Abstract
Spontaneous fluctuations of neural activity may explain why sensory responses vary across repeated presentation of the same physical stimulus. To test this hypothesis, we recorded electroencephalography in humans during stimulation with identical visual stimuli and analyzed how prestimulus neural oscillations modulate different stages of sensory processing reflected by distinct components of the event-related potential (ERP). We found that strong prestimulus alpha- and beta-band power resulted in a suppression of early ERP components (C1 and N150) and in an amplification of late components (after 0.4 s). Whereas functional inhibition of sensory processing underlies the reduction of early ERP responses, we found that the modulation of non-zero-mean oscillations (baseline shift) accounted for the amplification of late responses. Distinguishing between these two mechanisms is crucial for the understanding of how internal brain states modulate the processing of incoming sensory information.
Acknowledgments
This work was supported by the HSE Basic Research Program and the Russian Academic Excellence Project ‘5-100’ (LI, VVN), and by a grant from the German Research Foundation (DFG) to NAB (BU2400/9-1).
We thank Francesco Di Russo and Volodymyr B Bogdanov for help with designing the stimuli, Johanna Rehder for help with the literature search, and Charles Schroeder, Saskia Haegens, and Omri Raccah for comments on the manuscript.
Introduction
The brain generates complex patterns of neural activity even in the absence of tasks or sensory input. This activity is referred to as “spontaneous”, “endogenous”, or “prestimulus”, as opposed to the activity triggered by and, thus, following experimental events. Numerous studies have shown that such spontaneous neural activity can explain a substantial amount of the trial-by-trial variability in perceptual and cognitive performance (e.g., Haegens et al., 2011; Myers et al., 2014; Iemi et al., 2017) and that abnormalities in spontaneous neural activity can be biomarkers for neuropathologies (e.g., in schizophrenia, Parkison’s disease, and Autism Spectrum Disorder; Uhlhaas and Singer, 2010; McCarthy et al., 2011; Simon and Wallace, 2016) and aging (Voytek et al., 2015). Yet, the mechanisms by which spontaneous neural activity impacts the processing of sensory information in humans remain unknown.
This study aims to clarify how spontaneous fluctuations of prestimulus brain states, reflected by the power of low-frequency oscillations (8–30 Hz), affect the trial-by-trial variability in the amplitude of sensory event-related potentials (ERPs). The mechanisms underlying the effect of prestimulus power on ERP amplitudes are currently unknown, partly because previous studies have been inconsistent regarding the latency and even the directionality of this effect. Specifically, several studies found that prestimulus α-band power suppresses the amplitude of early ERP components (< 0.200 s: Rahn and Başar, 1993; Roberts et al., 2014; Becker et al., 2008; Başar and Stampfer, 1985; Jasiukaitis and Hakerem, 1988), whereas other studies found that prestimulus α-band power enhances the amplitude of late ERP components (> 0.200 s: Dockree et al., 2007; Becker et al., 2008; Roberts et al., 2014; Başar and Stampfer, 1985; Jasiukaitis and Hakerem, 1988; Barry et al., 2000). In this study, we addressed this issue by considering how prestimulus power affects the mechanisms of ERP generation at different latencies: namely, additive and baseline shift mechanisms.
ERP components occurring during the early time window (< 0.200 s) are thought to be generated by an activation of sensory brain areas adding on top of ongoing activity (additive mechanism; Bijma et al., 2003; Shah et al., 2004; Mäkinen et al., 2005; Mazaheri and Jensen, 2006). Invasive studies in non-human primates demonstrated that early ERP components are associated with an increase in the magnitude of multi-unit activity (MUA) in sensory areas (Kraut et al., 1985; Schroeder et al., 1990, 1991; Schroeder, 1998; Shah et al., 2004; Lakatos et al., 2007), presumably as a result of membrane depolarization due to excitatory synaptic activation (Schroeder et al., 1995). Non-invasive studies in humans showed that early ERP components (e.g., C1) are associated with an increase in the hemodynamic fMRI signal in visual areas (Di Russo et al., 2001, 2003), which may reflect an additive sensory response. Low-frequency neural oscillations are thought to set the state of the neural system for information processing (Jensen and Mazaheri, 2010; Mathewson et al., 2011; Spitzer and Haegens, 2017), which in turn may modulate the generation of additive ERP components. In particular, numerous studies have demonstrated that states of strong ongoing α- and β-band oscillations reflect a state of functional inhibition, indexed by a reduction of neuronal excitability (e.g., single-unit activity: Haegens et al., 2011; Watson et al., 2018; MUA: van Kerkoerle et al., 2014; Becker et al., 2015; ongoing γ-band power: Spaak et al., 2012; hemodynamic fMRI signal: Goldman et al., 2002; Becker et al., 2011; Scheeringa et al., 2011; Harvey et al., 2013; Mayhew et al., 2013) and of subjective perception (e.g., conservative perceptual bias: Limbach and Corballis, 2016; Iemi et al., 2017; Craddock et al., 2016; Iemi and Busch, 2018; lower perceptual confidence: Samaha et al., 2017b; lower visibility ratings: Benwell et al., 2017b). Accordingly, prestimulus low-frequency oscillations may exert an inhibitory effect on the additive mechanism of ERP generation: i.e., states of strong prestimulus power may suppress the activation of sensory areas, resulting in an attenuation of the amplitude of additive ERP components during the early time window (Fig 1).
While early ERP components are likely to be generated primarily through an additive mechanism, late ERP components can have a contribution from both additive and baseline shift mechanisms (where “baseline” denotes the mean offset of the signal, rather than prestimulus activity). According to the baseline shift mechanism, the slow ERP component, which becomes visible during the late time window (> 0.200 s), is thought to be generated by a modulation of ongoing oscillatory power, rather than by an additive response (Nikulin et al., 2007, 2010a; Mazaheri and Jensen, 2008, 2010; Dijk et al., 2010). The effect of the baseline shift mechanism on the relationship between prestimulus power and ERP amplitude has never been tested. In fact, it is generally assumed (and not even questioned) that neural oscillations are symmetrical around the zero line of the signal. Accordingly, trial averaging is expected to eliminate non-phase-locked oscillations due to phase cancellation (assuming a random phase distribution over trials), thereby resulting in a signal baseline with a zero mean. It follows that a modulation of zero-mean oscillations by stimuli/tasks leaves the signal baseline unaffected (Fig 1 A).
In contrast to this traditional view, recent studies (Nikulin et al., 2007, 2010a; Mazaheri and Jensen, 2008; Dijk et al., 2010; Schalk, 2015; Cole and Voytek, 2016) proposed that neural oscillations do not vary symmetrically around the zero line of the signal, but rather around a non-zero offset/mean (Fig 1 B).Accordingly, trial averaging does not eliminate non-phase-locked oscillations with a non-zero mean. As a consequence, any amplitude modulation of oscillations with a non-zero mean is expected to change the signal baseline (baseline shift), and thus will affect the ERP amplitude. Specifically, during the event-related desynchronization (ERD) of low-frequency oscillations, the power suppression is expected to cause a slow shift of the signal baseline toward the zero line. Subtracting the prestimulus non-zero baseline from the post-stimulus signal creates a slow shift, which mirrors the spatio-temporal profile of the ERD. In particular, an ERD of oscillations with a negative non-zero-mean is expected to generate an upward slow shift of the ERP (and viceversa). The idea that the ERD contributes to the generation of the slow ERP component, implies that the larger the ERD, the stronger the slow ERP component. Accordingly, we predicted that states of strong prestimulus power would yield a strong ERD (Min et al., 2007; Becker et al., 2008; Tenke et al., 2015; Benwell et al., 2017a), resulting in an enhancement of the slow ERP component during the late time window.
