The Anesthetic Cascade: A Theory of How Anesthesia Suppresses Consciousness

During the past 15 yr, major progress was made to establish that anesthetics achieve immobilization largely by actions on the spinal cord. These investigations were guided by demonstrations that the thalamus and cortex are more sensitive to anesthetics than the spinal cord7and that transmission is blocked and spinal motor neurons are depressed by inhaled anesthetics.8,9Rampil and Laster10provided supporting evidence further implicating action at the level of the spinal cord by their demonstration that movement responses to noxious stimulation could occur despite an electroencephalogram made essentially isopotential by isoflurane, after removal of cortex and thalamus,11and after hypothermic transection of the spinal cord.12Concordant evidence was provided by the findings that neither freezing the cerebral hemispheres13nor causing forebrain ischemia14nor using brain-preferential anesthetic procedures15blocked movement. This body of evidence supports the conclusion that immobility may be achieved by action at the level of the spinal cord.5However, the blockade of ascending somatosensory transmission by spinal anesthesia alters the excitability of reticulothalamocortical arousal mechanisms, reflecting the interactions between the different levels of the nervous system.16–18 

Much subsequent research focused on ligand- and ion-gated channels, examining whether specific receptors mediated immobility. Numerous neurotransmitters or ionic sites of action are plausible candidates for such effects. Genetic engineering provides animals with specific modifications that allow discrete testing of particular candidates for producing immobility. After a review of the extensive literature on the results of interfering with ion channels, Sonner et al.  19concluded that no action on a single receptor can explain how inhaled anesthetics achieve immobility and concurrent actions on many receptors are implausible. Their evidence indicates that among the candidates not yet convincingly ruled out are glycine receptors, serotonin type 2 (5-HT^2A) receptors, sodium channels, and N -methyl-d-aspartate (NMDA) receptors, whereas potassium channels, α²adrenoreceptors, and α-amino-3-hydroxy-5-methyl-4-isoxa-zole propionic acid (AMPA), γ-aminobutyric acid (GABA), opioid, serotonin type 3 (5-HT³), or acetylcholine receptors are probably irrelevant.

There is adequate reason to believe that the amnestic effects of anesthetics are mediated by some processes other than those that block awareness. Abundant findings unequivocally demonstrate that memory can be blocked at levels of anesthetic that do not suppress awareness. As cited previously, for many anesthetics the dose to suppress consciousness significantly exceeds that required to prevent memory storage.3Gajraj et al.  20used repeated arousals from periods of unconsciousness produced by target-controlled propofol infusion during hip or knee replacement under spinal anesthesia in 12 patients. The auditory evoked potential (AEP) index, a measure derived from the midlatency auditory evoked response (MLAER), increased sharply at each awakening, thereby indicating recovery of the MLAER. The MLAER is depressed with loss of consciousness and recovers on return of consciousness (ROC; see AEP monitors in section entitled Electrophysiologic Monitors of the Effects of Anesthetics). However, at the end of the procedures, none of the more than 120 such awakenings was recalled by these patients.

As well as anesthetics, a class of “nonimmobilizing” agents predicted to have anesthetic action by virtue of their lipid solubility can suppress learning and produce amnesia without causing LOC or surgical immobility.21,22Dutton et al.  23studied one of these agents to ascertain whether these effects might result from blockade of perception of external stimuli required for learning to take place. They studied alterations of the vertex-recorded MLAER in rats during inhalation of isoflurane, desflurane, or nitrous oxide that produced sedation, as well as inhalation of the nonimmobilizer compound 2N that failed to produce sedation. All the volatile anesthetics increased the latency and depressed the amplitude of particular components of the MLAER recorded at the vertex. These components arise from the medial geniculate in the thalamus and the primary auditory cortex. Because compound 2N did not significantly alter either of these indices, it was concluded that the site at which 2N interfered with learning was not a component generator of the MLAER vertex response. This finding implicates some anatomical substrate later in the processing stream than the primary auditory cortex.

In a series of elegant studies using quantitative electroencephalographic as well as positron emission tomographic (PET) evaluation of alterations of brain activity, correlated with blockade of memory during conscious sedation, Veselis et al. ,24–28presented evidence leading to the same conclusion, showing that the effects of anesthetics on episodic memory are related to actions primarily on the dorsolateral prefrontal cortex. The anterior cingulate, medialis dorsalis nucleus in the thalamus, and parietal association areas may also be affected. We know of no studies in humans that have probed the extent to which the effects of these agents on memory may involve the communications of various neocortical regions with each other, as reflected in studies of coherence. Effects on interactions with the limbic system must also be evaluated, possibly awaiting technically difficult and demanding studies that require use of functional magnetic resonance imaging or intracerebral electrodes. Finally, the studies of the sort thus far cited have only evaluated episodic but not implicit memory that is not manifested as explicit recall but is usually demonstrable only using associative or priming paradigms. A substantial literature indicates that implicit memory formation is possible even though consciousness is not present.

The evidence thus far summarized supports the conclusions that anesthetics act on different regions of the nervous system to produce immobilization, amnesia, and absence of awareness, and that immobilization very likely is mediated by effects on the spinal cord while amnesia is probably mediated by effects on the dorsolateral prefrontal cortex. This theoretical article is focused on the most significant, perhaps more intriguing, and less well-understood problem of how anesthetics act to suppress consciousness, which we define as the absence of awareness. Although our primary concern is the attempt to explain the neurophysiologic processes by which this remarkable property is mediated, the current status of understanding of the underlying molecular mechanisms will first be summarized.

A number of neurochemical/neuropharmacologic theories of how anesthetics suppress consciousness have been proposed. This approach seeks to identify specific receptors and molecular sites for anesthetic action, with the goal of identifying the processes within the brain at the neuronal level that are critical for anesthesia. With the Meyer-Overton demonstration of correlation between anesthetic action and solubility of anesthetic drugs in fat-like solvents, attention was directed toward membranes as likely sites of drug action. Initially, many concluded that general anesthetics acted by extensively disrupting the lipid portions of neuronal membranes. It was thought that nonpolar amino acid side groups become buried within protein interiors in the process of protein folding and that polar interactions with solvent water would be avoided in such hydrophobic pockets. The lipid-like characteristics of these pockets would dissolve anesthetics, which might bind in them.

As experimental evidence contradicted the hypothesis of lipophilicity, the work of Franks and Lieb29shifted attention from lipids to proteins as anesthetic targets. Many subsequent studies on the effects of anesthetics in the central nervous system focused on nerve membranes, ion channels and their regulatory mechanisms, and especially on synaptic transmission. Franks and Lieb30proposed that at the molecular level, anesthetics almost certainly act by binding directly to proteins rather than by perturbing lipid layers in cell membranes. They concluded that although at high enough levels general anesthetics might act nonspecifically on a wide variety of neuronal sites, at clinical concentrations they are much more selective and might act by binding to only a small number of targets in the central nervous system. These were considered likely to be postsynaptic ligand-gated ion channels, because voltage-gated ion channels are very resistant to clinical concentrations of general anesthetics. Some agents seem to be effective at both inhibitory and excitatory synapses, and both inhibition of excitatory synapses and potentiation of inhibitory synapses may be involved in anesthetic effects. Although some agents might act at excitatory synapses (such as ketamine at NMDA receptors), others suggested that anesthetic potentiation of inhibitory synaptic receptors (primarily GABA) best matches the pharmacologic profile of a wide variety of general anesthetic agents. General anesthetics potentiate GABA by increasing flow of Cl ions in open channels, and it has been suggested that anesthetics act at the GABA receptors at the lipid–protein interface.31,32Franks and Lieb33later suggested that inhalation anesthetics act by inactivating proteins that are transmitter-gated receptors for GABA, serotonin and acetylcholine, and possibly glutamate receptors. Urban and Friederich34believed that the evidence demonstrates that both voltage-gated and ligand-gated ion channels may be affected by clinically relevant concentrations of general anesthetics.

A number of explanations of a more theoretical nature have been proposed to account for the effects of anesthetics on consciousness. In a conceptually different approach to this problem, Flohr35proposed that the conditions for consciousness to appear can be defined by a specific computational structure that gives rise to functional states that are identical with states of consciousness. Such functional states are higher-order representations by which an information processing system represents its own inner state. Flohr reviewed the long philosophical tradition of this proposition and proposed that higher-order representations are instantiated by complex, rapidly formed cell assemblies that depend critically on the function of the NMDA synapse. The NMDA synapse is both voltage and transmitter dependent. This group argued that the NMDA receptor serves as a Hebbian coincidence detector, modulating the synaptic efficacy as a function of amount of use, and the activation state of the cortical NMDA synapses determines the size and complexity of representational structures that can be built up. If the receptor were disabled, the representations could not be produced, which they considered equivalent to LOC. They proposed that the common mode of anesthetic action is the disruption of NMDA-dependent computational processes. They asserted that inhibition of such processes is the necessary and sufficient condition for agents to have anesthetic properties, reviewing the extensive literature showing that many substances antagonize or modulate membrane functions relevant to activation of the NMDA receptor channel complex. An excellent review of research on how different classes of inhaled anesthetics act on specific molecular targets and ion channels to affect neural networks has recently been provided by Campagna et al.  6 

Challenging the view that consciousness arises as the result of discrete interactions via  synaptic connections in specific interconnected neural networks, which are blocked by anesthetics, some have attempted to explain the molecular basis of anesthetic action using concepts from quantum physics. Hameroff36has suggested that nonpolar amino acid side groups are interiorized by folding of target proteins, creating hydrophobic pockets to which anesthetics probably bind. This retards electron mobility required for protein dynamics and may thereby potentiate the activity of inhibitory and inhibit the action of excitatory receptors, especially in thalamocortical projections. However, he questioned whether the mere presence of anesthetic molecules in hydrophobic pockets would be sufficient to explain anesthesia. He has presented a quantum-theoretical model proposing that consciousness arises from the coherent superposition of states described by wave functions, generated by endogenous van der Waal London forces in hydrophobic pockets of certain brain proteins, collapsing when an objective threshold of entropy is reached. Anesthetics are proposed to act by preventing quantum states in such pockets.

