Validation of a new approach for distinguishing anesthetized from awake state in patients using directed transfer function applied to raw EEG

Objective quantification of how EEG signals change in relation to subjects’ states of consciousness has a long history [1‐3]. Recently, measures quantifying properties such as complexity [4], functional and effective connectivity [5], and information content [6] in signals recorded with electroencephalogram (EEG) have been used successfully for objectively distinguishing between conscious and apparently unconscious states in humans. Generally, an apparent loss of consciousness is related to changes in such EEG signal properties or a combination of them [7], but capturing the relevant changes in a way that makes the measures useful for bedside monitoring of the level of consciousness in patients is not straightforward.

Recently, we published results indicating that the directed transfer function (DTF)—a measure that can be used to quantify sensor space directed connectivity from EEG recordings of spontaneous brain activity—changed abruptly as patients undergoing surgical propofol anesthesia apparently lost and regained consciousness [8]. The observed changes could be used to classify the state of individual patients as awake or anesthetized with 98% accuracy with a temporal resolution of 1 s. Importantly, these results were obtained using EEG data recorded in a normal clinical setting (general anesthesia for surgery), and without requiring any data cleaning, indicating that the DTF may be useful for developing objective, real-time monitors of patients undergoing anesthesia.

The previous study was done on a population of patients undergoing a single anesthetic protocol during surgery—propofol anesthesia with the opioid analgesic remifentanil. Thus, it is conceivable that the differences between the awake and anesthetized state observed in that study merely reflected changes related to the specific anesthetic agent rather than the changes related to general anesthesia, or loss of consciousness, in general. If so, the DTF-based method may not be fit to distinguish between states of consciousness more generally. However, qualitatively similar findings have been reported when using DTF to assess brain connectivity in groups of patients suffering from disorders of consciousness [9], and healthy individuals falling asleep [10‐12]. Taken together, this suggests that the DTF calculated from EEG may consistently change between conscious and unconscious states, regardless of how the change in state of consciousness came about.

Here, we report from a follow-up study, designed to test whether the DTF-based approach presented in our previous study can also be used to distinguish between awake and anesthetized states in patients undergoing general anesthesia caused by sevoflurane. Thus, we test whether the changes observed in the DTF derived connectivity parameters in the propofol study, were also apparent for patients undergoing sevoflurane anesthesia, and whether the changes could once again be used to successfully classify the state of the patients in accordance with the clinician’s judgement of their state of wakefulness.

The final data material comprised 7 patients (3 female) with a median age of 48 years (range 41–56 years). One patient (patient #6) was excluded from the analysis because large portions of the data were corrupted. During the induction phase, the patients were intravenously given 350 µg (range 250–450 µg) of the analgesic drug fentanyl, and the percentage of sevoflurane in the end-tidal volume was measured and adjusted (measured range 2.6–5.8%, mean 4.2%). After the patients became unresponsive no further doses of fentanyl were given. The median sevoflurane concentration during the maintenance phase of anesthesia was measured to 2.0% (range 1.7–2.4%) of end-tidal volume. Throughout the surgical procedure, physiological variables followed the expected development during anesthetic induction, maintenance, and emergence (Fig. 1).

From each of the patients, continuous EEG was recorded throughout the clinical procedure (median length: 161 min, range 115–247 min). Every recording contained segments from before (median length: 9 min, range 7–16 min), during (median length: 148 min, range 105–232 min), and after anesthesia (median length: 3 min, range 2–5 min). In the automatic artefact scanning process, 13% (range 2–39%) of all epochs from the wakeful periods were marked as artefactual, while 6% (range 4–17%) of epochs were marked from the anesthesia period. In addition, a variable number of channels (median: 2, range 0–4) channels were marked as artefactual.

The DTF analysis showed a qualitative difference between the ‘awake’ and ‘anesthetized’ states. The mLDTF topography was heterogeneous in the ‘awake’ state, but more homogenous in the ‘anesthetized’ state (Fig. 2b). A relatively strong apparent source of information outflow could be seen located over the posterior midline channels, whereas this region was far less distinct in the plot for the ‘anesthetized’ state. Qualitatively speaking, this looks similar to the changes observed in our previous work with patients undergoing propofol anesthesia (Fig. 2c), and lends support to our previous finding that the awake state is associated with a more heterogeneous pattern of outgoing information flow than the anesthetized state.

