Neuroinformatics (2022) 20:991–1012

https://doi.org/10.1007/s12021-022-09581-8

ORIGINAL ARTICLE

Separating Neural Oscillations from Aperiodic 1/f Activity: Challenges

and Recommendations
Moritz Gerster1,2,3 · Gunnar Waterstraat2 · Vladimir Litvak4 · Klaus Lehnertz5,6,7 · Alfons Schnitzler8 · Esther Florin8 ·
Gabriel Curio2,3 · Vadim Nikulin1,2,3

Accepted: 25 February 2022 / Published online: 7 April 2022

© The Author(s) 2022

Abstract
Electrophysiological power spectra typically consist of two components: An aperiodic part usually following an 1/f power
law P ∝ 1∕f 𝛽 and periodic components appearing as spectral peaks. While the investigation of the periodic parts, commonly
referred to as neural oscillations, has received considerable attention, the study of the aperiodic part has only recently gained
more interest. The periodic part is usually quantified by center frequencies, powers, and bandwidths, while the aperiodic
part is parameterized by the y-intercept and the 1/f exponent 𝛽 . For investigation of either part, however, it is essential to
separate the two components. In this article, we scrutinize two frequently used methods, FOOOF (Fitting Oscillations &
One-Over-F) and IRASA (Irregular Resampling Auto-Spectral Analysis), that are commonly used to separate the periodic
from the aperiodic component. We evaluate these methods using diverse spectra obtained with electroencephalography
(EEG), magnetoencephalography (MEG), and local field potential (LFP) recordings relating to three independent research
datasets. Each method and each dataset poses distinct challenges for the extraction of both spectral parts. The specific spectral
features hindering the periodic and aperiodic separation are highlighted by simulations of power spectra emphasizing these
features. Through comparison with the simulation parameters defined a priori, the parameterization error of each method
is quantified. Based on the real and simulated power spectra, we evaluate the advantages of both methods, discuss common
challenges, note which spectral features impede the separation, assess the computational costs, and propose recommenda-
tions on how to use them.

Keywords 1/f exponent · FOOOF · IRASA · Neural oscillations · Spectra · EEG/MEG

Introduction spectral peaks, also referred to as the periodic part of the

spectrum (Buzsáki & Draguhn, 2004; Engel et al., 2001;
Analysis of macroscopic electromagnetic brain activ- Schnitzler & Gross, 2005). The full spectrum, however,
ity (e.g., by EEG and MEG) has long been focusing on also consists of a continuous component whose analysis
the investigation of ‘rhythmic’ neural oscillations. In the has, so far, seen less attention. This aperiodic or ‘arrhyth-
frequency domain, neural oscillations appear as distinct mic’ part of the spectrum (Freeman & Zhai, 2009; Miller

4
* Moritz Gerster Wellcome Centre for Human Neuroimaging, UCL Queen
[email protected] Square Institute of Neurology, London, UK
5
1 Department of Epileptology, University of Bonn Medical
Research Group Neural Interactions and Dynamics,
Centre, Bonn, Germany
Department of Neurology, Max Planck Institute for Human
6
Cognitive and Brain Sciences, Leipzig, Germany Helmholtz Institute for Radiation and Nuclear Physics,
2 University of Bonn, Bonn, Germany
Neurophysics Group, Department of Neurology, Charité
7
– Universitätsmedizin Berlin, corporate member of Freie Interdisciplinary Center for Complex Systems, University
Universität Berlin and Humboldt-Universität zu Berlin, of Bonn, Bonn, Germany
Berlin, Germany 8
Institute of Clinical Neuroscience and Medical Psychology,
3
Bernstein Center for Computational Neuroscience, Berlin, Medical Faculty, Heinrich-Heine University Düsseldorf,
Germany Düsseldorf, Germany

