencies in a dose-dependent fashion. Each subplot shows data from a single subject. The contours were generated to include 90% of the eigenvalues in each subject at each ketamine concentration. Note stabilization occurring across all frequencies in a dose-dependent fashion. The most marked stabilization can be observed at higher frequencies (20 Hz and above).
Classification of ketamine dose based on stability is more accurate than classification based on spectra
Ketamine administration can variably affect many features of EEG signals such as power in different frequencies, phase relationships between oscillations at different frequencies, pairwise interactions among channels, stability, etc. Thus, we wanted to determine whether stability or spectral features were better able to track the effects of ketamine. One way to approach this question is to train a classifier to distinguish between the EEG signals obtained at different concentrations of ketamine on either the stability or the spectra and determine which classification is more reliable.
Classifier performance
The criticality-based random forest classifier was more accurate relative to the PSD-based classifier by more than an order of magnitude across all evaluated metrics. The out-of-bag (OOB) error rate of the criticality classifier was 0.019, while the OOB error rate of 0.217 for the PSD classifier. This result indicates that the criticality classifier achieves higher accuracy in dose prediction and generalizes better to unseen data (Fig. 7A). A separate classifier was trained on criticality data derived from time-shuffled surrogates and achieved an OOB error rate similar to that of the PSD classifier (0.169). Altogether these results suggest that network-level information about the dynamics within the EEG, rather than spectra of individual channels or pairwise correlations between them, is a key feature that is reliably affected by ketamine in a dose-dependent fashion.
Criticality-based classifier predicts ketamine concentration more reliably than Spectrum-based classifier. A, Out-of-bag error rates for multi-class Random Forest classifiers trained on Criticality versus Power Spectral Density (PSD). The error rate is the pooled estimate across all ketamine concentrations. B, Variable Importance (quantified as mean decrease in Gini impurity) for Criticality Classifier versus PSD classifier. C, D, Confusion Matrices for Criticality and PSD classifier, respectively. Note that the criticality classifier makes fewer errors at each ketamine concentration. E, Variance explained plotted as a function of number of principal components (PCs) for changes in the distribution of criticality indices and of the spectra (teal and orange respectively) for all subjects. F, Median (across time in each subject) projection onto first principal component of the criticality dataset by dose. Each colored line represents a subject; overlain boxplots show interquartile interval (whiskers show 1.5 times the interquartile range). G, Same as F but for the spectra.
An analysis of the classifiers’ confusion matrices revealed that the criticality classifier made fewer errors at all dose conditions (Fig. 7C) compared with the PSD classifier (Fig. 7D). Fisher's exact test was performed on contingency tables at each dose derived from the confusion matrices, with Benjamini–Hochberg FDR correction applied to resulting p values (0.0 μg/ml: FDR adj. p = 1.8939 × 10−7, 0.2 μg/ml: FDR adj. p = 8.8011 × 10−43, 0.4 μg/ml: FDR adj. p = 1.0967 × 10−30, 1.0 μg/ml: FDR adj. p = 2.8137 × 10−67). The lowest dose of ketamine is naturally the hardest to distinguish from wakefulness and was most error prone for both classifiers. In this most difficult classification, the stability-based classifier made approximately eight times fewer errors than the spectral classifier. In the highest ketamine dose, the improvement was >50-fold.
Variable importance
To identify the key features driving classifier performance, the variable importance scores were examined for both classifiers. Variable importance was calculated using mean decrease in Gini impurity (see Materials and Methods). For the criticality classifier, the most influential features were criticality index values close to 0, which correspond to highly stabilized modes, and those near criticality, particularly in the range of 0.8–1 (Fig. 7B). This finding aligns with the theoretical relevance of these criticality index ranges, with near-critical indices being associated with responsiveness and very stable indices being associated with ketamine administration.
For the PSD classifier, frequencies within the delta/theta range and the alpha/beta range showed the highest mean decreases in Gini impurity (Fig. 7B), consistent with previous studies indicating that ketamine affects power spectra in these bands (Vlisides et al., 2018). However, spectral features provided less predictive power for dose classification compared with the criticality index features.
