Elsevier

NeuroImage

Volume 54, Issue 2, 15 January 2011, Pages 860-874
NeuroImage

A better oscillation detection method robustly extracts EEG rhythms across brain state changes: The human alpha rhythm as a test case

https://doi.org/10.1016/j.neuroimage.2010.08.064Get rights and content

Abstract

Oscillatory activity is a principal mode of operation in the brain. Despite an intense resurgence of interest in the mechanisms and functions of brain rhythms, methods for the detection and analysis of oscillatory activity in neurophysiological recordings are still highly variable across studies. We recently proposed a method for detecting oscillatory activity from time series data, which we call the BOSC (Better OSCillation detection) method. This method produces systematic, objective, and consistent results across frequencies, brain regions and tasks. It does so by modeling the functional form of the background spectrum by fitting the empirically observed spectrum at the recording site. This minimizes bias in oscillation detection across frequency, region and task. Here we show that the method is also robust to dramatic changes in state that are known to influence the shape of the power spectrum, namely, the presence versus absence of the alpha rhythm, and can be applied to independent components, which are thought to reflect underlying sources, in addition to individual raw signals. This suggests that the BOSC method is an effective tool for measuring changes in rhythmic activity in the more common research scenario wherein state is unknown.

Research highlights

►BOSC is a method to identify oscillatory activity in time series data. ►BOSC sets a power and a duration threshold to exclude non-rhythmic epochs. ►The power threshold is derived from an estimate of the background spectrum. ►The duration threshold is based on the desired number of cycles at each frequency. ►In this study BOSC reliably extracts the human occipital alpha oscillation.

Introduction

In 1929, Hans Berger was the first person to describe rhythmic oscillations of electrical potential recorded from the human scalp. This included a 10-Hz signal present over the occipital cortex which became known as the alpha rhythm (Berger, 1929). The alpha rhythm occurs at frequencies of 8–12 Hz, is most prominent during behavioural states of relaxed wakefulness with eyes closed, and is replaced by low-voltage, faster activity with eye opening. While alpha was long thought to represent a brain idling state, its role in neural processing now appears much more complex. Alpha power increases are elicited in tasks requiring top-down inhibition or highly selective processing, such as ignoring specific stimuli while attending to others (Freunberger et al., 2009, Klimesch et al., 2007, Worden et al., 2000). In addition, alpha power has been shown to increase with memory load during the retention interval in a modified Sternberg task (Jensen et al., 2002).

In the years since the discovery of alpha, many other electroencephalographic (EEG) oscillations have been described at various frequencies and locations, and have been associated with distinct brain processes and behavioural states. EEG oscillations represent the collective activity of rhythmically synchronized neuronal networks (reviewed in Niedermeyer and Lopes da Silva, 2005) and are thought to be essential for both ongoing modulation of behaviour and offline processing, for example, occurring during sleep or rest (Buzsaki and Draguhn, 2004).

Despite the fact that EEG oscillations have been studied for over 80 years, identifying oscillatory activity in EEG traces is still highly variable across studies (van Vugt et al., 2007). The primary method in the field is to calculate frequency spectra using either Fourier or wavelet transforms of the signal, and to visually inspect the resulting spectra for peaks in specific frequency bands. Each method has its own particular strengths and weaknesses. The Fourier transform efficiently decomposes any stationary signal as a series of sine waves of different frequencies and amplitudes (Brigham, 1974). Unfortunately, brain signals are rarely (if ever) stationary, and are often dynamic on time scales that are too fast to allow for an adequate representation of all frequency components using traditional Fourier-based computational methods. Wavelet transforms, on the other hand, fit more temporally discrete pieces of the signal using a family of mathematical functions (known as “wavelets”) that vary both in frequency and time windows (Kemerait and Childers, 1972, Schiff et al., 1994). Wavelet methods avoid the problem of nonstationarity by allowing spectral characteristics to change over time. Unfortunately, the presence of a spectral peak at a given frequency using either method does not necessarily imply underlying oscillatory activity at that frequency because non-oscillatory, large-amplitude artefacts and transient signals can produce power changes at specific frequencies. In addition, some functionally relevant oscillations are highly transient, and depending on the size of the analysis time window they can be obscured by other, larger amplitude, oscillatory activity or they can be overshadowed by the overall power in the background spectrum.