In more recent work, Woolf and Hameroff37have proposed that rudimentary visual consciousness depends on quantum computation in microtubules in the cytoplasmic interiors of cortical pyramidal dendrites interconnected by gap junctions, forming a horizontal syncytium or “hyperneuron” spanning visual cortical areas. These interactions are critically dependent on cholinergic action on pyramidal cell dendrites and on γ-aminobutyric acid–mediated (GABAergic) interneurons interconnected by electrotonic gap–junction connections. They suggest that it is at this level that the anesthetic action interferes. Jibu38has attempted to explain the action of anesthetics within the framework of condensed matter physics. A new phase of condensation of massive photons (called tunneling photons ) in the perimembranous region is proposed to control the lateral diffusion of molecules in the nerve cell membrane, determining the molecular biologic functioning of the cell by modulating chemical reactions. Jibu proposes that anesthetic molecules trespassing into the perimembranous region break the order of this condensation and thereby disrupt cell functions that require maintenance of ordered chemical reactions. Although such proposals remain theoretically controversial and without experimental support, they are mentioned here for completeness and to illustrate that theories of the molecular mechanisms of anesthesia still cover a wide range.

As is evident from consideration of this body of research and contemporary theoretical speculations, anesthetics may act on a variety of neurotransmission processes, and the effects are anatomically widespread throughout the brain. Ion channel proteins do not function by themselves but are integrated with other proteins into a membrane patch, patches are integrated into a nerve cell, nerve cells are integrated into networks, and local networks are integrated into functional units. It is not sufficient for a theory of anesthesia to be based on demonstration of some primary effects of an anesthetic drug on neuron membranes or receptors or on some particular molecular or cellular site of action. Evidence on that level of discourse cannot explain why such actions result in the disappearance of consciousness.

As Mashour39has pointed out, perceptual processing is accomplished by relatively discrete and functionally specialized neural ensembles that are anatomically dispersed. Abundant evidence suggests that synchronization within and among these regions may play a critical role in integrating such dispersed information into a unified perception. Cognitive binding at all levels plays a crucial role in the generation of conscious experience. A deeper understanding of how anesthetics disrupt consciousness might require examination of whether they interfere with this binding process at some critical levels. In view of the hierarchical organization of the central nervous system and the cascade of consequences of action at any level, it seems evident that to judge the relevance of any particular effects at any level of observation, whether molecular, ion channel, membrane, cellular, or network, we must identify the neuronal networks underlying the interlocking constituents of general anesthesia and how these components interact with one another. The primary goal of this article is to describe these networks and explain how their interaction produces perception and awareness.

One approach to understanding the critical mechanisms by which general anesthetics suppress awareness is to seek for invariant changes in the human brain as patients lose and regain consciousness under the effects of a variety of anesthetic agents. Such information is increasingly available from a variety of electrophysiologic instruments that have been developed to enable continuous intraoperative quantification of the depth of anesthesia, independent of any particular agent. Numerous studies support the belief that variables extracted by computer analysis from the electroencephalogram and the AEP are well correlated with changes in consciousness. Although they are based on quite different techniques, the methods have in common their reliance on electrophysiologic assessment of brain activity, reflecting effects that are anatomically widespread throughout the brain. This evidence clearly establishes that several dimensions of brain electrical activity sensitively and reliably reflect effects of the administration of a wide variety of agents that result in LOC. This article combines what is known about processes underlying these various aspects of brain electrical activity with some recent findings in cellular neurophysiology to conceptualize how anesthetics suppress awareness.

The instruments currently being used to monitor the depth of hypnosis fall into two categories. One class of such devices quantifies electroencephalographic variables extracted from the spontaneous electrical activity of the resting brain, quantitative electroencephalographic monitors. Some other monitoring approaches measure the AEPs elicited by stimulation (AEP monitors).

Based on visual inspection of electroencephalographic tracings, early studies found that anesthesia induction was accompanied by an increase of fast, high-frequency oscillations of voltage waves in frontal regions, spreading to more posterior regions with increased sedation and LOC. Conversely, slow low-frequency waves appeared in posterior regions and migrated forward. This topographic pattern of frontal predominance or “anteriorization” of quantitative electroencephalographic power with LOC during anesthesia was first reported in monkeys40and has since been consistently replicated in humans.

Early attempts to use brain electrical activity in patients to evaluate the depth of anesthesia relied on global features extracted from the power spectrum provided by the fast Fourier transform. This transform describes any brief waveshape that is represented as a series of numerical values at successive time points, such as a digitized segment of a quantitative electroencephalographic tracing, QEEG (t), as the sum of a number of sinusoidal waveforms across the frequency range from low to high frequencies. The amount or “amplitude” of each frequency fⁱcontained in the original tracing is represented by a coefficient, aⁱ, of the corresponding term in the time series, so that

The waveshape of the original tracing can be accurately reconstructed by addition of the set of waveforms fⁱin this “expansion,” multiplying each oscillation by the corresponding coefficient aⁱ. A precise and concise representation of the original tracing is therefore provided by the set of amplitude coefficients a¹to aⁿ. It has become conventional to square each term in this representation, thereby converting amplitude to power, thus obtaining the “power spectrum.”

In attempts to summarize the most relevant information about a particular quantitative electroencephalographic sample that was provided by this analysis, measures have been devised, such as the median frequency or the spectral edge frequency, which respectively define those frequencies below which one finds 50% or 95% of the quantitative electroencephalographic power. Such measures have been related to clinical signs or compared to analgesic endpoints (see review elsewhere).41The use of these relatively crude features is based on the hypothesis that with increasing depth of anesthesia, the distribution of power in the quantitative electroencephalogram steadily shifts toward lower frequencies. Median frequency can be conceptualized as akin to the center of mass of the area of the curve outlining the power spectrum. Such methods attempt to summarize the information in the whole spectrum by a single number representing only that point in the spectrum that satisfies an algebraic definition. However, a great variety of more comprehensive descriptors of the electrical activity of the brain can be constructed by mathematically combining quantitative features extracted from an array of electrodes using such spectral analysis methods.

The quantitative electroencephalogram measures currently being used in commercial monitors to estimate the depth of anesthesia include the following.

The Bispectral Index® (BIS®) monitor (Aspect Medical, Newton, MA) relies on spectral analysis of the electroencephalogram recorded from one or two electrode positions on the forehead. The constituent measures contributing to the BIS value are derived from this analysis and include power in selected frequency bands and certain elements of the “bispectrum,” which represent the covariance of power in particular low frequencies with the power at particular high frequencies. This monitor combines a variety of such descriptors of the quantitative electroencephalogram into a multivariate number, the BIS, using a proprietary algorithm.42This number is then scaled to range between 0 and 100. High correlations between this number and the state of consciousness have been reported in numerous studies.43–45The performance of BIS has been evaluated as the input to a closed-loop system for propofol administration that was considered to be clinically acceptable in a comparison with standard practice–controlled administration.46 

The Patient State Analyzer, PSA® monitor (Physiometrix, Inc, N. Billerica, MA) computes a different multivariate quantitative electroencephalographic index called the Patient State Index  to assess the depth of anesthesia.47,48The initial version of this instrument analyzed the quantitative electroencephalogram from four brain regions in an anterior/posterior array, and a later version uses a number of electrodes on the forehead. Extracted features describe not only the local quantitative electroencephalographic power spectrum at each electrode but also some dynamic relations among these regions, such as power gradient and coherence. These features are combined into a proprietary discriminant function to assess the probability that the patient is “awake,” a number scaled to range from 0 to 100. Significant relations of the Patient State Index with “level of consciousness” and sensitivity to change in state have been demonstrated.47–49 

The dimensionless Narcotrend® index (NCI; MonitorTechnik, Bad Bramstedt, Germany) is a new index based on quantitative electroencephalographic pattern recognition,50classifying the raw electroencephalographic epochs into six different stages from A (awake) to F (increasing burst suppression to electrical silence), rescaled into an index from 100 (awake) to 0 (electrical silence). This instrument derives from previous studies on automatic sleep staging of the electroencephalogram into five stages, with an added sixth stage of isopotential electroencephalography.51After artifact rejection, electroencephalographic segments are classified based on similarity to analyses of prototypic waveshapes in the time and frequency domain, results are entered into a discriminant analysis, probabilities of similarity are calculated, and the result is expressed as the index from 0 to 100.