To summarize, states of strong prestimulus power are expected: i) to suppress the amplitude of additive ERP components during the early time window (via functional inhibition); and ii) to amplify the late ERP component, generated by a event-related modulation of non-zero-mean oscillations (via baseline shift). To test these predictions, we recorded electroencephalography (EEG) in human participants during rest and during stimulation with identical high-contrast checkerboard stimuli and analyzed the relationship between ERPs, ongoing and event-related oscillations. To anticipate, the effects of prestimulus power on early and late ERP components are consistent with the functional inhibition and baseline shift mechanisms, respectively. Taken together, these results largely resolve apparent inconsistencies in previous literature and specify two distinct mechanisms by which prestimulus neural oscillations modulate visual ERP components.
Results
Event-related potentials
The experiment included stimulation trials with high-contrast checkerboard stimuli presented in the lower (LVF; Fig 2 A, column 1) or upper (UVF; Fig 2 A, column 2) visual field with equal probability, and fixation-only trials without any checkerboard stimulus (F; Fig 2 A, column 3). All trials included a change of the central fixation mark at the time of stimulus presentation (see Methods for details). For each participant we quantified the ERP at the electrode with peak activity between 0.055 and 0.090 s after stimulus onset, reflecting the C1 component which indicates initial afferent activity in primary visual cortex (Di Russo et al., 2001).
In stimulation trials, the C1 component peaked on average at 0.079 s (SEM = 0.001) and at 0.078 s (SEM = 0.001) for LVF and UVF stimuli respectively, with a maximum at occipito-parietal electrodes (Fig 2 B, column 1 and 2). The comparison of C1 amplitudes at individual peak electrodes on LVF (M = 10.157 μV; SEM = 0.918) and UVF (M= −10.567 μV; SEM = 1.058) trials revealed the expected polarity reversal, confirming that this component originates from initial afferent activity in early visual areas (Di Russo et al., 2001, 2003; Bao et al., 2010). Following the C1, we observed a N150 component peaking between 0.100 and 0.200 s relative to stimulus onset, with an occipital topography and a consistently negative polarity for both LVF and UVF stimuli. The N150 was followed by a late deflection in the time range between 0.200 and 0.600 s relative to stimulus onset, with a parietal topography and consistent positive polarity for both LVF and UVF stimuli.
As expected, in F trials no C1 component was detected in the ERP at individual C1-peak electrodes for LVF and UVF stimuli (mean amplitude in the C1 time window: −0.044 μV, SEM = 0.196). F trials showed a late positive deflection with similar timing and topography as in stimulation trials (Fig 2 B, column 3).
Event-related oscillations
For each participant we estimated the event-related synchronization (ERS) and desynchronization (ERD) at frequencies between 5 and 30 Hz and at each electrode and time point of the poststimulus window (0–0.900 s). For the group-level statistical analysis, we used cluster permutation test to determine at which time, frequency and electrode the ERS/ERD was significantly different from 0 across participants. For LVF stimulation trials, the statistical test revealed a negative cluster (p < 0.001) indicating ERD, spanning time points from 0 to 0.900 s relative to stimulus onset, frequencies between 6 and 30 Hz, and all 64 electrodes (Fig 2 C, column 1). The minimum t statistic was found at 20 Hz, 0.234 s, and at electrode P7. A positive cluster (p = 0.041) indicating ERS spanned time points from 0 to 0.900 s relative to stimulus onset, frequencies between 5 and 8 Hz, and all 64 electrodes. The maximum t statistic was found at 5 Hz, 0.097 s, and at electrode P7.
For UVF stimulation trials, the statistical test revealed a negative cluster (p < 0.001) indicating ERD, spanning time points from 0 to 0.900 s relative to stimulus onset, frequencies between 6 and 30 Hz, and all 64 electrodes (Fig 2 C, column 2). The minimum t statistic was found at 20 Hz, 0.234 s, and at electrode P7. The statistical test also found two positive clusters indicating ERS. The first positive cluster (p = 0.032) spanned time points from 0 to 0.900 s relative to stimulus onset, frequencies between 5 and 8 Hz, and all 64 electrodes. Within this cluster, the maximum t statistic was found at 5 Hz, 0.152 s, and at electrode P2. The second positive cluster (p = 0.034) spanned time points from 0.49 to 0.900 s relative to stimulus onset, frequencies between 13 and 30 Hz, and 62 electrodes. Within this cluster, the maximum t statistic was found at 17 Hz, 0.648 s, and at electrode FT7.
For F trials, the cluster permutation test revealed one negative cluster (p < 0.001) indicating ERD, spanning time points from 0 to 0.900 s relative to stimulus onset, frequencies between 5 and 30 Hz, and all 64 electrodes (Fig 2 C, column 3). Note that “stimulus onset” in F trials refers to the time when a C1-eliciting stimulus would be presented in stimulation trials (see Methods for details). In F trials, the minimum t statistic was found at 20 Hz, 0.214 s, and at electrode PO8.
Evidence for functional inhibition
The functional inhibition mechanism predicts that states of strong prestimulus power reflect neural inhibition, resulting in reduced amplitudes specifically of early ERP components generated in early sensory areas by the additive mechanism.
For LVF stimulation trials, a statistical test comparing ERP amplitudes during the early time window (< 0.200 s) on trials with strong versus weak prestimulus power found a significant negative cluster (p = 0.015). This indicates that the ERP amplitude in a time range containing the C1 (0.043 to 0.121 s) was weaker (i.e. less positive) on trials with strong prestimulus power between 8 and 28 Hz, and at all 64 electrodes (Fig 3 A, column 1). The minimum t statistic was found at 10 Hz, 0.078 s, and at electrode P4. Furthermore, the statistical test found a significant positive cluster (p < 0.001), indicating that the ERP amplitude in a time range containing the N150 (0.090 to 0.200 s) was weaker (i.e., less negative) on trials with strong prestimulus power between 5 and 24 Hz, and at all 64 electrodes (Fig 3 A, column 1). The maximum t statistic was found at 9 Hz, 0.098 ms, and at electrode Pz.
For UVF stimulation trials, the statistical test during the early time window found two significant positive clusters. The first cluster (p < 0.001) indicated that the ERP amplitude in a time range containing the C1 (0.02 to 0.113 s) was weaker (i.e. less negative) on trials with strong prestimulus power between 5 and 22 Hz, and at all 64 electrodes (Fig 3 A, column 2). The maximum t statistic was found at 13 Hz, 0.082 ms, and at electrode PO4. The second cluster indicated that the ERP amplitude in a time range containing the N150 (0.125 to 0.200 s) was weaker (i.e., less negative) on trials with strong prestimulus power between 5 and 25 Hz, and at all 64 electrodes. The maximum t statistic was found at 10 Hz, 0.177 ms, and at electrode F2.
For F trials, the statistical test during the early time window found no significant clusters (< 0.200 s, Fig 3 A, column 3).
