Exploring the reliability and sensitivity of the EEG power spectrum as a biomarker

https://doi.org/10.1016/j.ijpsycho.2020.12.002Get rights and content

Abstract

Introduction

The slope of the electroencephalography (EEG) power spectrum (also referred to as 1/f noise) is an important consideration when calculating narrow-band power. However, psychophysiological data also suggest this slope is a meaningful signal itself, not merely background activity or noise. We present two different methods for quantifying the slope of the power spectrum and assess their reliability and sensitivity.

Methods

We used data from N = 60 participants who had EEG collected during rest, a videogame task, and a second period of rest. At all phases of the experiment, we calculated the “spectral slope” (a regression-based method fit to all datapoints) and the “aperiodic slope” (estimated with the fitting oscillations with 1/f algorithm FOOOF). For both methods we assessed: their reliability, their sensitivity to the transition from rest to task, their sensitivity to changes during the videogame task itself, and the agreement between the two measures.

Results

Across resting phases, both spectral and aperiodic slopes showed a high degree of reliability. Both methods also showed a steepening of the power spectrum on-task compared to rest. There was also a high degree of consistency between the two methods in their estimate of the underlying slope, but FOOOF explained more variance in the power spectra across regions and type of activity (rest versus task).

Conclusion

The slope of the power spectrum is a highly reliable individual difference and sensitive to within-subject changes across two different methods of estimation. Moving forward, we generally recommend the use of the FOOOF algorithm for its ability to account for narrow-band signals, but these data show how regression-based approaches produce similar estimates of the spectral slope, which may be useful in some applications.

Introduction

A large body of literature has documented the relationship between the oscillatory activity in the brain and different behavioral or cognitive states (Buzsáki and Draguhn, 2004; Engel et al., 2001). For instance: studies showing increased delta power during non-rapid eye movement sleep is associated with memory consolidation (Marshall et al., 2006); midline frontal theta power as a correlate of cognitive control (Cavanagh and Shackman, 2015; Gevins et al., 1997; Pathania et al., 2019) and sustained attention/performance (Sauseng et al., 2005); or occipital alpha power linked to numerous aspects of visual perception (Griffiths et al., 2019; Hanslmayr et al., 2012; Michelmann et al., 2016).

Notwithstanding these and numerous other findings, there has also been a recognition that these canonical frequency bands are somewhat arbitrary and may not fully capture the information contained in the frequency-domain. For example, there are individual differences in the precise bandwidths of some canonical frequencies (Grandy et al., 2013; Haegens et al., 2014; Köster et al., 2019), which, in the case of alpha, are both heritable (Smit et al., 2006) and linked to cognitive ability (Grandy et al., 2013). Not only do such effects logically entail variation in adjacent frequency bands across individuals (Klimesch, 2012), but it appears likely that frequency boundaries even vary within individuals during cognitive tasks (Benwell et al., 2019; Cohen, 2014; Wutz et al., 2018). Further, and most importantly for present purposes, the power in any single frequency band can also be distorted by the underlying slope of the power-spectrum as a whole.

The background pattern which produces this slope in the spectral density has been referred to as “1/f noise” (Voytek and Knight, 2015), and reflects the fact that (in many time-series), power tends to decrease at higher frequencies, in the form S(f)~1/fα. In this equation, spectral power, S(f), is a function of the frequency, f, and the rate of decay in the power spectrum is determined by a constant, α. In the case of a truly random process (i.e., white noise), all frequencies are equally represented and therefore the slope of the power spectrum is flat, a = 0. In human EEG, low frequencies are dominant, and the slope of power spectrum is negative, α < 0.

Although often called 1/f noise, it is important to recognize that the slope of the power-spectrum is not necessarily noise and can be a signal itself. Additionally, there are different approaches, historically, for characterizing the background slope in the power spectrum. It can be done by fitting a line to all the data-points in power spectrum, as shown in Fig. 1B, or by algorithmically estimating the slope underneath the canonical frequency band components, as shown in Fig. 1C. We refer to these two different approaches as the “spectral slope” and the “aperiodic slope” (Haller et al., 2018), respectively. Below, we will contrast these two methods in more detail, but first we will briefly review the literature suggesting that the spectral slope is a meaningful psychophysiological signal.

There is emerging evidence to support the idea that the spectral slope is a signal rather than an anatomical or mathematical artifact. For instance, individual differences in the spectral slope mediate age-related changes in visuospatial working memory (Voytek et al., 2015), the spectral slope is more negative under anesthesia than in conscious states (Gao, 2015), more positive during wakefulness compared to sleep (Freeman et al., 2003; Freeman and Zhai, 2009) and the spectral slope is a better predictor of schizophrenia than canonical frequency band patterns (with more positive slopes associated with pathology; Peterson et al., 2017).

