Author response: An electrophysiological marker of arousal level in humans

Article Figures and data Abstract eLife digest Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Deep non-rapid eye movement sleep (NREM) and general anesthesia with propofol are prominent states of reduced arousal linked to the occurrence of synchronized oscillations in the electroencephalogram (EEG). Although rapid eye movement (REM) sleep is also associated with diminished arousal levels, it is characterized by a desynchronized, ‘wake-like’ EEG. This observation implies that reduced arousal states are not necessarily only defined by synchronous oscillatory activity. Using intracranial and surface EEG recordings in four independent data sets, we demonstrate that the 1/f spectral slope of the electrophysiological power spectrum, which reflects the non-oscillatory, scale-free component of neural activity, delineates wakefulness from propofol anesthesia, NREM and REM sleep. Critically, the spectral slope discriminates wakefulness from REM sleep solely based on the neurophysiological brain state. Taken together, our findings describe a common electrophysiological marker that tracks states of reduced arousal, including different sleep stages as well as anesthesia in humans. eLife digest Electroencephalogram (EEG for short) is a widespread technique that helps to monitor the electrical activity of the brain. In particular, it can be used to examine, recognize and compare different states of brain consciousness such as sleep, wakefulness or general anesthesia. Yet, during rapid eye movement sleep (the sleep phase in which dreaming occurs), the electrical activity of the brain is similar to the one recorded during wakefulness, making it difficult to distinguish these states based on EEG alone. EEG records brain activity in the shape of rhythmic waves whose frequency, shape and amplitude vary depending on the state of consciousness. In the EEG signal from the human brain, the higher frequency waves are weaker than the low-frequency waves: a measure known as spectral slope reflects the degree of this difference in the signal strength. Previous research suggests that spectral slope can be used to distinguish wakefulness from anesthesia and non-REM sleep. Here, Lendner et al. explored whether certain elements of the spectral slope could also discern wakefulness from all states of reduced arousal. EEG readings were taken from patients and volunteers who were awake, asleep or under anesthesia, using electrodes placed either on the scalp or into the brain. Lendner et al. found that the spectral slope could distinguish wakefulness from anesthesia, deep non-REM and REM sleep. The changes in the spectral slope during sleep could accurately track the degree of arousal with great temporal precision and across a wide range of time scales. This method means that states of consciousness can be spotted just from a scalp EEG. In the future, this approach could be embedded into the techniques used for monitoring sleep or anesthesia during operations; it could also be harnessed to monitor other low-response states, such as comas. Introduction General anesthesia is a reversible, pharmaceutically induced state of unconsciousness, while sleep is internally generated and cycles between rapid (REM) and non-rapid eye movement sleep (NREM; Brown et al., 2010; Franks and Zecharia, 2011). Both sleep stages and anesthesia are characterized by a behaviorally similar state of reduced physical arousal (Brown et al., 2010; Franks and Zecharia, 2011; Murphy et al., 2011). Definitions of arousal vary and include e.g. autonomic, behavioral or mental arousal. For this study, we followed an updated version of the framework by Laureys et al. that defined consciousness on two axes – content (awareness) and level (arousal; Boly et al., 2013; Laureys, 2005). While the conscious content is low in NREM sleep and propofol anesthesia, it is high in wakefulness and dreaming states like REM. The arousal level is low during anesthesia and in all sleep states including REM. Both NREM sleep stage 3 (also called slow-wave sleep) and general anesthesia with propofol exhibit similar electrophysiological features, such as an increase in low frequency activity and the occurrence of prominent slow oscillations (<1.25 Hz; Brown et al., 2010; Franks and Zecharia, 2011; Murphy et al., 2011; Prerau et al., 