1. Neuroscience
Download icon

Modality-specific tracking of attention and sensory statistics in the human electrophysiological spectral exponent

  1. Leonhard Waschke  Is a corresponding author
  2. Thomas Donoghue
  3. Lorenz Fiedler
  4. Sydney Smith
  5. Douglas D Garrett
  6. Bradley Voytek  Is a corresponding author
  7. Jonas Obleser  Is a corresponding author
  1. Max Planck UCL Centre for Computational Psychiatry and Ageing Research, Max Planck Institute for Human Development, Germany
  2. Center for Lifespan Psychology, Max Planck Institute for Human Development, Germany
  3. Department of Cognitive Science, University of California, San Diego, United States
  4. Eriksholm Research Centre, Oticon A/S, Denmark
  5. Neurosciences Graduate Program, University of California, San Diego, United States
  6. Halıcıoglu Data Science Institute, University of California, San Diego, United States
  7. Kavli Institute for Brain and Mind, University of California, San Diego, United States
  8. Department of Psychology, University of Lübeck, Germany
  9. Center of Brain, Behavior, and Metabolism, University of Lübeck, Germany
Research Article
  • Cited 0
  • Views 704
  • Annotations
Cite this article as: eLife 2021;10:e70068 doi: 10.7554/eLife.70068

Abstract

A hallmark of electrophysiological brain activity is its 1/f-like spectrum – power decreases with increasing frequency. The steepness of this ‘roll-off’ is approximated by the spectral exponent, which in invasively recorded neural populations reflects the balance of excitatory to inhibitory neural activity (E:I balance). Here, we first establish that the spectral exponent of non-invasive electroencephalography (EEG) recordings is highly sensitive to general (i.e., anaesthesia-driven) changes in E:I balance. Building on the EEG spectral exponent as a viable marker of E:I, we then demonstrate its sensitivity to the focus of selective attention in an EEG experiment during which participants detected targets in simultaneous audio-visual noise. In addition to these endogenous changes in E:I balance, EEG spectral exponents over auditory and visual sensory cortices also tracked auditory and visual stimulus spectral exponents, respectively. Individuals’ degree of this selective stimulus–brain coupling in spectral exponents predicted behavioural performance. Our results highlight the rich information contained in 1/f-like neural activity, providing a window into diverse neural processes previously thought to be inaccessible in non-invasive human recordings.

Introduction

Non-invasive recordings of electrical brain activity represent the aggregated post-synaptic activity of large cortical neuronal ensembles (Buzsáki et al., 2012). Frequency spectra of electrophysiological recordings commonly display quasi-linearly decreasing power with increasing frequency (in log/log space; see Figure 1; see Miller et al., 2009; Voytek et al., 2015). In humans, this decrease is super positioned by several peaks, of which the most prominent is typically in the range of alpha oscillations (~8–12 Hz; see Buzsáki et al., 2013). The overall decrease in power as a function of frequency, f, reflects aperiodic as opposed to oscillatory activity; the steepness of such 1/fχ spectra can be captured by the spectral exponent χ, wherein smaller values reflect flatter spectra.

Figure 1 with 2 supplements see all
Electroencephalography (EEG) spectral exponents track anaesthesia-induced E:I changes.

(A) Normalized EEG spectra averaged across five subjects and five central electrodes (inset) displaying a contrast between rest and propofol (left) and ketamine anaesthesia (right). Spectral parameterization yielded aperiodic fits that estimated the spectral exponent (dashed lines) and full fits that included oscillatory spectral peaks (transparent lines). (B) Pairwise scatter plots depicting subject-wise averaged EEG spectral exponents during awake rest, propofol (left) and ketamine (right). Coloured dots represent spectral exponents of 5 s snippets, black horizontal bars single subject means. p-Values based on 1000 random permutations of data.

Importantly, inter-individual differences in the steepness of human electroencephalography (EEG), power spectral densities (PSD), estimated by the spectral exponent χ, are related to chronological age, behavioural performance, and display stable inter-individual differences (Dave et al., 2018; Donoghue et al., 2020; Sheehan et al., 2018; Voytek et al., 2015; Waschke et al., 2017). Intra-individual variations in EEG spectral exponents have been reported as a function of overall arousal level and activation (Colombo et al., 2019; Lendner et al., 2020; Podvalny et al., 2015). Based on computational models and invasive recordings of neural activity, it has been demonstrated that electrophysiological spectral exponents capture the balance of excitatory and inhibitory neural activity (E:I; Gao et al., 2017), with recent causal, optogenetic work showing that lower exponents are mechanistically linked to increased E:I balance (Chini et al., 2021). Although unknown at present, it is plausible that differences in non-invasive electrophysiological spectral exponents might also reflect variations in E:I balance. Additionally, it remains unclear if specific, behaviourally relevant intra-individual E:I variations that accompany the allocation of attentional resources in non-human animals (Ferguson and Cardin, 2020; Kanashiro et al., 2017; Ni et al., 2018) can be inferred in a similar manner in humans. An attentional shift towards a given sensory modality is suggested to entail desynchronized activity (i.e., reduced low-frequency oscillations and increased high-frequency power) in cortical areas that process information from the currently attended domain (Cohen and Maunsell, 2011; Harris and Thiele, 2011). These shifts in desynchronization likely trace back to an attention-related change in E:I balance towards excitation (Harris and Thiele, 2011; Zagha and McCormick, 2014), which is thought to manifest in a reduction of spectral exponents (Gao et al., 2017; Waschke et al., 2019). If EEG spectral exponents indeed represent a sensitive approximation of E:I balance, selective attention should also result in a topographically specific decrease of exponents. Such attention-related decreases in EEG spectral exponents should take place over and above potential changes in alpha oscillations which have long been assumed to capture dynamics of sensory cortical inhibition in the service of selective attention (Jensen and Mazaheri, 2010; Klimesch et al., 2007).

A conjecture of the present study thus is that 1/f-like EEG activity captures changes in the E:I balance of underlying neural populations. Such a non-invasive approximation of variations in human E:I would be of great value, enabling investigations of processes previously inaccessible using non-invasive imaging techniques. This includes the role of E:I in sensory processing and perception (Wehr and Zador, 2003; Wörgötter et al., 1998), selective attention (Harris and Thiele, 2011) and ageing (Luebke et al., 2004), and not least in disease (Cummings et al., 2009; Dani et al., 2005; Lisman, 2012).

We here will demonstrate the sensitivity of EEG spectral exponents to variations in E:I in two ways. First, predicated on the differential effects of propofol and ketamine (Concas et al., 1991; Deane et al., 2020; Franks, 2008) on E:I (im-)balance, and building on invasive as well as modelling work, we compare EEG spectra of a previously published dataset (Sarasso et al., 2015) between quiet wakefulness, propofol, and ketamine anaesthesia and demonstrate that EEG spectral exponents are valid markers of broad intra-individual variations in E:I balance. Considering the EEG spectral exponent as a viable estimate of E:I balance, we then sought to test the extent to which it could be modulated also through a cognitive manipulation.

Second, we investigated the critical role of E:I balance in attention. We analysed data from human participants who performed a challenging multisensory detection task during which they had to attend to one of two concurrently stimulated sensory modalities (auditory vs. visual) to detect faint target stimuli. Our results demonstrate that despite constant multisensory input, selective attention entails a modality-specific reduction of spectral exponents (spectral flattening) that is in line with an increased E:I ratio.

When aiming to explain the 1/f spectral exponent of the EEG and its relation to E:I balance, it is important to also consider potential other factors that might alter spectral exponents. This is relevant especially since 1/f-like processes are not limited to the brain, but are ubiquitous in nature (Brown et al., 2002; Coensel et al., 2003; Keshner, 1982; Mandelbrot and Wheeler, 1983). Given reports of behaviourally relevant synchronization of oscillatory brain activity with oscillatory sensory inputs (Breska and Deouell, 2017; Henry et al., 2014; Lakatos et al., 2008; Spaak et al., 2014), this raises the question of whether a similar link might exist between the 1/f profiles of neural responses and sensory inputs. Indeed, even speech signals show a pronounced 1/f shape within their amplitude modulation (AM) spectrum (see e.g., Attias and Schreiner, 1997). Hence, 1/f-like sensory inputs might hold ecological relevance. However, despite first non-human in vitro and in vivo evidence for such a neural tracking of 1/f sensory signals (Nozaki et al., 1999; Qu et al., 2019; Yu et al., 2005), it is unclear at present if the 1/f structure of sensory inputs is tracked by 1/f-like non-invasively recorded neural activity in humans.

How would 1/f characteristics of sensory input map onto 1/f of neural activity, if not by changing endogenous E:I balance? Most parsimoniously, a spectral-exponent concordance between sensory statistics and EEG activity might trace back to the superposition of postsynaptic potentials, the main source of EEG signals (Buzsáki et al., 2012). Postsynaptic potentials in sensory cortical areas could temporally align and scale with the magnitude of sensory stimuli and hence result in a mimicking of the stimulus spectrum, as has been suggested for oscillatory signals and steady state evoked potentials (e.g., Norcia et al., 2015) and broadband signals or speech (Lalor et al., 2009). It thus appears plausible that the spectral exponent of the human brain will capture and prove not only endogenous changes in E:I balance but also reflect statistical features of the sensory input. If such non-oscillatory neural tracking of 1/f sensory features (i.e., an alignment of neural and sensory spectral exponents) is non-invasively detectable in humans, this would greatly extend the toolset of perceptual neuroscience to including a wide range of naturalistic (1/f-like) stimuli.

Thus, in addition to neurochemical and attentional effects on EEG spectral exponents that likely capture changes in E:I balance, we investigated the link between environmental 1/f input and 1/f-like brain activity. We tested the single trial relationship between 1/f sensory features and 1/f EEG spectral exponents. Revealing modality-specific and topographically distinct links of stimulus exponents and EEG spectral exponents that explain inter-individual differences in task performance, we present evidence for the behaviourally-relevant neural tracking of 1/f-like sensory features.

Results

Establishing the validity of the EEG spectral exponent as a non-invasive marker of E:I

Previous invasive animal work has demonstrated the sensitivity of spectral exponents to anaesthesia-related changes in the balance of excitatory and inhibitory neural activity (E:I balance; Gao et al., 2017). To expand these results and test, the sensitivity of non-invasively recorded human EEG to physiological changes in E:I balance, we contrasted spectral exponents of human EEG recordings between quiet wakefulness and anaesthesia for two different general anaesthetics: propofol and ketamine. Both anaesthetics exert widespread effects on the overall level of neural activity (Taub et al., 2013) as well as on oscillatory activity in the range of alpha and beta (8–12 Hz; ~ 15–30 Hz). Importantly, however, propofol is known to commonly result in a net increase of inhibition (Concas et al., 1991; Franks, 2008) whereas ketamine results in a relative increase of excitation (Deane et al., 2020; Miller et al., 2016). In accordance with invasive work and single cell modelling (Chini et al., 2021; Gao et al., 2017), propofol anaesthesia should thus lead to an increase in the spectral exponent (steepening of the spectrum) and ketamine anaesthesia to a decrease (flattening). Based on previous results, the effect of anaesthesia on EEG spectral exponents is expected to be highly consistent and display little topographical variation (Lendner et al., 2020). For simplicity, we focused on a set of five central electrodes that receive contributions from many cortical and subcortical sources (see Figure 1) but report topographically resolved effects in the supplements (see Figure 1—figure supplement 2). Here, propofol anaesthesia led to an overall increase in EEG power which was especially pronounced in the alpha-beta range. Ketamine anaesthesia decreased the frequency of alpha oscillations and supressed power in the beta range. Importantly, however, EEG spectral exponents that were estimated while accounting for changes in oscillatory activity increased under propofol and decreased under ketamine anaesthesia in all participants (both ppermuted < 0.0009, Figure 1). These results replicate previous invasive findings and support the validity of EEG spectral exponents as markers of overall E:I balance in humans.

Can the EEG spectral exponent track attentional focus and sensory statistics?

The findings outlined above demonstrate the ability of the EEG spectral exponent to non-invasively track intra-individual changes in E:I balance. Under the presumption that the spectral exponent can thus be used as a viable marker of E:I balance in general, we then sought to test the extent to which the spectral exponent could be modulated through cognitive manipulation, rather than through pharmacological intervention. Accordingly, using an audio-visual decision-making task, we examined selective attention-related variations in the EEG spectral exponent. In addition, expanding on the idea of 1/f-like sensory signals being tracked by cortical activity, we used these data to test a link between the spectral exponents of audio-visual stimuli and recorded EEG signals.

Behavioural performance in a multimodal detection task

Participants (N = 24) performed a challenging multisensory task during which participants had to detect brief time periods during which the AM of the presented white noise switched from aperiodic to sinusoidal (Figure 2A). In detail, participants attended either auditory or visual noise stimuli, which were always presented simultaneously and displayed AM spectra with spectral exponents between 0 and 3 (Figure 2B). While training and adaptive adjustments of difficulty (see methods for details) ensured that the task was challenging but doable in both modalities (average accuracy ≥ 70 %), participants performed better during visual compared to auditory trials (t23 = 5.8, p = 7 × 10–6, Cohen’s d = 1.18; Figure 2C).

Task design and behavioural performance.

(A) Participants were simultaneously presented with auditory and visual amplitude modulated (AM) noise and had to detect periods of sinusoidal AM (grey box) in the luminance variations of a visually presented disk (left) or in auditory presented white noise (right) by pressing a button. Example visual stimuli correspond to spectral exponents of 0 (left) and 2 (right), auditory stimuli to spectral exponents of 2 (left) and 1 (right). (B) Frequency spectra for four sets of AM spectra (left) demonstrate the identical flat spectra (white noise), further visualized by artificially offset spectra in the inset. AM spectra displayed spectral exponents between 0 and 3 (right). (C) Auditory accuracy (70 %) was significantly lower than visual accuracy (73 %). Dots represent single subject data, horizontal lines the mean, and vertical lines ± one standard error.

Modality-specific attention selectively reduces the EEG spectral exponent

To further explore the sensitivity of EEG spectral exponents to specific experimental manipulations, we investigated changes in selective attention, which have been proposed to coincide with a relative increase of excitatory neural activity in sensory cortices of the attended domain. Average EEG spectral exponents for central and occipital regions of interest (see insets in Figure 3) were controlled for neural alpha power and a set of additional nuisance variables (see Methods for details and control analyses for a separate analysis of alpha power) and compared between auditory and visual attention using a 2 × 2 repeated measures analysis of variance. In addition to a main effect of attentional focus (F1,92 = 12.98, p = 0.0005, partial eta squared = 0.124), this analysis revealed an interaction between attentional focus and ROI (F1,92 = 12.91, p = 0.0005, partial eta squared = 0.123). Notably, EEG spectral exponents at occipital electrodes strongly decreased (spectra flattened) under visual compared to auditory attention (t23 = 7.4, p = 1 × 10–7, Cohen’s d = 1.52; see Figure 3B), while this was the case to a much lesser extent at central electrodes (t23 = 2.6, p = 0.01, Cohen’s d = 0.54).

Figure 3 with 1 supplement see all
Electroencephalography (EEG) spectral exponents track the focus of selective attention.

(A) t-values of the average difference between EEG spectral exponents during auditory and visual attention. Note that exponents were controlled for neural alpha power and other confounding variables before subject-wise differences were calculated. A central negative and an occipital positive cluster are clearly visible. (B) Average central (lilac) and occipital EEG spectral exponents (teal) for auditory and visual attention (residuals shown). Horizontal bars denote the grand average. While visual attention as compared to auditory attention was associated with a decrease of spectral exponents at central and occipital sites, this decrease was more pronounced at occipital electrodes. This interaction of ROI × Attention is captured by the cross-over of lilac and teal lines. (C) Grand average spectra for auditory (blue) and visual attention (red), shown for a central (left) and an occipital ROI (right). Insets display enlarged versions of spectra for low and high frequencies, separately.