Taken together, the results on early ERP components show that ERP amplitude in stimulation trials is attenuated during states of strong prestimulus power, regardless of the polarity of the components. This provides evidence for the functional inhibition mechanism underlying the modulatory effect of prestimulus power on the early ERP components.
Evidence for the baseline shift mechanism
The baseline shift mechanism predicts that states of strong prestimulus oscillations with a non-zero mean are followed by strong post-stimulus power suppression (ERD), resulting in greater ERP amplitudes specifically during the late time window. To demonstrate that the late ERP component was generated by a baseline shift, it is necessary to establish that: i) the ongoing oscillations have a non-zero mean; ii) the non-zero mean and the late ERP component have opposite polarity; and that iii) the ERD magnitude is associated with the amplitude of the late ERP component.
To demonstrate the non-zero mean property of ongoing oscillations, we computed the Baseline Shift Index (BSI: Nikulin et al., 2007, 2010a) and the Amplitude Fluctuation Asymmetry Index (AFAI: Mazaheri and Jensen, 2008). For each participant we estimated BSI and AFAI from resting-state oscillations for each electrode and frequency between 5 and 30 Hz, and then tested whether these indexes were significantly different from 0 across participants using a cluster permutation test. For AFAI, we found a significant negative cluster (p-value < 0.001) between 5 and 30 Hz with a parietal, occipital and central topography (Fig 4 A/B), indicating a stronger modulation of the troughs relative to the peaks, resulting in a negative mean. For BSI, we found a significant negative cluster (p-value < 0.001) between 5 and 21 Hz with a parietal, occipital and central topography (Fig 4 A/B), similar to AFAI. BSI < 0 indicates that strong oscillatory power corresponds to a more negative value of the low-pass filtered signal, as expected in the presence of oscillations with a negative mean. Taken together, the results on BSI and AFAI provide evidence for a non-zero (negative) mean of resting-state low-frequency oscillations. It is important to note that the late ERP component had a positive polarity in all trial types (Fig 1), which is expected as a result of ERD of oscillations with a negative mean (Nikulin et al., 2007, 2010a; Mazaheri and Jensen, 2008).
Next, we analyzed the relationship between the ERD magnitude and the ERP amplitude during the late time window (> 0.200 s). We compared the amplitude of the late ERP between groups of trials of weak and strong ERD estimated at each frequency and electrode. For the group-level statistical analysis, we used cluster permutation test to determine significant ERP differences across ERP time points, and ERD electrodes and frequencies. The statistical test in LVF stimulation trials revealed one significant positive cluster (p < 0.001), indicating that the late ERP (0.200–0.900 s) was greater during states of stronger ERD at frequencies between 5 and 30 Hz, and at all 64 electrodes (Fig 5 A/B, column 1). The maximum t statistic was found at 8 Hz, 0.266 s, and at electrode POz. The statistical test in UVF stimulation trials revealed two significant positive clusters, indicated that the late ERP (cluster 1: 0.336–0.900 s; cluster 2: 0.200 to 0.328) was greater during states of strong ERD at frequencies between 5 and 30 Hz, and at all 64 electrodes (Fig 5 A/B, column 2). The maximum t statistic was found at 19 Hz, 0.258 s, and at T8 electrode. The statistical test in F trials revealed one significant positive cluster (p < 0.001), indicating that the late ERP (0.488–0.900 s) was greater during states of strong ERD at frequencies between 5 and 30 Hz, and at all 64 electrodes (Fig 5 A/B, column 3). The maximum t statistic value was found at 13 Hz, 0.637 s, and at electrode PO8.
Taken together, these results indicate that states of stronger ERD were associated with a more positive deflection of the late ERP component, consistent with the baseline shift mechanism.
After demonstrating that the ERD magnitude correlates with the late ERP amplitude, we tested whether the ERD magnitude was, in turn, correlated with prestimulus power. To this end, we compared the ERD magnitude (at the subject-specific C1 electrode) between groups of trials of weak and strong prestimulus power estimated for each frequency and electrode. For the group-level statistical analysis, we used a cluster permutation test to determine significant differences across ERD time points, prestimulus-power frequencies, and electrodes.
The statistical test revealed a large significant cluster in each trial type (p < 0.001), spanning frequencies between 5 and 30 Hz, time points between 0 and 0.900 s, and all 64 electrodes (Fig 6: LVF trials in A, column 1; UVF trials in A, column 2; and F trials in A, column 3), indicating that at low frequencies strong prestimulus oscillations were associated with strong ERD.
After demonstrating that the late ERP amplitude correlates with the ERD magnitude, and that the ERD magnitude in turn correlates with prestimulus power, we tested whether prestimulus power was directly correlated with the amplitude of the late ERP component. To this end, we compared the late ERP amplitude between groups of trials with weak and strong prestimulus power estimated for each frequency and electrode. For the group-level statistical analysis, we used cluster permutation test to determine significant differences across ERP time points, prestimulus-power frequencies, and electrodes.
The statistical test during the late time window revealed a significant, sustained, and positive cluster in each trial type, indicating that the late ERP component was amplified during states of strong prestimulus power.
In LVF stimulation trials, the significant positive cluster (p < 0.001) spanned time points from 0.402 to 0.900 s, frequencies between 5 and 25 Hz, and all 64 electrodes (Fig 3 A/B, column 1). The maximum t statistic was found at 11 Hz, 0.676 s, and at electrode POz.
In UVF stimulation trials, the significant positive cluster (p = 0.004) spanned time points from 0.445 to 0.900 s relative to stimulus onset, frequencies between 5 and 15 Hz, and all 64 electrodes (Fig 3 A/B, column 2). The maximum t statistic was found at 5 Hz, 0.648 s, and at electrode CP1.
In F trials, the significant positive cluster (p < 0.001) spanned time points from 0.484 to 0.900 s relative to stimulus onset, frequencies between 5 and 23 Hz, and all 64 electrodes (Fig 3 A/B, column 3). The maximum t statistic was found at 7 Hz, 0.781 s, and at electrode POz.
Taken together, these results show that: i) the late ERP component is likely generated by a baseline shift during the ERD of non-zero mean oscillations; ii) states of strong prestimulus power are followed by strong ERD, which manifests as an enhancement of the late ERP component.
Evidence against a confound by sleepiness
To ensure that the relationship between prestimulus power and ERP amplitude was not simply an epiphenomenon of time-varying variables such as sleepiness, we analyzed the scores of a subjective sleepiness questionnaire that participants filled in at the end of every block (Karolinska Sleepiness Scale, KSS: Kaida et al., 2006a, b). First, at the single-subject level, we computed the correlation between prestimulus oscillatory power and KSS rating. At the group-level, we tested whether these correlations were significantly different from 0. We found significant positive clusters for frequencies below 18 Hz and with a widespread topography in each trial type (Fig S1 A; for details see SI). This result indicates that the stronger the prestimulus power, the higher the subjective sleepiness. Within each participant we removed the contribution of sleepiness to the trial-by-trial estimates of oscillatory power and repeated the power-ERP analysis with these corrected power estimates. The results of this re-analysis (Fig S1 B) were virtually identical to the ones obtained with raw power estimates (Fig 3 B), suggesting that the effects we observed were not confounded by sleepiness.