The mLDTF topographies from each individual, in both awake and anesthetized states, can be seen in Fig. 3. These topographies show the median LDTF values for each patient, across all epochs in a particular state (‘awake’ or ‘anesthesia’). Permutation tests indicate that the qualitative differences observed between conditions were statistically significant (p = 0.016; 2⁷ = 128 permutations) for the patients undergoing sevoflurane anesthesia. A significant change was also observed when comparing the awake with the anesthetized condition in the data from the patients undergoing propofol anesthesia (p = 0.016; 2⁷ = 128 permutations). This was also the case when pooling together the data from the two experiments, indicating that the differences observed here might be similar to those observed in our previous study (p = 4.0*10⁻⁴; 2¹⁴ = 16,384 permutations). In all tests, the natural grouping of patient labels (anesthesia vs awake) was the grouping with the largest overall difference between the groups of all possible permutations, indicating the strongest possible evidence for differences between groups given the number of samples. Furthermore, when comparing within conditions, between experiments (propofol anesthetized vs sevoflurane anesthetized and propofol awake vs sevoflurane awake), the test yielded non-significant results (p > 0.05 for all 100 runs; 2⁷ = 128 permutations). Thus, there was little or no evidence for differences between mLDTF values in the patient groups from the two studies, in either the anesthetized or the awake state, and we cannot reject the hypothesis that the mLDTF topographies in comparable behavioral states were not different between the two experiments.

The mLDTF patterns remained relatively stable over time within states, but of the patterns transitioned abruptly when the patients’ state of wakefulness changed (Fig. 4). This was the case both when patients transitioned from wakefulness to anesthesia (Fig. 4, left column) as well as when the patient regained consciousness after anesthesia (Fig. 4, right column). However, there was no distinct change in the mLDTF pattern around the time when the anesthesia delivery was stopped (Fig. 4, middle column). Importantly, the abrupt change in mLDTF pattern was clearly reflected in the algorithm’s ‘classification confidence’ (red lines in Fig. 4), which appears to transition between two distinct states right around the transitions between wakefulness and anesthesia.

From the LDTF values, the algorithm’s confidence (C) in classifying a patient as awake in a given one-second epoch was calculated to form a basis for classification. Varying the threshold value required for classifying a patient as awake resulted in a maximum overall accuracy of 96.8% (SE = 0.63%) for threshold values of 0.1 < C < 0.17 (Fig. 5). While sensitivity monotonically increases, and specificity monotonically decreases, for larger cut-off values, the classification accuracy remained high for a broad range of threshold values, only dropping below 96% for C < 0.01 and C > 0.9. However, when we optimized for maximized sum of sensitivity and specificity, the range of valid threshold values shrank significantly, and moved towards lower values of C. Specifically, the optimal threshold value was C = 0.001, yielding an accuracy of 95.1% (SE = 0.001), with sensitivity of 98.4% (SE = 0.80%) and specificity of 94.8% (SE = 0.10%). The results of the classification using the optimal threshold value is shown in Fig. 6 to give an impression of the temporal stability of the mLDTF patterns and the accuracy of the classification for every patient.

This study shows that our DTF-based measure yields similar results when applied to EEG recordings obtained from patients undergoing general anesthesia with sevoflurane and with propofol. Like we found with propofol anesthesia in our previous paper [8], the source strength topographies went from having relatively stronger apparent sources of information outflow in the posterior region of the head in the ‘awake’ state, to becoming far more homogeneously distributed in the ‘anesthetized’ state during sevoflurane anesthesia (Fig. 3). The differences between conditions (awake vs. anesthesia) were significant for both anesthetics (propofol and sevoflurane), and there was no significant difference between the changes observed with the two anesthetics. Taken together, this indicates a common change being captured by the LDTF when comparing the awake and anesthetized conditions, using different anesthetics, under slightly different conditions, and using slightly different analysis. Finally, the changes in DTF-based parameters could be successfully used to classify the patients’ states as awake or anesthetized in accordance with the clinician’s judgement with a 1-s temporal resolution.

Even though these results are similar to what was found in our previous study, there were some differences in the analysis that should be mentioned. In addition to adding an automatic artefact-scan (only used for visualization/reporting), we changed two aspects in the analysis: in the present study we (1) focused on the theta rather than alpha frequency band, and (2) optimized the classification using the parameter C—the ‘confidence of classifying an epoch as coming from the awake distribution’.