Autocorrelations of EEG signals are often performed to establish the presence of oscillatory activity, but this analysis method also has limitations: the fundamental frequency will dominate the autocorrelation function, and it is difficult to rigorously compute a significance value for rhythmicities that may be apparent in the autocorrelation plot. Further, the baseline for deciding when bandpass power reflects an oscillation varies from study to study. In some cases, the observation of a peak in the spectrum and/or autocorrelation is sufficient, whereas in other cases, a direct comparison to some baseline condition is used. In either approach, the determination of rhythmicity depends upon the variable characteristics of the signals themselves. Finally, pre-whitening (normalizing power across frequency) can help to overcome the frequency bias due to the colored-noise form of the EEG spectrum, but it can be overly conservative, overcorrecting when peaks (potentially reflecting oscillations—the signal of interest) are present. What is needed is a method of detecting oscillations that derives detection thresholds in a way that is consistent across frequency, electrode (brain region), task, electrophysiological state and species.

With these issues in mind Caplan et al. (2001) introduced a new method to detect oscillatory activity in EEG signals which we term the Better OSCillation detection method, or BOSC. This method is designed to take into account the functional form of “background,” non-rhythmic portion of the signal and to reveal segments of the recording that deviate significantly from the spectral characteristics of the background. In BOSC, one calculates a power threshold (PT) and a duration threshold (DT) for oscillatory episode detection by modeling the known functional form of the background power spectrum. That is, at a given frequency, BOSC detects increases in power, above PT, of a specific minimum duration, DT, thereby rejecting increases in spectral amplitude that are non-repeating (Fig. 1). Briefly, an average wavelet power spectrum is calculated, and this spectrum is modeled as colored noise (power scaling as 1/frequency) with the possible addition of peaks at some frequencies that potentially reflect the presence of oscillations (Fig. 1D). The colored noise spectrum is a basic property of EEG as well as other natural autocorrelated signals (Schlesinger and West, 1988). This spectrum is then fit by linear regression in log-log space. In previous applications of BOSC, the power threshold (PT) has been set to the 95th percentile of the theoretical probability distribution (with χ2(2) form; e.g., Fig. 1E) of power values at a given frequency and the duration threshold (DT) has been set at each frequency f to 3 complete cycles (3/f). Oscillations are only detected when both PT and DT are exceeded. The qualitative outcome of application of the method is robust to the exact choice of threshold values; Caplan et al. (2001) characterized how variations in PT and DT affect the frequency specificity and conservatism of oscillation detection. The BOSC method also lends itself to a useful quantitative measure: the proportion of time (within a trial or other time segment) that detected oscillations are present, which we have termed Pepisode (Caplan and Glaholt, 2007, Caplan et al., 2001, Caplan et al., 2003, van Vugt et al., 2007).

This BOSC method has been used successfully to identify and quantify theta oscillations in the human neocortex that are correlated with memory encoding and retrieval and sensorimotor integration (Caplan and Glaholt, 2007, Caplan et al., 2001, Caplan et al., 2003). It has also been applied to oscillatory activity from intracranial recordings in the human hippocampus (Ekstrom et al., 2005). Previous work has demonstrated that the BOSC method provides standardized detection criteria across frequency, electrode and task. Ideally, this method can be generalized to extend to all types of EEG recordings, providing standardized oscillation detection criteria across electrophysiological state and species (Hughes et al., 2009). Thus, we sought to more systematically evaluate the BOSC method on a well-known and easy-to-induce oscillation that occurs during wakefulness: Alpha. An ancillary goal was to determine whether the BOSC method is compatible with independent component analysis (ICA) which has been increasingly adopted as means of inferring activity of presumed underlying sources (Makeig et al., 1997). This would only work if the power spectrum of an independent component also exhibited a colored-noise functional form. Our final goal was to determine whether state changes that occur during wakefulness present a challenge to the method. Because the method relies on estimating the background spectrum, if the background spectrum were to change sufficiently across state, this would very likely result in false positives or false negatives or a bias across frequency. We show in the present study that the BOSC method is robust to state changes and reveals features that are entirely consistent with previous research on the alpha rhythm.