Discussion
Numerous studies suggested that spontaneous fluctuations of prestimulus brain states can account for the trial-by-trial variability in sensory responses. Specifically, the power of low-frequency neural oscillations (8–30 Hz) was found to be associated with ERP amplitude in visual, auditory, and somatosensory modality (e.g. Becker et al., 2008; Jones et al., 2009; De Blasio and Barry, 2013; Roberts et al., 2014). However, the results have been mixed: several studies found decreased ERP amplitude during states of strong prestimulus power (Rahn and Başar, 1993; Roberts et al., 2014; Becker et al., 2008; Başar and Stampfer, 1985; Jasiukaitis and Hakerem, 1988); while others studies found increased ERP amplitude during states of strong prestimulus power (Dockree et al., 2007; Becker et al., 2008; Roberts et al., 2014; Başar and Stampfer, 1985; Jasiukaitis and Hakerem, 1988; Barry et al., 2000). Therefore, the precise mechanism by which prestimulus oscillations modulate sensory responses constitutes a continuing subject of debate in neuroscience. We addressed this issue by considering different mechanisms of ERP generation and how they may depend on prestimulus oscillations: namely, the additive and baseline shift mechanisms. First, early ERP components are thought to reflect neural activation in sensory areas adding to prestimulus activity. Accordingly, we predicted that early ERP components would be attenuated during states of strong prestimulus low-frequency power, as suggested by physiological inhibition accounts (Haegens et al., 2011). Second, the late ERP component is likely generated by a post-stimulus modulation of ongoing oscillations (ERD) via baseline shift (Nikulin et al., 2007). Accordingly, we predicted that the late ERP component would be enhanced during states of strong prestimulus power, which are also associated with strong ERD. These predictions were confirmed by the data.
Functional inhibition mechanism
The results on the early ERP components in this study (< 0.200 s: C1/N150) confirm and extend findings from past literature in the visual and auditory modalities. Specifically, previous studies in the visual modality found a negative relationship between prestimulus α-band power and the amplitude of the visual N1P2 (i.e. amplitude difference between N1 and P2 components Rahn and Başar, 1993), N1 (Roberts et al., 2014) and N175 components (Becker et al., 2008). A similar pattern of results was found for the N100 in the auditory modality (Başar and Stampfer, 1985; Jasiukaitis and Hakerem, 1988). It is important to note that previous results (e.g. Başar and Stampfer, 1985; Becker et al., 2008) that have been used to support the functional inhibition account could actually have been caused by a baseline shift. In the current study, we leverage the fact that the C1 has a well-known polarity reversal as a function of the visual field of the stimulus. By showing that the absolute amplitude of the C1 component is diminished by stronger prestimulus power, regardless of polarity, we can rule out a baseline shift which would affect both polarities in the same direction (e.g, a net increase or decrease of voltage). This provides the first conclusive evidence for the functional inhibition effect of prestimulus oscillations on the early ERP amplitude.
Unlike in the visual and auditory modality, the relationship between prestimulus power and early ERP components in the somatosensory modality (e.g. N1) may be non-linear (inverted U-shaped: Zhang and Ding, 2009; Ai and Ro, 2014; Forschack et al., 2017) or vary across early components (i.e. negative for M50 and positive for M70, P35 and P60: Jones et al., 2009; Nikouline et al., 2000). Similarly, the relationship between prestimulus power and somatosensory perceptual performance has been found to have an inverted U-shape (Linkenkaer-Hansen et al., 2004), or to be linear (Haegens et al., 2011; Craddock et al., 2016). Taken together, these findings suggest that in the somatosensory domain distinct functional mechanisms may map onto low-frequency oscillations.
Importantly, several studies report a positive relationship between prestimulus α-band power and the amplitude of the visual N100 (Jansen and Brandt, 1991; Brandt, 1997), N1P2 (Brandt et al., 1991; Brandt and Jansen, 1991; Barry et al., 2000) and P200 (Jansen and Brandt, 1991). Thus, these studies appear inconsistent with the current results and other studies in the visual and auditory modality. However, a direct comparison is difficult for several reasons. First, some of these studies delivered visual stimuli to participants with eyes closed (Brandt and Jansen, 1991; Brandt et al., 1991; Brandt, 1997; Jansen and Brandt, 1991). Instead, the majority of previous studies (including ours) delivered visual stimuli to participants with eyes open. It is known that oscillatory power in low frequencies has different spectral (Barry et al., 2007) and functional (Kaida et al., 2006a) properties depending on whether subjects’ eyes are open or closed: thus, these inconsistencies may be due to the eyes-open/closed difference. Second, unlike our study, which analyzed a broad frequency band, 64 electrodes, and an extensive post-stimulus time window (0–0.900 s), most previous studies only analyzed a narrow frequency band, few electrodes and a single time point. Therefore, it is possible that the inconsistent effects in previous studies were due to this selective (and under-sampled) analysis of EEG data. Third, some previous studies lack of: i) sufficient description of the EEG analysis (e.g., Brandt, 1997), ii) adequate statistical power (due to low number of participants or trials: e.g., Brandt et al., 1991; Brandt and Jansen, 1991; Brandt, 1997), and iii) quantitative statistical testing (qualitative reports: e.g., Brandt, 1997). Accordingly, these reasons make it difficult to compare these studies to the current one.
The present results have implications for the role of low-frequency oscillations in perceptual decision-making and in the top-down control over sensory processing (e.g., by spatial attention). In fact, numerous studies have found that weak prestimulus α-band power increases observers’ hit rates for near-threshold stimuli (Ergenoglu et al., 2004; van Dijk et al., 2008; Chaumon and Busch, 2014). More recently, studies have demonstrated that this effect is not, in fact, due to more accurate perception, but due to a more liberal detection bias (Limbach and Corballis, 2016; Iemi et al., 2017; Craddock et al., 2016; Iemi and Busch, 2018) and a concomitant increase in confidence (Samaha et al., 2017b) and subjective visibility (Benwell et al., 2017b). Unfortunately, conventional signal detection theory cannot be used to distinguish between alternative kinds of bias (Morgan et al., 2013; Witt et al., 2015). Specifically, a change in bias could be due to a change in the observer’s deliberate decision strategy without any change in sensory processing (decision bias). Alternatively, a change in bias could be due to a change in the subjective appearance of stimuli (perceptual bias): liberal perceptual bias during states of weak prestimulus power could result from increased neural excitability amplifying both neural responses to sensory stimuli (thereby increasing hit rates) and responses to noise (thereby increasing false alarm rates). Interestingly, the present finding that the C1 is amplified during states of weak prestimulus power, indicates that even the earliest visual evoked responses are modulated by prestimulus oscillations. Even though we could not study an equivalent amplification of responses to noise using the present paradigm, this finding supports a perceptual bias mechanism more than a decision bias mechanism.
Furthermore, many experiments have noted a relationship between the topography of α-band power and the focus of covert spatial attention (e.g., Samaha et al., 2016). However considerable debate exists as to whether this preparatory α-band modulation (and hence spatial attention) is capable of modulating feed-forward visual input (e.g., the C1 component). Our results show a clear impact of spontaneous α and, γ-band power on C1 amplitudes, supporting the idea that attention-related low-frequency modulation can affect the earliest stages of sensory processing. However, it is possible that attention-related and spontaneous oscillations have different effects on the amplitude of the C1 component. This question is a candidate for future investigation, ideally, by using stimuli such as those employed here, which generated robust C1 responses.