The topographical specificity of this attention-induced spectral flattening was qualitatively confirmed by a cluster-based permutation test on the relative (z-scored) single-subject EEG spectral exponent differences between auditory and visual attention, revealing a central, negative cluster (p = 0.04) and an occipital positive cluster (p = 0.01). Thus, the selective allocation of attentional resources to one modality results in a flattening of the EEG power spectrum over electrodes typically associated with this modality, especially for occipital electrodes during visual attention.

1/f-like stimulus properties are tracked by modality-specific changes in EEG spectral exponent

Next, our goal was to delineate how changes in the EEG spectral exponent might be driven by sensory input statistics, instead of endogenous changes to E:I balance. In addition to attention-related variations in brain dynamics, we probed the link between spectral exponents of sensory stimuli and EEG activity. This approach departs from the focus on EEG spectral exponents as a non-invasive approximation of E:I and aims to test how the alignment of sensory input with sensory cortical activity could shape EEG spectral exponents to mimic the spectra of sensory signals.

Electrode-wise linear mixed effect models of EEG spectral exponents were used to test the relationship between the AM spectral exponents of presented stimuli and recorded EEG activity (see model details in Supplementary file 1). Of note, the impact of early sensory-evoked responses on estimates of 1/f stimulus tracking (i.e., the linear relationship between EEG spectral exponents and AM stimulus spectral exponents) was excluded by limiting analyses to the time period between 600 ms after stimulus onset and target onset. Single trial parameterizations of power spectra provided excellent fits for the vast majority of trials (mean R2> 0.84 at all electrodes, see Figure 1—figure supplement 2b). Note that reliable spectral parameterization is key to disentangle oscillatory from aperiodic activity, enabling us to analyse EEG spectral exponents which otherwise would be confused with a mix of low frequency and high frequency power. A main effect of auditory stimulus exponent, capturing the trial-wise positive linear relationship between auditory AM spectral exponents and EEG spectral exponents (i.e., neural tracking) was found at a set of four central electrodes (all z > 3.5, all pcorrected< 0.02; see Figure 4A). Similarly, a positive main effect of visual stimulus exponent was present at a set of three occipital electrodes (all z > 3.7, all pcorrected< 0.01; see Figure 4A). Hence, EEG spectral exponents displayed a topographically resolved tracking of stimulus exponents of both modalities. Additionally, standardized single subject estimates of stimulus tracking at both electrode clusters identified by the mixed model approach were extracted after controlling for covariates also used in the final mixed model and revealed qualitatively similar results for both auditory (t23 = 5.1, p = 0.00004, Cohen’s d = 1.03) and visual stimulus tracking (t23 = 3.9, p = 0.0008, Cohen’s d = 0.79; see Figure 4A).

Figure 4 with 7 supplements see all
Electroencephalography (EEG) spectral exponents track stimulus spectral exponents.

(A) Topographies depict t-values for the main effect of stimulus spectral exponent, taken from a mixed model of EEG spectral exponents. White dots represent electrodes with significant effects after Bonferroni correction. Auditory stimulus tracking (upper row) clusters at central electrodes, visual stimulus tracking (lower row) at occipital electrodes. Line plots show single subject tracking estimates (standardized betas), dark grey lines represent negative betas, black lines the grand average. (B) Topographies show t-values for the interaction of attentional focus and stimulus spectral exponent, taken from the same mixed model as in A. Positive clusters appear over central (auditory) and occipital (visual) areas and represent improved tracking during selective attention to the modality in question. Line plots visualize single subject effects of selective attention on stimulus tracking for the clusters found in A (insets). Note that t-values for the auditory stimulus × attention interaction were inverted to remove the sign change caused by zero-centred effect coding of attentional focus. In this way, positive t-values represent evidence for an increase in stimulus tracking if attention is directed towards this modality.

Stimulus tracking in EEG spectral exponent interacts with attentional focus

To further investigate how the just-observed neural tracking of stimulus spectral exponents might interact with the putatively E:I-balance-mediated changes in attentional focus, we tested for stimulus exponent× attention interactions on the observed EEG spectral exponent, within a mixed-model framework.

The interaction of auditory stimulus exponents and attentional focus surfaced as a positive central cluster, and the interaction of visual stimulus exponents and attentional focus yielded a positive occipital cluster (see Figure 4B).

As can be discerned from the topographies in Figure 4B, selective attention likely improved the tracking of stimulus spectral exponents over sensory-specific areas, separately for each sensory domain. To extract single subject estimates of the stimulus × attention interaction, we controlled for several covariates, focusing on EEG spectral exponents averaged within the clusters that displayed significant tracking (see above). Paired t-tests comparing standardized regression coefficients that represent the strength of neural stimulus tracking revealed a significant increase of auditory stimulus tracking under auditory attention (t23 = 2.4, p = 0.03, Cohen’s d = 0.49). Visual stimulus tracking did not significantly improve under visual attention, though the direction of the effect was the same (t23 = 1.0, p =0 .33, Cohen’s d = 0.20). Thus, modality-specific attention was selectively associated with an improvement in the neural tracking of auditory AM stimulus spectra.

Evoked responses and spectral power estimates do not index the spectral exponent of audiovisual stimuli

To test whether conventional estimates of sensory evoked EEG activity tracked the spectral exponent of stimuli, we analysed evoked potentials (ERPs) at the single-trial level. This analysis did not reveal significant ERP clusters for either auditory (p > 0 .3) or visual (p > 0.2) spectral AM tracking (see Figure 4—figure supplement 1 for ERPs). The absence of such effects is visualized in Figure 4—figure supplement 4 which displays ERPs time-courses for four bins of increasing auditory as well as visual AM spectral exponents. In contrast to 1/f EEG exponents, conventional metrics of sensory evoked EEG activity were thus insensitive to the AM spectral exponents of presented stimuli.

Furthermore, to estimate the specificity of the link between EEG spectral exponents and stimulus exponents, we also tested the tracking of stimulus exponents within low-frequency (1–5 Hz) and alpha power (8–12 Hz). To this end, we inverted linear mixed models to either include auditory or visual stimulus exponent as the dependent variable and low-frequency power, alpha power, or EEG spectral exponents as predictors of interest (covariates kept constant across models). Model comparisons revealed better model fits based on EEG spectral exponents throughout (see Figure 4—figure supplement 6). Only when modelling visual stimulus exponents, alpha power explained significantly more variance than EEG spectral exponents at one parietal electrode that was not part of the significant tracking cluster reported above (Figure 4A).

The extent of modality-specific spectral-exponent tracking predicts behavioural performance

Next, to investigate the link between individual levels of neural stimulus tracking and behavioural performance, we computed between-subject correlations of standardized stimulus tracking betas (auditory and visual) and common metrics of behavioural performance (accuracy and response speed, separately for both modalities) using a multivariate partial least squares analysis (PLS; see Methods for details; McIntosh et al., 1996). This model revealed one significant latent variable (p = 0.03) that captured the low-dimensional latent space of the correlation between neural stimulus tracking and behavioural performance, which displayed a clear fronto-central topography (Figure 5A).

Stimulus tracking explains inter-individual differences in performance.

(A) Results of a multivariate neuro-behavioural correlation between stimulus tracking and performance using partial least squares analysis (PLS). The topography depicts bootstrap ratios (BSR) of the first latent variable and can be interpreted as z-values. Within the smaller topography, BSRs are thresholded at a BSR of 2 (p < 0.05). Bar graphs represent the correlation (Spearman rho) between auditory (blue) and visual stimulus tracking (red) with performance (accuracy as % correct and response speed; RS in s–1) in both modalities. Vertical lines denote 95 % bootstrapped confidence intervals. (B) Scatter plots of latent correlations between latent auditory (upper panel) and visual tracking (lower panel) with latent performance, respectively. Auditory stimulus tracking was positively linked with auditory performance but negatively linked with visual performance. Visual stimulus tracking was positively linked with visual performance. Headphones and eye symbolize auditory and visual performance, respectively.

At the overall latent level, both auditory (rho = 0.49, p = 0.016) and visual stimulus tracking (rho = 0.43, p = 0.035) were significantly correlated with performance across subjects (Figure 5B). In detail, auditory stimulus tracking was positively correlated with auditory response speed (rho = 0.30, CI = [0.03, 0.64]) but negatively correlated with visual accuracy (rho = –0.28, CI = [–0.62, –0.06]) and response speed (rho = –0.36, CI = [–0.73, –0.17]). Analogously, visual stimulus tracking was positively linked with visual accuracy (rho = 0.31, CI = [0.15, 0.68]) and response speed (rho = 0.35, CI = [0.11, 0.72]; see Figure 5A), but showed no significant relationship with auditory performance.

Taken together, participants who displayed stronger neural tracking of AM stimulus spectra also performed better. However, the behavioural benefit of stimulus tracking was modality-specific: performance in one sensory domain (e.g., auditory) only benefited from tracking within that domain, but not in the other (e.g., visual). Instead, visual detection performance was slower and less accurate in individuals who displayed strong neural tracking of 1/f-like auditory stimulus features.

Control analysis: no stimulus-exponent effects on evoked potentials or oscillatory power

To compare established metrics of early sensory processing, we contrasted evoked EEG activity of different attention conditions. Event-related potentials (ERPs) after noise onset were compared between auditory and visual attention for electrodes Cz and Oz, separately. While there was no significant attention-related ERP difference found at electrode Cz, early ERP components at electrode Oz were increased during visual attention (~.08 –.15 s post-stimulus, pcorrected < 0.05; see Figure 3—figure supplement 1).

Also, we subjected single trial alpha power (after partializing for EEG spectral exponents) from the ROIs shown in Figure 3 to the same ANOVA (Attention (2) × ROI (2)) as EEG spectral exponents. This analysis revealed a main effect of attention (F = 20.3, p = 0.00002, partial eta squared = 0.180) as well as an interaction of attention and ROI (F = 20.4, p = 0.00002, partial eta squared = 0.181). Hence, alpha power increased from visual to auditory attention in a region-specific manner, with stronger increases over occipital cortical areas (see spectra in Figure 3C).

Importantly, this attention-related change in alpha power took place over and above the observed attentional modulation of EEG spectral exponents for at least three reasons: First, spectral parameterization separates oscillatory and non-oscillatory components, dissecting alpha power from the spectral exponent (Donoghue et al., 2020). Second, and to account for shared variance between alpha power and EEG spectral exponents that might occur due to fitting issues, each measure had been partialed out from each other to account for collinear contributions of the two. Third, attention-related modulations of alpha power and EEG spectral exponents displayed no significant correlation across participants (all electrode-wise pFDR > 0.15).

The presence of modality-specific, attention-dependent changes despite these controls strongly suggests that neural alpha power, as well as EEG spectral exponents, track distinct changes in neural activity that accompany attentional shifts, but they are not affected by the spectral content of the stimuli themselves.

Exploratory analysis: estimating temporal response functions to link stimulus spectra and resulting EEG spectra

To directly test to what degree spectral exponent changes in the stimulus material become reflected in the spectral exponent of the EEG response via a time-resolved, evoked-response–like ‘neural tracking’ (Obleser and Kayser, 2019), we also employed an alternative analysis framework to estimate the link between sensory input and electrophysiological activity in the time domain. Specifically, we used temporal response functions (TRF; Crosse et al., 2016; Lalor and Foxe, 2010; Wöstmann et al., 2016) to capture the temporally resolved multivariate linear relationship between EEG and continuous sensory input.

After estimating a modality-specific impulse response function (see Figure 4—figure supplement 3), this function can be used to either reconstruct stimulus input (backward model) or predict neural recordings (forward model) from the combination of estimated TRFs and EEG data or stimulus time series. Importantly, auditory response functions appeared unreliable and displayed an uncommon shape with only one peak at a lag of 200 ms whereas visual response functions appeared stable. In line with his observation, backward models only yielded significant stimulus reconstruction for the visual domain (see Figure 4—figure supplement 4). Given the absence of reliable auditory TRFs and backward model predictions, we modelled EEG time-series based on a canonical TRF (grand average visual TRF) and stimulus data. This simplistic model indeed produced EEG data whose spectral exponents were positively linked with the spectra of presented stimuli (Figure 4—figure supplement 5) both for the auditory (β = 0.37, SE = 0.08, t = 4.3, p < 0.0001) and the visual domain (β = 0.59, SE = 0.07, t = 8.9, p = 0).

Hence, while TRF-based models were unable to empirically demonstrate a significant link between 1/f-like stimulus input and 1/f-like EEG activity given the data analysed in the current manuscript, these results illustrate the general feasibility and potency of such an approach. However, future work is needed to empirically demonstrate that a positive link between the spectral exponents of sensory input and EEG activity traces back to the temporal alignment of both. Overall, and in given our data, a spectrally based approach that directly links spectra of presented stimuli with those of recorded EEG appears as the more powerful approach to detect an alignment of input and neural activity.

Discussion

We have presented evidence for the sensitivity of the EEG spectral exponent to a variety of influences — neurochemical, cognitive, and sensory in nature. Jointly, these results underscore the value of 1/f-based neural measures such as the spectral exponent for studying the neural bases of perception and behaviour. Specifically, we here have shown that the EEG spectral exponent (1) reflects systemic, anaesthesia-induced changes in brain state closely linked to E:I balance, (2) captures focal attention-related changes in E:I brain state, and (3) simultaneously is able to track sensory stimuli by way of stimulus-specific spectral exponents. Furthermore, when presenting stimuli in the auditory and visual domain in parallel, stimulus tracking retained its modality-specificity in the spatial distribution of EEG spectral exponents. Not least, EEG spectral exponents explained inter-individual variance in behavioural performance, underlining the functional relevance of 1/f processes in human brain activity.

EEG spectral exponents as a non-invasive approximation of E:I balance

As hypothesized, our initial analysis of a general-anaesthesia dataset confirmed that propofol leads to steeper EEG spectra, while ketamine causes flattening (see Figure 1). Importantly, the spectral parameterization approach used here enabled us to separate changes in spectral exponents from other prominent anaesthesia-driven alterations such as boosted oscillatory power in the alpha-beta range (propofol) or decreased alpha frequency (ketamine). While these changes in oscillatory activity might capture anaesthesia-related effects on connectivity and link to the loss of consciousness (Purdon et al., 2013), the present study focused on changes in EEG spectral exponents with the goal of evaluating its sensitivity to changes in E:I balance. The observed increase of spectral exponents during propofol anaesthesia is generally in line with previous results that relied on different measures and methodologies (Colombo et al., 2019; Waschke et al., 2019). Increased spectral exponents under propofol likely trace back to strengthened inhibition via increased activity at gamma-Aminobutyric acid receptors (GABAergic activity) and hence a reduced E:I balance as compared to quiet wakefulness (Concas et al., 1991; Franks, 2008). Furthermore, ketamine-induced decreases of spectral exponents likely depict the outcome of overall decreased inhibition that is caused by the blocking of excitatory N-methyl-d-aspartate (NMDA) receptors and an associated decrease in GABA release, resulting in an increased E:I balance (Behrens et al., 2007; Deane et al., 2020; Franks, 2008).

Our results build on the findings of previous studies (Gao et al., 2017; Lendner et al., 2020; Medel et al., 2020); replicate and validate these earlier results in non-invasive recordings; and offer crucial advancements by directly comparing the effect of two distinct anaesthetics on EEG spectral exponents. Hence, EEG spectral exponents pose a non-invasive approximation of intraindividual state changes in E:I balance.

