Abstract
Deep non-rapid eye movement sleep (NREM) – also called slow wave sleep (SWS) – and general anesthesia are prominent states of reduced arousal linked to the occurrence of slow oscillations in the electroencephalogram (EEG). Rapid eye movement (REM) sleep, however, is also associated with a diminished arousal level, but is characterized by a desynchronized, ‘wake-like’ EEG. This observation challenges the notion of oscillations as the main physiological mediator of reduced arousal. Using intracranial and surface EEG recordings in four independent data sets, we establish the 1/f spectral slope as an electrophysiological marker that accurately delineates wakefulness from anesthesia, SWS and REM sleep. The spectral slope reflects the non-oscillatory, scale-free measure of neural activity and has been proposed to index the local balance between excitation and inhibition. Taken together, these findings reconcile the long-standing paradox of reduced arousal in both REM and NREM sleep and provide a common unifying physiological principle — a shift in local Excitation/ Inhibition balance — to explain states of reduced arousal such as sleep and anesthesia in humans.
Significance Statement The clinical assessment of arousal levels in humans depends on subjective measures such as responsiveness to verbal commands. While non-rapid eye movement (NREM) sleep and general anesthesia share some electrophysiological markers, rapid eye movement sleep (REM) is characterized by a ‘wake-like’ electroencephalogram. Here, we demonstrate that non-oscillatory, scale-free electrical brain activity — recorded from both scalp electroencephalogram and intracranial recordings in humans — reliably tracks arousal levels during both NREM and REM sleep as well as under general anesthesia with propofol. Our findings suggest that non-oscillatory brain activity can be used effectively to monitor vigilance states.
Introduction
Sleep and anesthesia both present with a behaviorally similar state of diminished arousal(1) and shared neurophysiologic features, namely increased low frequency power(2, 3) and a reduction in effective connectivity(4, 5). It has been argued that the reduced arousal in both states stems from a common neuronal mechanism. Current definitions of arousal vary and include e.g. autonomic, behavioral or mental arousal. An updated framework has been proposed recently(6). Here, we use the term arousal in its relation to vigilance states.
Most studies comparing sleep and anesthesia concentrated on slow-wave sleep and oscillatory dynamics such as slow waves (< 1.25 Hz)(1, 7, 8) as an increased activity in this frequency band has been associated with reduced arousal(1, 3). REM sleep is also associated with decreased arousal but is characterized by a desynchronized, active pattern in the electroencephalogram (EEG) similar to wakefulness(8). This paradox challenges the notion that changes in oscillatory activity such as slow waves are the exclusive determinant of reduced arousal.
Non-oscillatory, scale-free neural activity constitutes an important index of brain physiology and behavior(9–11). In the frequency domain, the scaling law between the power and the frequency of non-oscillatory brain activity can be estimated from the exponential decay of the power spectral density(9) and has previously been used to assess a variety of cognitive and EEG phenomena(12–18). A variety of terms have been used to describe this power-frequency relationship, such as power-law distribution, scale-free behavior, 1/f electrophysiological noise, fractal/spectral exponent(12, 17, 19) or fractal dynamics(9, 20–22). The exponent of the 1/f power-law distribution, also called spectral slope, differs between rest and task activity(9, 10) and changes with aging(21). Fractal dynamics and neural avalanches have also been observed in long-range temporal correlations of band-limited signals(23), however, it is likely that these two phenomena may reflect distinct entities with a different neurophysiological basis(9). Here, we focus on the fractal 1/f dynamics of the background activity.
Computational simulations indicate that the spectral slope provides a surrogate marker for the excitatory to inhibitory (E/I) balance with more negative slope values indexing enhanced inhibition(10, 20, 22) (Fig. S1), while others have observed the reversed pattern(11).
For this study, we followed the framework of Laureys et al. that defined consciousness on two axis – content (awareness) and level (arousal)(24). While the conscious content is low in NREM sleep and GABAergic anesthesia, it is high in wakefulness and dreaming states like REM. The arousal level, on the other hand, is low in all sleep states including REM. We hypothesized that states of reduced arousal are characterized by a shift of the E/I balance towards inhibition indexed by more negative slopes. To test this prediction, we analyzed four independent datasets: Electrophysiological recordings during sleep using either scalp EEG (Study 1, n = 20) or combined scalp and intracranial EEG (Study 2, n = 10; coverage see Fig. S2a) as well as under general anesthesia with propofol combined with scalp EEG (Study 3, n = 9) or intracranial EEG (Study 4, n = 12; subdural grid electrodes (electrocorticography; ECoG) and stereotactically placed depth electrodes (SEEG); coverage see Fig. S2b).