Section snippets

Data collection

Twelve human participants recruited from the University of Alberta community were fitted with a 256-channel HydroCell geodesic sensor net (Electrical Geodesics Inc., Eugene, OR) with electrode impedances kept below 50 kΩ (Ferree et al., 2001). Data were amplified by the EGI NetAmps 300 amplifier with a 400-Hz anti-aliasing hardware filter; digitized at a rate of 250 Hz and acquired via Net Station software using a 24-bit A/D converter.

Experimental protocol

The task was presented in E-prime (Psychology Software Tools

Detecting alpha

In 11/12 subjects we found ICs with clear spectral peaks in the alpha (8–12 Hz) band. Fig. 2 illustrates the detection of alpha oscillations in a single independent component by BOSC. The component was chosen for further study based on an 8–12 Hz peak in the conventional spectrum (Fig. 2B) and a clear dipole-like topography with a maximum weight in electrodes overlying occipital cortex (Fig. 2A). Consistent with the peak in the conventional power spectrum, the BOSC method detected oscillations in

Discussion

Our findings support the BOSC method as a powerful tool for identifying oscillations in neural data. It does so by setting parameters in a manner that minimizes bias across frequency, electrode and task (as shown previously), electrophysiological state (shown here) and species (Hughes et al., 2009). The alpha rhythm was an ideal candidate for testing the validity of this method: It is large in amplitude, it dominates the recording when present, and it is easily evoked. In addition to evaluating

Conclusion

This study presented a test case for the BOSC method in detecting an oscillation that is strong, easy to elicit and well described in the literature. The results show that the method can detect alpha oscillations that are temporally restricted to periods when they are expected (eyes-closed condition) in both ICs and single electrodes. Furthermore, the detection of alpha oscillations is robust to the selection of the background window. These findings, along with recent validation of the method

Acknowledgments

We would like to thank the following people who contributed to this work: Leanna Cruikshank for technical assistance with data analysis and experiment programming, as well as data collection and helpful comments on the manuscript; Mayank Rehani and Michelle Chan for data collection; Chris Madan for technical assistance. Supported by grants to JBC from the Natural Sciences and Engineering Research Council (NSERC) of Canada and the Alberta Ingenuity Fund and grants to CTD from NSERC. CTD is an

References (28)

  • G. Buzsaki et al.

    Neuronal oscillations in cortical networks

    Science

    (2004)
  • J.B. Caplan et al.

    Distinct patterns of brain oscillations underlie two basic parameters of human maze learning

    J. Neurophysiol.

    (2001)
  • J.B. Caplan et al.

    Human theta oscillations related to sensorimotor integration and spatial learning

    J. Neurosci.

    (2003)
  • A.D. Ekstrom et al.

    Human hippocampal theta activity during virtual navigation

    Hippocampus

    (2005)
  • Cited by (90)

    • Training the spatially-coded SSVEP BCI on the fly

      2022, Journal of Neuroscience Methods
      Citation Excerpt :

      For example, considering alpha power in relation to power in the theta and beta frequency bands may yield a better indicator for alpha increases in response to closing the eyes (Nguyen et al., 2018). Alternatively, the power spectrum of a trial may be compared to a model of the background spectrum, and deviations of a given minimum duration may be taken as a voluntarily generated alpha power increase (Whitten et al., 2011). But of course the feedback method can be selected from the whole gamut of sufficiently reliable BCI paradigms.

    View all citing articles on Scopus
    View full text