Entropy, as a physical concept, is related to the amount of “disorder” in a system52and, in an information theoretical context, describes the irregularity, complexity, or unpredictability of a signal.53Entropy can be computed in the time domain, the frequency domain, or both. The S/5® Entropy Module (Datex-Ohmeda, Helsinki, Finland) computes the approximate entropy of the power spectrum to quantify the depth of anesthesia.54,55This method treats the quantitative electroencephalographic power spectrum obtained before induction of anesthesia essentially as the “ground state” of the organized signal system represented by the quantitative electroencephalogram and quantifies the shift from that state as negative entropy, or disorganization. Entropy measures have also been utilized to evaluate LOC induced by anesthetic as a phase transition.56 

Evoked potentials (EPs) are a series of oscillations in the brain electrical activity that are “time-locked” to the presentation of a stimulus, i.e. , the successive peaks in these oscillations appear at particular latencies after stimulus onset. EP waveshapes elicited by different stimulus modalities have been well studied, and the most probable neural generators whose activation corresponds to peaks at distinctive latencies have been identified. The latency of each peak in the EP waveshape reflects the time required for the neuronal encoded information about the stimulus to be transmitted to successive structures in the sensory pathway. As long as an individual subject remains in a stable state, transmission velocities are constant and latencies remain stable. Changes in latencies of EP components along a neural pathway caused by anesthetic therefore can potentially provide a sensitive indicator of anesthetic effects on the excitability of the corresponding brain structures. The EP measures in current use to monitor anesthesia include the following.

The AEP can be divided into several domains as a function of latency. The brainstem auditory evoked response extends from 0 to approximately 6 ms and is composed of a wave with five peaks, reflecting transmission in the lateral lemniscal pathway from the acoustic nerve to the inferior colliculus. The next domain in the AEP is the midlatency response or MLAER that extends from approximately 6 to 60 ms, reflecting transmission through the medial geniculate body in the thalamus to the primary auditory cortex.57The remaining components of the AEP are considered to be long-latency responses and are comprised of several latency regions that reflect increasingly complex cognitive processes engaging cortical association areas, the frontal cortex, and the hippocampus. It has long been known that the waveshape of the auditory evoked response simplifies and becomes smaller with increasing depth of anesthesia.

Although the brainstem auditory evoked response is essentially resistant to anesthetic effects and the long-latency responses are too variable to facilitate reliable and rapid evaluation, the MLAER has received substantial attention as a possible monitor of the depth of anesthesia. The MLAER has been shown to reflect reliably the level of anesthesia with a wide variety of anesthetic agents58–61and to detect awareness.62Methods have been devised to quantify the morphology of this region of the auditory EP, using features extracted from the MLAER61,63–66and incorporated into anesthesia monitors. One such measure, referred to as the Auditory Index , is implemented in an instrument called the Alaris AEP® Monitor (Danmeter, Inc., Odense, Denmark). This measure represents the morphology of the MLAER by calculating the total length of the curve describing the EP waveshape as if it were a string. The longer the string is, the lighter the anesthesia level is, and the shorter the string is, the deeper the level of anesthesia is.

Another method to use auditory evoked responses to evaluate anesthetic depth has been proposed, which is based on the 40-Hz steady state auditory evoked response (ASSR).67The ASSR arises mainly from the primary auditory cortex,68,69with a lesser contribution from subcortical generators in the thalamus or midbrain.70,71This method takes advantage of the fact that the responses of the brain to auditory stimuli become “entrained” as the repetition rate of the stimulation is increased. Essentially, the later components of the MLAER elicited by each stimulus become superimposed on the early peaks of the response to the subsequent auditory input. Thus, the ASSR provides a simple index of the ability of the brain to respond to rapid stimulation as it is impaired by anesthesia, which can be quantified by using spectral analysis with the fast Fourier transform to compute the power at the stimulus repetition rate.72,73 

As might be expected, investigations of the relative sensitivity and specificity of the detection of the absence and return of awareness, ease, and reliability of these different methods are already in progress. Gajraj et al.  20compared the performance of the BIS, the AEP index, the spectral edge frequency, and the mean frequency during repeated transitions from consciousness to unconsciousness. Sleigh et al.  74compared the BIS, spectral edge, and approximate entropy. Vanluchene et al.  75compared spectral entropy to the BIS and processed MLAER. Bonhomme et al.  76compared the ASSR and the BIS, and Struys et al.  77compared the BIS and MLAER. Kreuer et al.  78compared the Narcotrend® and BIS indices during propofol sedation. Schmidt et al.  79compared the Narcotrend®, the BIS, and the classic quantitative electroencephalographic variables: spectral edge frequency, median frequency, and relative power in the four quantitative electroencephalographic frequency bands, δ, θ, α, and β. A detailed evaluation of the relative virtues of various indices is beyond the scope of this article. These studies are cited only to point out that the concern is no longer whether such electrophysiologic descriptors have validity but rather the relative sensitivity and specificity of different measures and to provide readers interested in such comparisons with ready access to relevant reports.

Some researchers have investigated the feasibility of anesthesia monitors using heart rate variability and respiratory sinus arrhythmia80and electromyography of the frontal and orbicularis muscles, or the Facial Electromyographic Index.81These measures are also electrophysiologic and are mentioned in the interest of completeness. Although these indicators display sensitivity to anesthetics, they will not be further discussed in this article, which is devoted to more direct measures of brain electrical activity.

The proposition that aspects of the quantitative electroencephalogram and the EP reflect the depth of anesthesia has become well accepted. Quantitative electroencephalographic and AEP monitors to extract and quantify mathematical electrophysiologic descriptors have been incorporated into a wide variety of instruments and are increasingly being used routinely by many anesthesiologists as adjuncts to standard procedures for the evaluation of intraoperative depth of anesthesia. Numerous reports have been published confirming that selected measures of brain electrical activity change in a reliable way with the administration of a wide variety of agents that cause LOC. The magnitude of such changes can be used to ascertain whether the depth of anesthesia is sufficient to allow intubation and to permit surgical procedures to begin. Reversal of these changes gives a reliable indication that the patient is returning to a conscious and responsive state.

Reliable correlations between specified quantitative electroencephalographic and AEP parameters and clinical signs demonstrate that certain aspects of brain electrical activity are sensitive to the level of consciousness. These clinical monitors rely on the stability and sensitivity of brain electrical activity but differ with respect to the particular quantitative electroencephalographic rhythms or EP waveshapes selected for monitoring, the scalp regions from which recordings are collected, the precise nature of the features objectively extracted from the raw data, and the invariance or lack thereof that has been demonstrated for the effects of various anesthetic agents on the measure. They have in common that they provide practical clinical evidence that consciousness is a neurobiologic phenomenon that can be objectively quantified so that depth of anesthesia can be analyzed reliably using electrophysiologic variables. Although such derived measures correlate well with drug concentrations or clinical state, they provide relatively little insight into the underlying physiology of anesthesia or the brain mechanisms mediating consciousness.

They differ in that they focus on different aspects of brain electrophysiology. To try to infer the basis of anesthetic action from such condensed descriptors is reminiscent of the four blindfolded individuals in the folk tale who were trying to describe an elephant by each touching one of his four legs. However, by considering the neuroanatomical and neurochemical system which is the origin of the electrical activity being measured, and by comprehensively analyzing the set of neurophysiologic changes in processes generated by this system that take place in common when anesthetics suppress consciousness, much can be inferred about the brain mechanisms of anesthesia and perhaps about the processes that sustain consciousness. The intent of this article is to examine the neurophysiologic processes that underlie the quantitative electroencephalographic and EP measures that display reliable and sensitive alterations during anesthesia and to combine that information with a number of recent research findings to construct a neurophysiologic theory of anesthesia.

Recordings from the surface of the scalp have long been known to demonstrate rhythmic voltage oscillations derived from the electrical activity of the subjacent neuronal populations.82Such fluctuations of voltage in brain electrical activity, whether recorded from the scalp as the electroencephalogram or EP waves or as intracerebral local field potentials, reflect the nonrandom synchronization of postsynaptic potentials in enormous numbers of neurons. The various quantitative electroencephalographic and EP monitoring techniques rely on the fact that the spontaneous electrical activity, recorded from the scalp of a healthy person resting with closed eyes, is dynamically regulated by interactions within a homeostatic system that are mediated by many different neurotransmitters. The existence of such a homeostatic system is established by evidence that the power spectrum of the quantitative electroencephalogram can be predicted for repeated samples of adequate length, recorded from healthy, normally functioning individuals while at rest with closed eyes. The same power spectrum will be obtained reproducibly from a set of artifact-free samples of adequate length obtained shortly after each other, provided that the state of the brain does not change. The homeostatic regulatory system is discussed in the section entitled Homeostatic Regulation of Brain Electrical Activity.

Such instruments inherently depend on establishing a multivariate signal space of features extracted from the quantitative electroencephalogram, in which a reference region around the origin or “baseline” is defined as awake and functioning normally. This baseline is determined by statistically scaling and combining measures extracted from the patient and compared to some large reference database. The BIS, GABA, and other quantitative electroencephalographic indices may be conceptualized as the estimated length of a vector describing the distance between the momentary brain measure taken from a patient and the origin of the signal space in which they are contained. These indices differ with respect to the actual extracted measures chosen to be combined measures and also in that some use actual values of the extracted features, such as microvolts or phase angle or percent change, whereas others rescale to reflect the probability of the observation relative to the baseline.