Results
During a full night of sleep, the time-resolved spectral slope closely tracked the hypnogram (Fig. 1a). In the scalp EEG group (Study 1, n = 20; a baseline rest recording was available in n = 14), we observed a decrease from values of −1.87 ± 0.18 (mean ± SEM) during quiescent rest to −3.46 ± 0.16 in NREM (N3) and −4.73 ± 0.23 in REM sleep (Fig. 1b). These differences were significant across all scalp EEG channels (repeated-measures ANOVA: p < 0.0001, F1.94, 25.17 = 56.05, dRest-Sleep = 3.07). Furthermore, N2 sleep exhibited an average slope of −3.67 ± 0.10 that was also significantly below rest (n = 14; pRest-N2 < 0.0001; t13 = 7.97; dRest-N2 = 3.31; Fig. S3a). Post-hoc t-tests (uncorrected) revealed a significant difference between rest and N3 (pRest-N3 < 0.0001, t13 = 5.69, dRest-N3 = 2.49), between rest and REM (pRest-REM < 0.0001, t13 = 11.67, dRest-REM = 3.71) and between N3 and REM sleep (pN3-REM = 0.0007, t13 = 4.44, dN3-REM = 1.70).
If all the available wake periods before, during and after the sleep recordings were utilized for slope analysis (n = 20), it resulted in a higher variability across subjects during wakefulness (Fig. S3b), which can be explained by the fact that the subjects were already or still drowsy and data during state transitions was included. However, the overall pattern was remarkably similar (Fig. S3c). As this approach increased our available data, we used all wake trials (referred to as wake) for subsequent analysis.
To assess where on the scalp the slope tracks arousal states best, we calculated the Mutual Information (MI) between the time-resolved spectral slope and the hypnogram in all 20 subjects. We observed a significant positive cluster across all sensors, which peaked over frontal electrodes F3, Fz and F4 (Fig. 1b).
Cranial muscle activity has similar frequency characteristics in the 30-70 Hz range and might confound spectral slope estimates. Therefore, we controlled for any impact of muscle activity by repeating the analysis after local referencing (Laplacian, pSpearman < 0.001; pMI < 0.0001) and additionally utilized partial correlations that considered the slope of the electromyography (EMG) as a confounding variable (pSpearman < 0.001). All control analyses confirmed that the observed effect was not confounded by muscle activity (Fig. S4).
During REM sleep, power in the slow wave range (SO power; <1.25 Hz) was comparable to wakefulness corroborating the observation of a ‘wake-like’ EEG pattern in REM and the paucity of slow oscillations (p = 0.423, t18= −0.82, d = −0.25; Fig. 1a). The spectral slope, however, was significantly different between these states. To further quantify this effect, we trained a classifier (linear discriminant analysis; LDA) to discriminate between REM sleep and wakefulness using either the spectral slope or SO power (n = 18). The classifier performance was significantly better for the spectral slope compared to SO power when differentiating between REM and waking (78.75 ± 2.98 % (mean ± SEM) vs. 60.03 ± 3.72 %; p = 0.0023, t17 = 3.58, dSlope-SO power= 1.21, chance level: 50 %). When differentiating between N3 sleep and wakefulness, both spectral slope and SO power had a classifier performance that was significantly above the 50 % chance level (for slope p < 0.001 vs. for SO power p < 0.001) and comparable to each other (73.05 ± 2.97 % for spectral slope vs. 82.09 ± 2.13 % for SO power, p = 0.0423, t17 = −2.19, dSlope-SO power = −0.83). Likewise, when all three states were classified simultaneously, both SO power and the spectral slope performed well above chance (chance = 33%; SO: 64.94 ± 2.04%, mean ± SEM; t17 = 15.04, p < 0.001, d = 5.01; slope: 58.09 ± 2.35%; t17 = 10.55, p < 0.001, d = 3.52) and did not differ in the overall performance (t17 = −1.80, p = 0.0899, d = −0.63). This is due to the fact that SO power is advantageous to classify N3 sleep, while the slope is superior to detect REM sleep. Notably, significant classification is also possible when the spectral slope is estimated at lower frequencies (e.g. 1-20 Hz; 84.19% ± 2.46, paired t-test vs. chance (33%): p < 0.001, t17 = 20.64, d = 6.88). This effect is partly driven by an increase in low frequency power needed to correctly classify N3, and is equivalent to using SO power, but the 1-20 Hz ranges does not track wakefulness and REM, thus, reducing mutual information with the hypnogram (see also Fig. S7).