2017; Purdon et al., 2013). Moreover, propofol anesthesia has been linked to the emergence of a strong frontal alpha oscillation (8–12 Hz; Purdon et al., 2013) whereas spindles (12–16 Hz) typically appear in NREM sleep stage 2 (Prerau et al., 2017). In contrast, REM sleep is characterized by a desynchronized, active pattern in the electroencephalogram (EEG), which resembles wakefulness (Brown et al., 2010; Prerau et al., 2017). The additional defining features of REM sleep are therefore peripheral markers including muscle atonia as detected by electromyography (EMG) combined with rapid eye movements in the electrooculogram (EOG; Prerau et al., 2017). To date, it has been challenging to differentiate REM sleep from wakefulness in humans solely from the electrophysiological brain state (Pal et al., 2016). Recently, several lines of inquiry highlighted the importance of non-oscillatory, scale-free neural activity for brain physiology and behavior (Miller et al., 2009a; Gao et al., 2017; Voytek et al., 2015; Voytek and Knight, 2015; Miller et al., 2009b; He et al., 2010). The electrophysiological power spectrum is characterized by a 1/f signal drop-off, i.e. higher frequency activity exhibits reduced power as compared to low frequency activity. This scaling law between power and frequency can be estimated from the exponential decay of the power spectrum (He et al., 2010) and has previously been used to assess a variety of cognitive and EEG phenomena (Colombo et al., 2019; Lina et al., 2019; Miskovic et al., 2019; Pereda et al., 1998; Pritchard, 1992; Shen et al., 2003; Susmáková and Krakovská, 2008). Notably, this decay function mainly captures non-oscillatory brain activity, which is not characterized by a defining temporal scale, such as band-limited oscillations (He et al., 2010). Therefore, analyses of scale-free 1/f dynamics might prove especially helpful when analyzing brain states that are not characterized by prominent oscillations such as REM sleep in humans. We hypothesized that markers of 1/f activity, such as the spectral slope of the power spectrum, may provide an electrophysiological signature that distinguishes ‘paradoxical’ REM sleep (Siegel, 2011) from wakefulness. Importantly, 1/f dynamics can also be observed in a variety of other signals. For instance, long-range temporal correlations of neuronal oscillations (Linkenkaer-Hansen et al., 2001) or the size and duration of neuronal avalanches (Beggs and Plenz, 2003; Palva et al., 2013) also follow a power law but these scale-free behaviors likely have a different neurophysiological basis than the 1/f drop-off of the power spectrum (He et al., 2010). Recent findings suggested that 1/f dynamics differentiate wakefulness from general anesthesia (Colombo et al., 2019; Gao, 2016). For instance, using intracranial recordings in macaque monkeys, it had been shown that the spectral slope between 30 and 50 Hz reliably tracked changes in arousal level under propofol anesthesia from induction to emergence (Gao et al., 2017). Moreover, it has been reported that that the spectral slope between 1 and 40 Hz in human scalp EEG recordings discriminated states with conscious content, namely wakefulness and ketamine anesthesia, from states where no conscious report was possible, i.e. Xenon and propofol anesthesia (Colombo et al., 2019). Collectively, these studies implied that propofol anesthesia was accompanied by a steeper decay of the power spectrum (Colombo et al., 2019; Gao et al., 2017). With regard to sleep physiology, it had been observed that the spectral exponent of human scalp EEG becomes more negative during NREM sleep, when estimated e.g. in the 1 to 5 Hz (Shen et al., 2003), 3 to 30 Hz (Pereda et al., 1998) or 0.5 to 35 Hz frequency range (Miskovic et al., 2019). A similar pattern was observed in intracranial recordings with subdural grid electrodes in humans between 10 and 100 Hz (one subject; Freeman and Zhai, 2009) or 1 and 100 Hz (five subjects; He et al., 2010). Note that the 1/f background activity was estimated from frequency bands that were potentially influenced by simultaneously occurring low frequency oscillation i.e. delta (<4 Hz) or slow waves (<1.25 Hz) that might affect the degree of spectral tilt. General anesthetics like propofol, etomidate and barbiturates act on GABAergic receptors to enhance inhibition (Brown et al., 2011). Recently, computational