Similar results are observed when variable importance is calculated as permutation variable importance (see Materials and Methods), with no change in the overall pattern of feature importance noted for either classifier.
Origins of differences in classification performance
Improved performance of the stability classifier suggests that stability is more reliably affected by ketamine administration than spectral signatures of the EEG. To investigate this further, we subjected both the distributions of criticality indices and the spectra to dimensionality reduction. Based on the results of the classifiers, we hypothesized that ketamine-induced changes in criticality indices should be confined to lower dimensional space than the spectral estimates. This hypothesis was confirmed by estimating the fraction of total variance captured as a function of the number of principal components (Fig. 7E). For the criticality data, the first principal component (PC1) alone accounted for ∼90% of the total variance, indicating that the dataset is highly structured and that most of the variability can be summarized by a single dimension. In contrast, the PSD data required ∼50 principal components to cumulatively explain 90% of the variance, reflecting the complex and variable nature of ketamine-induced changes in the spectra of the EEG.
Dose–response relationship in PC1 projections
We then hypothesized that most of the variability in the distribution of criticality indices is due to ketamine administration. To test this hypothesis, we projected the data from all dose conditions onto the first PC. For the criticality data, this projection revealed a clear dose–response relationship. Specifically, PC1 scores increased with increasing ketamine doses in all subjects (Fig. 7F). While the specific features of the dose–response relationship varied across subjects, as expected from the variable effects of ketamine (Honey et al., 2008), there is a clear trend that is observed for each subject. In contrast, the PC1 projections for the PSD data did not exhibit a consistent dose–response relationship in most subjects (Fig. 7G). Additionally, the distributions of PC1 scores exhibited substantial overlap between ketamine concentrations. This implies that the largest component to the variability of the spectra is not related in a clear fashion to the ketamine dose.
Finally, we examined the moment-to-moment fluctuations in both criticality indices and spectra in each subject at each ketamine dose. The boxplots in Figure 7F,G show that the distribution of PC1 projection of criticality indices (F) is significantly tighter than that for spectral estimates (G). This implies that within each subject at a given ketamine concentration, the distribution of criticality indices remains relatively unchanged while the spectral estimates fluctuate from moment to moment. These moment-to-moment fluctuations in the spectra combined with a lack of a simple relationship between spectral changes and ketamine concentration give rise to the difficulty in reliably distinguishing between wakefulness and ketamine on the basis of conventional EEG measures. These problems are largely overcome by considering the stability of EEG dynamics.
Discussion
Consistent with prior theoretical (Magnasco et al., 2009) and experimental findings (Solovey et al., 2012, 2015; Alonso et al., 2014), we find that during normal wakefulness brain dynamics operate at a near-critical region between damped and unstable regimes. With increasing concentrations of ketamine, as subjects first experience mild dissociative symptoms, then progress to more intense hallucinations, and ultimately to unresponsiveness, brain dynamics exhibit progressively greater and greater stabilization—suggesting a strong dose–response relationship between ketamine administration and stability. We show that unlike conventional measures such as spectral composition of signals and their pairwise correlations, stability of the EEG depends on more global features of brain activity. Finally, using classifiers based on either spectral content or stability and principal component analysis, we demonstrate that the increase in stabilization is a more reliable neurophysiological marker of ketamine's effect on human EEG than spectral content of the signals.
The first practical implication of these findings is that stability analysis offers promise in monitoring the neurophysiological effects of ketamine in patients undergoing ketamine therapy for neuropsychiatric disease treatment. Ketamine has proven promising for the management of treatment-resistant depression (Corriger and Pickering, 2019) and other neuropsychiatric disease such as chronic pain (Yang et al., 2020) and addiction (Ivan Ezquerra-Romano et al., 2018). However, only ∼54% of patients with multidrug-resistant depression respond to ketamine treatment (DeWilde et al., 2015). It has been suggested that the degree of dissociation is positively correlated with response to treatment (Luckenbaugh et al., 2014). Other mechanistically distinct drugs that produce hallucinations also show promise in the treatment of diverse neuropsychiatric conditions (Nagele et al., 2015; Corriger and Pickering, 2019; Palhano-Fontes et al., 2019; Johnston et al., 2023). While ketamine has been used to treat depression for many years now, the optimal drug dose to achieve satisfactory outcomes remains unclear (Fava et al., 2020). The dose-dependent effect of ketamine on EEG stability may offer a tentative explanation for the highly heterogeneous responses to ketamine treatment. Even in our highly homogeneous study population, we observe clear individual differences in the dose–response characteristics of effect of ketamine on stability of the EEG. It is possible that customized dosing, guided by stability analysis, may allow clinicians to attain more reliable outcomes. The simplicity of our approach readily lends itself to near real-time estimation of EEG stability making it potentially suitable in clinical settings. Further study into the relationship between degree of stabilization, intensity of the dissociated state, and response to treatment is required to test this notion.