After a sufficient sample of artifact-free data has been collected, using validated automatic editing algorithms to remove contaminated signals that do not arise from brain electrical activity, quantitative electroencephalography uses computer analysis to describe the power spectrum in each scalp region. The many measures that can be calculated using spectral analysis include the total power and the amount in each of several frequency bands of the local activity in each region (absolute power), the percentage of power in a region that lies within each of these bands (relative power), and the relations among regions within or between the two hemispheres in total and within each band (coherence and symmetry). For such quantitative electroencephalographic analyses, the quantitative electroencephalogram is conventionally divided into frequency bands, approximately defined as δ (0.5–4 Hz), θ (4–8 Hz), low α (8–10 Hz), high α (10–12 Hz), β (12–25 Hz), and γ (25–50 Hz).

There are a variety of methods for quantitative electroencephalographic analysis. The quantitative electroencephalographic analysis method we have used in the studies reported here is called neurometrics . This term refers to using certain mathematical procedures described below, for probabilistically scaling and combining measures derived from an individual and compared to a particular set of normative equations that describe the resting state.

Figure 1illustrates average power spectra, extracted from the quantitative electroencephalogram recorded from the scalp over the left occipital cortical areas (10/20 electrode position O1) and obtained from a large group of healthy, normally functioning individuals at ages ranging from 5 to 97 yr of age. Similarly, regular evolution with age has been shown for power spectra from quantitative electroencephalograms recorded at every position of the International 10/20 Electrode Placement System.

These sets of quantitative electroencephalographic spectra have been described by algebraic equations in which the major variable is the age of the subject, multiplied by a different set of constant coefficients for every cortical region. These age-regression equations provide the mean value and SD of the distribution of many different variables extracted from the power spectrum of the quantitative electroencephalogram recorded from any scalp electrode position. The accuracy with which such developmental equations predict the composition of the quantitative electroencephalogram in healthy individuals has been widely investigated and confirmed in a large number of research studies. These precise descriptions are independent of the ethnic or cultural background of the subject, confirmed in many countries around the world.83Neurometric quantitative electroencephalography has excellent test–retest replicability within an individual and has been demonstrated to have extremely high specificity as well as extremely high sensitivity to a wide variety of cognitive, developmental, neurologic, and psychiatric disorders.84,85Because it is also exquisitely sensitive to changes of state, it offers an ideal method to examine the effects on the brain of agents that alter consciousness, such as pharmacologic agents and anesthetics.

After extraction of quantitative electroencephalographic variables from any recording, they are preferably subjected to a mathematical transformation to achieve gaussianity of the normative distributions.86,87Transforming a set of measures sampled in a reference population into gaussian distributions, sometimes referred to as normal distributions , legitimizes the rescaling of each such measure to a standard or Z score, such that

where M = mean value of the measure in the population, S = the value obtained from an individual subject, and sd = the SD of the reference sample.

Z transformation rescales a variable into units of SDs from the mean value of the reference distribution. Because the transformation to gaussianity has forced 100% of the reference distribution to correspond to the familiar bell curve, calculation of the Z score for an obtained measurement defines the percentage of the reference population that lies more SDs away from the mean than the observation in question. Integration of the remaining area establishes the probability, P , that the observation could be obtained in the reference population by chance alone and can be found using any set of statistical tables. For example, P = 0.35 if Z = 1.0, 0.10 if Z = 1.64, 0.05 if Z = 1.96, 0.01 if Z = 2.56, 0.001 if Z = 3.20, and so forth. Thus, use of these methods for quantitative electroencephalographic analysis allows electroencephalographic recordings to be evaluated statistically by transforming each quantitative variable from its initial physical dimensions (voltage, covariance, latency, and others) into the metric of probability. The probability of a mean Z score for a group of n individuals can be assessed by multiplying the mean Z value by the square root of n and estimating the probability of n^1/2Z.

A further advantage of the transformation of univariate quantitative electroencephalographic measurements into Z scores is that variables in the common metric of probability can be legitimately combined into multivariate descriptors of brain state, and differences among such brain states can be further evaluated by multivariate discriminant functions or other similarly powerful objective methods. Such analyses have shown quantitative electroencephalographic evaluations to have high specificity, with false-positive findings at the chance level, and also to have high sensitivity to a wide variety of cognitive, neurologic, and psychiatric disorders.83,86,87 

The EP monitoring techniques rely on the fact that each peak in an EP waveshape reflects the excitation of huge numbers of neurons within successive levels of the transmission pathways from sensory receptors to the cortex. The waveshapes of EPs from any region elicited by multimodal sensory stimuli are also predictable88and sensitive.89,90The latency of each EP peak is determined by conduction time and the amplitude by the excitability of these neuroanatomical structures, and both of these parameters are similarly regulated. As a result, the morphology and scalp distributions of auditory, visual, and somatosensory evoked potentials in the resting, normal individual are well known and are used widely in clinical neurology to assess the integrity of sensory pathways.91 

The neuroanatomical structure of the homeostatic system regulating the quantitative electroencephalographic power spectrum is depicted in the much simplified schematic diagram in figure 2. The system shown here is adapted from a more complete diagram constructed by Hughes and John,92which also indicates the putative neurotransmitters mediating some of the indicated neural interactions. In this figure, dotted arrows indicate primary multimodal sensory inputs; solid arrows indicate excitatory relations; and heavy, dashed arrows indicate inhibitory interactions.

This complex neuroanatomical homeostatic system, probably genetically determined, regulates baseline levels of (1) local synchrony,93(2) global interactions among regions,94and (3) periodic sampling of the signal space.95A critical role in the regulation of brain rhythms is played by the interaction between brainstem, limbic system, thalamus, and cortex. The spontaneous electroencephalogram is conventionally considered to consist of rhythmic oscillations in several broad frequency bands, referred to as δ, θ, α, β, and γ activity. The following processes are believed to generate these distinctive rhythms.

The sensory receptors encode information about the environment. Multimodal sensory inputs (dotted arrows) go to sensory specific relay nuclei in the thalamus, serving as gates between the receptors and the cortex. Pacemaker neurons distributed throughout thalamic regions oscillate in the frequency range of the α rhythm (8–12 Hz, with a mean frequency of approximately 10 Hz), regulating and synchronizing the excitability of the cells in the thalamocortical pathways. This modulation is further distributed throughout the cortex by cortico–cortical interactions. Small spatially distributed cortical areas seem to act as epicenters from which α activity spreads through cortical neuronal networks by interneuronal connections, generating the α rhythm that dominates the resting quantitative electroencephalographic power spectrum seen in recordings from many scalp regions. This rhythm is often at different mean frequencies and phase in these dispersed regions.

To clarify this explanation, a brief discussion of the structure of the thalamic neural networks is necessary. A much more detailed explanation has been provided by Steriade96and by Lopes da Silva.97In the thalamic relay nuclei, some types of neurons display rhythmic oscillations in the frequency range of 6–10 Hz. This oscillatory behavior seems to be an intrinsic property of these neurons. Alpha activity arises from the interaction between populations of these neurons in the thalamus and in certain areas of the cortex. In the thalamic nuclei where this activity can be recorded, three main types of neurons interact: thalamocortical relay (TCR) nuclei whose axons project to the cortex, reticular nucleus (RE) neurons that interact synaptically with the TCR cells and contribute GABAergic inhibitory feedback control, and local intrinsic neurons. The TCR neurons display two distinct modes: either functioning as relay cells that depolarize and produce spikes in response to an adequate input volley from sensory pathways or as oscillatory cells that produce rhythmic bursts of high-frequency spikes repeated in a rhythmic oscillatory pattern. The mode of activity is determined by the resting membrane potential of the TCR neuron and the strength of its synaptic interactions with the sensory inputs and the RE neurons. Hyperpolarization of the TCR neurons by the RE neurons blocks transmission by TCR neurons of sensory input to the cortex and enhances slow rhythmic oscillations.

Essentially in the center of the brainstem, mesial to the lemniscal auditory and somatosensory pathways, lies the brainstem reticular formation. All afferent sensory pathways send collaterals into this region. Based on a series of lesion and stimulation experiments, it has been known for many years that bilateral transection of this system led to long-lasting coma. Further, electrical stimuli delivered to this region would awaken a sleeping animal and desynchronize the electroencephalogram in widespread regions of the cortex.98From such early observations, this system was denoted as the ascending reticular activating system  (ARAS). Cholinergic influences of the ARAS, as a consequence of sensory stimulation, diminish the efficacy of the GABAergic RE neurons, resulting in removal of their hyperpolarizing influences and facilitating throughput to the cortex. A concomitant of strong activation by the ARAS is cortical “arousal,” desynchronization of the α oscillators, with the appearance of faster rhythms in the β frequency range (12–25 Hz). Evidence that such arousal could also be accomplished by stimulation of the intralaminar nuclei of the thalamus has led some to modify this view, referring to an “extended reticular–thalamic activating system.”99It has been proposed that normal conscious functioning requires a circulating flow of activation among the ARAS, the intralaminar nuclei, and the cortex. Activity within a subcortical reticular–intralaminar nucleus–RE system is proposed to be necessary for the state of consciousness, whereas interaction of this system with cortex is envisaged to provide the perceptual content of consciousness.100,101 

By GABAergic action, RE neurons of nucleus reticularis, a thin shell of cells surrounding much of the thalamus, can inhibit these thalamic pacemaker neurons (thick dashed arrows), slowing their rhythms. Unless opposed, these inhibitory influences of the nucleus reticularis can act to hyperpolarize the pacemakers and diminish sensory throughput from the TCR neurons to the cortical receiving areas, slowing the mean frequency of the oscillators and shifting the α rhythm toward θ (4–8 Hz).