simulations indicated that the spectral slope might provide a surrogate marker for the excitation to inhibition (E/I) balance with more negative spectral slopes (esp. in the 30 to 50 Hz range) indexing enhanced inhibition (Gao et al., 2017). This model was validated using intracranial recordings in macaques and rodents: A shift in E/I-balance towards inhibition by administrating propofol resulted in a steeper slope of the power spectrum. Likewise, the spectral slope in the rodent hippocampus varied across the depth of hippocampus, directly reflecting the ratio of excitatory to inhibitory cells in the underlying neuronal population. Moreover, a modulation of spectral slope was also observed as a function of the hippocampal theta cycle, likely reflecting rapid shifts in E/I-balance (Gao et al., 2017). A recent study that employed two-photon calcium imaging in mice provided additional insight into putative changes in E/I-balance during the sleep cycle. Cortical activity in mice was reduced during NREM sleep compared to wakefulness and, notably, even further reduced during REM sleep (Niethard et al., 2016). Crucially, the authors observed a selective increase in inhibitory interneuron activity (parvalbumin-positive interneurons; Niethard et al., 2016) during REM but not NREM sleep revealing an overall shift towards inhibition during REM sleep. In the present study, we assessed if 1/f spectral dynamics, in particular in the 30 to 50 Hz range, which is devoid of prominent low-frequency oscillatory activity (Gao et al., 2017), could track arousal states in humans both under anesthesia with propofol and during sleep in intracranial and scalp EEG recordings. Specifically, we hypothesized that the spectral slope should become more negative (i.e. the power spectrum steeper) in sleep and under anesthesia compared to wakefulness. Importantly, we also predicted that the spectral slope could discriminate wakefulness from NREM as well as REM sleep. Based on recent reports linking E/I-balance and electrophysiology (Gao et al., 2017; Niethard et al., 2016), we reasoned that the spectral slope, as a putative marker of E/I-balance, should facilitate the detection of REM sleep directly from the current brain state, without complementary information from additional EMG or EOG recordings. While previous studies that included lower frequency power in their slope estimates, found the slope of REM to be close to wakefulness (He et al., 2010), we specifically investigated if the aperiodic background activity in the 30 to 50 Hz range could reliably discriminate REM sleep from wakefulness and NREM sleep. Results We tested if non-oscillatory brain activity as quantified by the spectral slope of the electrophysiological power spectrum could discriminate different states of arousal in four independent studies: We obtained both (1) scalp EEG (n = 9) and (2) intracranial EEG (n = 12) under general anesthesia with propofol. Furthermore, we recorded (3) scalp EEG (n = 20) as well as (4) scalp EEG combined with intracranial EEG (n = 10) during a full night of sleep. We utilized both extra- and intracranial recordings to assess the precise spatial extent of the observed effects. In line with previous reports, the spectral slope was defined by a linear fit to the power spectrum in log-log space between 30 and 45 Hz (Gao et al., 2017). Critically, we carefully validated the chosen parameters in a series of control analyses as indicated below. 1/f spectral dynamics during propofol anesthesia We first tested if the spectral slope discriminates wakefulness and propofol anesthesia in humans in two experiments. In the first study, we recorded scalp EEG during general anesthesia for orthopedic surgery in otherwise healthy adults (Study 1, n = 9). In the second study, we obtained intracranial EEG in epilepsy patients who were implanted with intracranial electrodes for seizure onset localization while they underwent general anesthesia for electrode explantation (Study 2, n = 12; subdural grid electrodes (electrocorticography; ECoG) and stereotactically placed depth electrodes (SEEG; coverage see Figure 1—figure supplement 1a). In Study 1 (n = 9), we found that the time-resolved spectral slope closely tracked changes in arousal levels while subjects underwent propofol anesthesia (Figure 1a). Specifically, we observed a significant decrease of the spectral slope from wakefulness (−1.84 ± 0.30; mean ± SEM) to anesthesia (−3.10 ± 0.20) when averaged across all electrodes (Figure 1b; permutation t-test: p<0.0001, obs. t8 = 7.09, dWake-Anesthesia = 1.65). Figure 1 with 4 supplements see all Download asset Open asset The spectral slope tracks changes in arousal level under general anesthesia with propofol. (a) Time-resolved average of three frontal EEG channels (F3, Fz, F4) during anesthesia. Upper panel: Time-frequency decomposition. Dotted white line: Induction with propofol. Middle: Spectral slope (black; mean ± SEM). Lower panel: Slow frequency (<1.25 Hz; gray) and alpha (8–12 Hz; purple) baseline-corrected power (mean ± SEM). Note, elevated slow frequency activity is already present during wakefulness. While alpha frequency activity is steadily increasing in the first minutes of anesthesia, slow frequency activity exhibits a waxing and waning pattern which may reflect the premedication with a sedative. (b) Anesthesia in scalp EEG (n = 9). Upper panel: Spatial extent of spectral slope difference. Cluster permutation test: *p<0.05. Lower panel: Left - Power spectra (mean ± SEM); Right – Spectral slope. Wakefulness (red), anesthesia (blue) and grand average (black; all mean ± SEM). Permutation t-test: ***p<0.001. (c) Anesthesia in intracranial recordings (n = 12). Upper panel: Left – coronal, right – axial view of intracranial channels that followed (magenta) or did not follow (white) the EEG pattern of a lower slope during anesthesia compared to wakefulness. Lower panel: Left – Power spectra; Right – Spectral slope. Wakefulness (red), anesthesia (blue) and grand average (black; mean ± SEM). Permutation t-test: ***p<0.001. A cluster-based permutation test assessing the spatial extent of this effect on the scalp level resulted in a single large cluster that spanned all 25 electrodes without a clear peak (p<0.001; Figure 1b). To further examine the spatial distribution of the observed scalp EEG pattern and to assess subcortical contributions, we utilized intracranial recordings in Study 2 (n = 12). Again, we observed that the spectral slope was higher during wakefulness (−2.75 ± 0.15) than during anesthesia (−4.34 ± 0.11) when averaged across all electrodes (Figure 1c; permutation t-test: p<0.0001, obs. t11 = 9.93, dWake-Anesthesia = 3.57). This effect was present at the majority of recording sites (470 of 485 SEEG (96.9%); Figure 1c, Table 1). Notably, recordings from subdural grid electrodes (n = 4) showed the same pattern: The spectral slope decreased from wakefulness to anesthesia in the majority of recording sites (129 of 147 ECoG (87.75%); Figure 1—figure supplement 2a). Table 1 Anatomical distribution of stereotactically placed intracranial depth electrodes in Study 2 – Intracranial anesthesia (n = 12). Brain regionTotal number of electrodesElectrodes with state-dependent slope modulationALL485470 (96.9 %)Prefrontal Cortex (PFC)179175 (97.8 %) medial Prefrontal Cortex (mPFC)2727 (100 %) lateral Prefrontal Cortex (lPFC)147143 (97.3 %) Orbito-frontal Cortex (OFC)55 (100 %)Medial temporal Lobe (MTL)401 (95.0 %) Hippocampus2624 (92.3 %) Amygdala1313 (100 %)Cingulate Cortex2222 (100 %)Insula1313 (100 %)M1/Premotor4847 (97.9 %)Lateral Temporal Cortex (LTC)5050 (100 %)Parietal Cortex8478 (92.9 %)Visual Cortex4947 (95.9 %) Taken together, we observed a more negative spectral slope under anesthesia compared to wakefulness in both scalp as well as intracranial EEG (Figure 1b,c). Our results indicate that the spectral slope differentiates between wakefulness and general anesthesia in humans. This effect spanned all scalp and the majority of intracranial electrodes, hence, supporting the notion that propofol anesthesia induces a global, brain-wide state change (Brown et al., 2010). 