13
Vol.:(0123456789)
992 Neuroinformatics (2022) 20:991–1012

et al., 2009) has been related to the integration of underly- exponents during wakefulness and anesthesia. However,
ing synaptic currents (Buzsáki et al., 2012). Since the time in the presence of periodic components, this method is
series of the aperiodic part is typically self-similar across error-prone because larger periodic peaks will bias the
many temporal scales, it is also referred to as “fractal” or linear regression fit.
“scale-free” activity. The power spectral density (PSD) Irregular-resampling autospectral analysis (IRASA) (Wen
of the aperiodic component follows a power law P ∝ 1∕f 𝛽 & Liu, 2016) aims to separate periodic components from the
(Miller et al., 2009) and is sometimes called 1/f activity aperiodic part of the spectrum. Due to their fractal nature, ape-
for that reason. In this text, we will refer to the scaling riodic time series remain robust against resampling, whereas
exponent 𝛽 in this equation as 1/f exponent. periodic components are strongly affected by this procedure.
The investigation of neural oscillations has received IRASA takes advantage of this dichotomy and ‘removes’ the
much attention in electrophysiological studies (Buzsáki periodic parts from a spectrum. The–ideally–pure aperiodic
& Draguhn, 2004; Singer, 1999; Ward, 2003). However, part of the spectrum obtained with this method can then be
the standard analysis of assessing periodic power through used for fitting the 1/f exponent.
bandpass filtering is problematic because the pass-band Another method, ‘fitting oscillations & one over f’
comprises both periodic and aperiodic activity. If the (FOOOF) (Donoghue et al., 2020), aims at modeling the
power of aperiodic activity changes between two condi- periodic components: It iteratively applies Gaussian fits to
tions, analyzing neural oscillations in bandpass filtered all periodic components and hereby obtains a model of the
signals would hence be confounded by these changes in periodic part. This model of periodic activity is subtracted
the aperiodic part of the spectra. For that reason, estimat- from the spectrum to obtain an–ideally–pure aperiodic com-
ing the 1/f component before determining the power of ponent which can be used for fitting 𝛽 . In addition, the peri-
periodic activity has recently been suggested (Donoghue odic model allows for analyzing the periodic components
et al., 2021; Wen & Liu, 2016). (e.g., regarding center frequencies, bandwidths, and power)
Besides investigating neural oscillations, the investigation without the bias from aperiodic activity.
of the aperiodic component has recently gained considerable This article highlights and discusses the general chal-
interest (He, 2014; Kello et al., 2010). For example, the 1/f lenges of estimating 1/f exponents. In addition, we also dis-
exponent was shown to change with task (Ouyang et al., 2020; cuss method-specific challenges of FOOOF and IRASA, the
Podvalny et al., 2015; Waschke et al., 2021), age (Bódizs most commonly used algorithms for that purpose.
et al., 2021; Dave et al., 2018; Waschke et al., 2017; Cel- In the Methods section, we will introduce our simula-
lier et al., 2021; He et al., 2019; Schaworonkow & Voytek, tions, our datasets, and both algorithms FOOOF and IRASA.
2021; Voytek et al., 2015), psychoactive drug administra- We will analyze challenges by the example of FOOOF in
tion (Muthukumaraswamy & Liley, 2018; Stock et al., 2020; section FOOOF and by the example of IRASA in the sec-
Timmermann et al., 2019), and disease (Molina et al., 2020; tion IRASA. To aim for broad applicability of our assess-
Robertson et al., 2019; Veerakumar et al., 2019; van Heumen ment, we will apply these methods to simulations with known
et al., 2021; Ostlund et al., 2021, Karalunas et al. 2022) and it ground truth in addition to various electrophysiological signals
decreases with cortical depth (Halgren et al., 2021). Further- obtained from empirical EEG, gradiometer MEG, magnetom-
more, using computational modeling, (Gao et al., 2017) sug- eter MEG, source-reconstructed voxel activity from MEG, and
gested the 1/f exponent 𝛽 as an estimator of excitation–inhibi- subthalamic nucleus-(STN-)LFP data acquired by three inde-
tion (E–I) balance. Many studies comparing conscious states pendent research groups. We will discuss these challenges in
to unconscious states seem to support this concept. Conscious the section Discussion, and we will provide some guidance on
states are associated with increased excitation while uncon- how to use these methods in the Conclusion section.
scious states, such as NREM sleep and anesthesia, are associ-
ated with pronounced inhibitory processes. Experiments com-
paring the 1/f exponent in these two conditions did indeed find Methods
larger values for NREM sleep (Lendner et al., 2020; Miskovic
et al., 2019) and anesthesia (Colombo et al., 2019; Muthuku- Simulations
maraswamy & Liley, 2018;Waschke et al., 2021; Zhou et al.,
2021) compared to awake states. We simulate aperiodic 1/f activity by constructing a Fourier
But how to best estimate the 1/f exponent? This will power spectrum following a preset 1∕f 𝛽 power-law. The cor-
be the main question discussed in this study. One option responding phases of the Fourier spectrum are distributed
is to simply fit a straight line using (robust) linear regres- uniformly randomly. To add oscillations, we add Gaussian-
sion. (Gao et al., 2017) used this method in the frequency shaped peaks to the Fourier power spectrum with amplitudes
ranges apart from pronounced oscillatory peaks in elec- A and a spectral extent given by center frequencies fcenter and
trocorticography (ECoG) data and identified distinct 1/f variances 𝜎f2 . The corresponding time series, consisting of