Baseline shift mechanism
In this study we demonstrated that the late component of the visual ERP was generated by a modulation of non-zero mean oscillations via baseline shift. There are four requirements to demonstrate the baseline shift mechanism. First, the ongoing oscillations must have a non-zero mean. To this end, we estimated the non-zero-mean property of resting-state oscillations using AFAI and BSI. This analysis revealed that α- and γ-band oscillations were characterized by a negative non-zero-mean. The frequencies and electrodes of the significant cluster for AFAI were more extended relative to the cluster for BSI. This could be due to the fact that, unlike BSI, AFAI is biased by harmonics and thus it reflects both non-zero mean oscillations and the “comb-shape” of oscillations, which may yield amplitude asymmetries even when the signal has a zero mean (Nikulin et al., 2010a, b). Thus, AFAI is expected to be susceptible to more asymmetry-related features with larger spatial and spectral distribution compared to BSI.
Second, sensory stimuli must modulate the amplitude of ongoing oscillation. To test this requirement, we estimated the power modulation in the post-stimulus window relative to a prestimulus baseline (i.e., event-related oscillations: ERD/ERS). We observed a strong ERD in frequencies between 6 and 30 Hz in all three trial types. In F trials there were no robust early evoked components due to the lack of strong visual input, yet we observed an ERD following the same spatio-temporal dynamics as in stimulation trials (though of a lesser magnitude). In addition to the ERD, we also observed a strong ERS below 8 Hz in stimulation trials, but not in F trials, possibly reflecting a leakage from the robust evoked components measured during the early time window.
Third, the non-zero mean and the late ERP must have opposite polarity. Consistent with this requirement, our results showed that oscillations with a negative non-zero mean were associated with a late ERP component of positive polarity. Fourth, ERD magnitude must correlate with the amplitude of the late ERP component. Our results indicated that strong ERD of non-zero mean oscillations was associated with enhanced ERP amplitude during the late time window. Importantly, the late ERP component was characterized by a topography and time-course similar to the ones of the ERD, consistent with Mazaheri and Jensen (2008). Taken together, these findings confirm the four requirements necessary to demonstrate the baseline shift mechanism for the generation of the late ERP component.
A prediction of the baseline shift mechanism is that stronger ERD occurs during states of stronger prestimulus power, which generates a greater baseline shift. In the case of negative non-zero-mean oscillations, this process results in an enhancement of the late ERP component with positive polarity. To test this prediction, we analyzed how prestimulus power affects the ERD magnitude, and in turn the amplitude of the late ERP component. We found a positive relationship between prestimulus power and ERD magnitude, consistent with previous studies (Min et al., 2007; Becker et al., 2008; Tenke et al., 2015; Benwell et al., 2017a). Furthermore, we found a positive relationship between prestimulus power and the amplitude of the late ERP component. These results confirm and extend previous findings in visual and auditory modalities. Specifically, in the visual modality prestimulus α-band power was found to be positively correlated with the ERP amplitude in a late time window starting from 0.200 s relative to stimulus onset (0.550–0.800 s: Dockree et al., 2007; 0.220-0.310 s: Becker et al., 2008; 0.400 s: Roberts et al., 2014). A similar pattern of results was found on late ERP components in the auditory modality (0.250–0.800 s Jasiukaitis and Hakerem, 1988; 0.400 s Başar and Stampfer, 1985; 0.200–0.500 s Barry et al., 2000). Previous studies (e.g. Barry et al., 2000) were unable to explain the positive relationship between α-power and ERP amplitude, which appeared inconsistent with the functional inhibition account (Haegens et al., 2011). This study thus resolves this apparent inconsistency in previous literature, by demonstrating that this positive relationship can be accounted for by the baseline shift mechanism, rather than functional inhibition.
It may seem surprising that the effects of prestimulus power on the late ERP occurred after the peak of the classically-defined slow component at approximately 0.300 s relative to stimulus onset (Nikulin et al., 2007; Mazaheri and Jensen, 2008). While early ERP components are likely generated primarily through the additive mechanism (because ERD is negligible in this time window), late ERP components can have a contribution from both additive and baseline shift mechanisms. Functional inhibition of additive components in the initial part of the late time window might have canceled the amplification effect due to the baseline shift. This cancellation might explain the lack of a significant effect at the peak of the late component. In contrast, the ERP during the later time window (> 0.400 s) is more likely to show primarily baseline-shift-generated components and thus is more susceptible to the amplification effect of prestimulus power.
We conclude that the positive modulation of the late ERP component is directly produced by the modulation of ERD magnitude as a function of prestimulus power. This provides the first evidence that the effect of prestimulus oscillations on the late ERP component is due to the mechanism of baseline shift.
α- and β-band oscillations: a similar functional role?
The results of this study demonstrate a modulatory role of low-frequency oscillations on ERP amplitude. Both effects of prestimulus oscillations on early and late ERP components were characterized by a broad frequency range spanning the α- and β-band. Likewise, the ERD and the non-zero-mean property of oscillations were found for both the α-band and β-band. Specifically, α-band ERD was sustained in time while β-band ERD was more transient, consistent with previous studies (e.g., Salenius et al., 1997). This suggests that β-band ERD may also reflect the generation of late ERP component. One possible explanation for this multi-band effect can be the non-sinusoidal nature of neural oscillations (e.g., “comb-shape”: Cole and Voytek, 2016), where α- and β-band power covary. In this case the event-related power modulation would similarly affect α- and β-band activity. Because of such comodulation, baseline-shifts associated with α-band oscillations would also appear for β-band oscillations, resulting in similar BSI and AFAI for both frequency bands (Nikulin et al., 2010a).
The β-band effect may seem surprising since the majority of past literature focused solely on the α-band due to its high signal-to-noise ratio compared to other frequencies. However, the broad frequency range of the effects reported in this study is in line with studies reporting a temporal and spatial co-modulation of α- and β-band oscillations (Bastos et al., 2015; Lakatos et al., 2016; Michalareas et al., 2016). It is also consistent with recent studies reporting a similar relationship between α- and β-band prestimulus power, perceptual reports (Benwell et al., 2017b; Iemi et al., 2017; Samaha et al., 2017b, a; Iemi and Busch, 2018) and firing rate (Watson et al., 2018). Accordingly, it has been proposed that β-band oscillations exert an inhibitory function, similar to α-band oscillations (Spitzer and Haegens, 2017; Shin et al., 2017; Kilavik et al., 2013).
Conclusion
This study demonstrates that spontaneous fluctuations of oscillatory brain activity modulate the amplitude of visual ERP via two distinct mechanisms: i) functional inhibition of the early additive ERP components and ii) baseline shift affecting the late ERP component. Therefore, these findings show that neural oscillations have concurrent opposing effects on ERP generation. Distinguishing between these effects is crucial for the understanding of how neural oscillations control the processing of incoming sensory information in the human brain.
Materials and Methods
Participants
Previous studies on the relationship between neural oscillations and ERPs have typically reported samples of 7–19 participants (e.g. Jasiukaitis and Hakerem, 1988; Brandt and Jansen, 1991; Rahn and Başar, 1993; Nikulin et al., 2007; Mazaheri and Jensen, 2008; Becker et al., 2008; Dijk et al., 2010; Rajagovindan and Ding, 2011). To ensure a robust estimate of our neurophysiological effect and account for potential missing data (due to e.g., artifacts), we recruited a larger sample of 27 participants (mean age: 26.33, SEM = 0.616; 14 females; 3 left-handed). All participants had normal or corrected-to-normal vision and no history of neurological disorders. Prior to the experiment, written informed consent was obtained from all participants. All experimental procedures were approved by the ethics committee of the German Psychological Society. Two participants were excluded before EEG preprocessing because of excessive artifacts. One participant was excluded after preprocessing because no C1 component could be detected, unlike the rest of the sample. A total of 24 participants were included in the analysis.