It is a matter of debate whether the “anesthetic” doses of ketamine elicit a qualitatively distinct state than the “subanesthetic” dose. While we observe a robust dose-dependent stabilization of dynamics, the degree of stabilization between a high subanesthetic dose and an anesthetic dose of ketamine is relatively subtle. Our findings indicate that, at least through the lens of dynamical criticality, dissociative anesthesia is not fundamentally distinct from subanesthetic dissociation but is rather a more intense dissociated state in which responsiveness even to painful stimuli is greatly diminished.
Stabilization appears to occur across a broad range of frequencies, particularly in the alpha band and above. Power at alpha and above has been implicated in sensory integration (Mishra et al., 2007; Keil and Senkowski, 2018; Kaiser et al., 2019), and dramatic stabilization of modes in this range may speak to a reduction in the brain's ability to form unified percepts. It should be noted that this frequency-wide stabilization was not observed in prior work analyzing ECoG recordings under propofol and combined ketamine–medetomidine, which found stabilization to chiefly occur at higher frequencies (Alonso et al., 2014; Solovey et al., 2015). This is potentially due to attenuation of higher frequencies recorded by scalp EEG.
Our results elaborate on the critical brain hypothesis (Beggs, 2007; Hesse and Gross, 2014; Stepp et al., 2015) which maintains that the awake brain self-tunes to a near-critical state which allows for optimal information processing (Beggs and Plenz, 2003; Kinouchi and Copelli, 2006; Beggs, 2007; Magnasco et al., 2009; Shew et al., 2009; Rajan et al., 2010; Bellay et al., 2015; Tkačik et al., 2015). While the criticality hypothesis has gained significant experimental support, it also faces many challenges. The diversity of definitions of criticality and techniques used to identify signatures of criticality in the brain is a significant confounder. Much work has focused on identifying power-law distributions in various aspects of brain activity (Bullmore et al., 2001; Beggs and Plenz, 2003, 2004; Destexhe et al., 2003; Miller et al., 2009; Milstein et al., 2009; Dehghani et al., 2010; He et al., 2010; Cocchi et al., 2017; Colombo et al., 2019; Varley et al., 2020; Maschke et al., 2024), correlations in neuronal firing (Tkačik et al., 2015), and chaos (Skarda and Freeman, 1987; Wang et al., 2017; Toker et al., 2022; Maschke et al., 2024).
In our approach, motivated by work on the cochlear hair cell (Hudspeth et al., 2010), we sought a distinct, dynamical, signature of criticality—a Hopf bifurcation. Hopf bifurcation corresponds to a combination of parameters at which a fixed point loses its stability and, in its stead, a stable periodic solution arises (Strogatz, 2000). From the standpoint of our autoregressive model of the EEG, Hopf bifurcation occurs when the modulus of an eigenvalue is 1 (Solovey et al., 2012; Yan and Magnasco, 2012). Much like other definitions of criticality, dynamics near a Hopf bifurcation can endow the system with some attractive features such as compressive gain and infinitely sharp tuning (Eguíluz et al., 2000). While the behavior near a Hopf bifurcation is essentially nonlinear, previous work suggests that a linear autoregressive model fit to short windows of data can nevertheless be used to detect the behavior near the bifurcation (Solovey et al., 2012, 2015; Alonso et al., 2014). One of the potential significant limitations of this approach is that a time scale and order must be chosen a priori. To mitigate this potential confounder, we varied both of these parameters and did not observe any appreciate change in the goodness of fit to the EEG. Locally linear approaches (Costa et al., 2019) which allow the window size to be dynamically adjusted based on the data also converged on the conclusion that waking dynamics are found near a critical point and become more stabilized with onset of anesthesia.