13
Neuroinformatics (2022) 20:991–1012 993

both ‘neural’ oscillations and aperiodic activity, is then for visualization purposes (multi-taper estimation of sinusoi-
obtained by applying the inverse fast Fourier transform. The dal components “spectrum_fit” of MNE python (Gramfort
simulated time series either have a duration of 180 s at a et al., 2013). Note, that a notch filter should not be applied
sampling rate of fsample = 2400 Hz or are matched to the to frequency ranges used for FOOOF fitting. Therefore, we
empirical data to which a simulation might be compared. If exclude 50 Hz from the notch filter (see SI Fig. 1).
noted in the text, Gaussian white noise may be added to the
time series afterward. Since most algorithms to generate 1/f Dataset 2
activity lead to identical power spectra, the specific choice
of the algorithm has no impact on the present analysis. Dataset 2 contains EEG data from a 12-year-old boy with
absence epilepsy recorded at the Department of Epileptol-
Empirical Data ogy at the University of Bonn (Gerster et al., 2020). The uni-
versity’s ethics committee approved the study, and a parent
We compare the results from our simulations to three empir- gave written informed consent that the clinical data might
ical datasets. be used and published for research purposes. EEG data were
acquired at a sampling rate of fsample = 256 Hz (16-bit A/D
Dataset 1 conversion) within a bandwidth of 0.3–70 Hz from 19 elec-
trodes in a bipolar montage. The locations and nomencla-
Dataset 1 was re-analyzed from (Litvak et al., 2010, 2011) ture of these electrodes are standardized by the American
and contains MEG and LFP data of 14 Parkinsonian patients Electroencephalographic Society (Sharbrough, 1991). The
after bilateral implantation of subthalamic nucleus (STN) EEG was recorded over several hours and contains 5 absence
stimulation electrodes (Medtronic, Minneapolis, MN, USA seizures. In this study, we present 40 s of the bipolar EEG
with four platinum-iridium cylindrical surfaces of diameter channel “F3 − C3” during one absence seizure.
1.27 mm, length 1.5 mm, and center-to-center separation
2 mm) for deep brain stimulation (DBS). The joint ethics Dataset 3
committee of the National Hospital of Neurology and Neu-
rosurgery and the University College London Institute of Dataset 3 contains MEG recorded with gradiometers and mag-
Neurology approved the study, and all patients gave their netometers and LFP data from a Parkinsonian patient recorded
written informed consent. The patients were recorded three at the Universitätsklinikum Düsseldorf. The data were acquired
days after surgery when the electrode leads were still exter- using a whole-head MEG system with 306 channels (Elekta
nalized. The recordings were obtained during a Parkinso- Vectorview, Elekta Neuromag, Finland), and segmented
nian state OFF medication (after overnight withdrawal) and “1–3–3–1” electrode DBS-LFP (Abbott St. Jude Medical model
an ON medication state. MEG (275 channels, CTF/VSM 6172, contact height: 1.5 mm with 0.5 mm vertical spacing)
MedTech, Vancouver, Canada) and DBS-LFP were recorded during the ON- and OFF-medication state (after overnight with-
simultaneously during three minutes of resting-state at a drawal). The patient was recorded 1 day after surgery when
sampling rate of fsample = 2400 Hz. The LFP recordings the electrode leads were still externalized. The resting-state
were referenced to the right mastoid during recording and was recorded for 10 min at a sample rate of fsample = 2400
later re-referenced to a bipolar montage between adjacent Hz. The LFP recordings were referenced to the right mastoid
electrode contacts. This results in 3 bipolar LFP channels during recording and later re-referenced to a bipolar montage
per hemisphere. All data were bandpass filtered in hardware between adjacent electrodes. The data were offline band-pass
between 1–600 Hz. MEG source reconstruction was per- filtered between 0.3 Hz and 600 Hz and notch-filtered at 50,
formed with varying regularization by Linearly Constrained 100, …, 600 Hz power line noise (with a second-order IIR filter
Minimum Variance beamformer (Van Veen et al., 1997). of bandwidth 1 Hz). Note that notch filtering in the fitting range
Aside from the six LFP channels, the dataset contains three at 50 Hz is unproblematic with using IRASA. The patient gave
MEG channels per patient from voxels located in the sup- written consent to participate in the study, which was approved
plementary motor area (SMA), left primary motor cortex by the Ethics committee of the Universitätsklinikum Düssel-
(M1), and right M1. In this study, we draw examples from dorf. In this study, we analyze data from one gradiometer chan-
voxel data located in the supplementary motor area (SMA) nel, one magnetometer channel, and one LFP channel of the
of patients 5 and 6 and bipolarly recorded LFPs from the subject.
STN of patients 9 and 10. Details regarding the data record-
ing, processing, and inverse modeling can be obtained from Power Spectral Densities (PSDs)
the original publications of this dataset (Litvak et al., 2010,
2011). In this study, we further process this dataset by apply- We calculate the PSDs from the simulated and recorded time
ing a notch filter at 100, 150, …, 600 Hz power line noise series using the Welch algorithm. We use a segment length