Stimuli and Experimental Design
The experiment was written in MATLAB (RRID:SCR_001622) using the Psychophysics toolbox 3 (RRID:SCR_002881; Brainard, 1997; Pelli, 1997). The experiment included a resting-state session and a stimulation session, lasting approximately 1.5 h including self-paced breaks.
The resting-state session was divided in two recording blocks, each of which lasted 330 s, separated by a short self-paced break. In this session participants were required to keep their eyes open and fixated on a mark located at the center of the screen, to avoid movements and not to think of anything in particular.
In the stimulation session, participants were presented with visual stimuli specifically designed to elicit a robust C1 component of the visual ERP. The C1 is described as the earliest component of the visual ERP with a peak latency between 0.055 and 0.09 s and an occipital topography. The C1 component is regarded as an index of initial afferent activity in primary visual cortex, because of its early latency and polarity reversal with reference to V1 anatomy (Di Russo et al., 2001, 2003).
The stimuli consisted of full-contrast bilateral black-and-white checkerboard wedges. Each wedge corresponded to a radial segment of an annular checkerboard (spatial frequency = 5 cycles per degree) with inner and outer circle of 3 and 10° of eccentricity relative to a central fixation point, respectively. Each wedge covers 3.125% of the area of the annular checkerboard and spans 11.25° of visual angle (Vanegas et al., 2013).
In each stimulation trial, a pair of wedges was presented bilaterally either in the UVF or LVF with equal probability. UVF and LVF stimulus positions were located at polar angles of 25° above and 45° below the horizontal meridian, respectively. These asymmetrical positions for UVF and LVF stimuli ensure a stimulation of primarily lower and upper banks of the calcarine fissure, respectively (Aine et al., 1996; Clark et al., 1995; Di Russo et al., 2001, 2003), resulting in a polarity reversal of scalp potentials. A positive C1 component is obtained by LVF stimulation, while a negative C1 component is obtained by UVF stimulation (Di Russo et al., 2001, 2003).
The stimuli were presented for a duration of 0.100 s (Fu et al., 2010; Ding et al., 2014; Kelly et al., 2008) at full contrast (Hansen et al., 2016; Vanegas et al., 2013) on a gray background that was isoluminant relative to the stimuli’s mean luminance. The stimuli were presented at a viewing distance of 52 cm, on a cathode ray tube monitor operated at 100 Hz, situated in a dark, radio-frequency-interference (RFI) shielded room. Throughout the experiment, fixation distance and head alignment were held constant by a chin rest.
For each participant the stimulation session included 810 trials, divided into 9 recording blocks. In each block, 60 trials contained stimuli in either LVF or UVF with equal probability (stimulation trials), while 30 trials were stimulus-absent (fixation-only trials, F). Trial type and stimulation field were randomized across trials within each block. To ensure that the participants maintained the gaze to the center, we included a discrimination task at the central fixation mark, similar to previous studies (Di Russo et al., 2001; Chen et al., 2016). In stimulation trials the central fixation mark turned into either one of two equally probable targets (> or <) during stimulus presentation for a duration of 0.100 s. In F trials, the change of the central fixation mark occurred during a 0.100 s window between 1.8 and 2.4 s relative to trial onset. A proportion of correct discrimination responses of approximately 100% was expected if the gaze was maintained on the central fixation mark. This task also ensured that the participants remained alert throughout the experiment. Trials with incorrect discrimination performance were excluded from EEG analysis. Mean accuracy in the fixation task was 94.85% (SEM = 0.0109) and did not significantly differ between the trials types (p > 0.05), indicating that participants were able to maintain central fixation. Incorrect trials were discarded from further analysis (mean = 41.08; SEM = 8.6199). On average we analyzed 761.36 (SEM = 9.8239) trials per participant.
After target offset, the fixation mark was restored for a duration of 0.100 s. After this delay, the fixation mark turned into a question mark, which instructed the participants to deliver a response via a button press with their dominant hand. After the button press, the fixation mark was displayed again and a new trial started. The following stimulus presentation or fixation task occurred after a variable delay chosen from a uniform distribution between 1.8 and 2.4 s.
In addition to the fixation task, to further prevent eye movements, all participants were trained prior to EEG recording to maintain fixation on the central mark and their fixation ability was monitored throughout the experiment using the electro-oculogram (EOG). Moreover, we used a shape of the fixation mark specifically designed to maximize stable fixation (Thaler et al., 2013).
To control for an effect of sleepiness on the level of ongoing low-frequency power, we asked participants to report their level of sleepiness at the end of each block during resting-state and stimulation session. We used the Karolinska Sleepiness Scale (KSS) which has been validated as an indicator of objective sleepiness (Kaida et al., 2006a, b). The KSS scale consists of a nine-point Likert-type scale ranging from 1 (extremely alert) to 9 (very sleepy) that represents the sleepiness level during the immediately preceding 5 minutes. The scale was presented on the screen at the end of every block and participants were instructed to report how alert they felt during the immediately preceding block by pressing the corresponding number on the keyboard (1–9). After the button press, participants could take a self-paced break and the following block was initiated by a further button-press.
EEG recording and preprocessing
EEG was recorded with a 64-channel Biosemi ActiveTwo system at a sampling rate of 1024 Hz. Electrodes were placed according to the international 10-10 system (electrode locations can be found on the Biosemi website: https://www.biosemi.com/download/Cap_coords_all.xls). The horizontal and vertical electro-oculograms were recorded by additional electrodes at the lateral canthi of both eyes and below the eyes, respectively. As per the BioSemi system design, the Common Mode Sense (CMS) and Driven Right Leg (DRL) electrodes served as the ground. All scalp electrodes were referenced online to the CMS-DRL ground electrodes during recording. Electrodes impedance were kept below 20 mV. The raw data was recorded with ActiView (version 6.05).
The EEGLAB toolbox version 13, running on MATLAB (R2017b), was used to process and analyze the data (Delorme, 2004). In both resting-state and stimulation session, the data were re-referenced to the mastoids and down-sampled to 256 Hz. In the resting-state session, data were epoched from 0 to 330 s relative to the start of the recording block. In the stimulation session, data were epoched from −1.6 to 1.3 s relative to the onset of the stimulus presentation in stimulation trials or to the onset of the fixation task in F trials. In both sessions, the data were then filtered using an acausal band-pass filter between 0.25 and 50 Hz. We manually removed gross artifacts such as eye blinks and noisy data segments. In the stimulation session, we discarded entire trials when a blink occurred within a critical 0.5 s time window preceding stimulus onset, to ensure that participants’ eyes were open at stimulus onset. Note that “stimulus onset” in F trials refers to the time when a C1-eliciting stimulus would be presented in stimulation trials. Furthermore, we manually selected noisy channels on a trial-by-trial basis for spherical spline interpolation (Perrin et al., 1989). We interpolated on average 8.40 channels (SEM = 0.96) in 34.76 trials (SEM = 6.71). No channels were interpolated in the resting-state session. In both sessions, we transformed the EEG data using independent component analysis (ICA), and then we used SASICA (Semi-Automated Selection of Independent Components of the electroencephalogram for Artifact correction) (Chaumon et al., 2015) to guide the exclusion of IC related to noisy channels and muscular contractions, as well as blinks and eye movements. On average, we removed 7.9 (SEM 0.46) and 7.8 (SEM 0.60) out of 72 ICs in the resting-state and stimulation session, respectively.