While the EEG spectral exponent as a remote, summary measure of brain electric activity can obviously not quantify local E:I in a given neural population, the non-invasive approximation demonstrated here enables inferences on global neural processes previously only accessible in animals and using invasive methods. Future studies should use a larger sample to directly compare dose-response relationships between GABA-A agonists or antagonists (e.g., Flumanezil) and the EEG spectral exponent as well as common oscillatory changes.

Next, we capitalized on the sensitivity of the EEG spectral exponent to broad changes in E:I by investigating its ability to capture attention-related fluctuations in cortical E:I balance in a dedicated EEG experiment.

EEG spectral exponents track modality-specific, attention-induced changes in E:I

While the EEG spectral exponent here proved sensitive to systemic, exogenously driven, drug-induced alterations in E:I, we also sought to test whether it can also be leveraged to detect non-drug-induced but endogenous, attention-related variations in E:I balance. Invasive animal work suggests that the allocation of attentional resources should result in a topographically specific shift towards desynchronized activity and an increased E:I ratio (Harris and Thiele, 2011; Kanashiro et al., 2017).

Here, we observed a topographically-specific pattern of reduced EEG spectral exponents through modality-specific attention. As indexed by a significant interaction between region of interest and attentional focus, visual attention led to a flattening of EEG spectra that was especially pronounced over occipital EEG channels (Figure 3). The absence of a comparable effect for auditory attention at central electrodes potentially traces back to the varying sensitivity of EEG recordings to different cortical sources. Central electrodes capture auditory cortical activity but are positioned far away from their dominant source ( Huotilainen et al., 1998; Stropahl et al., 2018). Occipital electrodes, however, are sensitive to visual cortex activity and are directly positioned above it (Hagler et al., 2009). Further exaggerated by the scaling of volume conduction with distance, this likely results in a reduced signal to noise ratio for auditory compared to visual cortex activity (Piastra et al., 2020).

Despite these differences in the sensitivity of EEG signals, our results provide clear evidence for a modality-specific flattening of EEG spectra through the selective allocation of attentional resources. This attention allocation likely surfaces as subtle changes in E:I balance (Börgers et al., 2005; Harris and Thiele, 2011). Importantly, these results cannot be explained by observed attention-dependent differences in neural alpha power (8–12 Hz, Figure 3), which have been suggested to capture cortical inhibition or idling states (Cooper et al., 2003; Pfurtscheller et al., 1996). The employed spectral parameterization approach enabled to us to separate 1/f like signals from oscillatory activity and hence offered distinct estimates of spectral exponent and alpha power that would otherwise have been conflated (Donoghue et al., 2020). By accounting for the shared variance of EEG spectral exponents and alpha power while modelling the impact of modality-specific attention, we further controlled for a potential conflation of both and present evidence for their distinct involvement in the allocation of attentional resources.

How could attentional goals come to shape spectral exponents and alpha oscillations? Both attention-related changes in EEG activity might trace back to distinct functions of thalamo-cortical circuits. On the one hand, bursts of thalamic activity that project towards sensory cortical areas might sculpt cortical excitability in an attention-dependent manner by inhibiting irrelevant distracting information (Klimesch et al., 2007; Saalmann and Kastner, 2011). On the other hand, tonic thalamic activity likely drives cortical desynchronization via glutamatergic projections and, with attentional focus, results in boosted representations of stimulus information within brain signals (Cohen and Maunsell, 2011; Harris and Thiele, 2011; Sherman, 2001).

Our findings of separate attentional modulations of both, EEG spectral exponents and alpha power, point towards the involvement of both thalamic modes in the realization of attentional states. Recently, momentary trade-offs between both modes of thalamic activity have been suggested to give way to attention-related modulations of alpha power and E:I balance, as captured by EEG spectral exponents (Kosciessa et al., 2021). Here, task difficulty remained constant throughout the experiment an fluctuations between both modes might not follow momentary demand (Kosciessa et al., 2021; Pettine et al., 2021) but varying sensory-cognitive resources.

Additionally, attention-related modulations of both alpha power and EEG spectral exponents appeared uncorrelated across individuals – further evidence that they reflect separate neural sources. Future studies that combine a systemic manipulation of E:I (e.g., through GABAergic agonists) with the investigation of attentional load in humans are needed to specify with greater detail how thalamic activity modes drive alpha oscillations and EEG spectral exponents. Specifying potential demand- and resource-dependent trade-offs between different modes of attention-related modulations of cortical activity and sensory processing will offer crucial insights into the neural basis of adaptive behaviour.

EEG spectral exponents track the 1/f features of sensory stimuli

Over and above the demonstrated anaesthesia- and attention-related effects and their putative relation to E:I balance, we also examined if and how EEG spectral exponents contain information about stimulus characteristics of incoming sensory information. This would reflect a separate neural mechanism that the spectral exponent can come to reflect: A neural tracking of stimulus statistics.

For both modalities, the tracking of stimulus spectral exponents, based on single trials, represented a strong effect, as indicated by standardized effect size estimates (Cohen’s d > 0.79). This set of results is especially striking given that conventional sensory-evoked responses or estimates of low-frequency power showed no link with the spectral exponent of sensory stimuli (see Figure 4—figure supplements 1 and 6). Furthermore, potential influences of single trial performance on stimulus tracking were ruled out by a control analysis that accounted for single-trial behavioural performance and replicated the results summarized above (see Figure 4—figure supplement 7).

The neural tracking of stimulus spectra displayed distinct central and occipital topographies for auditory and visual stimuli, respectively (Figure 4). Of course, current source density transforms of sensor-level EEG topographies as used here do not represent definitive evidence for specific cortical sources. Yet, the spatial distinctiveness of both tracking patterns strongly suggests separate cortical origins. Furthermore, the topographies of stimulus tracking strongly resemble those of sensory processing in auditory and visual sensory cortical areas, respectively (Iemi et al., 2019; Waschke et al., 2019).

These results are conceptually in line with findings from extracellular recordings in ferrets demonstrating the tracking of different AM spectra along the auditory pathway (Garcia-Lazaro et al., 2006; Garcia-Lazaro et al., 2011), and also extend previous work that analysed oscillatory human brain activity during the presentation of 1/f stimuli (Hermes et al., 2015; Teng et al., 2018). However, by presenting first evidence for a linear relationship between 1/f-like stimulus features and 1/f-like EEG activity in humans, our results argue for a sensory-specific tracking of AM stimulus exponents within sensory cortical areas at the level of single trials.

Of note, auditory stimulus tracking increased significantly when participants focused their attention on auditory stimuli. A weaker if not statistically significant effect was discernible in the visual domain (Figure 4B). Hence, the selective allocation of attentional resources yielded improved tracking of 1/f-like sensory features. Due to too short, stimulus-free inter-trial intervals, we were unable to analyse if the degree of attention-induced reduction in EEG spectral exponents directly reflected the magnitude of stimulus tracking. Future research is needed to further investigate the precise link between individual averages of EEG spectral exponents, their attention-related change, and the tracking of environmental 1/f distributed inputs.

Neural processes potentially driving 1/f stimulus tracking in the EEG

What might constitute the mechanism that, at the level of sensory neural ensembles, gives rise to the observed link between sensory stimuli and the spectral shape of the EEG? First, it is important to emphasise that the representation of stimulus spectra in the EEG does not trace back to an alignment of true oscillatory neural activity and oscillatory stimulus features, commonly referred to as ‘entrainment’ in the narrow sense (Obleser and Kayser, 2019) entrainment in the narrow sense would be contingent on the presence of true endogenous oscillatory activity within recorded EEG that progressively synchronizes its phase to the phase of exogenous stimulus oscillations. Here, presented stimuli were stochastic in nature and without clear sinusoidal signals, preventing narrow-sense oscillatory entrainment from taking place. Neural tracking of the statistical properties of random noise time-series might emerge via the temporal alignment of high amplitude stimulus periods with high amplitude neural activity periods, a mechanism similar to the one implied in the generation of steady-state evoked potentials (SSEPs; Norcia et al., 2015). SSEPs are commonly studied by presenting rhythmic sensory input to participants which is assumed to evoked trains of evoked responses at the frequency of presentation, resulting in peak in the EEG spectral at that frequency. However, the stimuli and analysis approaches used in the current study suggest a mechanism of neural tracking that goes beyond common steady state responses to a single presentation frequency.

First, phase and amplitude time-courses of stimuli were dissimilar across trials, preventing phase-locked evoked activity, a hallmark of SSEP generation (Vialatte et al., 2010). Second, and unlike most steady state experiments, the tracked stimulus information (AM spectral exponents) was not constant across time but only accessible via temporal integration. Finally, we excluded EEG signals during the first 600 ms after stimulus onset from all tracking analyses and additionally revealed a null-effect of stimulus spectral exponents on single trial evoked potentials. Hence, the observed neural tracking of AM spectral exponents does not emerge via a neural adaptation to constant amplitude spectra or trial-wise differences in evoked responses.

Importantly, the temporal alignment of broadband sensory input with human brain activity has been studied in the context of ‘neural tracking’ using multivariate linear models and might be able to explain the link between stimulus and EEG spectral properties we observe (Lalor and Foxe, 2010; Wöstmann et al., 2016). Here, a linear relationship between time-courses of stimulus features and neural responses is assumed to capture their temporal alignment, commonly referred to as ‘entrainment in the broad sense’ (Obleser and Kayser, 2019).

As outlined above, we estimated auditory and visual TRFs to test whether forward modelling of EEG data would result in EEG spectra that mimicked properties of stimulus spectra. However, auditory TRFs were unreliable (see Figure 4—figure supplement 1). Visual TRFs on the other hand enabled significant stimulus reconstruction and were used within a simplified proof-of-concept model to predict EEG signals that indeed mimicked the spectral properties of stimuli (Figure 4—figure supplement 4). The non-predictiveness of auditory TRFs potentially traces back to an insufficient signal-to-noise ratio and limited training data. In general, EEG spectral exponents might also capture the consequences of non-linear interactions between stimulus input and neural response by focusing on their spectral representation across a wide frequency range. Such non-linear links of stimulus and response are by design inaccessible to TRF approaches that rely on the linear relationship of both time series.

Although spectral-based approaches of neural stimulus tracking clearly displayed higher power in context of the analysed dataset, we deem it probable that both approaches eventually capitalize on the same aspect of central neural processing: the temporal alignment of high amplitude/salience stimulus events with high amplitude neural activity. While this does not correspond to entrainment in the narrow sense or SSVEP-like superposition of oscillatory activity or ERPs, 1/f AM spectra might evoke trains of evoked responses with similar spectral exponents. Indeed, a simple proof-of-concept model based on real stimulus data resulted in EEG spectra whose exponents were positively linked with the exponents of stimuli (Figure 4—figure supplement 5). Hence, time- and spectrally based approaches of stimulus tracking might indeed capture similar aspects of postsynaptic neural activity that align with sensory input during early processing. Importantly, however, future studies are needed to further test the relationship between temporal neural tracking using TRF approaches and spectral tracking as put forward in the current manuscript.

Neural stimulus tracking explains inter-individual differences in performance

We used partial least squares to investigate the multivariate between-subject relationship of neural stimulus tracking to behavioural performance. While the statistical power of this between-subject analyses clearly is inferior to other analyses of the current manuscript, the resulting positive correlations (non-parametric) between stimulus tracking and performance cannot be explained by potential outliers (see Figure 5). Importantly, this effect was confined to each modality; while individuals who displayed high auditory tracking also displayed fast responses in auditory trials, they exhibited slower and less accurate responses on visual trials. Furthermore, participants who showed strong tracking of visual stimuli performed especially fast and accurate on visual but not auditory trials. This specificity of behavioural benefits through stimulus spectral tracking to each modality argues against the idea of attention-dependent sensory filters that entail bi-directional effects (i.e., auditory attention = visual ignoring; Lakatos et al., 2013; Obleser and Kayser, 2019).

The sample size of N = 24 is modest (although it does not stand out as small when compared with usual practices of the field), and we have employed non-parametric correlations combined with a two-stage permutation approach to not rely on unwarranted assumptions. Nevertheless, data from a larger cohort of individuals would be ideal to test this between-subject relationship, better estimate its effect size, and test the generalizability of observed neuro-behavioural correlations. Furthermore, introducing within-subject manipulations of difficulty in a comparable design might allow for the investigation of within-subject effects of neural stimulus tracking on perceptual performance.

Of note, the topography of the between-subject neuro-behavioural correlations does not display peaks at the central or occipital regions that were found to show significant stimulus tracking. However, a difference between both topographies is plausible for at least two reasons. First, the sensory stimulus tracking topographies represent fixed effects, by definition minimizing between-subject variance. In contrast, the between subject correlation of stimulus tracking and performance seeks to maximise between-subject variance, increasing the probability that a non-identical topography may be found. Second, although stimulus exponents are significantly tracked at sensor locations that point to early sensory cortices, our multisensory task required more abstract, high-level representations of sensory input for accurate performance. Indeed, the fronto-central topography of the between subject correlation is suggestive of sources in frontal cortex, which have been shown to track multi-sensory information (Ghazanfar and Schroeder, 2006; Senkowski et al., 2007). Furthermore, prefrontal cortex activity that gives rise to highly similar frontal topographies (Figure 5) has been found to represent information about the frequency content of auditory, visual, and somatosensory stimuli (Spitzer and Blankenburg, 2012). Thus, the positive link between neural stimulus tracking and performance at fronto-central electrodes points to the behavioural relevance of higher-level stimulus features represented in a supramodal fashion.

Limitations and next steps

First, attention-dependent changes in EEG spectral exponents might trace back to altered sensory-evoked responses. We argue that such a link is unlikely since differences in evoked responses were limited to the visual domain, occipital electrodes, and an early time-window (80–150 ms post noise onset, see Figure 3—figure supplement 1) that was well detached from the time-window used to extract single trial EEG spectra (starting at 600 ms post noise onset).

Additionally, TRF-based models of EEG rather speak to a steepening of spectra via increases of sensory processing that accompany attentional focus, contrary to the observed decreases (see Figure 4—figure supplement 5). However, this does not rule out entirely a remaining conflation of selective-attention effects and sensory-processing signatures in the EEG spectral component, as no trials without sensory input were included. Although sensory input was comparable across different attention conditions (auditory and visual stimuli simultaneously), future studies are needed to further specify the link between modality-specific attention and EEG spectral exponents in the absence of sensory input.

Second, one reason for the difference in attentional improvement of stimulus tracking between modalities might lie in the difficulty of the task. Although auditory and visual difficulty were closely matched, we found significantly lower performance for auditory compared to visual trials (see Figure 2). Although we deem it unlikely that the observed difference of 3 % in accuracy (70% vs. 73%) might be indicative of a meaningful difference in performance, we cannot rule out the possibility that participants needed more cognitive resources to perform the auditory task and neurally track stimulus spectra. Due to these increased demands, the effects of selective attention might have been able to amplify stimulus tracking more strongly as compared to the potentially less demanding visual condition. Future studies should investigate the role of parametric task demands for stimulus tracking, and attentional improvements thereof, by additionally recording fluctuations in pupil size during constant light conditions as a proxy measure of demand-related fluctuations in arousal (Yerkes and Dodson, 1908; Zekveld et al., 2010).

Conclusion

The present data show that the EEG spectral exponent represents a non-invasive approximation of intra-individual variations in states of E:I balance – may these E:I states be driven globally (here, by central-acting anaesthetics) or more focally by re-allocation of selective attention. In addition to these links between E:I and EEG spectral exponents, we highlight the sensitivity of the EEG spectral exponent to aperiodic, 1/f-like stimulus features simultaneously in two sensory modalities and in relation to behavioural outcomes. These findings pose a tightening link from invasive non-human animal physiology to human cognitive neuroscience. They set the stage for a new line of experiments, using non-invasive approximations of aperiodic neural activity to study intra-individual variations in brain dynamics and their role in sensory processing and behaviour.