In parallel with these processes, a mesolimbic system receives multimodal inputs, from the ARAS in the brainstem and collaterals of afferent sensory pathways as well as via  the inferotemporal cortex, and distributes this activity to a system comprised of the entorhinal cortex, hippocampus, amygdala, septum (phasically inhibited by the hippocampus-thick dotted arrows), and anterior cingulate cortex. Representations of relevant past experience, stored in this system, are activated by associative linkages. This readout from the “endogenous system” is transmitted via  the non–sensory-specific, intralaminar, diffuse projection nuclei and nucleus medialis dorsalis, and by the anterior cingulate gyrus, to the axodendritic synapses of pyramidal neurons in upper cortical layers (layer 1). This nonsensory specific input produces the late secondary positive peaks that appear after the primary component of the EP at latencies that depend on the sensory modality. Strong activation of this system can generate widespread θ rhythms. A distinctive output from this system is often observed as “frontal midline θ” during performance of cognitive tasks such as delayed matching or mental arithmetic.

Extreme depression of thalamic gates releases some cortical cells from the influences of sensory specific input that, together with diminished activation of the cortex by the ARAS, results in the production of a very slow rhythm called δactivity  (0.5–4 Hz). Note that the cortex can inhibit the ARAS by descending pathways via  the striatum (heavy dotted arrows).

The ARAS receives inputs via  collaterals (double dotted arrows) of afferent activity from the sensory pathways. Activation of this system by incoming stimuli causes the brainstem reticular formation to inhibit the nucleus reticularis, opposing the GABAergic inhibitory action of nucleus reticularis by acetylcholine and releases its inhibitory actions on the thalamus. The frequency of the thalamic oscillators is increased into the 10- to 12-Hz region. Thalamic gates are opened so that afferent inputs from the “exogenous system” are transmitted from the sensory specific thalamic nuclei via  the projection pathways to axosomatic synapses of pyramidal neurons in lower layers of the cortex (layer 5). Information about complex, multimodal environmental events is fractionated, and “fragments of sensation” are distributed across the cortex into cell assemblies of feature extractors. Cortical activity is desynchronized in some regions thus activated, sometimes referred to as event-related desynchronization , and corticocortical interactions generate the β rhythm (12–25 Hz). This sensory-specific input also produces the primary or early positive components of the EP recorded from the scalp.

Information about current complex environmental stimuli and relevant previous experience, arising from the anatomically distinct exogenous and endogenous systems, converges on synapses in layers 1 and 5 of the anatomically extensive sheets of cortical pyramidal neurons. Experiments using direct electrical stimulation of pyramidal neurons with micropipettes,102as well as studies in brain slices combining direct electrical stimulation of specific and nonspecific thalamic nuclei with visualization of cortical responses using voltage sensitive dyes,103have directly demonstrated that the pyramidal neurons act as comparators, detecting temporal coincidence of inputs to layer 1 and layer 5 synapses. Direct stimulation of the soma of the pyramidal neuron or stimulation of a specific thalamic relay nucleus (ventrobasal) caused a moderate activation in layer 5 of the corresponding sensory cortex. Direct stimulation of the apical synapse or stimulation of nonspecific nucleus of the diffuse projection system (centralis lateralis) caused a moderate activation in cortical layer 1. When both somatic and apical synapses or ventrobasal plus centralis lateralis were stimulated concurrently, that is when exogenous and endogenous inputs were coincident, corticothalamic discharges were markedly enhanced, and activity at the γ frequency back-propagated to the cortical regions where coincidence had occurred. This feedback from the corticothalamic volley binds the distributed fragments and causes coherent corticothalamocortical loops to reverberate at the frequency of the γ rhythm (25–50 Hz). Such reverberation has been proposed by Thompson and Varela,104Rodriguez et al. ,105and Llinas et al.  103to play an essential role in perception.

These neurophysiologic interactions that regulate brain electrical activity are mediated by a wide variety of neurotransmitters. Brain functions are critically dependent on the availability of these substances and the processes that control their synthesis and metabolism, a complex topic beyond the scope of this article. The monitoring strategies implemented in quantitative electroencephalographic and EP anesthesia monitors are based on detecting electrophysiologic correlates of disruptions of this homeostatic regulation by different anesthetic agents, which exert their effects at various points of the system. Some of these distinctive effects are illustrated in this article.

Based on animal experiments, John106,107proposed that perception depended on the coincidence at the cortical level between exogenous, sensory-specific input of information about the environment and endogenous, non–sensory-specific readout from a representational system encoding memories of the relevant past. Confirmation in human subjects of this hypothesis was first provided by Libet.108In awake neurosurgical patients, Libet stimulated the cortex electrically to coincide with the time of arrival of the non–sensory-specific, secondary component of the EP waveshape, usually present in the response of the somatosensory cortex that was evoked by a mild electrical shock to the wrist, and found that such time-locked brain stimulation could block subjective awareness of the wrist shock. Similar findings during brain surgery were reported shortly thereafter by Hassler,109who recorded the cortical evoked responses to mild wrist shock, electrical shock to the sensory specific ventrobasal nucleus, and electrical shock to the non–sensory-specific nucleus centralis lateralis. The waveshape of the cortical response evoked by wrist shock consisted of an early component like the response to ventrobasal nucleus alone plus a later component like that to centralis lateralis alone. Hassler then showed that the perception by a patient of wrist shock could be blocked by later stimulation of centralis lateralis in the thalamus, appropriately delayed to disrupt the late cortical response.

Based on intracerebral recordings, several authors have reported that a brief period of γ activity with zero phase lag appears coherently between prefrontal cortex and parietal cortex of human subjects during performance of perceptual tasks.110,111Varela104has proposed that such spatially extensive, phase-locked γ oscillations may be the physiologic concomitant of perception. Coincidence detection may serve to convert fragmented sensory attributes of the complex stimuli to fragments of percepts, thus identified and encoded in distributed feature detecting cell ensembles. Coherent corticothalamic discharge of dispersed synchronized populations of cortical pyramidal neurons, and the resulting back-propagation from those neurons whose thalamocortical exogenous and endogenous projections resulted in coincidence detection, may bind these fragments together into a unified resonating system, which is the perceptual content of consciousness.103,112–114Such perception has been referred to by some as the “remembered present.”115A model of this process is shown in figure 3.

The blue arrows in figure 3depict afferent input to the exogenous system, propagating from the eyes and ears to the primary relay nuclei in the thalamus and thence to the axosomatic synapses of pyramidal neurons, in layer 5 of the auditory and visual cortices. Collaterals from these exogenous inputs go to the endogenous system via  the reticular formation and the mesolimbic system, activating episodic memories by associative mechanisms. The gold arrows in the figure depict the readouts of the multimodal aspects of these relevant memories, which are projected via  the non–sensory-specific nuclei of the diffuse projection system to the apical dendrites of the pyramidal neurons, in layer 1 of the multimodal sensory cortices. Coincidence between the inputs from the exogenous and the endogenous systems, converging on selected neurons in the sensory cortices, causes a coherent corticothalamic volley to impinge on the thalamic cells from which the ascending coincident influences arose. This volley causes a back-propagation and a thalamocortical–thalamic reverberation in the γ frequency range, depicted by the green arrows. Persistence of this reverberation engages the prefrontal cortex, and these simultaneous reverberating, coherent circuits become phase locked and resonant, binding the unimodal elements of sensation into a multimodal conscious perception. As has been pointed out by Mashour,39as the functions of binding dispersed elements of sensation into a unified perception provide the framework for integrating multiple simultaneous processes into a unified conscious experience, so might a paradigm of unbinding provide a framework for understanding the actions that underlie anesthesia.

In figure 4Aare shown the positions of the 19-electrode array located at positions defined according to a convention standardized for clinical electroencephalography, the International 10/20 Electrode Placement System. As in many of the head maps illustrated later in this article, the head is displayed as if seen from above, with the face at the top.

In figure 4Bis a recording made from such an array, recorded from an awake subject resting with closed eyes. This recording is contaminated by some artifactual potentials generated by movements of the eyeballs and the head, but an artifact-free segment can be seen between the arrows in the central region of this time period. To perform reliable intraoperative electrophysiologic monitoring based on accurate quantitative electroencephalographic measurements, such devices must include algorithms for automatic detection and removal of all artifactual contamination.

Figure 5illustrates the prototypic effects of anesthesia on quantitative electroencephalographic samples before and after LOC, induced by infusion of propofol. After computation of the absolute power spectra, the amount of power at each frequency has been Z-transformed, and the results are depicted as Z spectra. The Z spectra are illustrated for every frequency at every electrode position of the 10/20 System, labeled according to the usual convention and shown above each graph. The two horizontal lines are at Z =±2 SDs, the interval that represents the 0.05 probability level of significant departure from the reference distribution. The vertical lines in each small graph are at the strongest peak in the power spectra, and the number corresponding to the Z score for that frequency is indicated to the right of the electrode identification.