We changed the frequency band used for DTF analysis from the alpha band, which we used in our previous study [8] to the theta band, to reduce potential confounds caused by changes in the alpha band power related to closing of the eyes (patients had their eyes open in the awake state but closed in the anesthetized state). The patterns of DTF derived connectivity observed in the two bands were relatively similar in our previous analyses leading us to believe that it would be possible to successfully use the theta band for the purpose of distinguishing between the states. Indeed, this change seems to yield a small benefit in reducing one potential confound, without causing any clear disadvantages or detrimental effects to the classification. Similarly, the choice of optimizing the threshold parameter, C, seemed to improve the analysis compared to use the more naïve approach used for classification from our previous paper. Interestingly, this optimization had quite a strong effect on the quality of classification in the current paper. Even though the accuracy of classification would have been similar using the naïve approach, the sensitivity was improved when using the optimized threshold (see Fig. 5). In fact, the sensitivity would have been worse than in our previous paper if we had not used a different threshold (only ~ 70% of the epochs from awake patients would have been correctly classified, in contrast to the 95% in our previous paper). However, using the optimized threshold, C = 0.001, instead yielded classification results that were slightly better than in our previous study [8].

Another difference from the previous study was that the patients were asked about whether they experienced anything during the anesthesia and surgical procedure, to investigate whether any of them would have recollection of waking up, dreaming, or otherwise being conscious. However, to avoid inducing traumatic memories, the depth of the questioning was kept to a minimum. None of the patients reported memories of waking up during the surgery, but two patients (#1 and #8) reported having had simple dreams during the anesthesia. However, their timelines of typical outgoing information flow (mLDTF) did not show any sign of awake-like patterns during the anesthesia. There might be several reasons for this, but we do not have sufficient data to make strong conclusions here. For example, our measure may be insensitive to the dream state, the dreams may have occurred during emergence or have been confabulated, or the patients may have had brief periods of waking up that were later interpreted as dreaming or forgotten. Determining the real causes of this sort of observations requires a different type of protocol, better suited for experiments outside of clinical surgery setting. In other studies, the presence of dreams has been seen to be quite common during general anesthesia, being reported to occur up to 60% of the time with multiple anesthetics commonly used in clinics, including sevoflurane and propofol [20, 21]. Whether or not an anesthesia monitor should distinguish between states with and without dreams depends on the purpose of the monitor. Whereas such a distinction may not be relevant for monitors intended to help clinicians determine whether a patient is in a state of general anesthesia suitable for surgery, it may be central for monitors intended to distinguish brain states with and without consciousness defined as experience, including dreams [22, 23].

Of course, the most profound difference between this study and the previous was the fact that the patients underwent a different type of anesthesia (sevoflurane rather than propofol). This fact was used to investigate whether the DTF-method can successfully classify the state of wakefulness in patients using an anesthetic with assumed distinct mechanism of action [24, 25], but a comparable endpoint for the patient: unresponsiveness and apparent unconsciousness due to the general anesthetic [26]. The two anesthetics, propofol and sevoflurane, share some mechanisms of action: they both potentiate GABA_A and glycine receptors, and inhibit voltage gated potassium channels and acetylcholine receptors [24, 25]. However, they also differ in certain respects. Most notably, sevoflurane potentiates two-pore potassium channels and inhibits serotonin receptors, while propofol potentiates kainate receptors. Furthermore, sevoflurane has been reported to have a stronger inhibiting effect on AMPA and NMDA receptors. In short, the two anesthetics have complex and different interactions with ion channels and receptors affecting several neuronal populations, thus profoundly altering the neuronal activity patterns in the brain [27].