Materials and methods

Pre-processing and analysis of EEG data under different anaesthetics

Request a detailed protocol

To test the effect of different central anaesthetics on 1/f EEG activity, we analysed a previously published openly available dataset (Sarasso et al., 2015). Sarasso and colleagues recorded the EEG of healthy individuals during quiet wakefulness and after the administration of different commonly used central anaesthetics including propofol and ketamine. Details regarding the recording protocol can be found in the original study and a recently published re-analysis (Colombo et al., 2019; Sarasso et al., 2015). We analysed EEG (60 channels) recordings from 10 participants who either received propofol or ketamine infusion (5/5). EEG data were re-referenced to the average of all electrodes, down-sampled to 1000 Hz, and filtered using an acausal finite impulse response bandpass filter (0.3–100 Hz, order 127). Next, to increase the number of samples per condition, recordings were split up into 5 s epochs. Since the duration of recordings varied between participants, this resulted in different numbers of epochs per anaesthetic and participant (propofol: 98 ± 88 epochs; ketamine: 69 ± 29 epochs). The power spectrum of each epoch and electrode between 1 and 100 Hz (0.25 Hz resolution) was estimated using the Welch method (pwelch function). The spectral parameterization algorithm (version 1.0.0; Donoghue et al., 2020) was used to parameterize neural power spectra. Settings for the algorithm were set as: peak width limits: [1 – 8]; max number of peaks: 8; minimum peak height: 0.05; peak threshold: 2.0; and aperiodic mode: ‘fixed’. Power spectra were parameterized across the frequency range 3–55 Hz.

To statistically compare EEG spectral exponents between quiet wakefulness (resting state) and anaesthesia despite the low number of participants (four per anaesthesia condition), we focused on five central electrodes (see inset in Figure 1) and employed a permutation-based approach. After comparing average spectral exponents of resting state and anaesthesia recordings using two separate paired t-tests, we permuted condition labels (rest vs. anaesthesia) and repeated the statistical comparison 1,000 times. Hence, the percentage of comparisons that exceed the observed t-value represents an empirically defined p-value. Note that spectra were normalized before visualization (Figure 1) and non-normalized power spectra can be found in the supplements (Figure 1—figure supplement 1).

EEG spectral exponents during a multisensory detection task

Request a detailed protocol

To investigate the dynamics of 1/f EEG activity during varying selective attention and the processing of sensory stimuli with distinct 1/f features, we recorded EEG from 25 healthy undergraduate students (21 ± 3 years old, 10 male) while they performed a challenging multisensory detection task. All participants gave written informed consent, reported normal hearing and had normal or corrected to normal vision. All experimental procedures were approved by the institutional review board of the University of California, San Diego, Human Research Protections Program. Due to below-chance level performance, one participant had to be excluded from all further analyses.

Task design and experimental procedure

Request a detailed protocol

The novel multisensory design used in the current study required participants to focus their attention to one modality of concurrently presented auditory and visual noise stimuli to detect brief sinusoidal amplitude variations of the presented noise (Figure 2A). Participants were asked to press the spacebar as fast and accurately as possible whenever they detected such a sinusoidal AM (target) in the currently attended sensory domain. The experiment was divided into 12 blocks of 36 trials (432 trials total). At the beginning of each block participants were instructed to detect targets embedded in either auditory or visual noise stimuli. The to be attended modality alternated from block to block and was randomized across participants for the first block. Prior to each trial, the central white fixation cross changed its colour to green and back to white to indicate the start of the next trial. After 500 ms the presentation of noise in both modalities started simultaneously. Trials lasted between 4 and 4.5 s, ended with the central fixation cross reappearing on the screen, and were separated by silent inter-trial intervals (uniformly sampled between 2 and 3.25 s). After each experimental block, participants received feedback in the form of a percentage correct score and were asked to take a break of at least 1 min before continuing. Participants were seated in a quiet room in front of a computer screen. The experiment, including EEG preparation, lasted approximately 2.5 hr.

To ensure that the task was challenging but executable for all participants to a comparable degree, we combined training with an adaptive tracking procedure. In detail, participants performed four practice trials of each modality during which target stimuli were clearly detectable. Subsequently, participants performed 12 blocks of 36 trials each during which difficulty was adjusted by changing the modulation depth of presented targets to keep performance constantly around 70 % correct.

Stimulus generation

Request a detailed protocol

Auditory and visual stimuli of different AM spectra were built in three steps: First, 30 s segments of white noise (sampling frequency 44.1 kHz) were generated and high-pass filtered at 200 Hz. Second, four random time-series of the same duration but differing 1/fχ exponent (χ = 0, 1, 2, or 3) were generated using an inverse Fourier transform and lowpass filtered at 100 Hz. Finally, separate multiplication of the white noise carrier with the modulators of different spectral exponents resulted in four signals that only varied in their AM but not in their long-term frequency spectra (see Figure 2B). The same noise was used for auditory and visual stimuli after root mean square normalization (auditory) or down-sampling to 85 Hz and scaling between 0.5 and 1 (visual). Noise stimuli presented during the experiment were cut out from the 30 s long time-series. Importantly, the AM spectra of cut out noise snippets do not necessarily overlap with the AM spectra of the longer time-series they were cut from. This difference between global and local spectra resulted in a wide distribution of AM spectra that were presented throughout the experiment. AM exponents were uncorrelated between modalities across trials. Auditory noise was presented as amplitude modulated white noise over headphones whereas visual noise was shown as luminance variations of a visually presented disk. Targets consisted of short sinusoidal AMs (6–7.5 Hz, 400 ms) and modulation depth was varied throughout the experiment to keep performance around 70 % correct. All stimuli were generated using custom Matlab code. Auditory stimuli were presented over headphones (Sennheiser using a low-latency audio soundcard (Native Instruments)). Visual stimuli were presented on a computer screen (85 Hz refresh rate). Both auditory and visual stimuli were presented using MATLAB and Psychophysics toolbox (Brainard, 1997). To later analyse the relationship between EEG activity, behaviour and the AM spectra of presented stimuli, single trial stimulus spectra were extracted (1–30 Hz, 0.1 Hz resolution, pwelch in MATLAB) and parameterized to fit 1/f exponents (Donoghue et al., 2020). Settings for the algorithm were set as: peak width limits: [0.5–12]; max number of peaks: infinite; minimum peak height: 0; peak threshold: 2.0; and aperiodic mode: ‘fixed’. Power spectra were parameterized across the frequency range 1–25 Hz.

EEG recording and pre-processing

Request a detailed protocol

64-channel EEG was recorded at a sampling rate of 1000 Hz using the brainamp and the actichamp extension box (active electrodes; Brainproducts). Artifacts representing heartbeat, movement, eye blinks or saccades and channel noise were removed using independent component analysis based on functions from the fieldtrip and EEGlab toolboxes (Delorme and Makeig, 2004; Oostenveld et al., 2011). Components were rejected based on power spectra, time-series, topography and dipole fit. Continuous EEG signals were referenced to the average of all channels and filtered between 0.05 and 100 Hz (acausal FIR filter, order 207). Data were segmented into trials between –1 and 5 s relative to trial start (noise onset) and baseline corrected to the average of 1 s prior to trial start. Trials containing artifacts were removed based on visual inspection (5 ± 7 trials rejected). EEG time-series were transformed to scalp current densities using default settings of the fieldtrip toolbox (ft_scalpcurrentdensity). Single trial power spectra between 1 and 100 Hz (0.5 Hz resolution) were calculated using the welch method (Welch, 1967). To minimize the impact of early sensory-evoked potentials, these spectra were based on the EEG signal between 600 ms after noise stimulus onset and the appearance of a target sound. Trials during which the target appeared within 500 ms after trial start were excluded (1.5% ± 0.3% of trials). Furthermore, mirrored versions of single trial data were appended to the beginning and end of each trial before calculating spectra, effectively tripling the number of samples while not introducing new information (i.e., ‘mirror padding’). To approximate EEG activity in a state of awake rest despite the absence of a dedicated resting state recording, we calculated spectra based on the 500 ms before the fixation cross changed its colour, signalling the start of the next trial (same settings as above). These single trial ‘resting-state’ spectra were averaged to reveal one resting state spectrum per participant and electrode. To estimate 1/f spectral exponents of EEG activity as well as oscillatory activity, single trial and average resting state spectra were fed into the spectral parameterization algorithm (Donoghue et al., 2020) and exponents were fit between 3 and 55 Hz using Python version 3.7. Trials where fits explained less than 20 % of variance in EEG spectra were excluded from all further analyses (0.4 % ± 0.6 % of trials). Note that single trial parameterization provided good fits in all subjects and at all electrodes. Figure 4—figure supplement 2 provides an overview of fit statistics and examples of two representative single trial fits. To control for the influence of alpha-oscillations (8–12 Hz), we extracted single-trial, single-electrode power estimates from spectral parameterization results if an oscillation was detected within the alpha frequency range (see clear alpha-range peak in Fig. S4). For trials where this was not the case, spectral power was averaged between 8 and 12 Hz as a substitute.

Statistical analysis

Attention tracking

Request a detailed protocol

To test whether the allocation of attentional resources to one sensory domain is accompanied by a selective flattening of the EEG power spectrum over related sensory areas, we followed a two-step approach. First, we used multiple linear regression to control single trial, single electrode EEG spectral exponents for a number of covariates. Specifically, and for every participant, we controlled EEG spectral exponents for the influence of auditory stimulus exponents, visual stimulus exponents, alpha power, and trial number. Next, we averaged the residuals per electrode and attention condition (auditory vs. visual attention) across trials, resulting in 2 × 64 EEG spectral exponent estimates per participant. The topographical pattern of the average difference between auditory and visual attention EEG spectral exponents is visualized in Figure 3A. Following our hypothesis of sensory specific, attention-related flattening of EEG spectra, we averaged EEG spectral exponent residuals across a set of fronto-central (FC1, FC2, Fz, C1, C2, Cz) and parieto-occipital (P03, PO4, POz, O1, O2, Oz) electrodes to contrast activity from auditory and visual sensory areas, respectively. We modelled EEG spectral exponent as a function of ROI (central vs. occipital), attentional focus (auditory vs. visual) and their interaction, including a random intercepts and random slopes for all predictors within a linear mixed model (fitlme in MATLAB). Subsequently, analysis of variance (ANOVA) was used to statistically evaluate the main effect of attentional focus as well as the interaction of attentional focus and ROI. Note that attentional focus only varied between blocks of 36 trials, in contrast to stimulus exponents which varied on a trial-wise basis. Hence, a single-trial modelling approach as described below for stimulus tracking is not warranted here as it would artificially increase statistical power. Furthermore, such an approach would have complicated the statistical evaluation of the hypothesized interaction of ROI with attentional focus.

As an additional control analysis, we tested the subject- and electrode-wise differences of EEG spectral exponents between auditory and visual attention against zero using a cluster-based permutation approach (Maris and Oostenveld, 2007). Finally, we used paired t-tests to compare attention effects between ROIs and resolve the interaction effect.

Stimulus tracking

Request a detailed protocol

To test the link between EEG spectral exponents and AM spectral exponents in the auditory and visual domain on the level of single trials, we used single-electrode linear mixed effect models. Model fitting was performed iteratively and hypothesis-driven, starting with an intercept only model and gradually increasing model complexity to find the best fitting model (Tune et al., 2018; Waschke et al., 2019). After every newly added fixed effect, model fits were compared using maximum likelihood estimation. Once the final set of fixed effects was determined, a comparable procedure was used for random effects. All continuous variables were z-scored across participants (but within electrodes) before entering models, rendering the regression coefficients β direct measures of effect size. The winning model included fixed effects for auditory stimulus exponents, visual stimulus exponents, attentional focus, trial number and resting state EEG spectral exponent (between subject factor) as well as a random intercept for all predictors and random slopes for attentional focus (see Supplementary files 1 and 2). To additionally test the role of selective attention for the tracking of the presented sensory stimuli, we modelled separate interactions between auditory as well as visual stimulus exponents with attention, respectively. As models were fit for single electrodes (64 models), we corrected the resulting p-values for multiple comparisons by adjusting the family-wise error rate using the Bonferroni-Holm correction (Groppe et al., 2011; Holm, 1979).

To arrive at single subject estimates of stimulus tracking, we chose a stepwise regression approach since including random slopes for auditory or visual stimulus exponents within the final mixed models did not improve model fit but led to convergence issues due to model complexity. Hence, on the level of single participants and electrodes, we regressed EEG spectral exponents on attentional focus, trial number, and stimulus exponents of one modality (e.g., visual). The z-scored residuals of this multiple regression were used in a second step where they were regressed on the z-scored stimulus exponent of the remaining sensory domain (e.g., auditory). The resulting beta coefficients were averaged across the electrodes that showed significant stimulus tracking within the mixed model approach, representing single subject estimates of stimulus tracking (see Figure 3). By limiting data to trials from one attention condition, single subject estimates of stimulus tracking for different targets of selective attention were calculated following a similar approach.

Model comparisons

Request a detailed protocol

To compare the reported tracking of stimulus exponents in EEG spectral exponents with other established spectral measures of EEG activity and estimate the specificity of the reported effects, we performed formal model comparisons. To this end, we inverted all models and instead of modelling EEG spectral exponents used auditory or visual stimulus exponents as dependent variables. Predictors were identical to the previously reported models (see Supplementary files 1 and 2 for details) but additionally included either single trial estimates of alpha power (8–12 Hz), low-frequency power (1–5 Hz), or EEG spectral exponents. Note that alpha power estimates were extracted using the same spectral parameterization approach that was used to estimate spectral exponents. Trials without a detected oscillation in the alpha range were excluded from all models to render likelihood comparisons interpretable (11.2% ± 3.4 % of trials excluded). Since oscillations were only seldomly detected in the low-frequency range, we instead used single trial power averaged across this range as a predictor. For each electrode, 4 likelihood ratio tests were performed, one for each stimulus modality and one for each predictor (low-frequency or alpha power), always testing against the respective EEG spectral exponent-based model.

Temporal response functions

Request a detailed protocol

To explore an alternative model of stimulus tracking that operates in the time-domain, we estimated multivariate temporal response functions (mTRFs) of EEG recordings to continuous auditory and visual sensory input. Here, we used the amplitude envelope onsets of auditory as well as visual stimulus time-courses to estimate TRFs. To this end, we filtered the absolute values of their analytic signals below 20 Hz, calculated the first derivative and applied half-wave rectification, down-sampling to 250 Hz, and trial-segmentation, following common standards in the field (Crosse et al., 2016; Fiedler et al., 2019). EEG data were down-sampled to 250 Hz and cut into trial segments. EEG and stimulus data from all trials were used to train forward or backward multivariate linear models using ridge regression and a set of different regularization parameters (10–3–103) as implemented in the mTRF toolbox (Crosse et al., 2016).

The forward model approach resulted in auditory and visual response functions per electrode and subject (see Figure 4—figure supplement 3). Next, using backward models and leave-one-subject-out cross-validation, we reconstructed stimulus time-courses based on response functions and recorded EEG. To this end, response functions and model constants were averaged across all but one left-out subject after which these TRFs and EEG data from the hold-out subject were used to reconstruct stimulus time-courses. Reconstructed and observed stimuli were correlated on a single trial basis. For each participant, we separately tested the across-trial distributions of correlations against zero using one sample t-tests (see Figure 4—figure supplement 4).