Figure 5Adepicts the Z spectra recorded from a typical patient at baseline, with a maximum deviation at approximately 12 Hz, perhaps reflecting the effect of a mild sedative premedication. Figure 5Bdepicts the mean Z spectra for a group of 15 patients shortly after LOC due to infusion of propofol, with a maximum deviation at approximately 1–3 Hz in every lead. In figure 5A, all preinduction baseline Z-spectral values lie between −0.1 and 1.7, within the 95% confidence interval of the normal range in each graph (i.e. , Z =±2.). After LOC, the Z scores at the frequency of 2 Hz lie between 3.6 and 6.8, far outside the normal range.

Note that in figures 5–7, all Z scores were calculated relative to the reference distributions of the same measures extracted from quantitative electroencephalographic recordings obtained from baseline recordings from 164 resting patients, before premedication. Note also that the significance of mean Z scores, averaged across a group of n individuals, can be estimated by multiplying the average Z value by the square root of the sample size.

Marked effects of anesthetic agents on a wide variety of quantitative electroencephalographic variables can be detected in all scalp electrodes during induction of anesthesia and maintenance at surgical plane with many different agents. Representative quantitative electroencephalographic effects of such agents are depicted on group averaged interpolated topographic Z score maps in figure 6. The Z scores were calculated relative to the distribution of these quantitative electroencephalographic variables in this population of 164 patients before the onset of induction.

Each map depicts interpolated mean Z scores of the indicated feature from 164 patients, averaged across all anesthetic protocols used for induction and surgical maintenance. Statistical significance for all maps in figure 6is encoded by using a single color palette. Different scaling is necessary to avoid saturation of the color encoding the actual statistical significance because the magnitude of anesthetic effects varied on different variables. The appropriate scaling constant for each variable is provided by the numbers in the rows marked ± Z, under rows 2 and 4 of the mean Z score maps. Each number specifies the Z values to be assigned to the positive and negative extrema of the color palettes used for the corresponding variable in the preceding two rows of head maps. Recall that statistical significance of a mean Z score is estimated by multiplying the mean value by the square root of the sample size, or 12.8 with n = 164.

The six columns of maps depict (1) absolute power in the θ band, (2) absolute power in the α band, (3) absolute power in the γ band, (4) interhemispheric coherence between symmetrical positions over the two hemispheres in the δ band, (5) interhemispheric coherence in the β band, and (6) interhemispheric coherence in the γ band.

The first four rows of maps depict row 1, induction (while counting on the operating table at the onset of induction); row 2, LOC (immediately after cessation of counting with loss of lash reflex); row 3, surgical plane (during maintenance at surgical plane just before onset of weaning); and row 4, ROC (immediately on return of response to loud verbal command). The last two rows encode the F value obtained using analysis of variance to compare row 5, induction versus  LOC, and row 6, surgical plane to ROC. The numbers below each map in row 5 and row 6 indicated the value to be assigned to F, encoded by the horizontal color bar, 0 to *.

Note that the maps in row 1 already show a significant shift of these variables from the baseline, which would otherwise be encoded as the dull hues close to black on the color bar. These changes reflect the slowly increasing effects of the administration of the inducing agents, altering brain state with increasing sedation but still compatible with the patient systematically counting backward aloud, a cognitive activity that itself is accompanied by shifts from the resting baseline. The dramatic changes between rows 1 and 2 occur very abruptly with cessation of counting and the disappearance of the lash reflex. Note that many of the changes brought about with LOC caused by the set of inducing agents, seen in row 2, remain substantially unchanged during maintenance at surgical plane by quite a different set of substances, as seen in row 3. Finally, some but not all of these effects reverse with ROC, as some of the maps in row 4 again become similar to those in row 1, during induction.

The states and agents represented in figure 6include induction with methohexital, etomidate, or propofol and maintenance with total intravenous concentration, gases (desflurane, isoflurane, or sevoflurane), or nitrous/narcotic. The mapped features were selected from a much larger set to illustrate the great variety of quantitative electroencephalographic variables that change invariantly and reversibly with the LOC and ROC under the influence of agents commonly used during surgical anesthesia.

These data show clearly that monitors of the depth of anesthesia using quantitative electroencephalographic variables have a rich panoply of effects on which to base their assessment of the depth of anesthesia. The task of development of anesthesia monitoring algorithms is to select the most robust set of variables from among these many candidates that are both necessary and sufficient to provide the most accurate and sensitive assessments of patient state across the widest range of chemical agents, with the most rapid and reliable detection of the changes relevant for the optimal clinical management of each case.

Coherence provides a measure of the functional interaction or coupling between two brain regions. Administration of anesthetic agents causes marked changes in coherence, which are of particular interest because of the hypothetical role of coherence in binding that has been proposed in the theoretical formulation presented above in the discussion illustrated by figure 3.

Figure 7presents the results of more extensive computations of coherence as well as power that were carried out on the 164 patients whose Quantitative electroencephalographic data were partially presented above. All data shown in figure 7represent mean Z scores relative to the distribution of the resting baseline values in the 164 patients before onset of induction; the significance of these mean Z scores can be estimated by multiplying the average Z value by 12.8.

Figure 7Aillustrates that the frontal lead F3 and the occipital lead O1 demonstrate an increase in power and a relative change in the power gradient at LOC in all frequency bands except γ, showing enhanced anteriorization, i.e. , the increase in mean Z scores for δ, α, and β activity is greater for the frontal than for the occipital lead. Changes in θ power essentially paralleled δ and are not illustrated. These power changes gradually diminish during emergence.

Figure 7Bshows that coherence of γ, but not of δ, α, or β, between F3 and O1 increases during the mental task of counting backward during induction. Shortly after LOC, coherence in all bands decreases significantly and remains diminished during maintenance at surgical plane. As emergence begins, although there is no marked increase in power, coherence of β and γ returns to the levels seen while counting during induction.

These data suggest that anterior and posterior brain regions become functionally uncoupled with the LOC caused by the action of anesthetics and remain uncoupled during surgical maintenance. Consciousness returns, as evidenced by the opening of eyes on a loud command, as coupling between prefrontal cortex and sensory regions is restored to previous levels in β and γ frequencies. The most striking coherence changes as emergence progresses are in γ, for which sharp increases in coherence begin to appear several minutes before opening of the eyes. While similar uncoupling occurs in all frequency bands, recoupling of δ and α occurs only several minutes after responsive consciousness is restored. These findings are particularly interesting in view of some of the reports cited above, in the discussion of the role of coincidence detection in binding, of prefrontal–parietal phase-locked γ coherence during performance of perceptual tasks.

We interpret these data to suggest that anesthesia interrupts and consciousness restores the corticothalamocortical reverberation resulting from detection by the pyramidal neurons of coincidence between the exogenous report of sensory specific inputs and the endogenous readout of episodic memories, endowing the sensations with meaning. Whether this correlation is unique to γ coherence or is also a property of coherent activity in the β range awaits inquiries specifically designed to answer that question.

In figure 8Ais depicted the normal waveshape of the auditory evoked response and the neuroanatomical structures presumed to generate the successive components. These include the brainstem auditory evoked response, with peaks I (acoustic nerve), II (cochlear nucleus), III (superior olivary complex), IV (trapezoid body), and V (inferior colliculus) and the early cortical or midlatency evoked response, with peaks No and Po (medial geniculate), Na, Pa, and Nb (early responses of primary auditory cortex), and P1 (also sometimes referred to as Pb ), and later components are believed to reflect activity in association areas and frontal cortex. In figure 8Bare shown group averaged MLAERs, obtained in surgical procedures using total intravenous anesthetic (n = 19), isoflurane (n = 17), desflurane (n = 6), and nitrous/narcotic (n = 20). Note the lengthening of the Pa–Nb–Pb interval with LOC caused by each of these agents and the diminished amplitude and longer latency of subsequent components. Note that the preinduction Pa–Pb interval is approximately 25 ms, which is the period of a 40-Hz oscillation (γ). In figure 8Care depicted the set of power spectra computed from a simulated series of EP templates; these simulated waveshapes were morphed using the MLAER waveshapes from preoperative baseline to loss of consciousness recorded during induction with desflurane, as shown in the middle panel. Note the gradual disappearance of the 40-Hz component from the resulting spectra. These results suggest that perhaps the change in the MLAER waveshape reflects blockade of a 40-Hz thalamocortical back-propagation.

As noted above, the 40-Hz SSEP has been explored as a possible index of anesthesia. The 40-Hz SSEP waveshape is a rhythmic oscillation at 40 Hz, reflecting the entrainment of the small early components of the MLAER by the much larger components of the subsequent cortical auditory responses to a series of closely spaced auditory stimuli. Many anesthetics disrupt the 40-Hz SSEP, seen as the progressive diminution of amplitude with increasing amount of anesthetic agent.

This effect is illustrated in figure 9, showing the changes in ASSR with stepwise titration of the amount of anesthetic, concomitant with clinical assessments of the depth of anesthesia using the Observer’s Assessment of Anesthesia and Sedation scale. Group grand averages (n = 15) were constructed by averaging 40-Hz auditory steady state evoked potential averages recorded from the 19 electrodes of the International 10/20 System, obtained from volunteer subjects during stepwise titration of anesthesia using propofol. Each array is depicted as if viewed from above, with the face or nose at the top and the left side of the subject on the left. The numbers at the top left of each array refer to the approximate minimum alveolar concentration (MAC) equivalent, but LOC and ROC were established using the usual clinical criteria. The three arrays in the left column, from top to bottom, depict the steady state evoked potential at increasing depths of anesthesia, pre, descending LOC (approximately 0.4 MAC), and 1.0 MAC, and the three arrays in the right column depict the steady state evoked potential at 1.4 MAC, 0.4 MAC, and ROC, defined by opening of the eyes in response to a loud command.