These results reveal that the spectral slope is a more powerful predictor of REM sleep than SO power and also reliably discriminates deep N3 sleep from wakefulness. Furthermore, classification based on the spectral slope provides comparable accuracy levels in discriminating REM from wakefulness as trained personnel, given that the inter-rater reliability between sleep scoring experts is typically about 80%(25). Finally, the discrimination between REM and waking using the spectral slope does not require simultaneous electrooculography (EOG) or EMG recordings but can be detected solely from the electrophysiological brain state.
In the intracranial recording group (Study 2, n = 10), the simultaneous EEG recordings (Fz, Cz, C3, C4, Oz) again displayed a more negative spectral slope for reduced arousal levels: From −2.99 ± 0.32 (mean ± SEM) in wakefulness the slope decreased to −3.69 ± 0.12 in NREM (N3) to −4.15 ± 0.29 in REM sleep (Fig. 1c). Again, these three states were significantly different in a repeated-measures ANOVA (p = 0.001; F1.97, 17.74 = 10.79, dWake-Sleep = 1.12). Post-hoc t-tests (uncorrected) showed a significant difference between wakefulness and REM (p < 0.001; t9 = 4.78; d = 1.19) and wakefulness and N3 (p = 0.026; t9 = 2.66; d = 0.97) but not between N3 and REM (p = 0.098; t9 = 1.84; d = 0.64).
The intracranial SEEG contacts that mirrored the observed scalp EEG pattern (more negative spectral slope in N3 and REM; 155 of 352 SEEG (44.03 %; significantly above chance; X2 = 8.20, p = 0.0042; chi-squared test); Fig. 1c) exhibited a clear anatomical distribution centered in the medial prefrontal cortex and medial temporal lobe structures (for Wake - N3 and Wake - REM see Fig. S5a, b; grid electrodes see Fig. S6a, b), hence, converging on the very same brain regions known to be the most relevant for sleep-dependent memory consolidation(26–29). Note that we did not specifically target any brain regions and in contrast to previous studies using grid electrodes(9, 22), the majority of our probes were stereo-tactically placed depth electrodes. Given the spatial heterogeneity of intracranial responses(30), the convergence on medial PFC nicely resembles the observed scalp pattern as observed at the overlying scalp EEG electrode Fz. The distinct intracranial spatial pattern combined with the bipolar referencing scheme again confirms that the results are not confounded by muscle activity.
To verify the chosen fit parameters, we reanalyzed the correlation and MI analysis between hypnogram and time-resolved slope as a function of different center frequencies and window lengths (Fig. S7). In addition, we explored a wide-range of fit parameters after discounting the oscillatory components from the PSD by means of irregular resampling (IRASA(26, 31)). All control analyses corroborated our findings and indicate that the spectral slope in the range from 30 - 45 Hz reliably tracks arousal levels and behavioral state transitions (Fig. S7). Given the relationship between spectral edge frequency and median frequency(32), we assessed the relationship between SO power and the PSD slope. We observed that the SO power explains 7.9 ± 0.01% (mean ± SEM) of the variance in the slope, however, a partial correlation with SO power as a confound does not change the correlation between slope and hypnogram (Fig. S7).
On the scalp level, the trough of a slow wave is associated with cortical ‘down-state’, while the peak reflects an ‘up-state’(33, 34). The spectral slope was able to reflect these rapid changes during sleep with a more negative 1/f slope observed at troughs compared to peaks (Fig. S8). This effect was most pronounced over frontal channels (cluster-based permutation test: p = 0.005, dTrough-Peak = −0.65).
Slow waves are detected in slow-wave sleep but are also observed during REM sleep(35) as well as wakefulness(36); albeit less prevalently. We detected a significantly higher number of slow waves during N3 sleep (SON3 = 28.79 ± 0.79 per minute; mean ± SEM) compared to REM sleep (SOREM = 2.16 ± 0.89 per minute; SON3-REM: p < 0.0001, t19 = 22.64, d = 7.05) and wakefulness (SOWake = 5.05 ± 0.51 per minute; SON3-Wake: p < 0.0001, t19 = 25.32, d = 6.92; Extended Data Fig. 9c). Interestingly, the averaged slope at the through of the slow waves was significantly different between arousal states: −3.40 ± 0.09 in slow-wave, −4.00 ± 0.18 in REM sleep and −2.26 ± 0.12 in wakefulness (mean ± SEM) mirroring our observation of the overall slope differences (Fig. S9c; uncorrected for multiple testing: Wake-N3: p < 0.0001, t18 = 7.07, d = 2.38; Wake-REM: p < 0.0001, t18 = 9.67, d = 2.55, N3-REM: p = 0.01, t19 = 2.73, d = 0.91). Therefore, the spectral slope is able to discern arousal even during slow wave events.