Event-related potentials
The aim of this study was to examine the influence of prestimulus oscillatory power on ERP amplitude. We used visual stimuli to specifically elicit a robust C1 component of the visual ERP, which reflects initial afferent activity of the primary visual cortex (Di Russo et al., 2001). For each participant we identified the electrode and time point with peak activity between 0.055 and 0.090 s after stimulus onset (peak C1 activity: Di Russo et al., 2001; Bao et al., 2010), separately for LVF and UVF trials. In F trials, no C1 component of the visual ERP is expected. To quantify the ERP for this trial type and to enable comparison with the stimulation trials, we averaged the EEG data across the subject-specific electrodes with peak C1-activity in the stimulation trials. We baseline corrected single-trial ERP estimates by subtracting the prestimulus signal baseline averaged across a 0.500 s prestimulus window.
Event-related oscillations
We used time-frequency analysis to obtain a measure of ongoing oscillatory power and to estimate event-related oscillations (ERO). We first computed the stimulus-evoked, phase-locked activity (ERP) by averaging the EEG signal across trials. Then, we subtracted the average ERP from single-trial EEG signal. We applied this procedure separately for LVF, UVF, and F trials. This procedure ensures that the resulting ERO estimate does not contain stimulus-evoked activity (Kalcher and Pfurtscheller, 1995). We applied a wavelet transform (Morlet wavelets, 26 frequencies, frequency range: 5–30 Hz, number of cycles increasing linearly from 3 to 8, time window: −1 – 1.3 s relative to stimulus onset) to the EEG signal. This procedure was performed separately for each electrode and trial type (LVF, UVF, and F). We then quantified ERO as follows: where Ppost is a the time course of post-stimulus oscillatory activity and μ(Ppre) is the average ongoing power in a prestimulus window between −0.600 and −0.100 s relative to stimulus onset. This window for baseline correction was chosen based on Mazaheri and Jensen (2008) to circumvent the temporal smearing due to the wavelet convolution. ERO < 0 indicates the presence of an event-related desynchronization (ERD), indicating stimulus-induced power attenuation. ERO > 0 indicates the presence of an event-related synchronization (ERS), indicating stimulus-induced power enhancement. This procedure was performed separately for each frequency, electrode and participant.
Hypothesis testing
In this study, we tested two mechanisms underlying the modulation of ERP amplitude by prestimulus power: namely, functional inhibition and baseline shift. Functional inhibition implies that the generation of early, additive ERP components is inhibited if stimulation occurs during a state of strong prestimulus activity. In other words, positive and negative early ERP components are expected to become less positive and less negative, respectively, during states of strong prestimulus power.
The baseline shift mechanism implies that states of strong prestimulus oscillations with a non-zero mean are followed by strong power suppression (ERD), which in turn results in an enhancement of the late ERP component.
Prerequisites of the baseline shift mechanism
Previous studies (Nikulin et al., 2007; Mazaheri and Jensen, 2008) proposed the following prerequisites for linking ERO to ERP generation (baseline shift mechanism): i) the ongoing oscillations must have a non-zero mean; ii) sensory stimuli must modulate ERO magnitude; iii) the non-zero mean and the late ERP component must have opposite polarity; iv) ERO magnitude is associated with the amplitude of the late ERP component.
Estimation of non-zero-mean property of resting-state neural oscillations
The aim of the resting-state session was to estimate the non-zero-mean property of ongoing oscillations, which is known to be a critical requirement for generation of the late ERP via baseline shift (Nikulin et al., 2007; Mazaheri and Jensen, 2008). To this end, we used two analytical methods: namely, the Baseline Shift Index (BSI: Nikulin et al., 2007, 2010a) and the Amplitude Fluctuation Asymmetry Index (AFAI: Mazaheri and Jensen, 2008). In each participant, we estimated BSI and AFAI for resting-state oscillations for each electrode and frequency between 5 and 30 Hz.
Following Nikulin et al. (2007), to quantify BSI we first band-pass filtered the EEG signal using a 4th-order Butterworth filter centered at each frequency of interest ±1 Hz. Then, we extracted oscillatory power using the Hilbert transform. Additionally, we low-pass filtered the EEG signal using a 4th-order Butterworth filter with a 3-Hz cut-off frequency. The baseline shifts are low-frequency components, because the amplitude modulation of 8-30 Hz frequency oscillation can be detected only at frequencies considerably lower than 8 Hz. Thus, the low-frequency components are extracted by low-pass filtering the artifact-cleaned data at 3 Hz, based on previous studies (Nikulin et al., 2007, 2010a). We quantified the BSI as the Spearman correlation coefficient (ρ) between the low-passed EEG signal and the band-passed power, separately at each frequency and electrode. Accordingly, BSI = 0 indicates no relationship between oscillatory power and low-passed signal, as expected for zero-mean oscillations. BSI > 0 indicate that strong oscillatory power is correlated with an increase of the low-passed signal, as expected for positive-mean oscillations; instead, BSI < 0 indicate that strong oscillatory power is correlated with a decrease of the low-passed signal, as expected for negative-mean oscillations.
The amplitude modulation of oscillations with a non-zero mean affects the amplitude of peaks and troughs differently. If the peaks are larger than the troughs relative to the zero line, the ERD will make the electric field go to zero and thus reduce the peaks more strongly than the troughs. It follows that, in this case, any amplitude modulation is expected to produce larger variance for amplitude values of peaks than troughs. The different modulation of peaks and troughs can be captured by the Amplitude Fluctuation Asymmetry index (AFAI). Following Mazaheri and Jensen (2008), to quantify AFAI we first band-pass filtered the EEG signal using a 4th-order Butterworth filter centered at each frequency of interest ±1 Hz, similarly to BSI computation. Then, we identified the time points of peaks and troughs in the band-passed data. These time points were then used to obtain the signal values of peaks and troughs in the non-band-passed (broadband) signal. We quantified the AFAI as the normalized difference between the variance of the peaks and troughs of the signal as follows: where Sp and St refer to the peak and trough values, respectively, estimated in the broadband signal, based on the band-passed signal at a specific frequency.
Accordingly, an AFAI = 0 indicates that the peaks and troughs are equally modulated (as for a signal that is symmetric relative to the zero line), as expected for zero-mean oscillations. An AFAI ≠ 0 indicates amplitude asymmetry: namely, positive values indicate a stronger modulation of the peaks relative to the troughs (i.e. positive amplitude asymmetry or positive mean) and negative values indicate a stronger modulation of the troughs relative to the peaks (i.e. negative amplitude asymmetry or negative mean).
Note that the sign of the BSI and AFAI predicts the polarity of the late ERP component: specifically, if the sign is negative (oscillations with a negative mean) and positive (oscillations with a positive mean), event-related power suppression (ERD) will lead to a positive and negative deflection in the ERP, respectively (Nikulin et al., 2007, 2010a; Mazaheri and Jensen, 2008).