1/f spectral dynamics discriminate wakefulness, NREM and REM sleep Having established that the spectral slope differs significantly between wakefulness and propofol anesthesia, we next examined if this state-dependent modulation generalized to other forms of decreased arousal, such as sleep. We specifically sought to determine if the spectral slope could discern wakefulness from different sleep stages. We analyzed two datasets obtained during a full night of sleep. In Study 3, we obtained polysomnography recordings from 20 healthy subjects, which included scalp EEG, as well as electrocardiography (ECG), electromyography (EMG) and electrooculography (EOG). To determine the precise spatial extent and subcortical contributions, we again recorded intracranial EEG in a separate cohort for Study 4 (n = 10; electrode coverage see Figure 1—figure supplement 1b). Critically, we combined intracranial EEG with polysomnography (scalp EEG, ECG, EMG, EOG) to enable comparable sleep staging across both the scalp and intracranial studies. We observed that the time-resolved spectral slope closely tracked the technician-scored hypnogram (Figure 2a). To quantify this effect, we compared spectral slope estimates across wakefulness, N3 and REM sleep. In Study 3, we obtained a separate baseline eyes-closed recording during rest in 14 out of 20 subjects. In this subset, we observed prominent slope differences between quiescent rest (−1.87 ± 0.18; mean ± SEM), N3 sleep (−3.46 ± 0.16) and REM sleep (−4.73 ± 0.23; Figure 2b). These differences were significant when averaged across all scalp EEG channels (repeated-measures ANOVA permutation test: p<0.0001, obs. F1.94, 25.17 = 56.05, dRest-Sleep = 3.07). Notably, N2 sleep exhibited an average slope of −3.67 ± 0.10 that was also significantly below rest (Figure 2—figure supplement 1a; permutation t-test: pRest-N2 <0.0001; obs. t13 = 7.97; dRest-N2 = 3.31). Permutation t-tests revealed a significant difference between rest and N3 (pRest-N3 <0.0001, obs. t13 = 5.69, dRest-N3 = 2.49), between rest and REM (pRest-REM <0.0001, obs. t13 = 11.67, dRest-REM = 3.71) and between N3 and REM sleep (pN3-REM = 0.0001, obs. t13 = 4.44, dN3-REM = 1.70). Importantly, while some overlap of absolute spectral slope values between rest and sleep existed when comparing across individuals (Figure 2—figure supplement 1a), we observed a consistent individual decrease of – 2.06 ± 0.21 (mean ± SEM) between rest and all sleep stages (Figure 2—figure supplement 1b; Rest-N1 = −1.95 ± 0.26, Rest-N2 = −1.81 ± 0.23, Rest-N3 = −1.59 ± 0.28, Rest-REM = −2.86 ± 0.25). Figure 2 with 11 supplements see all Download asset Open asset The spectral slope tracks changes of arousal level in sleep. (a) Time-resolved average of three frontal EEG channels (F3, Fz, F4) during a night of sleep. Upper panel: Expert-scored hypnogram (black), wake (pink), REM (light green). Upper middle: Time-frequency decomposition. Lower middle: Spectral slope (black; mean ± SEM). Lower panel: Slow frequency (<1.25 Hz) power (gray; mean ± SEM). (b) Sleep in scalp EEG. Upper panel: Left: Slope difference between sleep and rest (n = 14). Cluster permutation test: *p<0.05. Right: Mutual Information (MI) between the time-resolved slope and hypnogram (n = 20). Cluster permutation test against surrogate distribution created by random block swapping: *p<0.05. Lower panel: Left - Power spectra (n = 14; mean ± SEM); Right – Spectral slope (n = 14). Rest (magenta), NREM stage 3 (blue), REM sleep (green) and grand average (black; mean ± SEM). Repeated measures ANOVA permutation test: ***p<0.001. (c) Sleep in intracranial EEG (n = 10). Upper panel: Left – coronal, right – axial view of intracranial channels that followed (magenta) or did not follow (white) the EEG pattern of a lower slope during sleep (REM/N3). Lower panel: Left – Power spectra (mean ± SEM); Right – Spectral slope of simultaneous EEG recordings (Fz, Cz, C3, C4, Oz). Wakefulness (red), NREM stage 3 (N3; blue), REM sleep (green) and grand average (black; mean ± SEM). Repeated measures ANOVA permutation test: ***p<0.001. Including all available wake periods (before, during and after the night of sleep in all 20 subjects) increased the variance (Figure 2—figure supplement 1c), which can be explained by the fact that subjects were still drowsy and data during state transitions was included. However, the overall pattern was remarkably similar (Figure 2—figure supplement 1b,d). As this approach increased the available amount of data, we utilized all wake trials (referred to as wake) for subsequent we assessed the spatial of where the slope tracks the we the Mutual Information (MI) between the time-resolved spectral slope and the is to assess the between a and data and In we also all analyses based on linear which comparable results (Figure 2—figure supplement We observed that the spectral slope closely tracked the hypnogram at all electrodes as indicated by a permutation test (n = average = ± (mean ± SEM); all to a Figure 2b). This effect frontal