To test if the state-dependent modulation of the spectral slope was sleep-specific or generalized to other forms of decreased arousal, we analyzed two datasets obtained during general anesthesia with propofol. Under propofol anesthesia the time-resolved spectral slope again closely tracked changes in arousal level (Fig. 2a). In both scalp and intracranial EEG, we observed a more negative spectral slope under anesthesia compared to wakefulness (Fig. 2b, c): In the scalp EEG group (Study 3, n = 9), we found a decrease from −1.81 ± 0.29 (mean ± SEM) during wakefulness to −3.10 ± 0.19 under anesthesia. This difference was significant (paired t-test: p < 0.0001, t8 = 7.73, dWake-Anesthesia = 1.71) and in a cluster-based permutation test, the effect formed one single cluster spanning all 25 electrodes (p < 0.001).
In the intracranial recordings (Study 4, n = 12), we observed a spectral slope of −2.75 ± 0.15 during wakefulness and −4.34 ± 0.11 under anesthesia. Again, this difference was significant (paired t-test: p < 0.0001, t11 = 9.93, dWake-Anesthesia = 3.57) and could be detected in the majority of electrodes (470 of 485 SEEG (96.9 %); Fig. 2c). Patients who were implanted with surface grid in addition to depth electrodes (n = 4) showed the same pattern: The spectral slope decreased from wakefulness to anesthesia in the majority of the recording sites (129 of 147 ECoG (87.75 %); Fig. S6c).
These findings demonstrate that the spectral slope reliably differentiates between wakefulness and general anesthesia in humans(22). Future studies will be needed to determine the reliability of this marker on larger cohorts to establish clinical usability. In both scalp and intracranial recordings, we observed a brain-wide decrease in the spectral slope, supporting the notion that propofol anesthesia induces a global brain-wide state of increased inhibition(8).
Discussion
Collectively, the results from these four studies provide five main advances. First, the spectral slope tracks changes in arousal levels in both sleep and anesthesia with high temporal precision from sub-second epochs to full night recordings. Note that the slope differences between wakefulness and states of reduced arousal show a similar pattern on the scalp level (Fig. 1b, 2b).
According to the framework proposed by Laureys et al., consciousness can be assessed on two axis – the content (e.g. awareness) and the level (e.g. arousal)(24), however, an updated framework has recently been proposed(6). Our definition of arousal is similar to what has been described as vigilance. However, our neurophysiological investigations did not set out to test one specific framework, but we do interpret our findings in light of previously published definitions. Hence, we assume that our marker does not track conscious thoughts, content or awareness, but indexes a vigilance state. While the arousal level is reduced in all three states, conscious content is thought to fluctuate during sleep, mostly in the form of dreams during REM(37). Thus, measures such as the Perturbational Complexity Index(38) that might track the level of consciousness are decreased in slow-wave sleep and GABAergic anesthesia but are maintained to a certain degree during REM sleep and ketamine anesthesia, both states associated with vivid dreams(37–39). These measures are unable – unlike the spectral slope – to reliably differentiate arousal levels, e.g. wakefulness and REM. Previous studies in rodents identified markers of reduced arousal in sleep and under general anesthesia, namely fronto-parietal theta and high-gamma connectivity(39, 40). In several control analyses we found that the spectral slope was superior to fronto-parietal theta connectivity in tracking sleep stage dependent dynamics (p < 0.0001, t19 = 7.01; d = 2.22) and in reliably differentiating REM and slow wave sleep (Fig. S10). Our dataset did not have a sufficient number of electrodes in the parietal lobe to extend the analysis to the high-gamma band since this is an infrequent site for epilepsy.