Autoregressive modeling and other related techniques provide an indirect measure of dynamical stability. To obtain experimental confirmation of the stabilization, a perturbation must be applied. A dynamically critical system poised near a Hopf bifurcation is expected to respond to a perturbation by producing spatiotemporally extensive activation. Indeed, threshold stimuli associated with behavioral report elicit extensive brain-wide activation patterns involving the prefrontal cortex in primates (van Vugt et al., 2018). These activity patterns switch during spontaneous perceptual fluctuations elicited by ambiguous stimuli (Kapoor et al., 2022; Dwarakanath et al., 2023). In awake rodents suprathreshold visual stimuli elicit traveling waves of activity that percolate across the cortex (Aggarwal et al., 2022) but fail to do so under conventional and dissociative anesthetics (Aggarwal et al., 2019, 2022, 2024). Coordinated patterns of neuronal activity normally elicited by auditory stimuli in rodents fail to form under anesthesia (Filipchuk et al., 2022). In primates a single action potential elicits a traveling wave (Nauhaus et al., 2009) in good agreement with a model of a dynamically critical network poised at a Hopf bifurcation (Yan and Magnasco, 2012). These findings using physiological stimuli are consistent with the theory that normally critical dynamics become dampened during loss of consciousness.
Recent work investigating statistical criticality metrics under diverse anesthetics have suggested that brain criticality may be unaffected by dissociative drugs such as ketamine but greatly affected by conventional anesthetics (Wang et al., 2017; Colombo et al., 2019; Varley et al., 2020; Maschke et al., 2024). Our results, when coupled with this prior work, suggest that statistical and dynamical criticality may each point to distinct aspects of brain activity. Statistical criticality measures are correlated with the presence or absence of “conscious content.” In contrast, dynamical criticality seems to align with brain states characterized by a reduced responsiveness to external stimuli.
What could be the origin of this distinction? Sensory perception involves a delicate balance between feedforward propagation of sensory information through the thalamus and the feedback signaling from higher-order cortical regions (Heekeren et al., 2004; Summerfield et al., 2006; Hardstone et al., 2021; Semedo et al., 2022). Conventional anesthetics disrupt both feedforward thalamocortical signaling and corticocortical feedback signaling (Hentschke et al., 2017; Murphy et al., 2019). One consequence of this fact is that the primary sensory cortex seems to ignore thalamic inputs and stimulus evoked activity becomes essentially similar to that observed in the absence of stimulus (Filipchuk et al., 2022). The suppression of both feedback and feedforward interactions appears to be associated with loss of both statistical (Colombo et al., 2019; Varley et al., 2020; Maschke et al., 2024) and dynamical criticality (Alonso et al., 2014; Solovey et al., 2015).
The effect of ketamine on feedforward signaling is less clear as most animal studies combine ketamine with medetomidine (a conventional anesthetic; Filipchuk et al., 2022). Some evidence, however, suggests that ketamine attenuates visual evoked responses in mice (Aggarwal et al., 2024). Furthermore, evidence suggests that under ketamine spontaneous waves of activity propagate in the feedback direction from higher-order cortical areas toward primary sensory cortices (Vesuna et al., 2020) and that these spontaneous waves of activity resemble those evoked by sensory stimuli in the awake state (Aggarwal et al., 2024). Other hallucinogens also appear to depress feedforward signaling (Michaiel et al., 2019). Altogether, these findings suggest that in states associated with sensory hallucinations, feedback signaling is preserved, and possibly enhanced (Stoliker et al., 2024), whereas feedforward signaling is suppressed.