Given non-significant stimulus reconstruction of auditory backward models, we could not definitely conclude that phase locking explains the mapping of EEG and stimulus spectra we report. To nevertheless test this hypothesis at least on a conceptual level, we decided to base the prediction of EEG data on a canonical response function (grand average visual TRF) instead of subject- and electrode-wise TRFs. Hence, we predicted one channel EEG data solely based on one canonical TRF per attention condition and trial-wise stimulus time-courses of both modalities. To extract estimates of simulated EEG spectral exponents, we applied the same approach outlined above for real EEG data. The results of this simplistic model were evaluated using a linear model that included main effects for stimulus exponent (four bins of equal size) and attention (auditory/ visual) and their interaction.

Neuro-behavioural correlation

Request a detailed protocol

To investigate whether inter-individual differences in neural stimulus tracking relate to inter-individual differences in performance, we analysed correlations between single subject estimates of stimulus tracking and different metrics of behavioural performance using a multivariate PLS approach (Krishnan et al., 2011; McIntosh et al., 1996). In brief, so-called ‘behavioural PLS’ begins by calculating a between-subject correlation matrix linking brain activity at each electrode with behavioural measures of interest. The size of this rank correlation matrix is determined by the number of electrodes, brain variables and behavioural variables [size = (Nelectrodes× Nbrain variables)× Nbehavioural variables]. In the present study, we used two brain variables with 64 electrodes each (auditory and visual tracking betas) and four behavioural variables (auditory and visual accuracy and response speed). Next, this correlation matrix is decomposed using singular value decomposition (SVD), which results in Nbrainvar × Nbehavvar latent variables (eight in our case).

This approach produces two crucial outputs: (1) A singular value for every latent variable, representing the proportion of cross-block covariance accounted for by that latent variable, and; (2) a pattern of weights (n = number of electrodes) or saliencies representing the correlation strength between stimulus tracking and the used behavioural measures. The multiplication (dot product) of these weights with electrode-wise tracking estimates yields so-called ‘brain scores’, which here reflect the between-subject relationship of stimulus tracking and performance where positive brain scores indicate that individuals with stronger tracking display better performance. Statistical significance of brain scores and latent variables was tested through permutations of behavioural measures across individuals (5,000 permutations). Additionally, the robustness of weights (saliencies) was estimated using a bootstrap procedure (5,000 bootstraps, with replacement). The division of each weight by the corresponding bootstrapped standard error yields bootstrap ratios, which estimate the robustness of observed effects on an electrode-wise basis. Bootstrap ratios can be interpreted as a pseudo-Z metric. Crucially however, because multivariate PLS is run in a single mathematical step that includes (and weights the importance of) all elements of the brain-behaviour matrix, multiple comparisons correction is neither typical nor required (McIntosh et al., 1996). Furthermore, bootstraps were used to estimate 95 % CIs for observed neuro-behavioural correlations.

Of note, behavioural performance metrics were calculated separately for auditory and visual attention trials. Since difficulty was adaptively adjusted throughout the experiment, leading to vastly different target modulation depths across participants, we controlled auditory and visual accuracy for the final modulation depth of the respective domain and used the residuals as a measure of performance. To furthermore exclude influences of learning and exhaustion of response speed (reaction time–1) we controlled single trial response times for trial number and used averaged residuals as subject-wise indicators of response speed. Note that such an approach is especially warranted since we were specifically interested in between-subject relationships and hence the association of correlation matrices between neural tracking and performance. In accordance with such a reasoning, and to prevent outliers in our small sample to obscure results, all PLS analysis were performed using spearman correlation.

Control analyses

Request a detailed protocol

To test the impact of attentional focus and AM stimulus spectra on sensory evoked activity, which might potentially confound differences in EEG spectral exponents, we compared ERPs after noise onset. First, we compared noise onset ERPs between auditory and visual attention using a series of paired t-tests, separately for electrodes Cz and Oz. We corrected for multiple comparisons by adjusting p-values for the false discovery rate (Benjamini and Hochberg, 1995). Next, on the level of subjects, voltage values were correlated with stimulus spectral exponents (separately for auditory and visual stimuli) across trials, per electrode, time-point and frequency (ft_statfun_correlationT in fieldtrip). On the second level, the resulting t-value time-series were tested against zero using a cluster-based permutation approach (Maris and Oostenveld, 2007), separately for auditory and visual stimuli. Finally, to rule out task difficulty as a potential confound of stimulus tracking, we re-ran stimulus tracking mixed models including a main effect of single trial performance (correct vs. incorrect) as well as interactions between single trial performance and auditory and visual stimulus spectral exponents, respectively, to control stimulus tracking for performance and difficulty.

Data availability

Data and code has been deposited on OSF and is available via https://osf.io/wyzrg/.

The following data sets were generated
The following previously published data sets were used
    1. Massimini M
    2. Laureys S
    (2017) Zenodo
    Rest EEG recordings in healthy subjects during wakefulness, sleep and anesthesia with ketamine, propofol, and xenon.
    https://doi.org/10.5281/zenodo.806176

References

  1. Book
    1. Attias H
    2. Schreiner CE
    (1997)
    Temporal Low-Order Statistics of Natural Sounds
    MIT Press.
    1. Coensel BD
    2. Botteldooren D
    3. Muer TD
    (2003)
    1/f Noise in Rural and Urban Soundscapes
    ACTA ACUSTICA UNITED WITH ACUSTICA 89:10.
    1. Keshner MS
    (1982) 1/f noise
    Proceedings of the IEEE 70:212–218.
    https://doi.org/10.1109/PROC.1982.12282

Decision letter

  1. Maria Chait
    Reviewing Editor; University College London, United Kingdom
  2. Barbara G Shinn-Cunningham
    Senior Editor; Carnegie Mellon University, United States
  3. Bradley R Postle
    Reviewer; University of Wisconsin-Madison, United States
  4. Jonathan Z Simon
    Reviewer; University of Maryland, United States

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Acceptance summary:

This manuscript reports on two separate investigations. In the first, the authors provide novel evidence from two anaesthesia challenges that the slope of the 1/f structure of the power spectrum of the EEG fluctuates in a manner that tracks the presumed excitation: inhibition (E:I) balance of the tissue generating the EEG signal. Next they show that fluctuations in this slope also covary in systematic and modality- and stimulus-specific ways with behavioral performance on a multimodal attention task. These observations have potential foundational implications for how this previously unappreciated component of the EEG can be interpreted in terms of brain physiology and function.

Decision letter after peer review:

Thank you for submitting your article "Modality-specific tracking of attention and sensory statistics in the human electrophysiological spectral exponent" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by Maria Chait as Reviewing Editor and Barbara Shinn-Cunningham as the Senior Editor. The following individuals involved in the review of your submission have agreed to reveal their identity: Bradley R Postle (Reviewer #1); Jonathan Z Simon (Reviewer #2).

The Reviewing Editor and reviewers have discussed the reviews with one another. We agree that the balance between excitation and inhibition in the cortex is an important and timely topic. While the approach you use to uncover the role of E/I in attention is interesting, unfortunately, for the reasons outlined below the data as they stand do not support the study's conclusions. We feel that this is potentially addressable with a revision and the RE has drafted this to help you prepare a revised submission.

Essential revisions:

(1) The results of Experiment 1, whilst compelling, require a delicate interpretation. In particular, it is difficult to make a clear distinction between different anaesthetics in terms of their effect on brain activity (see references provided by Rev3). Given this and the low N, results thus do not fully support the strong conclusions offered by the authors. We encourage the authors to revise based on the specific comments from Reviewers 1 and 2 including:

(a) addressing the pattern of spectral effects (propofol mostly enhancing frequencies below ~20-30 Hz and spreading α; ketamine suppressing α while enhancing lower and higher frequencies),

(b) justifying why modelling this as a 1/f change is appropriate, and

(c) quantifying the differences between the different awake spectra (there appears to be a large difference in the awake spectra between the anaesthetics conditions).

Please also acknowledge the limitations associated with small N and existing literature highlighted by Reviewer #3.

(2) The authors interpret the findings of Experiment 2, where changing the value of the spectral exponent in the stimulus resulted in a similar change in the value of the spectral exponent of the response, but only for the selectively attended modality, as originating from an attention-driven change in E/I balance. However, an alternative interpretation of the findings is that these effects reflect attention-driven changes to temporal tracking of the stimulus waveform. These concerns are potentially addressable in a revision but it would require an entirely new data analysis involving a thorough investigation of potential temporal tracking of the stimulus waveform and an unambiguous result. There will need to be a visual temporal analysis (a la VESPA) and auditory temporal analysis (a la AESPA) for both the attended and unattended conditions. The part of the response explained would need to be subtracted out first, and then the "spectral-exponent-tracking" analysis would need to be performed on the residual. There may be additional subtleties that arise in that process. Given the successes of AESPA/VESPA/TRFs in the literature, this should be considered a simpler explanation of the observed response patterns than dependence on E:I balance. It's the residual (true response minus response explained by this mechanism) that would still need an explanation, and that might be argued to be explainable by E:I balance.

Reviewer #1 (Recommendations for the authors):

Figure 1, it's really hard to see how the slopes change in the way that the authors state. For propofol, visual inspection suggests that the biggest change is a broadening of the α oscillation, such that its inflection starts at a lower frequency and then because the peak is also 'less pointy,' the purple line simply has to fall at a higher rate to catch up with the gray line by ~30 Hz. For ketamine, at the lowest frequencies (lower than α bump) the slope of the green line simply is steeper than the gray, and then again the biggest difference seems to be that the α bump is abolished with ketamine, and so the gray line is then steeper than the green line for the same reason that purple appears to be steeper than gray in propofol plot. Additionally, there's a lot of jitter with ketamine in the 20-60 Hz range. I realize that visual inspection isn't a rigorous way to analyze these data, but on the other hand it's generally preferable for a figure to clearly illustrate the point that the authors are trying to convey. Perhaps the authors should consider accompanying the 'raw' spectra shown here with the same data decomposed into oscillatory vs. aperiodic components, the way that it is done in the Donoghue et al., (2020) paper?

The Discussion section is largely a repetition of what was written in the Intro and/or a restatement of the results with little additional interpretation and contextualization. For example, although it's important to show that α and aperiodic components of the EEG are statistically dissociable, this is only a step toward understanding more fundamental questions such as (a) what are the functions that periodic vs. aperiodic components support? and (b) what underlying factors that give rise to them?

Here are some more specific comments about the Discussion.

"Jointly, these results underscore the importance of 1/f brain activity for perception and behaviour." Don't the authors really mean: "underscore the utility of parameters of 1/f brain activity for studying the neural bases of perception and behavior"? At the end of the day, the major take-home of this paper is that the slope of the 1/f spectrum is a valid index of E:I balance, but it's E:I balance, per se, that is 'important for perception and behavior,' not the slope itself.

"… these results cannot be explained by attention-dependent differences in neural α power (8-12 Hz, Figure 3), commonly interpreted as a marker of top-down guided sensory inhibition." Idling is an important alternative to inhibition that should be acknowledged.

"First, it is important to emphasise that the representation of stimulus spectra in the EEG likely does not trace back to an alignment of oscillatory neural activity and oscillatory stimulus features, commonly referred to as "entrainment" in the strict sense; the presented stimuli were stochastic in nature and without clear sinusoidal signals. However, neurally tracking the statistical properties of random noise time-series might emerge via a mechanism similar to the one implied in the generation of steady-state evoked potentials (SSEPs)." Both of these seem like important points that merit more elaboration. That is, the word "entrainment" tends to be used carelessly and so more detailed and explicit argumentation about why this is NOT an instance of entrainment would be valuable. With regard to SSEPs, specifying some details about this 'implied mechanism' would be helpful. More generally, although entrainment and evoked responses are precisely specified processes that can be shown to be true or not, the same is not true for "tracking," which is just a loose concept that can't be tested and falsified. Can the authors either specify what they mean by "tracking" or else replace it with a more rigorously defined process?

Reviewer #2 (Recommendations for the authors):

P. 4, last paragraph: It is somewhat disconcerting to learn in the Results section that the first study uses a publicly available dataset and the second is wholly separate and from data acquired by the authors. This would be be less startling if it were mentioned in the introduction.

Lines 159-160: As written, this sentence seems to implies that the new results of this paper aren't actually new but merely a confirmation of an old result. It would easier on the reader to more clearly distinguish the previous results (with very strong connections to E:I balance?) from the new findings (where the connection to E:I balance is less direct).

Figure 1B: Would the authors consider using the same vertical scale in both graphs? The overall numbers between the two sets are close enough in value that having two different scales can be distracting.

Figure 1B: The inset graphs are missing axis limits (or scale), and there is no definition of their error bars.

L. 196 and elsewhere: incorrect formatting of numbers in scientific notation, e.g. 7e-6 instead of 7 x 10-6.

L. 189 and following: The description of the stimuli, especially the auditory stimuli is confusing. The phase "to detect regular (i.e., sinusoidal) amplitude variations in streams of amplitude modulated white noise", in the auditory literature would be understood as analogous to "to detect tone pips in noise", but that is not what is meant here. Figure 2 indicates rather that the stimulus temporarily changes from non-sinusoidal amplitude modulated white noise to sinusoidal amplitude modulated white noise.

Figure 2C: Please explain what the circles and lines represent (I presume individual subjects with lines representing identities, but I need toask after seeing Figure 3B).

Figure 3B: Please explain what the circles and lines represent. Do the lines connect the different tasks of the same individuals? The systematic progression of the slopes of the lines seems to indicate that they do not.

Lines 234-235: Getting R2 > 0.84 is a real achievement-it speaks very highly of the importance of the spectral exponent.

L. 383: the phrase "and hence" is confusing here. Maybe "even though they"?

L. 419 and Supplemental Figures: There are two supplemental figures labeled as S4 and none as S5. This reference appears to be to the 5th supplemental figure.

Lines 445-475: This section appears to be where the possibility of temporal tracking is meant to be addressed, but it does not accomplish this (instead only justifying that steady-state analysis does not apply here, which is true). Note also to be careful with the word "stationary". A "stationary process" is one with a fixed spectrum and random phases, which seems to be a good description of the stimulus envelopes/contrasts used here.

Lines 576-577. What does "normalized" mean here? Standard usage is a multiplicative rescaling, not mean-centering. [On the other hand, if the mean-centering was performed on the logarithm (or in dB), then that is equivalent to a multiplicative rescaling of the original waveform.]

L. 616 and following: Regarding the visual modulation, why is the acoustic noise, which had been high-passed at 200 Hz before its modulation, downsampled to 85 Hz (which throws away all the carrier information), instead of just applying the 1/fX modulation directly (downsampled to 85 Hz)? Why the extra complication? Or am I just confused by the multiple uses of the word "noise"?

L. 739 and following: I very much appreciate the careful analysis methods employed here.

Figure S1 caption: This caption would be much clearer if it stated that the graphs and data were identical to that shown in Figure 1 except without normalization. (In its current form it seems almost like an example of an item in a change-blindness study.)

Figures 1B and S1B. There seems to be a lot of inter-subject variability in the Awake case between the subjects who used Propofol vs Ketamine (which should have nothing to do with the awake case). Is that an artifact of changes in the axis scaling (or normalization)? It shouldn't matter since the important statistics are changes within subject, but it is a little disconcerting.

Reviewer #3 (Recommendations for the authors):

The authors based their correlation analysis on 24 participants. While the authors do argue that bigger sample size and cross-validation could strengthen the results, the authors could do more with the data they have.