13
994 Neuroinformatics (2022) 20:991–1012

of 1 s which corresponds to a frequency resolution of 1 Hz, 1.9 with increments of 0.05, yielding 17 resampling fac-
and the Hann-windowed segments overlap by 50%. Please tors hset = {1.1, 1.15, ..., 1.9}. In addition, the time series
note that other segment lengths can be used depending on is downsampled by all inverse resampling factors 1∕hi ,
the properties of the data. However, for FOOOF, the PSDs with hi ∈ hset . For each of the 17 pairs of up- und downsam-
should be sufficiently smooth to avoid fitting noise peaks. pled spectra (Fig. 1b), the geometric mean of the PSD is
IRASA receives time series as input and calculates the PSDs calculated (Fig. 1c). For illustration purposes in Fig. 1, we
internally. For IRASA, the PSD resolution should be suf- use a very small hset = {1.3, 1.6, 2}. Finally, the median is
ficiently high. We, therefore, use a segment length of 4 s calculated from all 17 geometric means, yielding the aperi-
(corresponding to a resolution of 0.25 Hz), Hann windows, odic component (Fig. 1d). The compound oscillatory part of
and 50% overlap. the spectrum is obtained by subtracting the aperiodic com-
ponent from the original PSD. After applying IRASA, the
Irregular‑Resampling Auto‑Spectral Analysis slope 𝛽 can be obtained by fitting the aperiodic component
(IRASA) in double logarithmic space in the predefined fitting range.
As parameters, IRASA requires the fitting range, the
Irregular-resampling auto-spectral analysis (IRASA) aims resampling factors hset , and the segment length for the PSD
at separating periodic components from the aperiodic part calculation. In this study, we vary the fitting range and the
of the signal (Wen & Liu, 2016). In contrast to FOOOF, hset but keep the segment length at 4 s. IRASA’s Python
the algorithm requires time series as input (Fig. 1a) and implementation used for this article was adapted from the
does not explicitly model the signals’ spectra. The input YASA toolbox (Vallat, 2019) and is published along with
time series is upsampled by a set of predefined resam- the complete code for this study on GitHub at https://g​ ithub.​
pling factors hi ∈ hset . By default, hset ranges from 1.1 to com/​moritz-​gerst​er/​oscil​lation_​and_1-​f_​separ​ation.

Fig. 1  Algorithms for 1/f estimation. IRASA: a) Simulated time series. f) FOOOF applies an initial linear fit (dashed-blue) to the PSD
series. b) PSDs of resampled time series on the y-axis and frequen- in log–log space and g) subtracts the obtained linear trend from the
cies on the x-axis. In this figure, the time series is upsampled by the spectrum. h) A Gaussian model (dotted-green) is fitted to the largest
resampling factors hi of the hset = {1.3, 1.6, 2} and downsampled by peak exceeding the thresholds (dashed-grey) and removes it. The rela-
1∕hi. c) The geometric mean of all resampling pairs (hi, 1∕hi) is cal- tive threshold is recalculated from the peak-removed flattened spec-
culated. d) The aperiodic component (orange) is the median of the trum (pink). The procedure is repeated until no peak exceeds the rela-
geometric means. A final fit (dashed-blue) estimates the y-intercept tive threshold. d) Subtraction of all Gaussian models from the original
and the 1/f exponent 𝛽 . FOOOF: e) A PSD is calculated from the time PSD yields the aperiodic component, which is then finally re-fit