Interaction between event-related potentials and oscillations
To provide evidence that the baseline shift mechanism generates the late ERP component, we analyzed the relationship between ERP and ERO (ERS/ERD) across trials, as proposed by Mazaheri and Jensen (2010). According to the baseline shift mechanism, states of strong ERD should result in an enhanced late ERP component. To this end, we first identified trials with particularly weak and strong ERO, and then tested how these trials differed in the ERP amplitude during the late time window (>0.200 s). Specifically, we computed a trial-by-trial estimate of ERO magnitude at each electrode and frequency, averaged across the post-stimulus time window (0–0.900 s) (see Event-related oscillations). We also computed a trial-by-trial estimate of the late ERP component at the subject-specific C1-peak electrode (see Event-related potentials) (> 0.200 s). We baseline corrected single-trial ERP estimates by subtracting the prestimulus signal baseline averaged across a 0.500 s prestimulus window. Then, for each frequency, and electrode, trials were sorted from weak to strong ERO, divided into 5 bins (Linkenkaer-Hansen et al., 2004; Lange et al., 2012; Baumgarten et al., 2014; Iemi et al., 2017), and the amplitude of the late ERP component was calculated for each bin. The binning was done separately for each trial type (LVF, UVF, and F) and participant. Furthermore, to enable a comparison of the late ERP component across bins in each participant, the number of trials in each bin was equated by removing the trials recorded at the end of the experiment. To test the hypothesis, we then compared the amplitude of the late ERP component between strongest and weakest ERO bins (see Statistical Testing for more details).
Influence of prestimulus oscillations on event-related potentials and oscillations
We analyzed how prestimulus oscillatory activity influences ERP and ERO across trials. In this analysis, we identified trials with particularly weak and strong prestimulus oscillations, and then tested how these trials differed in the amplitude of the early and late ERP components and in the ERO magnitude. Specifically, we first computed a trial-by-trial estimate of oscillatory power with a Fast Fourier Transform (FFT) during a 0.500 s prestimulus window for each electrode and frequency. The FFT is advantageous because, unlike wavelet convolution, the results of an FFT computed over the prestimulus period cannot be influenced by signals occurring in the post-stimulus window. We also computed a trial-by-trial estimate of ERP components during the early time window (< 0.200 s) and ERP components during the late time window (> 0.200 s) at the subject-specific electrode of C1 peak activity (see Event-related potentials). We baseline corrected single-trial ERP estimates by subtracting the ERP averaged across a 0.500 s prestimulus window. In addition, we computed a trial-by-trial estimate of the ERO (ERD/ERS) at the subject-specific electrode of C1 peak activity and at each frequency and time point in the post-stimulus time window (0–0.900 s) (see Event-related oscillations). In this analysis, both ERP and ERO were quantified at the subject-specific electrode of C1 peak activity to enable comparison between the effects of prestimulus power on ERP and ERO. Then, for each frequency, and electrode, trials were sorted from weak to strong prestimulus power and divided into 5 bins (Linkenkaer-Hansen et al., 2004; Lange et al., 2012; Baumgarten et al., 2014; Iemi et al., 2017). For each bin we calculated the ERO magnitude and the amplitude of the early and late ERP components. The binning was done for each trial type (LVF, UVF, and F) and participant. Furthermore, to enable a comparison of ERO and ERP across bins in each participant, the number of trials in each bin was equated by removing the trials recorded at the end of the experiment. We then compared the ERO magnitude and the amplitude of the early and late ERP components between bins of strongest and weakest prestimulus power (see Statistical Testing for more details).
Because ongoing oscillatory activity (Kaida et al., 2006a, b; van Dijk et al., 2008; Mathewson et al., 2009; Benwell et al., 2017b, 2018) and ERP amplitude (Megela and Teyler, 1979; Budd et al., 1998; Truccolo et al., 2002; De Munck et al., 2004) may co-vary over the course of an experiment as a function of time-varying variables such as sleepiness, their correlation could be epiphenomenal. To rule this out, we asked participants to report their level of sleepiness at the end of each experimental block using the KSS questionnaire (see Stimuli and Experimental Design for more details; Kaida et al., 2006a). We then estimated how prestimulus power was related to KSS ratings throughout the stimulation session. Specifically, we computed a trial-by-trial estimate of prestimulus power for each electrode and frequency using an FFT on the 0.500 s prestimulus window. We obtained a trial-by-trial estimate of KSS scores by assigning each trial within a block with the KSS score collected at the end of the block. We then used Generalized Linear Modeling (GLM) to predict KSS ratings from prestimulus power at the single-trial level. For each participant, electrode and frequency, we fit a regression model of the following form: where KSS is the subjective sleepiness ratings obtained with the KSS questionnaire, P the prestimulus power at each frequency and electrode, β1 the estimated correlation coefficient indicating the contribution of P in explaining variability in KSS, and ε the residual errors. To remove the sleepiness-related time-varying changes in ongoing power, we recomputed a trial-by-trial measure of prestimulus power as follows: where P is the original power estimates and β1 the estimated GLM coefficient reflecting the sleepiness-power relationship. We then repeated the binning analysis on the early and late ERP amplitudes described above, with this new trial-by-trial estimate of power where sleepiness-related time-varying changes were ruled out (PKSS–c). If the relationship between prestimulus power and ERP is not determined by sleepiness affecting both variables, this new binning analysis would replicate the results of the analysis performed on raw power estimates.
Statistical testing
In the resting-state session, within each subject, we first computed the AFAI and BSI at each frequency and electrode. For the group-level statistical inference, we then tested whether the AFAI and BSI were significantly different from 0 across the sample of participants.
In the stimulation session, within each subject, we first computed: i) the difference in the late ERP component between the weakest and strongest ERD bins; ii) the difference in ERP between the weakest and strongest prestimulus power bins (separately for the early and late time window); iii) the difference in ERD between the weakest and strongest prestimulus power bins. For the group-level statistical inference, we computed the t statistics of these differences (ΔV) against the null hypothesis that there was no difference between the bins.
In both sessions, a non-parametric cluster based permutation test was used to determine significant effects while addressing multiple comparisons (Maris and Oostenveld, 2007).
We obtained a distribution of the variables of interest (i.e. AFAI /BSI for resting-state session and ΔV for stimulation session) under the null hypothesis by randomly permuting their signs 1000 times across participants. On each iteration, we tested the resulting variables with a two-tailed t-test against zero and assessed the sum of the t-values within the largest contiguous cluster of significant frequency-electrode (in the resting-state session) or time-frequency-electrode points (in the stimulation session) (cluster p-value = 0.05), resulting in a distribution of t-sums expected under the null hypothesis. A final p-value was calculated as the proportion of t-sums under the null hypothesis larger than the sum of t-values within clusters in the observed data. Because: i) the cluster permutation test is based on sampling all time points; and ii) the ERP signal comprise early, fast components and late, slow components having different statistical properties, we decided to test for significant effects separately during the early (< 0.200 s) and late (> 0.200 s) time window. To correct for multiple comparisons due to running the test twice on the same time series, we divided the final permutation p value by 2 (final p-value = 0.025, bonferroni corrected) and considered effects significant only if their p values was below this threshold. We performed the statistical test separately for positive and negative clusters as recommended by Maris and Oostenveld (2007) for a two-sided cluster permutation test. We focused the statistical analysis on all electrodes, on frequencies from 5 to 30 Hz and between 0 and 0.900 s relative to stimulus onset.