It is interesting to note, that while sensory evoked potentials during REM sleep (Sharon and Nir, 2018) are attenuated, consistent with sensory disconnection, responses to nonphysiological magnetic stimulation remain at near wakefulness levels (Massimini et al., 2005). Similar discordance between natural and magnetic stimuli is observed under ketamine anesthesia (Casali et al., 2013). In contrast, during slow wave sleep, responses to both natural and magnetic stimuli are suppressed and simplified (Massimini et al., 2005). Similar suppression to both magnetic and natural stimuli is observed with sedation (Ferrarelli et al., 2010), conventional anesthesia, and disorders of consciousness (Casali et al., 2013). Recent work suggests that statistical signatures of criticality (Maschke et al., 2024) can be used to predict responses to magnetic stimulation. When statistical signatures of criticality are present, responses to magnetic stimulation tend to be extensive and complex. Conversely, when statistical criticality features are lost, the responses to magnetic stimulation become damped and simplified. This may point to the possibility that the measures of statistical criticality are dominated by the feedback rather than feedforward activity. This supposition is supported by the fact that most analyses of statistical criticality measures (Colombo et al., 2019; Varley et al., 2020; Maschke et al., 2024) utilize low-pass filtered signals. Numerous studies suggest that feedback signals tend to be carried by lower frequency brain oscillations than the feedforward signaling (Bosman et al., 2012; Bastos et al., 2015; Mejias et al., 2016; Michalareas et al., 2016; Aggarwal et al., 2022). In contrast to statistical criticality, our findings here suggest that dynamical stability of the EEG tracks responsiveness to sensory stimuli and may therefore predominantly reflect feedforward signaling. Indeed, while we observe stabilization of brain activity across different frequency bands, the most dramatic effect occurs at higher frequencies (Fig. 6).
This study has some notable limitations. We used a relatively small sample size which inadvertently was severely biased toward males. EEG is a highly indirect measure of brain activity and suffers from several significant limitations (Nunez and Srinivasan, 2006). For instance, while ketamine is known to produce high-frequency oscillations in brain activity, these are likely greatly attenuated in the EEG signals. That being said, EEG is the only readily available and noninvasive monitor of brain activity that can be deployed in both research and clinical settings. Future work should focus on verifying the conclusions of this study in a more diverse and large sample. While we verified that our conclusions are not strongly affected by the specific parameters of the autoregressive model, it is clearly a much-simplified model of brain activity. While using a simple model has some obvious disadvantages, the notable advantage of this approach is that the calculation can be performed in near real time, thereby making it potentially useful for guiding ketamine infusions in a real-world setting. Finally, we did not quantitatively measure dose-dependent changes in responsiveness under ketamine in this study. Consistent with previous work, we do show that at the highest ketamine dose, majority of the subjects become unresponsive to verbal commands and gentle tactile stimulation. Furthermore, while our study did not directly quantify responsiveness, prior work has demonstrated that subanesthetic S-ketamine, an enantiomer of ketamine, suppresses signal detection, specifically vigilance and speed of information processing in a dose-dependent fashion (Passie et al., 2005). Future work using larger subject population should address the relationship between stabilization and signal detection at the level of an individual. Quantification of changes in responsiveness under subanesthetic ketamine concentrations, however, is a challenging task. Performance on psychophysics tests, normally used to quantify responsiveness, may be unreliable in a ketamine-induced dissociated states. Performance on psychophysics tests may be affected by motor coordination, motivation, and intense hallucinations. This is likely to complicate unequivocal interpretation of the standard psychophysics approaches. Nonetheless, future work could incorporate more sophisticated psychophysics approaches that control for all of these confounders to more directly link stability and responsiveness in the subanesthetic regime.
In conclusion, here we showed that EEG signals are stabilized in a dose-dependent fashion by ketamine. Independent of the potential theoretical implications of these findings, the ability to reliably detect and quantify the dose-dependent effects of ketamine on the basis of noninvasive brain recordings may prove useful clinically both in the setting of psychiatry and in the operating room.
Footnotes
This work was supported by National Institute of Neurological Disorders and Stroke (NINDS; NS113366; to A.P.), the Hearst Foundation Fellowship 2022 (to D.G.D.), and National Institutes of Health (T32HL07953; to D.G.D.). We thank Ethan Blackwood, Claudia Heymach, and Andrzej Wasilczuk for useful comments on the manuscript and also Will Carspecken and Woo Yul Byun for their help in performing the experimental recordings.
The authors declare no competing financial interests.
- Correspondence should be addressed to Alex Proekt at proekt{at}gmail.com.