For example, they can employ a leave-one-out linear regression approach, or use k-folds

With regards to the ERP analysis, the authors appear to be using a cluster-permutation approach to assess any differences between the conditions. Here they do have to keep in mind that such a mass-univariate approach is biased towards longer-sustained responses that have a wide scalp distribution, than the rather more focal discrete ERP components. Please see refer to the following discussion on this topic.

https://projects.iq.harvard.edu/files/kuperberglab/files/fieldskuperberg_psychophysiology_2020.pdf

Finally, why I am intrigued by the idea of the slope of 1/f as being something rather important, I am still not convinced that it could be a residual of other factors in the EEG, such as changes in slow frequency power, or evoked responses. I think it would be interesting to see how much unique variance the change 1/f can contribute relative to the other measures of the EEG.

https://doi.org/10.7554/eLife.70068.sa1

Author response

Essential revisions:

(1) The results of Experiment 1, whilst compelling, require a delicate interpretation. In particular, It is difficult to make a clear distinction between different anaesthetics in terms of their effect on brain activity (see references provided by Rev3). Given this and the low N, results thus do not fully support the strong conclusions offered by the authors. We encourage the authors to revise based on the specific comments from Reviewers 1 and 2 including:

(a) addressing the pattern of spectral effects (propofol mostly enhancing frequencies below ~20-30 Hz and spreading α; ketamine suppressing α while enhancing lower and higher frequencies),

(b) justifying why modelling this as a 1/f change is appropriate, and

(c) quantifying the differences between the different awake spectra (there appears to be a large difference in the awake spectra between the anaesthetics conditions).

Please also acknowledge the limitations associated with small N and existing literature highlighted by Reviewer #3.

We agree that the effects of anaesthetics on synaptic brain activity are complex and are not yet fully understood. While there are of course additional details and complexities relating to the action of anaesthetics, in this work, we sought to highlight a particular and salient aspect of neural activity, that is the shift in 1/f-structure that can be putatively related to the balance of excitatory and inhibitory activity. We also note that our analysis exists in the broader context of several other related studies that have examined state-related shifts in 1/f-like electrophysiological activity, for example in connection with sleep (Lendner et al., 2020; Bódizs et al., 2021), and also in other investigations of anesthesia (Gao et al., 2017; Colombo et al., 2019; Medel et al., 2020).

It is important to emphasize that this analysis and argument is not the central thesis of our manuscript, but rather is intended as a framing of the main experimental results. While there is ample evidence for predominant and distinct effects of propofol and ketamine on cortical E/I balance (Concas et al., 1991; Zhang et al., 2009; Ching et al., 2010; Brown et al., 2011; Deane et al., 2020) we realize that the initial version of our manuscript could have offered a more nuanced view on the topic. Hence, we have revised sections of introduction and discussion to reflect the complexity of anaesthesia-related dynamics, referencing work suggested by Reviewer 3. Furthermore, we explicitly note the limitation that comes with the small N of study 1 which might at least partially persist despite our use of non-parametric, permutation-based methods and the strong consistency of effects.

We appreciate the reviewers’ thoughtful comments on the specific spectral effects of anaesthesia and related modelling choices as well as interpretations. We have expanded relevant sections of results and discussion paragraphs to also address spectral differences in addition to the spectral exponent (a). Importantly, and with regard to (b), we now elaborate in greater detail on the spectral parameterization that is performed, also providing an additional visualization (see response to R1). In brief, we of course acknowledge the diverse effects of anaesthesia on brain activity which manifests in oscillatory as well as aperiodic aspects of the brain signal. As noted above, however, the goal of this analysis was to test for the potency of EEG spectral exponents as a potential marker of E:I balance by contrasting the spectral exponent of propofol and ketamine anaesthesia recordings against awake rest. While there might also be differences in oscillatory power between anaesthesia conditions, it is unclear how they link to E:I balance. More importantly, the analysed spectral exponents represent the 1/f-like aperiodic part of the spectrum after controlling for oscillatory peaks. Hence, we employ a theoretically guided and methodically supported focus on the EEG spectral exponent instead of an extensive characterization of the effects anaesthesia exerts on EEG spectra.

Nevertheless, the updated version of the manuscript offers more details on the observed spectral changes while explicitly noting the reason and suitability of our focus on spectral exponents, grounded in more recent work in this area.

Finally, we compared resting-state spectra between anaesthesia conditions, finding no significant difference between both (pperm = .75). The absence of such a group difference also becomes apparent in Author response image 1 which directly compares the awake state spectral exponents from both anaesthesia groups. In the updated version of the manuscript, we report this comparison of awake spectral exponents and have updated figure 1 to have constant y-axis limits (see response to R2), making it easier to visually compare spectral exponents of awake recordings.

Author response image 1
EEG PSD exponents do not significantly differ during awake rest.

EEG PSD exponents are shown for two groups of subjects, one that later received propofol and another that received ketamine anaesthesia. Single dots represent spectral exponents from 5 second long snippets, horizontal lines depict individual averages. Despite higher inter individual differences in the propofol group, no significant difference between these awake EEG PSD exponents of both groups was found (p = .75).

(2) The authors interpret the findings of Experiment 2, where changing the value of the spectral exponent in the stimulus resulted in a similar change in the value of the spectral exponent of the response, but only for the selectively attended modality, as originating from an attention-driven change in E/I balance. However, an alternative interpretation of the findings is that these effects reflect attention-driven changes to temporal tracking of the stimulus waveform. These concerns are potentially addressable in a revision but it would require an entirely new data analysis involving a thorough investigation of potential temporal tracking of the stimulus waveform and an unambiguous result. There will need to be a visual temporal analysis (a la VESPA) and auditory temporal analysis (a la AESPA) for both the attended and unattended conditions. The part of the response explained would need to be subtracted out first, and then the "spectral-exponent-tracking" analysis would need to be performed on the residual. There may be additional subtleties that arise in that process. Given the successes of AESPA/VESPA/TRFs in the literature, this should be considered a simpler explanation of the observed response patterns than dependence on E:I balance. It's the residual (true response minus response explained by this mechanism) that would still need an explanation, and that might be argued to be explainable by E:I balance.

We thank the reviewers for these suggestions. However, we believe that the perceived issues trace back to a misunderstanding of our original analyses for which we apologize and that we try to resolve step by step below. In brief, we report separate main effects of attention and stimulus spectral exponents on EEG PSD exponents, which we interpret as capturing different neural processes. Importantly, we have also re-analysed all data from experiment 2 using the suggested temporal response function approach and report detailed results after clarifying misunderstandings.

“changing the value of the spectral exponent in the stimulus resulted in a similar change in the value of the spectral exponent of the response, but only for the selectively attended modality”:

This is a misunderstanding. The linear relationship between stimulus spectral exponents and EEG spectral exponents (i.e., stimulus tracking) was present for both modalities, and only partially depended on attentional focus (Figure 4A). While we only found significant auditory stimulus tracking for auditory attention trials (see figure 4B), a similar attention×tracking interaction was not present for the visual domain where stimulus tracking was not significantly affected by attentional focus.

“The authors interpret the findings […] as originating from an attention-driven change in E/I balance.”:

This is not our intended interpretation of the stimulus tracking finding and we apologize if the initial version of our manuscript was not clear enough in differentiating interpretations of the main effects of attention (Figure 3) and stimulus tracking (Figure 4). In the original version of our manuscript, we interpreted the main effect of stimulus tracking as potentially originating from the alignment of high amplitude stimulus periods with high amplitude periods of local field activity (originally Discussion section “1/f stimulus tracking as a sign of non-oscillatory steady state-like activity”, now titled “Neural processes potentially driving 1/f stimulus tracking in the EEG”). Importantly, this assumed temporal alignment of stimulus features and neural activity in early sensory cortices can equally account for a positive link between stimulus and EEG spectra (spectral exponent tracking) and temporally resolved stimulus-EEG correlations (temporal response functions, TRF) as used in the speech tracking approaches suggested by reviewer 2. Hence, both techniques offer complimentary views on the same assumed neural process: changes in brain activity that trace back to the alignment of neural firing and features of sensory input. While we had discussed this idea in the context of and in contrast to SSEPs, we realize that a link with common approaches in the area of speech tracking might have been even more fitting. In the updated version of the manuscript, we explicitly link our findings to common TRF approaches.

“However, an alternative interpretation of the findings is that these effects reflect attention-driven changes to temporal tracking of the stimulus waveform.”:

We assume that the reviewers are referring to stimulus tracking effects (figure 4), specifically to the interaction of attention and stimulus input (figure 4B) and want to highlight that we interpret the significant interaction of attentional focus and stimulus tracking exactly along these lines (see lines 541 and following in the Discussion section).

We wish to point out that we interpret the separate main effects of attention and stimulus tracking to result from different neural processes. In contrast to the main effect of stimulus exponents on EEG PSD exponents, the main effect of attention (figure 3) in our view might capture local changes in cortical E/I balance towards excitation that are unrelated to attention dependent changes in stimulus processing. An important role of stimulus processing in this attention-dependent change of 1/f-like EEG activity appears unlikely for at least three different reasons. (1) All relevant statistical models contained attentional focus, stimulus exponents, and the interaction of stimulus exponent with attentional focus as predictors, controlling for shared influences on EEG spectral exponents. (2) All EEG spectra and hence all following analyses were based on a post-stimulus time-window that did not include the evoked response to noise onset and attention-related differences therein. (3) TRF-based EEG modelling confirmed empirically found attention-dependent changes in ERPs which led to an increase instead of a decrease in PSD exponents of modelled EEG data, tracing back to increased low frequency power (see below for details). Hence, it appears improbable that attention-related changes in sensory processing drive the observed reduction of EEG PSD exponents. If anything, such changes in stimulus processing might render our results to underestimate the true size of attention-related spectral flattening.

Temporal response function analyses:

We agree with the reviewers that a promising alternative approach to analyse the tracking of stimulus spectra in EEG data lies within the estimation of multivariate temporal response functions (mTRF), the widely used descendants of AESPA and VESPA. In brief, such an approach estimates the time-resolved linear relationship between one or more stimulus timeseries and concurrent neural activity. After estimating the impulse response function which captures the time-lagged weights that reveal the common sensory response if applied to neural data, this function can be used to either reconstruct stimulus input (backward model) or predict neural recordings (forward model) from the respective other. The correlation of predicted and observed signals offers an intuitive measure of model fit or predictive accuracy.

Here, we used the amplitude envelope onsets of auditory as well as visual stimulus signals to estimate TRFs. To this end, we filtered the absolute values of the analytic signals below 20 Hz before applying half-wave rectification, down-sampling to 250 Hz, and trial-segmentation, following common standards in the field (Crosse et al., 2016; Wöstmann et al., 2017). EEG data were down-sampled to 250 Hz and cut into trial segments. We then used EEG and stimulus data from all trials to train forward or backward multivariate linear models using ridge regression and a set of different regularization parameters (10-3–103) as implemented in the mTRF toolbox (Crosse et al., 2016). This approach resulted in auditory and visual response functions per electrode and subject. As can be discerned from figure 4 supplement 2, visual response functions displayed a common pattern of deflections that is reminiscent of sensory-evoked responses as well as a common occipital topography and were highly consistent across individuals. Auditory response functions, however, only displayed one peak at a lag of approximately 150-200 ms with an uncommon centro-parietal topography and high inconsistency across individuals (compare size of CIs). These results cast doubt on the suitability of the mTRF approach to predict stimuli or EEG signals given the data analysed in the present manuscript.

We continued by utilizing a leave-one-subject-out cross-validation approach to reconstruct stimulus time-courses based on response functions and recorded EEG. Thus, response functions and model constants were averaged across all but one hold-out subject before these response functions and EEG data from the hold-out subject were used to reconstruct stimulus signals. Reconstructed and observed stimuli were correlated on a single trial basis. For each participant, we separately tested the across-trial distributions of correlations against zero using one sample t-tests. While this procedure revealed significant stimulus reconstruction in the visual domain, this was not the case for auditory stimuli (see with figure 4 supplement 3). Given that such backward models are considerably more powerful compared to forward models (predicting EEG) due to the number of used features (64 electrodes across time), these results render interpretable reconstructions of EEG data based on auditory stimuli unlikely. The difference in response function stability and stimulus reconstruction power between auditory and visual modalities might trace back to variations in SNR between central and occipital electrode locations with lower SNR and decreased selectivity for auditory cortical activity at central electrodes as compared to visual cortical activity at occipital sites. Additionally, the employed design always presented stimuli from both modalities concurrently – a major difference to previous studies that showed tracking of noise amplitudes in the EEG (Lalor et al., 2009). In such a multisensory setup, visual inputs might have dominated and superposed auditory inputs at least partially. Considering these results, our finding of significant links between single trial auditory stimulus exponents and EEG PSD exponents speaks to the power of this spectrum-based approach. Single trial EEG spectral exponents might be sensitive to non-linear links between stimulus and postsynaptic activity that temporal domain response function approaches miss.

Given the instability of auditory TRFs and the failure to reconstruct stimulus time-series using backward modelling, we decided to employ a exemplary forward modelling approach to predict EEG signals from TRFs and stimuli. Instead of relying on subject-specific temporal response functions, we utilized two canonical response functions: the grand average visual response functions at electrode Oz from auditory and visual attention conditions (Figure 4- Figure supplement 2). We used these canonical response functions to predict single trial EEG data based on auditory or visual stimuli and calculated spectra of the predicted EEG signals. These spectra were parameterized using the same approach applied to empirical EEG data.

This modelling approach disregards individual differences as well as topographical information. However, it allowed to theoretically test whether the temporal tracking of sensory input in the EEG might result in EEG spectra that mimic the shape of stimulus spectra. As can be discerned from figure 4 supplement 4, predicted spectra displayed the commonly observed 1/f decay in power with increasing frequency, interrupted by a peak around 15 Hz. Although the centre frequency of this peak exceeded the usual range of α oscillations, we deem the spectra of modelled data a satisfactory approximation of human EEG signals.

Importantly, the spectral exponent of predicted EEG signals was positively linked with the spectral exponent of presented amplitude modulation spectra, both for the visual and the auditory domain (betavis = 0.59, SE = 0.07, t = 8.9, p < 0.0001; betaaud = 0.37, SE = 0.08, t = 4.3, p < 0.0001). These findings illustrate that the temporal alignment of sensory cortical activity with stimulus intensity fluctuations, sometimes referred to as phase-locking or entrainment in the broad sense (Obleser and Kayser, 2019), could in theory account for links between stimulus and EEG spectra we report. However, as outlined above, the data presented in the current manuscript resulted in unreliable individual TRFs and non-significant stimulus reconstruction for the auditory domain. Hence, our data failed to support this hypothesis empirically. This disparity of model and data might be rooted in the insufficient SNR of used EEG recordings or the limited duration of used sensory stimuli. Additionally, common speech tracking approaches rely on at least 20 min of sensory presentation per participant whereas the analysed data only contained roughly 13 min total per participant (13.2 ± 0.27 min). This limited amount of data might be insufficient to train accurate TRFs in the auditory domain.

Nevertheless, the modelling results provide a proof of concept, illustrating that both temporal tracking and exponent tracking analyses eventually might capture the same sensory-related neural process. The used spectral exponent tracking approach, however, appears more powerful in the context of the analysed data as it is able to unearth the effect of stimulus exponents on EEG data were TRF approaches are not. Overall, these results speak to a model in which the temporal alignment of sensory cortical postsynaptic activity with sensory input results in EEG signals that are temporally locked to time courses of stimulus features, resulting in EEG spectra that mimic stimulus spectra. Importantly, PSD-based approaches appeared more powerful in the context of our data, rendering the empirical link between both approaches a hypothesis to be formally tested in future work.