To round out this overview of the effects of anesthetics on the brain, it is useful to examine recent findings obtained in functional brain imaging studies.

Using evidence from PET scans to visualize changes in brain regional cerebral metabolic rate (rCMR) during general anesthesia with halothane and isoflurane, as reflected by uptake of radioactively labeled (F¹⁸) 2-deoxyglucose, Alkire et al.  116addressed two questions: (1) Did the regional effects of halothane and isoflurane, examined with this method, show commonalities of action on brain metabolism that help to understand the mechanism by which these agents produce unconsciousness, and (2) will the common effects of these two substances be suppression of thalamocortical activity? Subjects were 11 healthy volunteers who each underwent two separate PET scan procedures separated by at least 1 week. The baseline awake regional cerebral metabolic rate of glucose was assessed in the first scan. The second scan assessed metabolism during a period of unconsciousness, established by unresponsiveness to verbal commands, prodding, or shaking, induced either with halothane (n = 5) or isoflurane (n = 6) general inhalation anesthesia. The mean expired halothane and isoflurane concentrations were 0.7 ± 0.2 and 0.5 ± 0.1%, respectively.

For all subjects, general inhalational anesthesia induced both a global reduction and specific regional reductions of brain glucose metabolism. No brain regions were found that increased their metabolism during anesthesia. The mean global or whole-brain rCMRglu was reduced by 42% (± 13%) during isoflurane and 40% (± 10%) during halothane anesthesia. Effects common to both agents were observed and included a significant reduction of rCMRglu in the cuneus, thalamus, midbrain reticular formation, dorsolateral prefrontal cortex, medial frontal gyrus, inferior temporal gyrus, cerebellum, and occipital cortex.

The results were interpreted to arise from hyperpolarization block of the TCR neurons in thalamic networks and a transition from the state enabling thalamocortical throughput of sensory information to an inactivated state in which thalamocortical gates are essentially closed. These thalamic networks are discussed above in the section on the generation of α rhythms. These authors envisaged that anesthetics cause neuronal changes coincident with LOC by three possible mechanisms or a combination thereof: (1) direct hyperpolarizing effects on thalamic and cortical cell membrane potentials117; (2) suppression of midbrain/pontine areas involved with regulating arousal, removing excitatory inputs to the thalamocorticothalamic loops by inhibiting glutamatergic and cholinergic neurotransmission118; (3) enhancement of GABAergic synaptic neurotransmission mediating inhibitory circuitry within thalamocortical loops119; and (4) that different anesthetics might use only one of these proposed mechanisms, but some agents might use various combinations of these different mechanisms. No matter which process is initiated by the agent, these authors suggest that any substance or event that pushes thalamocortical cells toward hyperpolarization, no matter what the mechanism, will cause LOC.

Their initial hypothesis, based on neuroanatomical and neurophysiologic evidence, was that thalamic and midbrain reticular formation activity is needed to maintain function in a specific set of neurons essential for consciousness. They argue that these findings place the most important site of anesthetic action not in the reticular formation itself, but rather in the thalamic gated regions regulated by the reticular arousal centers. They proposed, consonant with the extended reticular–thalamic activating system theory proposed by Newman and Baars,99that those regions that generate consciousness might have been identified in their experiments.

Changes in regional cerebral blood flow (rCBF), determined using H²O¹⁵PET imaging, can be used to visualize changes in brain oxygen utilization by neural circuitry, reflecting the effects of centrally acting drugs. This method was used by Veselis et al.  25,26to study the effects of infusions of two different concentrations of midazolam on memory and rCBF in 14 healthy volunteers. Midazolam has an agonistic action at the GABA^Areceptor complex and is widely used clinically to accomplish temporary sedation and amnesia. Subjects were randomly assigned to receive computer-assisted continuous midazolam infusion intended to produce either a “low” or a “high” effect. The initial effect site target concentrations were 7.5 ± 1.7 mg midazolam (serum concentration = 74 ± 24 ng/ml) for the low-effect group and 9.7 ± 1.3 mg (serum concentration = 129 ± 48 ng/ml) for the high-effect group.

Subjects were then classified into low- and high-effect groups based solely on analysis of the quantitative electroencephalograms obtained during PET scanning. These effects were identified using spectral analysis of the quantitative electroencephalographic activity during the PET imaging, depending on the presence or absence of a clear peak at 14 Hz in the power spectrum arising from spindle activity. Effects were localized using statistical parametric mapping with respect to standard stereotaxic coordinates. Relative changes in rCBF were calculated after the absolute changes in global rCBF were determined. Participants were asked to memorize a list of 16 words presented verbally at baseline, before onset of drug infusion, and during midazolam infusion. The degree of memory loss was measured by the number of words retained. Participants in both groups experienced memory loss during midazolam administration. From the 16-word list presented during infusion, the mean numbers of words recognized at the end of the study day were 5 ± 4.5 (P < 0.002) and 1.6 ± 1.7 (P < 0.001) for the low-effect and high-effect groups, respectively.

The rCBF changes in the low-effect group were a subset of those observed in the high-effect group. Regions where the decreased in rCBF reached significance at the P < 0.001 level included the insula, the thalamus, the cingulate cortex, multiple areas of the prefrontal cortex, and the parietal and temporal association areas of the cortex. Conversely, rCBF was increased in the occipital areas. The authors concluded that midazolam sedation alters normal function of brain regions associated with arousal, attention, and memory.

Levels of midazolam as high or slightly higher than were used in the high effect group in this study can be expected to cause deep sedation and unresponsiveness, with severe obtunding of consciousness. The depression of the thalamus can be attributed to the agonistic enhancement of GABA action by nucleus reticularis on the TCR neurons, which would block throughput to the axosomatic synapses on the lower layers of the cortex. The observed effects also indicate interference with the outflow of the mesolimbic system to the prefrontal cortex and other neocortical regions via  the anterior cingulate and the medialis dorsalis nucleus of the thalamus. This would block input to the axodendritic synapses of the pyramidal neurons in upper layers of the cortex. The net effect would be to prevent coincidence detection by the pyramidal neurons of converging inputs from the exogenous and endogenous systems onto the cortex. Because prefrontal–parietal interactions are involved in focusing attention and cognitive processing, these effects are compatible with interference with attention and memory retrieval. The apparent increased rCBF in the occipital areas may in fact indicate that the occipital areas were relatively unaffected by midazolam, in view of the fact that the global rCBF decreased; that is, the images depict relative values, not absolute values, i.e. , percent of total.

If one knows that there are a number of specific voltages at particular points in space, together with certain electrical parameters of the space, it is possible to describe the potential field that will be generated in space by that set of voltage points, and the solution is unique. However, the converse is not true, i.e. , if one knows the distribution of potentials in space, or the field, especially if the number of generators is more than one, there are essentially an infinite variety of spatial distributions of voltage sources that will generate the same field. The mathematical problem of locating within the brain the most probable neuroanatomical sources of electroencephalogram or EP surface potentials is referred to as the inverse problem . Until recently, this problem was considered mathematically insolvable. By making a number of reasonable assumptions and imposing a number of reasonable constraints, several methods have been developed recently for this purpose.

One three-dimensional electroencephalographic source localization method that has been developed superimposes the computed generators of electroencephalographic voltages recorded from the scalp on images of brain slices taken from a Probabilistic MRI Atlas constructed at Montreal Neurologic Institute (Montreal, Quebec, Canada). This method is referred to as variable resolution electromagnetic tomography  (VARETA).120The solutions place the “proportional” positions of scalp electrodes in the International 10/20 System in registration with the “proportional” Probabilistic MRI Atlas to coregister the calculated three-dimensional quantitative electroencephalographic source locations with neuroanatomical images, providing a “virtual functional MRI” in so-called Talairach space. Methods like VARETA solve the formidable problem posed by the uncertainty of mathematical solutions to the problem, if approached with no constraints, by two major conceptual breakthroughs: (1) neuroradiologists have constructed “masks” that identify those voxels within white matter or cerebrospinal fluid space that can arbitrarily be excluded from candidacy as putative voltage sources, and (2) regularization parameters define the amount of “smoothness” that must be imposed on the solutions, thereby arbitrarily restricting the permitted gradients around voxels where sources are located. Numerous studies have found images computed using these various three-dimensional methods to localize sources of quantitative electroencephalographic abnormalities to be in good correspondence with images of neuropathology obtained using computed tomography, magnetic resonance imaging, and PET imaging techniques. Furthermore, a large quantitative electroencephalographic normative database has been used to compute the expected distribution in more than 3,500 brain voxels, which are cubic volumes 7 mm on a side, of the sources of power from 0.37 to 19 Hz. The statistical significance of deviations from the expected source strength of any voxel at any frequency can now be encoded in color, yielding functional neuroanatomical images that can be interpreted like a neurometric quantitative electroencephalographic map.