These alterations in neuronal activity are reflected in changes in EEG patterns. For example, as the molecular targets and effects of sevoflurane and propofol both partly overlap and partly differ, there are both similarities and differences also between their effects on large-scale measures of brain function [28]. For example, both sevoflurane and propofol have been reported to induce coherent frontal alpha oscillations and slow oscillations in EEG of humans [29], but sevoflurane has also be shown to be unassociated with the typical anteriorization of alpha rhythms [30]. Sevoflurane anesthesia has also been seen to increase the coherence in the theta frequency range (4–7 Hz), relative to comparable levels of propofol anesthesia [29]. In addition, propofol and sevoflurane have been shown to differentially suppress the relative glucose metabolic rate in several brain regions [31]. Furthermore, it is well known that anesthetics affect somatosensory evoked EEG potentials in humans and animals [32], but sevoflurane affects these potentials more strongly, in a dose-dependent fashion, than comparable doses of propofol [33]. These findings indicate that certain large-scale properties of the EEG do indeed change differentially in response to the two anesthetics. The fact that both the molecular, cellular, and large-scale brain effects of the two anesthetics differ in so many ways, increases the value of testing our method with both compounds, and enhances the significance of the remarkable similarity in their effects on DTF. Thus, since it is not obvious that EEG-derived measures such as the DTF would yield so similar results in patients undergoing anesthesia with as mechanistically different agents as propofol and sevoflurane, our results suggest that they have some substantial large-scale effects in common that are captured by the DTF derived measure.

In recent years, results from a broad range of studies seem to be converging on the “conclusion that a common neural correlate of anesthetic-induced unresponsiveness is a consistent depression or functional disconnection of lateral frontoparietal networks, which are thought to be critical for consciousness of the environment” [26]. DTF is one of several measures that can be used to quantify aspects of large-scale connectivity between time series such as EEG signals [34]. And, in addition to our own previous study [8], at least four studies have previously investigated how DTF-based connectivity measures are affected by changing states of consciousness [9‐12]. Each one of them reported changes in apparent brain connectivity related to distinct physiological states such as different stages of sleep and disorders of consciousness (DOC). Thus, changes in DTF may capture properties related to loss of consciousness in general, not just specific features of general anesthesia caused by sevoflurane (this study) or propofol [8].

In addition to the DTF, several other measures of connectivity have been suggested as a relevant markers for changes in the state or level of consciousness, including measures of transfer entropy [5, 35], directed coherence [36, 37], and Granger causality measures [38, 39]. Specifically, measures of the connectivity between frontal and parietal regions were found to differ between conscious and unconscious states, both for humans undergoing various forms of anesthesia [5, 35] or falling asleep [36], and for patients suffering from DOC [40]. For example, in a study using directed coherence—a measure closely related to DTF—the directionality of frontal–parietal functional connectivity covaried with NREM sleep stages and wakefulness [36]. However, approaches such as these (ours included), trying to quantify network properties based on passive observation of brain activity, are unlikely to justify strong conclusions about the underlying brain mechanisms [41]. Perturbational approaches, however, have indicated that a balanced interconnectivity between distant brain regions is likely to be crucial for maintaining a normal capacity for consciousness [4, 42‐45]. Thus, it is plausible that reports from observational studies of altered connectivity related to loss of consciousness also often reflect relevant underlying changes.

The fact that filtering, rejection of artefacts, and other data cleaning techniques were deliberately avoided in this study (in order to simulate a setting relevant for real-time monitoring of brain states in the clinic) may have further distorted the resulting connectivity matrices and yield faulty estimations of brain connectivity patterns. However, since the DTF is known to be robust to noise [46], and the median was used as a descriptive statistic for the mode of the distributions [47], this type of problems should be minimized [48]. Nevertheless, we remain agnostic regarding how the results reported here (regarding scalp level inferred connectivity) are related to the underlying changes in neural, effective connectivity within the brain. In fact, even with properly cleaned EEG data, the degree to which of sensor space estimates of connectivity are relevant for characterizing the underlying neural connectivity is disputed [49‐51].

Furthermore, it is possible that our DTF measure may be influenced by changes in muscle activity related to anesthesia, as has previously been shown for the bispectral (BIS) index which is one of several methods used for monitoring depth of anesthesia [52]. Propofol is known to decrease muscle tone in the absence of neuromuscular blocking drugs [53], and it is uncertain to what extent such effects may contribute to the observed changes in DTF reported here and in our previous paper [8] as it may not be not possible with our methods to tease apart muscle relaxing effects from other effects anesthetics have on the brain and body. Similarly, sevoflurane is known to cause immobility [54], depress the excitability of motor neurons [55], and enhance the effect of neuromuscular blockers [56]. However, the degree to which sevoflurane affects muscle tonus and EMG contamination of the EEG signal under the conditions used in our study remains to be determined. Thus, we cannot exclude the possibility that the apparent changes in brain connectivity observed here, may at least partly result from changes in EMG contamination in the two states, rather than reflecting consciousness related changes within the brain. This is an issue that is hard to control for without additional pharmacological interventions that were not feasible in the clinical setting our data were obtained from. We did, however, avoid neuromuscular blockers when possible to minimize anesthesia related changes in EMG (only one patient (#7) received an additional neuromuscular blocker (cisatrakurium, 14 mg i.v.) due to problems related to intubation). However, to investigate this issue more directly, we are now initiating a follow-up study in which we will test the effect of pure neuromuscular block in the absence of anesthetics on measures assumed to track the depth of anesthesia.