13
Neuroinformatics (2022) 20:991–1012 995

Fitting‑Oscillations‑&‑One‑Over‑F (FOOOF) In this study, we keep FOOOF at the default parameters if

not stated otherwise and input PSDs with a spectral resolu-
FOOOF was introduced to parameterize neural power spectra tion of 1 Hz.
as a combination of an aperiodic component and peaks repre-
senting oscillatory processes (Donoghue et al., 2020). The
Python-based toolbox works as outlined in Fig. 1. First, the PSD
of the time-series of interest (Fig. 1a) is calculated and input into FOOOF
the algorithm, Fig. 1e. Next, FOOOF calculates an initial robust
linear fit of the spectrum in double logarithmic space, Fig. 1f, Challenge 1: The Spectral Plateau Disrupts the 1/f
and subtracts the result from the spectrum, Fig. 1g. In this flat- Power Law
tened spectrum, a relative threshold is calculated based on the
standard deviation (SD) of the spectrum, Fig. 1h. The relative The 1/f power law is sometimes called “scale-free” because
threshold is set to two times the SD, by default log-transformed power is typically approaching a linear
threshrel = 2 ⋅ SD. Optionally, FOOOF also allows setting an function across an extended range of log-transformed fre-
additional absolute threshold for the peak heights, but it is set to quencies. For electrophysiological PSDs, however, this con-
0 by default (threshabs = 0). A Gaussian function is fitted to the cept should be exercised carefully. For example, (He et al.,
largest peak of the flattened PSD exceeding both thresholds and 2010) measured different values for the slope β for frequency
then subtracted from the spectrum. Note, that this fit is not ranges 0.01–0.1 Hz and 1–100 Hz and found a small plateau
applied to negative peaks in the spectrum subceeding both in the range from 0.1 Hz to 1 Hz. In many studies, these
thresholds. Therefore, spectral dips caused by notch filtering low-frequency ranges < 0.1 Hz are eliminated by a hardware
should be avoided. This procedure is iterated for the next largest high-pass filter. However, this finding underlines the impor-
peak after subtracting the previous peak until no peaks are tance of selecting a representative frequency range to fit the
exceeding the thresholds. The oscillatory components are finally 1/f slope.
obtained by fitting a multivariate Gaussian to all extracted peaks In addition to the aforementioned low-frequency plateau,
simultaneously. After the iterations, the initial fit is added back one regularly encounters a high-frequency spectral plateau
to the flattened peak-free PSD, which results in the aperiodic (or flattening) in spectra of electrophysiological data. Such
component of the PSD, Fig. 1d. Afterward, this aperiodic com- plateaus might be due to the presence of Gaussian noise
ponent is fitted again, leading to the final fit with y-intercept and which appears as a horizontal line with a slope 𝛽 = 0 in
slope 𝛽 as parameters. The fitted Gaussian functions are param- double-logarithmic space and disrupts the 1/f power law.
eterized by center frequency fcenter(“CF”), amplitude Af (“PW”), The origin of such white noise is often due to EMG artifacts
and bandwidth 2 ⋅ 𝜎f2 (“BW”). and electronic noise of the recording system (Waterstraat
The algorithm can be used with the following parameters: et al., 2015b). It has been shown in EEG (Scheer et al., 2006;
“peak_width_limits” allows setting the minimum and maxi- Waterstraat et al., 2015a) and MEG (Waterstraat et al., 2021)
mum peak width limits of the Gaussian fits. The default is set that extremely low-noise recording devices can shift this
to 0.5 Hz and 12 Hz, respectively. “max_n_peaks'' determines high-frequency plateau into the kHz range – leaving a wider
the maximum number of peak fitting iterations. The default is unaffected frequency range for fitting the spectra. In conven-
set to infinity. “min_peak_height” is the absolute threshold and tional data, however, spectral plateaus are regularly present
corresponds to the smallest peaks that will be fitted in the units and will be discussed in this section because this can pose
of the input data. The absolute threshold is set to 0 by default. a severe challenge for estimating the aperiodic exponent: it
“peak_threshold” is the relative threshold in SD multiples that shrinks the frequency range at which the 1/f exponent may
peaks must exceed to be fitted and defaults to “peak_thresh- be examined.
old” = 2 SD. The peak fitting stops when all remaining peaks are In Fig. 2a, a simulation of an aperiodic PSD with an expo-
below either of these two thresholds. “aperiodic mode” allows nent of 𝛽 = 2 is shown. By adding white noise, a plateau
for two modes of modeling: “fixed” and “knee” which allows can be observed starting at 100 Hz in the high-frequency
modeling a bend (i.e., a “knee”) in the PSD of the aperiodic range. Here, we define the onset of the plateau as the lowest
component. The default is “fixed.” Finally, FOOOF accepts a frequency of a 50 Hz frequency interval with a vanishing
fitting range for which the algorithm performs the given steps. exponent (i.e., flattening of the spectrum). Specifically, we
For a detailed description of this algorithm, we refer the reader apply FOOOF without periodic peak fitting to measure the
to the methods section of the original publication (Donoghue slope from 1–50 Hz. We then gradually shift this interval
et al., 2020). by 1 Hz towards higher frequencies and fit the slope again.