We have used VARETA to visualize the neuroanatomical regions that are the probable sources of changes in quantitative electroencephalographic activity that are invariant and reversible with the LOC and ROC independent of the anesthetic agents used for induction or maintenance of the anesthetized state.121While numerous quantitative electroencephalographic changes take place with anesthesia, as shown previously in figure 6, significant increases of absolute power at 3.5 Hz are perhaps the most reliable quantitative electroencephalographic changes that accompany LOC with anesthesia, and activity at this frequency is indicative of significant inhibition. Therefore, we chose to report findings at this frequency using this method. Representative VARETA images at 3.5 Hz are presented in figure 10.

Figure 10shows average VARETA images of groups of patients (n = 15 in each panel) in five stages verified by standard clinical practice: (1) at rest on the operating table before induction; (2) counting just before LOC, during induction with three different agents; (3) just after LOC; (4) just before ROC, during emergence from maintenance at surgical plane using three different agents; and (5) just after opening their eyes in response to a loud verbal command using their name. All images were based on the localization of quantitative electroencephalographic power at 3.5 Hz, one of the frequencies that displayed the most distinctive and consistent set of changes (heterogeneity of variance across states, homogeneity within each state across all agents). The frequency is at the margin between high δ and low θ and can be considered to reflect significant inhibitory effects.

The statistical significance of the changes in each voxel of these images is color coded relative to the mean value and SD (voxel Z score) of the distribution of resting power in that voxel at that frequency, calculated from the corresponding distribution of values in 176 patients in the preoperative resting state. Hues of red through yellow indicate increases, and hues of blue through turquoise indicate decreases from the mean value.

The results show 3.5-Hz power increased significantly (indicating relative inhibition) in a set of regions including the orbital prefrontal cortex, dorsolateral prefrontal cortex, anterior cingulate gyrus, basal ganglia, amygdala, postcentral gyrus, hippocampus, and thalamus. This set of structures is in good correspondence with those showing most marked changes in the rCMR and rCBF studies discussed above.26,116The depression of the thalamus can reflect inhibition by the GABA action of nucleus reticularis, itself released from inhibition due to depression of the midbrain reticular formation. Inhibition of the TCR neurons would block throughput to the axosomatic synapses on the lower layers of the cortex. The depression of amygdala, hippocampus, basal ganglia, and anterior cingulate indicates the suppression of mesolimbic system outflow to the prefrontal cortex and other neocortical regions. This would block input to the axodendritic synapses of the pyramidal neurons, in upper layers of the cortex.

Therefore, the results discussed in this section indicate that three different functional imaging methods, evaluating brain changes caused by a variety of different anesthetic agents, detected a common effect. This effect is prevention of coincidence detection by the pyramidal neurons of converging inputs from the exogenous and endogenous systems onto the cortex.

Based on the evidence presented above, we propose that the neurophysiologic effects that produce amnesia and loss of awareness due to the action of anesthetics occur in six steps:

Step 1: Depression of the brainstem reduces the influences of the ARAS on the thalamus and cortex.

Step 2: Depression of mesolimbic–dorsolateral prefrontal cortex interactions leads to blockade of memory storage.

Step 3: Further depression of the ARAS releases its inhibition of the nucleus reticularis of the thalamus, resulting in closure of thalamic gates (especially in the diffuse projection system) by hyperpolarizing GABA-mediated inhibitory action of the nucleus reticularis (θ increase), thereby blocking

Step 4: Thalamocorticothalamocortical reverberations and perception (γ decrease), so that

Step 5: Parietal–frontal transactions are uncoupled (γ coherence decreases), blocking cognition, and

Step 6: Prefrontal cortex is depressed to reduce awareness (increase of frontal δ and θ).

These steps are described in figure 11.

Although one may consider these six steps as a hierarchical sequence, it must be kept in mind that reciprocal pathways interconnect all of the neuroanatomical structures engaged in this cascade. There are many ways that amnesia and blockade of awareness can be accomplished. Agents that cause depression of the ARAS in the brainstem can reduce activating and arousal influences on the specific and nonspecific thalamic nuclei and the cortex. Depression of the ARAS can also release complex GABAergic inhibitory effects in the limbic system due to diminution of ARAS inputs that produce cholinergic inhibition of GABA. There seems to be a level of such agents that diminishes interactions of the mesolimbic circuits, including the amygdala, hippocampus–cingulate cortex with the dorsolateral prefrontal cortex, to block the storage of memory and achieve amnesia for ongoing events. A somewhat greater depression of the brainstem releases the nucleus reticularis from the inhibiting influence of the ARAS. This can lead to closure of thalamic gates due to the inhibitory action of nucleus reticularis, resulting in diminished cortical input. Decreased thalamocortical input may result either by loss of activation from the ARAS or by dynamic inhibition via  nucleus reticularis. Inhibition of either the cortex or the non–sensory-specific, diffuse thalamic projection nuclei blocks the corticothalamocortical reverberations hypothesized to be critical for awareness. Depression of the parietal cortex interrupt the prefrontal–parietal transactions critical for perception. Inhibition of the prefrontal cortex releases its modulation of nucleus reticularis, resulting in defocusing of attention and reducing activation of the systems mediating speech and movement. The level of vigilance can be lowered by influences from the striatum or substantia nigra imposed on the reticular formation, increasing the threshold for arousal.

We have attempted to determine how abruptly the brain state changes with LOC due to anesthesia. These studies have been limited by the temporal resolution of our equipment and the VARETA software currently available to us. Although they must be considered tentative and preliminary, the results lead us to an estimate on the order of 10–20 ms, for the change in state shown by the quantitative electroencephalographic alterations on LOC, as illustrated in figure 6, and in the underlying dispersed system of anatomical regions, illustrated in figure 10. It should be pointed out that such speed may make more plausible and attractive the notion of LOC as a “phase transition,” as proposed by Steyn-Ross et al.  122Even though these steps may occur almost instantaneously, we propose that steps 1–6 are arranged in a sequence that corresponds to their sequential activation and thus constitute a “cascade.” For example, it may be that the release of corticofugal inhibitory influences on the thalamus, as ascending reticular activating influences on the cortex (and perhaps inhibitory reticulohippocampal influences) are blocked in step 1, contributes to the behavioral excitation evident in delirium or stage II anesthesia. The blockade of memory by low doses of anesthetics shown by Veselis et al.  24may reflect blockade of long-term potentiation in the hippocampus and the depression of septal–hippocampal outflow to the prefrontal cortex as proposed in step 2. The closing of the gates of the thalamic diffuse projection system in step 3 nonetheless allow primary sensory thalamocortical inputs to continue, as reflected by preservation of the primary components of cortical evoked potentials in anesthesia or coma. Synchronization of interactions within cortical ensembles may generate some level of local γ activity even though interaction with other cortical regions is blocked. While uncoupling of corticocortical transactions is proposed in step 5, the implication is that neural processing of data encoding sensory fragments may continue although the formation of complex representations may not. Effects initiated at any level of the system rapidly propagate both upward and downward through the brain to modulate other parts of this interactive network. There is no unique neuroanatomical structure at which action is both necessary and sufficient for an agent to accomplish modulation of the level of awareness. Similarly, there is no unique molecular agent acting on a special cellular target that can initiate this process. However, by slow induction of anesthesia, it may be possible to demonstrate whether steps 1–6 listed above occur in the indicated sequence. The actions of anesthetic agents to achieve immobilization seem to require concentrations to be further increased, beyond the blockade of these interactions that is necessary to suppress awareness, to depress motor neurons in the spinal cord.

Particular importance must be attributed to the intraoperative observations of the effects of anesthetics on coherence in the various frequency bands of the quantitative electroencephalographic power spectrum. These effects arise from the rapid initiation of steps 4–6. It may not be possible experimentally to demonstrate separability of the processes postulated in these steps, because of their interdependence. It has been proposed that the lower the frequency is, the more remote the brain region is from which the influence arises,123,124evidence like that summarized previously and more fully elsewhere,112,113indicates that neural synchronization is present from extremely small to global scales and plays a critical role in assembling dispersed information into a seamless conscious experience. Using indwelling electrodes in epileptic patients, intracerebral electroencephalographic recordings reveal progressive suppression of γ activity in the hippocampus with increased concentrations of sevoflurane,125a finding buttressed by subsequent neurophysiologic studies in animals that showed septal–hippocampal inactivation and suppression of γ activity by both volatile and nonvolatile anesthetics.126We have elsewhere proposed that the process that generates γ activity in the hippocampus is the initial source of the endogenous input to the apical dendrites of the coincidence detector depicted in figure 3.127As Mashour39has emphasized in discussing and extending our findings, such evidence supports the concept of cognitive unbinding as a final common mechanism for anesthesia, occurring at many different levels ranging from convergence at the cellular level to interruption of synchronization within an ensemble assembling dispersed fragments of information within a system to binding the state of many systems into conscious awareness, and reversing the elements of brain interactions that produce cognitive binding.

Some of the original data contained herein was collected in collaboration with Lavern Gugino, Ph.D., M.D. (Associate Professor, Department of Anesthesia, Brigham and Women’s Hospital, Harvard Medical School, Boston, Massachusetts), and analyzed in collaboration with Robert Isenhart, B.S. (Assistant Research Scientist, Department of Psychiatry, New York University School, New York, New York). Some of the analyses performed on data reported herein were done in collaboration with Arnaud Jacquin, Ph.D. (Senior DSP Engineer, Everest Biomedical Instruments, Chesterfield, Missouri).