Furthermore, spectral exponent increased when data were modelled based on visual stimuli and the visual attention TRF and vice versa for auditory stimuli (betavis = .64, Se = .1, t = 5.6, p <.0001; betaaud = .78, SE = .14, t = 5.7, p < .0001). This spectral steepening likely traces back to the increased amplitude of low-frequency parts in the TRF that mimic early ERP effects and is in stark contrast to our empirical finding of topographically specific reduced spectral exponents through modality-specific attention. Thus, attention-related increases of sensory processing might indeed entail an increase in EEG PSD exponents which takes place in parallel with attention-related decreases of exponents that capture a shift in E/I balance. By analysing comparable distribution of stimulus exponents across both attention conditions and statistically controlling for single trial stimulus exponents, the analysis put forward within the present manuscript is able to isolate the attention-related decrease in spectral exponents.

Reviewer #1 (Recommendations for the authors):

Figure 1, it's really hard to see how the slopes change in the way that the authors state. For propofol, visual inspection suggests that the biggest change is a broadening of the α oscillation, such that its inflection starts at a lower frequency and then because the peak is also 'less pointy,' the purple line simply has to fall at a higher rate to catch up with the gray line by ~30 Hz. For ketamine, at the lowest frequencies (lower than α bump) the slope of the green line simply is steeper than the gray, and then again the biggest difference seems to be that the α bump is abolished with ketamine, and so the gray line is then steeper than the green line for the same reason that purple appears to be steeper than gray in propofol plot. Additionally, there's a lot of jitter with ketamine in the 20-60 Hz range. I realize that visual inspection isn't a rigorous way to analyze these data, but on the other hand it's generally preferable for a figure to clearly illustrate the point that the authors are trying to convey. Perhaps the authors should consider accompanying the 'raw' spectra shown here with the same data decomposed into oscillatory vs. aperiodic components, the way that it is done in the Donoghue et al., (2020) paper?

It is true that anaesthesia-related spectral changes are not limited to the spectral exponent but rather include several oscillatory alterations as well. Importantly, we account for these changes by parameterizing both oscillatory and 1/f-like signal properties of the EEG. While we understand the reasoning behind the reviewer’s comment on the link between the width of the α peak and the steepness of the power decay immediately above the α range, we want to emphasize that spectral parameterization was performed up to 60 Hz. Hence, a steeper decay after a wider α peak is not sufficient to drive a higher exponent estimate.

To improve the visual presentation of these effects, we followed the reviewer’s suggestion and now plot oscillatory as well as aperiodic fits in addition to average spectra. Note that spectra are normalized by dividing by the mean as suggested by R2. Furthermore, updated results are based on data from 5 participants in each group, including two datasets that we had previously missed (Author response image 2) .

Author response image 2
EEG PSD exponents track anaesthesia-induced E:I changes.

(A) Normalized EEG spectra averaged across 5 subjects and 5 central electrodes (inset) displaying a contrast between rest and propofol (left) and ketamine anaesthesia (right). Spectral parameterization yielded aperiodic fits that estimated the spectral exponent (dashed lines) and full fits that included oscillatory spectral peaks (transparent lines). (B) Pairwise scatter plots depicting subject-wise averaged EEG PSD exponents during awake rest, propofol (left) and ketamine (right). Coloured dots represent PSDexponents of 5 s snippets, black horizontal bars single subject means. P-values are based on 1000 random permutations.

On a more general note, regarding the specificity of anaesthesia effects on E/I and related shifts in EEG spectral exponents, we like to point the reviewer to our response to Reviewer 3 below. In brief, we argue that the assumption of differentially altered cortical E/I balance due to propofol or ketamine represents a simplification that is warranted given both previous findings and the goals of the current study. We neither wish to negate the complexity of anaesthesia-related changes in brain activity nor do we want to claim exclusive aperiodic changes due to anaesthesia and E/I. Instead, as nicely summarized by the reviewer above, our goal is to test the assumption of anaesthesia- and E/I-related changes in spectral exponents in non-invasive recordings.

The Discussion section is largely a repetition of what was written in the Intro and/or a restatement of the results with little additional interpretation and contextualization. For example, although it's important to show that α and aperiodic components of the EEG are statistically dissociable, this is only a step toward understanding more fundamental questions such as (a) what are the functions that periodic vs. aperiodic components support? and (b) what underlying factors that give rise to them?

We thank the reviewer for these helpful comments on the Discussion section.

We now discuss the presence of statistically unrelated attentional modulations of both α oscillations and EEG PSD exponents as a potential sign of distinct modes of thalamic firing that shape cortical activity in a demand- and resource-dependent manner. The relevant section is pasted below.

“Despite these differences in the sensitivity of EEG signals, our results provide clear evidence for a modality-specific flattening of EEG spectra through the selective allocation of attentional resources. […] Specifying potential demand- and resource-dependent trade-offs between different modes of attention-related modulations of cortical activity and sensory processing will offer crucial insights into the neural basis of adaptive behaviour.”

We also have re-worked large parts of the discussion aiming to offer our thoughts on potential mechanistic differences and similarities between approaches. Importantly, we now explicitly contrast temporal response function ideas against spectrally-based methods that we put forward in the current manuscript. In brief, we argue that both approaches might very well capture the same alignment of postsynaptic neural activity with stimulus input in different ways. However, while TRF approaches imply linearity and phase locking, the comparison of power spectra and their spectral exponents also incorporates potential non-linear aspects of stimulus response relationships. The relevant section is pasted below.

“Importantly, the temporal alignment of broadband sensory input with human brain activity has been studied in the context of “neural tracking” using multivariate linear models and might be able to explain the link between stimulus and EEG spectral properties we observe (Lalor and Foxe, 2010; Wöstmann et al., 2017). […] Importantly, however, future studies are needed to further test the relationship between temporal neural tracking using TRF approaches and spectral tracking as put forward in the current manuscript.”

Here are some more specific comments about the Discussion.

"Jointly, these results underscore the importance of 1/f brain activity for perception and behaviour." Don't the authors really mean: "underscore the utility of parameters of 1/f brain activity for studying the neural bases of perception and behavior"? At the end of the day, the major take-home of this paper is that the slope of the 1/f spectrum is a valid index of E:I balance, but it's E:I balance, per se, that is 'important for perception and behavior,' not the slope itself.

The suggested refinement makes sense and we have rephrased accordingly. However, we wish to point out that not all results of the current manuscript can be subsumed under the umbrella of “spectral exponents track E/I”. The alignment of 1/f-like aperiodic stimulus input and aperiodic EEG activity potentially does not represent E/I changes per se but rather an SSEP-like process of temporal stimulus tracking in sensory cortices.

"… these results cannot be explained by attention-dependent differences in neural α power (8-12 Hz, Figure 3), commonly interpreted as a marker of top-down guided sensory inhibition." Idling is an important alternative to inhibition that should be acknowledged.

We now acknowledge the idling hypothesis and cite relevant work.

“Despite these differences in the sensitivity of EEG signals, our results provide clear evidence for a modality-specific flattening of EEG spectra through the selective allocation of attentional resources. This attention allocation likely surfaces as subtle changes in E:I balance (Borgers et al., 2005; Harris and Thiele, 2011). Importantly, these results cannot be explained by observed attention-dependent differences in neural α power (8–12 Hz, Figure 3) which have been suggested to capture cortical inhibition or idling states (Cooper et al., 2003; Pfurtscheller et al., 1996). Also note that the employed spectral parameterization approach enabled to us to separate 1/f like signals from oscillatory activity and hence offered distinct estimates of spectral exponent and α power that would otherwise have been conflated (Donoghue et al., 2020).”

"First, it is important to emphasise that the representation of stimulus spectra in the EEG likely does not trace back to an alignment of oscillatory neural activity and oscillatory stimulus features, commonly referred to as "entrainment" in the strict sense; the presented stimuli were stochastic in nature and without clear sinusoidal signals. However, neurally tracking the statistical properties of random noise time-series might emerge via a mechanism similar to the one implied in the generation of steady-state evoked potentials (SSEPs)." Both of these seem like important points that merit more elaboration. That is, the word "entrainment" tends to be used carelessly and so more detailed and explicit argumentation about why this is NOT an instance of entrainment would be valuable. With regard to SSEPs, specifying some details about this 'implied mechanism' would be helpful. More generally, although entrainment and evoked responses are precisely specified processes that can be shown to be true or not, the same is not true for "tracking," which is just a loose concept that can't be tested and falsified. Can the authors either specify what they mean by "tracking" or else replace it with a more rigorously defined process?

Thank you for these perceptive comments. We have entirely re-structured the paragraph in question to focus more on the comparison of temporal response function approaches that link stimulus and EEG signals in the time domain with PSD exponent-based techniques employed in the current manuscript. Nevertheless, we have added a more detailed explanation of SSEPs and made it clearer why simple steady state responses cannot explain the observed effects.

Additionally, we now provide an explanation of entrainment in the strict sense and outline why it is an unlikely explanation for observed result.

“What might constitute the mechanism that, at the level of sensory neural ensembles, gives rise to the observed link between sensory stimuli and the spectral shape of the EEG? […] Hence, the observed neural tracking of AM spectral exponents does not emerge via a neural adaptation to constant amplitude spectra or trial-wise differences in evoked responses.”

Reviewer #2 (Recommendations for the authors):

P. 4, last paragraph: It is somewhat disconcerting to learn in the Results section that the first study uses a publicly available dataset and the second is wholly separate and from data acquired by the authors. This would be be less startling if it were mentioned in the introduction.

We apologize for this unnecessary surprise and now note the use of previously published data in the introduction section.

Lines 159-160: As written, this sentence seems to implies that the new results of this paper aren't actually new but merely a confirmation of an old result. It would easier on the reader to more clearly distinguish the previous results (with very strong connections to E:I balance?) from the new findings (where the connection to E:I balance is less direct).

We have rephrased to make clear that our approach aims at a non-invasive extension and generalization of previous findings.

Figure 1B: Would the authors consider using the same vertical scale in both graphs? The overall numbers between the two sets are close enough in value that having two different scales can be distracting.

We have changed the figure to use the same vertical scale in both panels. Note that spectra are normalized by dividing by the mean as suggested by R2. Furthermore, updated results are based on data from 5 participants in each group, including two datasets that we had previously missed (Author response image 2) .

Figure 1B: The inset graphs are missing axis limits (or scale), and there is no definition of their error bars.

These 45° plots had equal x-axis and y-axis limits and were used with the goal of further illustrating the consistency of anaesthesia effects on EEG PSD exponents. After careful consideration, we decided that they did not help the visibility of effects and have removed them in the updated version of the manuscript.

L. 196 and elsewhere: incorrect formatting of numbers in scientific notation, e.g. 7e-6 instead of 7 x 10-6.

We fixed that mistake.

L. 189 and following: The description of the stimuli, especially the auditory stimuli is confusing. The phase "to detect regular (i.e., sinusoidal) amplitude variations in streams of amplitude modulated white noise", in the auditory literature would be understood as analogous to "to detect tone pips in noise", but that is not what is meant here. Figure 2 indicates rather that the stimulus temporarily changes from non-sinusoidal amplitude modulated white noise to sinusoidal amplitude modulated white noise.

We rephrased this section and now note that “participants had to detect brief time periods during which the amplitude modulation of the presented white noise switched from aperiodic to sinusoidal.”

Figure 2C: Please explain what the circles and lines represent (I presume individual subjects with lines representing identities, but I need to to ask after seeing Figure 3B).

The reviewer is assuming correctly that dots represent single subject data and lines are taken to depict identities and we have added the relevant information in the figure caption.

Figure 3B: Please explain what the circles and lines represent. Do the lines connect the different tasks of the same individuals? The systematic progression of the slopes of the lines seems to indicate that they do not.

Circles represent single subject residual EEG PSD exponents at central (lilac) or occipital (teal) electrodes, averaged across auditory or visual attention trials. Lines identify individuals.

The systematic progression simply stems from the process of residualization which was performed within subjects, effectively zero-centering individuals. Note that this is not an issue, but desired as it allows us to zoom in on the within subject effects of attentional focus on EEG PSD exponents while controlling for stimulus exponents and other covariates. However, we happily provide an analogous visualization of raw EEG PSD in Author response image 3 . Here, the same attention dependent reduction of PSD exponents as well as its topographical dependence can be seen. However, between subject variance obscures the central within-subject effects.

Author response image 3
Raw EEG spectral exponents change with attention.

As in figure 3A, but based on raw EEG spectral exponents instead of model residuals. Individual average exponents are depicted by dots, grand averages by horizontal lines. Coloured lines denote subject identity. Exponents are shown for auditory (left) and visual attention (right) and for a cluster of central (lilac) and occipital electrodes (teal).

Lines 234-235: Getting R2 > 0.84 is a real achievement-it speaks very highly of the importance of the spectral exponent.

Thank you for this friendly comment.

L. 383: the phrase "and hence" is confusing here. Maybe "even though they"?

We have rephrased to:

“Central electrodes capture auditory cortical activity but are positioned far away from their dominant source (Huotilainen et al., 1998; Stropahl et al., 2018). Occipital electrodes, however, are sensitive to visual cortex activity and are directly positioned above it (Hagler et al., 2009)”

L. 419 and Supplemental Figures: There are two supplemental figures labeled as S4 and none as S5. This reference appears to be to the 5th supplemental figure.

Sorry for the confusion. We have corrected this mistake and refer to each supplemental figure by linking it with its parent figure in the updated version of the manuscript.

Lines 445-475: This section appears to be where the possibility of temporal tracking is meant to be addressed, but it does not accomplish this (instead only justifying that steady-state analysis does not apply here, which is true). Note also to be careful with the word "stationary". A "stationary process" is one with a fixed spectrum and random phases, which seems to be a good description of the stimulus envelopes/contrasts used here.

We agree that this section in its original form did not achieve what we were aiming for and have edited it to discuss potential neural mechanisms similar to what is commonly referred to as speech tracking. The relevant section is pasted below.

“Importantly, the temporal alignment of broadband sensory input with human brain activity has been studied in the context of “neural tracking” using multivariate linear models and might be able to explain the link between stimulus and EEG spectral properties we observe (Lalor and Foxe, 2010; Wöstmann et al., 2017). Here, a linear relationship between time-courses of stimulus features and neural responses is assumed to capture their temporal alignment, commonly referred to as “entrainment in the broad sense” (Obleser and Kayser, 2019).

As outlined above, we estimated auditory and visual TRFs to test whether forward modelling of EEG data would result in EEG spectra that mimicked properties of stimulus spectra. However, auditory TRFs were unreliable (see Figure 4 supplement 1). Visual TRFs on the other hand enabled significant stimulus reconstruction and were used within a simplified proof-of-concept model to predict EEG signals that indeed mimicked the spectral properties of stimuli (Figure 4 supplement 4). The non-predictiveness of auditory TRFs potentially traces back to an insufficient signal to noise ratio and limited training data. In general, EEG spectral exponents might also capture the consequences of non-linear interactions between stimulus input and neural response by focusing on their spectral representation across a wide frequency range. Such non-linear links of stimulus and response are by design inaccessible to TRF approaches that rely on the linear relationship of both time series.

Although spectral-based approaches of neural stimulus tracking clearly displayed higher power in context of the analysed dataset, we deem it probable that both approaches eventually capitalize on the same aspect of central neural processing: the temporal alignment of high amplitude/salience stimulus events with high amplitude neural activity. While this does not correspond to entrainment in the narrow sense or SSVEP-like superposition of oscillatory activity or ERPs, 1/f AM spectra might evoke trains of evoked responses with similar spectral exponents. Indeed, a simple proof-of-concept model based on real stimulus data resulted in EEG spectra whose exponents were positively linked with the exponents of stimuli (Figure 4 supplement 5). Hence, time- and spectrally-based approaches of stimulus tracking might indeed capture similar aspects of postsynaptic neural activity that align with sensory input during early processing. Importantly, however, future studies are needed to further test the relationship between temporal neural tracking using TRF approaches and spectral tracking as put forward in the current manuscript.”