13
996 Neuroinformatics (2022) 20:991–1012

Fig. 2  The spectral plateau disrupts the 1/f power law. The x-axis and spectrum from the subthalamic nucleus shows large oscillations that
the y-axis indicate frequency and PSD, respectively. a) Simulation of hinder the plateau onset’s precise detection. c) Adding oscillations of
an aperiodic PSD (black) with a plateau starting at 100 Hz (grey). The various powers and widths on top of different aperiodic ground truths
spectrum starts to deviate from the ground truth (dashed line) after yields the same 1/f estimation of 𝛽 ≈ 0.77 in FOOOF. The ground
around 10 Hz. Applying FOOOF yields smaller 1/f exponent esti- truths are 𝛽 = 1 (blue), 𝛽 = 1.5 (green), and 𝛽 = 2 (orange)
mates with larger upper fitting range borders. b) A Parkinsonian LFP

We repeat this procedure until the estimated slope reaches a extent alleviated if the analysis aims to study differences
value below 𝛽thresh = 0.05. between groups or experimental conditions such that rela-
We apply FOOOF in the frequency intervals 1-10 tive changes of the exponent are most important. How-
Hz, 1-50 Hz, 1-100 Hz, and 1-200 Hz which yields esti- ever, when the precise onset of the spectral plateau varies
mated 1/f exponents of 𝛽 = 1.97 , 𝛽 = 1.64 , 𝛽 = 1.17 , and across conditions, the upper fitting range border should
𝛽 = 0.70 , respectively. The spectral plateau gradually be chosen as low as possible to minimize this unequal
biases the estimated 1/f exponents towards smaller values bias. Even if the plateaus seem to be similar across con-
starting already at 10 Hz. This challenge might be to some ditions, a lower upper fitting range border will increase