Another possible limitation is that our choice of DTF as a measure was made for reasons not entirely grounded in theories of consciousness. This makes it more likely that the observed changes in connectivity are related to changes in state more generally (e.g. going from normal wakefulness to general anesthesia) rather than reflecting changes in the brain that are specifically and causally important for consciousness. However, taken together, our findings that similar patterns of DTF changes occur for at least two different types of anesthesia, combined with the previous findings of changes in DTF related to distinct states of consciousness, provide convergent evidence in support of a strong relation between consciousness and measures of brain connectivity [9‐12], although further studies are needed to clarify whether and how the observed changes in our DTF-based measure are related to changes in consciousness as such.

Finally, other efforts are ongoing to test if our previous findings [8] can be reproduced in different experimental settings and under different conditions in order to assess the generalizability of the DTF-based measure as a marker of consciousness. However, these efforts have shown that reproduction is not always straight-forward, and that the particular topography of apparent information outflow may be sensitive to parameter choice in the analysis (e.g. filter types, sampling frequency, reference position, spatial filter functions) as well as the state of the person having their EEG measured. Furthermore, before measures such as the one discussed here can be used as clinical monitors of consciousness or anesthesia, studies aimed at understanding how they behave in transitions between coarse ‘conscious’ and ‘unconscious’ states must be carried out. Although we have shown that the DTF parameters abruptly change near transitions between ‘awake’ and ‘anesthetized’ states, there is significant variance in the exact time course of the classification confidence. The transitions should be studied experimentally outside of the clinic to allow even more precise quantification of the time of LOC and ROC, as well as allowing for intermittent awakenings with immediate reports about conscious experience (i.e. waking up the sleeping or sedated person with sound and/or touch stimuli in order to ask if they had any experience immediately before waking up; see for example work by Noreika et al. [20]), to understand whether the measures do in fact correctly identify changes in state of consciousness. Additionally, the clinical applicability of the methods applied in this paper is still questionable—not only because of uncertainties related to its classification accuracy, but also regarding the feasibility of routinely applying 25 EEG channel recordings to patients undergoing anesthesia and the computational costs associated with online application of the method. It would also be advisable to rigorously investigate the effects of precise pre-processing choices (e.g. reference position, sampling rate, presence and type of filters, artefact rejection) on DTF-parameters. This is because different practitioners may apply the measures in slightly different ways, with different equipment, and making different choices for preprocessing. It is not unlikely that such differences would change the resulting DTF-parameters, which might affect the interpretation of the results. Thus, before drawing strong conclusions about the capacity of the method presented here and in our earlier publication to objectively track states of consciousness in humans, further studies are needed. In particular, studies with much larger sample sizes are required and investigations of the method’s sensitivity to changing parameters (e.g. number of channels, sampling rate, and model order) as well as its response to other anesthetics (e.g. ketamine, xenon, and nitrous oxide).

Following our previous study where a DTF-based approach was used to detect changes in apparent brain connectivity during general anesthesia with propofol, we performed a follow-up study to validate our method by testing whether similar changes occur during loss of consciousness caused by another anesthetic, sevoflurane, which has partly different mechanisms of action. We found that our DTF-derived connectivity parameters showed similar changes in patients undergoing sevoflurane anesthesia as were previously observed with propofol anesthesia. These changes could once again be used to successfully classify the patients’ state of anesthesia vs. wakefulness in accordance with the clinician’s judgement, with accuracies and sensitivities exceeding 96%, depending on the choice of cut-off value for our algorithm’s confidence in classification. These results indicate that certain changes in DTF caused by general anesthesia generalize across at least two different anesthetics with partially distinct mechanisms of action. This can be regarded as further evidence in favor of brain connectivity being related to the level of consciousness in humans, although our study does not yet exclude the possibility that effect on muscle activity (EMG) may contaminate our EEG results. Thus, further studies are required for better understanding how the observed alterations in DTF are related to changes in brain state and consciousness.