Second, the spectral slope provides a mechanistic explanation – a shift of the E/I balance towards inhibition – for the reduced arousal level in both slow-wave and REM sleep. The estimation of local E/I balance has been limited to invasive single cell recordings with a classification of neuron subtypes into excitatory and inhibitory cells(22). Recent computational simulations, however, demonstrated that local E/I balance can be inferred from changes of the spectral slope: An increase in inhibition results in a decrease of slope(11, 22). Our results of a decreased slope in slow-wave and REM sleep as well as under general anesthesia may be explained by an increase in inhibition. This interpretation is supported by results of single cell studies in animals that reported a reduction of multiunit or pyramidal cell activity during not only in slow-wave but also in REM sleep(41–44). Interestingly, REM exhibited a significantly lower slope than slow-wave sleep (Fig. S11). This result is in line with previous studies reporting a lower neuronal firing rate for REM sleep compared to slow-wave sleep(41, 43, 44) that was associated with an increase in inhibitory activity(41, 43). While these lines of research converge on the notion that the spectral slope tracks the E/I balance of the underlying population, it might also reflect changes in firing rate or synchronization. A testable hypothesis that arises from our observations is that cell-type specific causal manipulations by optogenetics (e.g. pyramidal and SOM interneurons) should bias the spectral slope in opposite directions.
Previous studies utilized a variety of different fit parameters and it is currently unclear what the ‘best’ range for slope fitting is(12–18). It had been suggested that fits to different frequencies might index different properties of the underlying population activity(9, 14, 16, 17). Our results that demonstrate that the range from 30-45 Hz best correlates (and exhibits significant mutual information) with the hypnogram, which is in line with recent modeling work indicating a similar range(22). Future studies involving single neuron recordings will be needed to unravel the precise relationship between population firing statistics and band-limited changes in the PSD slope. We believe that in particular comparative studies involving rodents(22, 45), primates(22) and humans combined with modeling work has the potential to integrate divergent findings into a coherent framework and to determine the neurophysiologic basis of the spectral slope. It will be of substantial interest to assess whether neurophysiological mechanisms are preserved across species, which greatly vary in anatomy, in particular in the prefrontal cortex(46, 47).
Third, the rapid changes in spectral slope observed over the course of a slow wave are in accordance with the notion that these oscillations orchestrate cortical activity during sleep by interleaving periods of neural silence with enhanced neural activity(41). This suggests that E/I balance and arousal level during slow wave sleep are not constant but wax and wane on a short time scale – whereas they seem to be more constant during REM sleep(41). This finding is in line with the active, maximal inhibition during REM sleep observed in single cell recordings of animal cortices(41, 43) and could explain why epileptic seizures during the night occur predominantly in NREM and rarely during REM sleep(48).
Fourth, our observations support the premise that anesthesia is a brain-wide state(8), whereas sleep exhibits network-specific activity patterns (e.g. between the PFC and the hippocampus)(49). This is especially relevant considering the theories of active memory processing in sleep(50, 51).
Fifth, the spectral slope can be reliably estimated from scalp EEG recordings, providing a potential tool that can be incorporated into intraoperative neuromonitoring, automatic sleep stage classification algorithms and tracking other states of reduced arousal such as epileptic seizures, coma and the vegetative or minimally conscious state.
Data Availability
Data generated and/or analyzed in the current study is available from the corresponding author upon reasonable request.
Code availability
Custom code used for analyzing the datasets of the current study is available from the corresponding author upon reasonable request.
Author Contribution
Conceptualization, J.D.L, R.F.H., M.P.W. and R.T.K.; Methodology, J.D.L. and R.F.H.; Software, J.D.L. and R.F.H.; Validation, J.D.L. and R.F.H.; Formal Analysis, J.D.L. and R.F.H.; Investigation, J.J.L., B.A.M., L.R. and P.G.L.; Resources, P.G.L, J.J.L., M.P.W. and R.T.K.; Data Curation, J.D.L, R.F.H., J.J.L., P.G.L., B.A.M., M.P.W. and R.T.K.; Writing – Original Draft, J.D.L.; Writing – Review & Editing, J.D.L., R.F.H., P.G.L, B.A.M., M.P.W. and R.T.K.; Visualization, J.D.L. and R.F.H.; Supervision, R.T.K.; Project Administration, P.G.L., J.J.L., M.P.W. and R.T.K.; Funding Acquisition, R.T.K.
Acknowledgement
This work was supported by Grant LE 3863/2-1 of the German Research Foundation (Deutsche Forschungsgemeinschaft (J.D.L.), a National Institute of Neurological Disorders and Stroke Grant R37NS21135 (R.T.K., J.D.L.), the Alexander von Humboldt Foundation (Feodor Lynen Program; R.F.H.), an intramural fellowship from the Dept. of Psychology, University of Oslo (R.F.H.), R01AG03116408 (M.P.W.), RF1AG05401901 (M.P.W.), RF1AG05410601 (M.P.W.) and F32-AG039170 (B.A.M.), all from the National Institute of Health. We thank Jie Zheng, Julia Kam and the EEG technicians at UC Irvine Medical Center for their assistance and all the patients for their participation.