Importantly, these data show that there is not a simple linear relationship between the spectral slope and cognitive states (i.e., positive is not necessarily worse and negative is not necessarily better). Thus, it is possible that the optimal spectral slope is dynamic, depending on the context, but that extreme ends of the spectrum are undesirable: extremely negative slopes are associated with sleep and anesthesia, whereas extremely positive slopes are associated with advanced age and pathology. To illustrate this point, we foreshadow some of our results by showing the average power spectrum at rest compared to during a videogame task in the current experiment, Fig. 1D. Note that the slope of the power spectrum generally becomes steeper on task (a broad-band change) and there is a noticeable reduction in α-band power (a narrow-band change). This non-linear relationship calls into question what physiological processes the spectral slope represents. Some authors have posited that the changes in signal complexity and the power spectrum are mediated by the contributions of excitatory and inhibitory cellular activity (Gao et al., 2017; Waschke et al., 2017).

As mentioned before, there are different methods for quantifying broad-band power, and we focus on two: the spectral slope and the aperiodic slope. Haller et al. (2018) have compellingly argued that the spectral slope (Fig. 1B) is problematic, as the spectral slope confounds power in canonical frequency bands with broad-band changes in the power-spectrum. To disentangle narrow-band power from broad-band power, Haller et al. recommend estimating the aperiodic slope, see Fig. 1C. This approach calculates the power within frequency bands, separate from the aperiodic slope, rather than measuring the total power in the spectrum (a more detailed description is provided in the Methods). Calculated in this way, task-specific activity in narrow bands can theoretically be dissociated from broad-band changes in the power spectrum. Estimating the aperiodic slope requires an algorithm (“fitting of oscillations and one-over f noise”; FOOOF) that extracts the periodic (narrow-band) and aperiodic (broad-band) components from the EEG power spectrum.

Thus, both the spectral slope and aperiodic slope provide methods for broad-band quantification of the power spectral density. Calculation of the aperiodic slope attempts to isolate a hypothetical background from the periodic components of the signal (e.g., changes in alpha or theta bands). In contrast, the spectral slope is sensitive to these periodic components and reflects the broadband balance of all frequencies in the power spectrum. These differences in the method of calculation create distinct strengths and weaknesses in the two different methods. For instance, FOOOF can delineate an aperiodic slope from narrow-band components, but this relies in several assumptions (e.g., that the aperiodic slope is indeed a separate signal, the number and width of narrow-band components that need to be estimated, etc). Conversely, the spectral slope is computationally simpler and treats aperiodic background and the periodic components as a single signal. However, this simplification conflates distinct neural processes (i.e., the aperiodic and narrow-band components). One way to ameliorate this problem is through deletion of the phasic components (e.g., removing data from alpha band, which has a prominent peak as shown in Fig. 1; see also Voytek et al., 2015). To date, however, there has not been a direct comparison of these measures (aperiodic slope versus spectral slope) and their sensitivity to different wakeful states. Establishing the psychometric properties spectral and aperiodic slopes is critical to their utility as physiological signals.

As an interim summary, the spectral slope appears to be an electrophysiological signal that is sensitive to different aspects of psychophysiological function. However, prior research has mostly explored extreme ends of psychological function (e.g., pathology versus healthly controls) or extreme ends of physiological change (e.g., young adults versus older adults). To date, relatively little is known about how the spectral slope changes as a function of different wakeful states (e.g., at rest versus on task, in relatively easy versus challenging situations). Further, there is uncertainty in the literature about the best method for measuring the spectral slope (shown in Fig. 1B) and how it differs from the aperiodic slope (shown in Fig. 1C).

In order to address these questions, we performed a secondary analysis of practice data from a previous learning study (Pathania et al., 2019). The aims of the present analysis were to establish (1) the reliability and (2) sensitivity of these two different measures of the power spectrum during different wakeful states. Specifically, we contrasted the spectral slope against the aperiodic slope during multiple periods of rest and multiple periods of playing a videogame. To measure reliability, we compared these different measures at two different rest periods in the same day (prior to the gaming task and after). To measure sensitivity, we first measured changes in the spectral density from rest to the videogame task (on average). Second, we looked at sensitivity within the task, by comparing blocks of different difficulties in the videogame.

Section snippets

Participants

Sixty participants (40 females, 20 male; median age = 21y, IQR = 20 to 21y), were recruited through undergraduate courses at Auburn University. The participants were included if they were right-handed with little to no experience playing the Microsoft Kinect ®, and excluded if they had any musculoskeletal or neurological impairments that might affect their performance (as self-reported by participants). The participants were instructed to avoid alcohol and nicotine consumption and to get an

Reliability: intraclass correlation across resting time points

Based on the two-way random-effects ICC estimates, there was generally a high degree of reliability between pretest and posttest for both spectral slopes and aperiodic slopes (see Table 1). ICC's for the spectral slopes ranged between 0.85 and 0.95, and ICC's for the aperiodic slopes ranged between 0.78 and 0.93. Example data for the central regions can also be seen in the Bland-Altman plots in Fig. 3. Both methods were relatively free of bias as the average difference between the repeated