Lines 576-577. What does "normalized" mean here? Standard usage is a multiplicative rescaling, not mean-centering. [On the other hand, if the mean-centering was performed on the logarithm (or in dB), then that is equivalent to a multiplicative rescaling of the original waveform.]

The reviewer is assuming correctly that mean-centering was performed on the logarithm. For the updated version of the manuscript, we performed multiplicative re-scaling before taking the logarithm of the spectrum.

L. 616 and following: Regarding the visual modulation, why is the acoustic noise, which had been high-passed at 200 Hz before its modulation, downsampled to 85 Hz (which throws away all the carrier information), instead of just applying the 1/fX modulation directly (downsampled to 85 Hz)? Why the extra complication? Or am I just confused by the multiple uses of the word "noise"?

Visual noise was down-sampled to match the maximal refresh rate of the used monitor. You are of course correct in stating that this down-sampling will get rid of carrier information. Importantly, since the visual stimulus consisted of a white disc that fluctuated in luminance according to the used amplitude modulation spectra, carrier information (i.e., frequency spectra) played a very limited role. In other words, while high pass-filtered white noise was used as the carrier signal in the auditory domain, luminance was used as carrier in the visual domain. We realize that our approach might be a bit idiosyncratic but are confident that it resulted in the desired outcome: amplitude modulation spectra with varying 1/fx in the visual domain.

L. 739 and following: I very much appreciate the careful analysis methods employed here.

Thank you.

Figure S1 caption: This caption would be much clearer if it stated that the graphs and data were identical to that shown in Figure 1 except without normalization. (In its current form it seems almost like an example of an item in a change-blindness study.)

We apologize for the confusion and agree that a focus on the difference to figure 1 makes much more sense. We have changed the figure to only show raw power spectra and changed the caption accordingly.

Figures 1B and S1B. There seems to be a lot of inter-subject variability in the Awake case between the subjects who used Propofol vs Ketamine (which should have nothing to do with the awake case). Is that an artifact of changes in the axis scaling (or normalization)? It shouldn't matter since the important statistics are changes within subject, but it is a little disconcerting.

To explore potential differences in awake rest EEG spectra between anaesthetic conditions, we ran analogous analyses to the ones used to test anaesthetic-depended spectral exponent changes. We observed no significant difference between awake rest spectral exponents of both anaesthetic conditions (pperm = .75). The impression of a difference to a large degree was driven by the difference in y-axis scaling which we have changed to equal axes in the updated version of the manuscript.

Reviewer #3 (Recommendations for the authors):

The authors based their correlation analysis on 24 participants. While the authors do argue that bigger sample size and cross-validation could strengthen the results, the authors could do more with the data they have.

For example, they can employ a leave-one-out linear regression approach, or use k-folds

We thank the reviewer for their comment regarding the potential to strengthen our between subject correlation analyses. However, neither leave-one out nor k-folds cross-validation appear as appropriate analyses choices in this case. One the one hand, leave-one out cross validation has been shown to not provide reliable but noisy and biased results based on samples with less than a couple hundred observations (Varoquaux, 2019). Analogously, foldwise cross validation further reduces the size of training sets by holding out a larger proportion of observations for later testing and hence requires even larger sample sizes to offer reliable results. In the light of these methodical considerations and to account for the small sample size, we chose to apply a two-stage permutation approach which has been developed for the reliable estimation and testing of brain-behaviour relationships in limited sample sizes and is being used widely (McIntosh et al., 1996; Krishnan et al., 2011). The employed partial least squares approach first tests the significance of observed brain-behaviour relationships in latent space by permuting data across participants (here: 1000 permutations). In a second step, a test statistic is estimated using a bootstrap approach (here: 1000 bootstraps). By comparing observed brain-behaviour relationships with mean and standard deviations of bootstrap results, a test-statistic that can be interpreted in line with a z-distribution is generated. In addition to this two-step procedure, we implemented a rank-based version of the entire approach, enabling us to non-parametrically test between-subject relationships of stimulus tracking and performance. Hence, combining a two-stage permutation approach with rank-based correlation, the used methods are well suited for the analyses of the present data.

Nevertheless, we interpret the reported across-subject correlations with great care and refrain from suggesting the prediction of behaviour based on the observed stimulus-related brain dynamics. Furthermore, we acknowledge the importance of generalization and out of sample prediction in the Discussion section and suggest approaches to achieve these goals in future studies.

With regards to the ERP analysis, the authors appear to be using a cluster-permutation approach to assess any differences between the conditions. Here they do have to keep in mind that such a mass-univariate approach is biased towards longer-sustained responses that have a wide scalp distribution, than the rather more focal discrete ERP components. Please see refer to the following discussion on this topic.

https://projects.iq.harvard.edu/files/kuperberglab/files/fieldskuperberg_psychophysiology_2020.pdf

While mass-univariate approaches might not be sensitive to very focal ERP differences, this does not represent an issue in our case for at least three reasons. First, as can be discerned from figure 4 supplement 1, ERPs of different stimulus conditions did not even show slight evidence for focal differences. Second, focal increases of visually-evoked ERPs between auditory and visual attention were detected with the chosen mass-univariate approach (see figure 3 supplement 1), rendering it unlikely that comparably focal stimulus condition differences in the same data might have been missed. Third, and most importantly, the time periods of early evoked responses (0–600 ms post-stimulus onset) were excluded from the estimation of power spectra and hence also the calculation of spectral exponents. Thus, even if the chosen statistical approach missed very focal and faint differences in ERPs, these were unable to affect estimates of EEG spectral exponents and the main results of our study.

Finally, why I am intrigued by the idea of the slope of 1/f as being something rather important, I am still not convinced that it could be a residual of other factors in the EEG, such as changes in slow frequency power, or evoked responses. I think it would be interesting to see how much unique variance the change 1/f can contribute relative to the other measures of the EEG.

We agree with the reviewer on the importance of differentiating the 1/f decay in power and the associated exponent from oscillatory signal parts and related peaks in the spectrum. Importantly, recent work by the second author of the present manuscript (Donoghue et al., 2020), put forward a parameterization approach that achieves exactly this. We have tried to make this clearer by including plots of fitted spectra that differentiate between oscillatory and aperiodic signals in the current manuscript. We wish to refer the reviewer to the original paper for simulation results and an at length discussion of potential links and conflations between oscillatory and aperiodic signal parts. In addition, we want to highlight that 1/f like signals have been argued to play a central role in sensory processing, different forms of cognition, development, and aging (He et al., 2010; Dave et al., 2018; Sheehan et al., 2018; Chini et al., 2021). Importantly, the relevance of 1/f-like parts of brain activity does not hinge on the specific methods employed in the current study but rather persist across imaging techniques from fMRI and M/EEG (He et al., 2010; Dave et al., 2018) towards invasive LFPs (Sheehan et al., 2018) and single unit recordings in non-human animals (Chini et al., 2021).

Regarding the reviewer’s comment on the unique variance accounted for by 1/f exponent changes, we assume that they are referring to variance in behaviour. While not strictly representing overt behaviour as such, selective attention exerted topographically-specific effects on EEG spectral exponents that took place over and above an observed change in α power. Hence, we report unique attention- and behaviourally-linked changes in brain activity that are specific to EEG spectral exponents. Furthermore, analysing inter-individual differences in behavioural performance, we performed a multivariate partial least squares approach to relate the tracking of stimulus exponents in EEG spectral exponents with detection performance. Testing the role of low-frequency power or α power in this analysis requires a significant association between EEG power in the respective frequency bands and stimulus spectral exponents which then can be used to explain inter-individual variance in behaviour. To follow the reviewer’s suggestion (and comments by R1), we calculated the linear relationship of low-frequency and α power with stimulus spectral exponents by reverting mixed models. We compared these models to models of same size that used EEG spectral exponents as a main predictor of interest and was identical otherwise. As can be taken from Figure 4-supplement 6, neither low-frequency nor α power explained significantly more variance in stimulus exponents than EEG spectral exponents for most conditions. Only α power led to increased model fit at one parietal electrode when modelling visual stimulus exponents. Importantly, this electrode did not overlap with the cluster of significant stimulus exponent tracking by EEG exponents and thus likely traces back to a different neural source.

https://doi.org/10.7554/eLife.70068.sa2

Article and author information

Author details

  1. Leonhard Waschke

    1. Max Planck UCL Centre for Computational Psychiatry and Ageing Research, Max Planck Institute for Human Development, Berlin, Germany
    2. Center for Lifespan Psychology, Max Planck Institute for Human Development, Berlin, Germany
    Contribution
    Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Resources, Supervision
    For correspondence
    waschke@mpib-berlin.mpg.de
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-1248-9259
  2. Thomas Donoghue

    Department of Cognitive Science, University of California, San Diego, La Jolla, United States
    Contribution
    Formal analysis, Methodology, Software, Supervision
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-5911-0472
  3. Lorenz Fiedler

    Eriksholm Research Centre, Oticon A/S, Snekkersten, Denmark
    Contribution
    Formal analysis, Software, Supervision
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7892-6917
  4. Sydney Smith

    Neurosciences Graduate Program, University of California, San Diego, La Jolla, United States
    Contribution
    Data curation, Investigation, Writing – original draft
    Competing interests
    No competing interests declared
  5. Douglas D Garrett

    1. Max Planck UCL Centre for Computational Psychiatry and Ageing Research, Max Planck Institute for Human Development, Berlin, Germany
    2. Center for Lifespan Psychology, Max Planck Institute for Human Development, Berlin, Germany
    Contribution
    Formal analysis, Investigation, Visualization, Supervision
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-0629-7672
  6. Bradley Voytek

    1. Department of Cognitive Science, University of California, San Diego, La Jolla, United States
    2. Neurosciences Graduate Program, University of California, San Diego, La Jolla, United States
    3. Halıcıoglu Data Science Institute, University of California, San Diego, La Jolla, United States
    4. Kavli Institute for Brain and Mind, University of California, San Diego, La Jolla, United States
    Contribution
    Conceptualization, Investigation, Methodology, Writing – review and editing, Software, Project administration, Resources, Supervision
    Contributed equally with
    Jonas Obleser
    For correspondence
    bvoytek@ucsd.edu
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-1640-2525
  7. Jonas Obleser

    1. Department of Psychology, University of Lübeck, Lübeck, Germany
    2. Center of Brain, Behavior, and Metabolism, University of Lübeck, Lübeck, Germany
    Contribution
    Conceptualization, Investigation, Methodology, Writing – review and editing, Project administration, Resources, Supervision
    Contributed equally with
    Bradley Voytek
    For correspondence
    jonas.obleser@uni-luebeck.de
    Competing interests
    Reviewing Editor for eLife
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7619-0459

Funding

Deutsche Forschungsgemeinschaft (Emmy Noether Programme)

  • Leonhard Waschke
  • Douglas D Garrett

H2020 European Research Council (ERC-CoG-2014-646696)

  • Jonas Obleser

Max Planck UCL Centre for Computational Psychiatry and Ageing Research

  • Leonhard Waschke
  • Douglas D Garrett

Whitehall Foundation (2017-12-73)

  • Bradley Voytek

National Science Foundation (BCS-1736028)

  • Bradley Voytek

National Institute of General Medical Sciences (R01GM134363-01)

  • Bradley Voytek

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

LW and DDG are supported by an Emmy Noether Programme grant from the German Research Foundation (to DDG), and by the Max Planck UCL Centre for Computational Psychiatry and Ageing Research. LW was supported by a G.-A. Lienert fellowship. BV is supported by the Whitehall Foundation Grant 2017-12-73, the National Science Foundation Grant BCS-1736028 and the National Institute of General Medical Sciences Grant R01GM134363-01. JO is supported by the European Research Council (ERC-CoG-2014–646696).

Ethics

Human subjects: All participants gave written informed consent, reported normal hearing and had normal or corrected to normal vision. All experimental procedures were approved by the institutional review board of the University of California, San Diego, Human Research Protections Program (UCSD IRB Protocol #150834. ).

Senior Editor

  1. Barbara G Shinn-Cunningham, Carnegie Mellon University, United States

Reviewing Editor

  1. Maria Chait, University College London, United Kingdom

Reviewers

  1. Bradley R Postle, University of Wisconsin-Madison, United States
  2. Jonathan Z Simon, University of Maryland, United States

Publication history

  1. Preprint posted: January 14, 2021 (view preprint)
  2. Received: May 5, 2021
  3. Accepted: October 18, 2021
  4. Accepted Manuscript published: October 21, 2021 (version 1)
  5. Version of Record published: November 11, 2021 (version 2)

Copyright

© 2021, Waschke et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 704
    Page views
  • 126
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Neuroscience
    Rawan AlSubaie et al.
    Research Article Updated

    Projections from the basal amygdala (BA) to the ventral hippocampus (vH) are proposed to provide information about the rewarding or threatening nature of learned associations to support appropriate goal-directed and anxiety-like behaviour. Such behaviour occurs via the differential activity of multiple, parallel populations of pyramidal neurons in vH that project to distinct downstream targets, but the nature of BA input and how it connects with these populations is unclear. Using channelrhodopsin-2-assisted circuit mapping in mice, we show that BA input to vH consists of both excitatory and inhibitory projections. Excitatory input specifically targets BA- and nucleus accumbens-projecting vH neurons and avoids prefrontal cortex-projecting vH neurons, while inhibitory input preferentially targets BA-projecting neurons. Through this specific connectivity, BA inhibitory projections gate place-value associations by controlling the activity of nucleus accumbens-projecting vH neurons. Our results define a parallel excitatory and inhibitory projection from BA to vH that can support goal-directed behaviour.

    1. Cell Biology
    2. Neuroscience
    Angela Kim et al.
    Research Article Updated

    Insulin-induced hypoglycemia is a major treatment barrier in type-1 diabetes (T1D). Accordingly, it is important that we understand the mechanisms regulating the circulating levels of glucagon. Varying glucose over the range of concentrations that occur physiologically between the fed and fuel-deprived states (8 to 4 mM) has no significant effect on glucagon secretion in the perfused mouse pancreas or in isolated mouse islets (in vitro), and yet associates with dramatic increases in plasma glucagon. The identity of the systemic factor(s) that elevates circulating glucagon remains unknown. Here, we show that arginine-vasopressin (AVP), secreted from the posterior pituitary, stimulates glucagon secretion. Alpha-cells express high levels of the vasopressin 1b receptor (V1bR) gene (Avpr1b). Activation of AVP neurons in vivo increased circulating copeptin (the C-terminal segment of the AVP precursor peptide) and increased blood glucose; effects blocked by pharmacological antagonism of either the glucagon receptor or V1bR. AVP also mediates the stimulatory effects of hypoglycemia produced by exogenous insulin and 2-deoxy-D-glucose on glucagon secretion. We show that the A1/C1 neurons of the medulla oblongata drive AVP neuron activation in response to insulin-induced hypoglycemia. AVP injection increased cytoplasmic Ca2+ in alpha-cells (implanted into the anterior chamber of the eye) and glucagon release. Hypoglycemia also increases circulating levels of AVP/copeptin in humans and this hormone stimulates glucagon secretion from human islets. In patients with T1D, hypoglycemia failed to increase both copeptin and glucagon. These findings suggest that AVP is a physiological systemic regulator of glucagon secretion and that this mechanism becomes impaired in T1D.