electrodes and = ± Figure 2b). of muscle activity overlap with the frequency range used for spectral slope we for muscle by the after et al., 2013). In we utilized correlations that the slope of the EMG as a These control analyses indicated that these the observed between the hypnogram and the spectral slope <0.0001; Figure 2—figure supplement Spatial of sleep state-dependent spectral slope We established that the spectral slope closely tracks the However, we observed differences between scalp electrodes (Figure the which brain to the observed at the scalp In a level using an exhibited the modulation (Figure 2—figure supplement To further the of and subcortical we obtained intracranial EEG recordings (Study n = which were combined with simultaneous scalp EEG recordings. we to the results from Study Again, we found that the slope decreased from wakefulness ± mean ± SEM) to N3 sleep ± and REM sleep ± Figure These three states were significantly different in a ANOVA permutation test obs. = = directly the pattern as observed in Study Permutation t-tests revealed a significant difference between wakefulness and REM obs. = = and wakefulness and N3 obs. = = but only between N3 and REM obs. = = we directly tested which intracranial SEEG the observed scalp EEG We observed the same pattern - a more negative spectral slope in N3 and REM sleep as compared to wakefulness - in of SEEG test against = Figure Figure 2—figure supplement Importantly, this revealed that medial and medial temporal see Table 2, Figure 2—figure supplement exhibit a significantly of electrodes sleep state-dependent slope modulation compared to their lateral - lateral p<0.0001, = – lateral temporal p<0.0001, = hence, on the same brain known to be the for et al., et al., et al., 2013; Murphy et al., Table 2 Anatomical distribution of stereotactically placed intracranial depth electrodes in Study 4 – Intracranial sleep (n = 10). Brain regionTotal number of electrodesElectrodes with state-dependent slope %)Prefrontal Cortex %) medial Prefrontal Cortex %) lateral Prefrontal Cortex %) Orbito-frontal Cortex %)Medial Temporal Lobe %) %) %)Cingulate (100 %)Lateral Temporal Cortex %)Parietal Note that we did not specifically brain and in to previous studies using subdural grid electrodes (Gao et al., 2017; He et al., 2010), the majority of our were stereotactically placed depth electrodes - N3 and - REM see Figure 2—figure supplement subdural grid electrodes see Figure 1—figure supplement the spatial of intracranial and was a on medial that the pattern observed at level (Figure 2—figure supplement and the scalp EEG electrode (Figure The spectral slope discriminates wakefulness from states of reduced arousal Our findings provide that the spectral slope reliably discriminates wakefulness from sleep. reports indicated that slow waves are a of decreased arousal states (Brown et al., 2010; Franks and Zecharia, 2011; Murphy et al., 2011). We directly compared well slow activity and spectral slopes estimates differentiate arousal states using both a linear and a general linear to discriminate different sleep states based on either the spectral slope or slow activity in subjects subjects had to be to wake Note that both the and the were on the same values that were used in the Both the as well as the provide a namely the of trials for and the explained variance quantified by for the a between different The the additional of the of the of the spectral slope and slow Data were with and were we directly tested if the spectral slope is in REM sleep from wakefulness. We found that was enhanced using the spectral slope compared to slow power ± (mean ± SEM), slow ± permutation t-test: obs. = = Figure This that the spectral slope a marker that discriminates REM sleep from wakefulness solely from the electrophysiological brain state. Note that is by the of the underlying sleep as a the between sleep is typically et al., the accurately the in of the Figure 3 Download asset Open asset of wakefulness from sleep and general anesthesia and general linear a – were and averaged across channels (a) Using the 1/f slope (n = two patients had to be to wake resulted in a higher of of wakefulness and REM compared to slow power (<1.25 Hz; ± (mean ± SEM), ± permutation t-test: observed = = line – level at obs. = = obs. = = (b) The of power and spectral slope resulted in comparable of wakefulness and NREM