Discussion

The main aim of this study was to observe the reliability and the sensitivity of two different measures of the EEG power spectrum: the spectral slope (fit by LMER) and the aperiodic slope (estimated by FOOOF). First, we measured the reliability of the spectral slope and the aperiodic slope for the two resting time points, pretest and posttest, using ICC(2,k). There was a high degree of test/re-test reliability for both measures across the two resting time points. Although both methods yielded

Conclusions

Notwithstanding these limitations, our results show that the behavior of the slope of the power spectrum is broadly consistent across two methods of measurement (the spectral slope and the aperiodic slope), but even small differences affected the statistical results with respect to on-task activity. These data reinforce the idea that the slope of power spectrum is a valuable physiological signal. The scope for the application of this technique is quite broad, as it can be applied in both

Declaration of competing interest

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Acknowledgments

The authors would like to express their gratitude to Dr. Mark Williams, Dr. Peter Fino, and Dr. Lee Dibble for their valuable time, guidance, and feedback for culmination of this study. We would also like to thank Braydon Kener and Ellen Clark Williams for their valuable assistance in data processing.

References (49)

  • S.D. Muthukumaraswamy et al.

    1/f electrophysiological spectra in resting and drug-induced states can be explained by the dynamics of multiple oscillatory relaxation processes

    NeuroImage

    (2018)
  • R. Oostenveld et al.

    The five percent electrode system for high-resolution EEG and ERP measurements

    Clin. Neurophysiol.

    (2001)
  • A. Pathania et al.

    Challenge, motivation, and effort: neural and behavioral correlates of self-control of difficulty during practice

    Biol. Psychol.

    (2019)
  • M. Pertermann et al.

    The modulation of neural noise underlies the effectiveness of methylphenidate treatment in attention-deficit/hyperactivity disorder

    Biological Psychiatry: Cognitive Neuroscience and Neuroimaging

    (2019)
  • P. Sauseng et al.

    Fronto-parietal EEG coherence in theta and upper alpha reflect central executive functions of working memory

    Int. J. Psychophysiol.

    (2005)
  • C.M. Smit et al.

    Genetic variation of individual alpha frequency (IAF) and alpha power in a large adolescent twin sample

    Int. J. Psychophysiol.

    (2006)
  • B. Voytek et al.

    Dynamic network communication as a unifying neural basis for cognition, development, aging, and disease

    Biol. Psychiatry

    (2015)
  • Bates, D., Mächler, M., Bolker, B., & Walker, S. (2014). Fitting linear mixed-effects models using lme4....
  • G. Buzsáki et al.

    Neuronal oscillations in cortical networks

    Science

    (2004)
  • M.X. Cohen

    Fluctuations in oscillation frequency control spike timing and coordinate neural networks

    J. Neurosci.

    (2014)
  • T. Donoghue et al.

    Electrophysiological frequency band ratio measures conflate periodic and aperiodic neural activity

    BioRxiv

    (2020)
  • A.K. Engel et al.

    Dynamic predictions: oscillations and synchrony in top–down processing

    Nat. Rev. Neurosci.

    (2001)
  • A. Fiol-Veny et al.

    Long range temporal correlations in EEG and depression

    Ann Depress Anxiety

    (2015)
  • W.J. Freeman et al.

    Simulated power spectral density (PSD) of background electrocorticogram (ECoG)

    Cogn. Neurodyn.

    (2009)
  • Cited by (24)

    • Spectral parameterization for studying neurodevelopment: How and why

      2022, Developmental Cognitive Neuroscience
      Citation Excerpt :

      Using longitudinal data, Schaworonkow and Voytek (2021) observed a progressive flattening of power spectra among infants across the first months of life. Psychometrically, the exponent has been shown to have high internal consistency in infancy and adolescence (Karalunas et al., 2021), as well as good test-retest reliability in adulthood (Pathania et al., 2021) and among children with autism spectrum disorder (Levin et al., 2020). While studies suggest good signal reliability, it is important to note that it would be difficult to compare aperiodic parameters across participants when different recording systems are used, or in different environments, due to the fact that measurement noise is also often 1/f in nature (He et al., 2010; Miller et al., 2009).

    • Stability of spectral estimates in resting-state magnetoencephalography: Recommendations for minimal data duration with neuroanatomical specificity

      2022, NeuroImage
      Citation Excerpt :

      For example, variability in the aperiodic slope and offset accounts for changes in the neural power spectrum along the lifespan (Cellier et al., 2021; Schaworonkow and Voytek, 2021), covaries with clinical outcomes in a number of disorders (Ostlund et al., 2021; Pani et al., 2021; Van Heumen et al., 2021; Wilkinson and Nelson, 2021), and predicts memory consolidation during sleep (Helfrich et al., 2021) and individual differences in visuomotor learning (Immink et al., 2021). Regarding the reliability and stability of these components, Pathania et al. (2021) found that estimates of aperiodic components are consistent over intervals of at least seven days. We wish to expand these investigations by determining the minimum data duration required to produce robust estimates of the periodic and aperiodic components of MEG resting-state source maps.

    View all citing articles on Scopus
    View full text