13
Neuroinformatics (2022) 20:991–1012 997

the signal-to-noise ratio of the exponent estimates if we Recommendations

define the 1/f ground truth as signal and the impact of
the plateau as (possibly Gaussian) noise. We, therefore, Scenario A - the power spectra have a plateau onset at higher
recommend choosing low upper fitting range borders and frequencies, and oscillations do not mask it:
determining the precise onset of the plateau across condi- Challenge 1 does not Apply
tions in order to estimate the exponents under equivalent Scenario B - the power spectra have a plateau onset at lower
conditions. frequencies, and the onset is easily discernible (because no
Yet, measuring the onset of the plateau can be diffi- or just small oscillations are present):
cult in practice if oscillatory peaks mask it. For example, Determine the precise plateau onsets across conditions.
the spectrum in Fig. 2b appears to have a spectral pla- Choose the upper fitting range border as low as possible to
teau onset at 120 Hz. The presence of the high-frequency increase SNR.
oscillation peaking at 360 Hz, however, produces a posi- Scenario C - the power spectra potentially have a plateau
tive slope between 160 and 320 Hz, potentially masking onset at lower frequencies, but oscillations mask the exact
a continued 1/f trend of the spectrum. Accordingly, one onset:
cannot exclude the possibility that the actual onset of the The upper fitting range border must be lower than the
plateau is at a higher frequency value (e.g., 200 Hz). On onset of the masking oscillation. If the remaining frequency
the other hand, the large oscillation ranging from 6 to range is too small (as in Fig. 2b), aperiodic fitting should
100 Hz, peaking in the beta range at 25 Hz, counteracts be avoided.
this effect: In theory, one also cannot exclude that the
actual flattening occurs already at 20 Hz. Challenge 2: Avoiding Oscillations Crossing Fitting
To demonstrate this effect, for Fig. 2c we simulate Range Borders
three power spectra with three different 1/f exponent
ground truths of 𝛽 = 1 (blue), 𝛽 = 1.5 (green), and 𝛽 = 2 When choosing the fitting range to model the aperiodic
(orange). Next, we add eight oscillations at 3 Hz, 5 Hz, component, oscillations crossing the fitting range borders
10.5 Hz, 16 Hz, 23 Hz, 42 Hz, 50 Hz, and 360 Hz and must be avoided for all investigated power spectra. FOOOF
tune the oscillation amplitude and width parameters in all assumes all oscillation peaks lying within the fitting range
three examples to match the recording of Fig. 2b (purple). because it does not fit partial Gaussian peaks. Consequently,
Finally, using FOOOF, we estimate the 1/f exponent in the estimated 1/f exponent error becomes large if the lower
the frequency range from 1-95 Hz in the three simulated or upper fitting range border overlaps with a spectral peak.
and the real PSD. Despite strongly diverging 1/f expo- In the upper panel of Fig. 3a, we simulate a PSD with a
nent ground truths, FOOOF estimates an 1/f exponent slope of 𝛽 = 2 and oscillation peaks at 5 Hz, 15 Hz, and 35
of about 𝛽 ≈ 0.77 in all four cases. The diverging ground Hz (black graph). We fix the upper fitting range border at
truths are apparent in Fig. 2c because the true aperiodic 100 Hz and measure the 1/f exponent for all lower fitting
component (which is shown in light grey and is invisible ranges from 1–100 Hz up to 80–100 Hz. The lower panel in
to FOOOF) has a plateau onset at high frequencies in the Fig. 3a indicates the absolute error of the estimated slope as
blue curve, at intermediate frequencies in the green, and a function of the lower fitting range border (red). The error is
at low frequencies in the orange curve. However, neither the absolute deviation from the ground truth ||𝛽truth − 𝛽FOOOF ||.
for FOOOF nor for the experimental observer, it is pos- Note that the error function resembles the oscillatory peaks,
sible to know which of these three scenarios best reflects with the greatest errors occurring approximately at the peak
the real spectrum in Fig. 2b. Therefore it is difficult to center frequencies. The FOOOF parameters are kept at the
determine at which frequency scale an 1/f estimate might default setting.
be valid. If the peaks do not completely lie within the fitting range,
Note that this challenge applies not only to cases where very error-prone fits are obtained, as shown in another exem-
the goal is to estimate the 1/f exponent but also when plary STN-LFP recording from a Parkinsonian patient (pur-
such an estimate is used in order to remove the aperiodic ple) in Fig. 3b. The fit from 30–45 Hz (turquoise), a fre-
component from the spectrum. While fitting a “shoulder” quency range commonly used for estimation of E–I balance,
allows for modeling such a plateau, the strongly varying measures the slope of the beta-to-gamma peak, not of the
“shoulder” onsets in Fig. 2c cannot be captured, given that aperiodic component. For this fit, we set the FOOOF param-
the three power spectra share the same appearance. Spe- eter for the maximum number of allowed peaks (max_n_
cifically, the oscillation power estimates based on FOOOF peaks) to 0 since this frequency range is usually chosen to
would be almost identical in all three spectra. However, in avoid oscillations altogether. Therefore, FOOOF just fits a
the simulation, the oscillation power increases consider- straight line without peak modeling. The 40–60 Hz range
ably from the blue to the green to the orange curve. (green) lies on top of the beta-to-gamma peak, too. Here, we

13
998 Neuroinformatics (2022) 20:991–1012

Fig. 3  Oscillations must not cross fitting range borders. a) Upper max_n_peaks = 0 (for 30–45 Hz); max_n_peaks = 1 (for 40–60 Hz);
panel: PSD of a simulated spectrum with β = 2 and oscillations at peak_width_limits = (1, 100) (for 1–45 Hz and 1–95 Hz). c) The
5 Hz, 15 Hz, and 35 Hz (black). The x-axis and the y-axis indicate simulated PSD in the middle panel (green) was tuned to match the
frequency and PSD, respectively. Lower panel: The exponent 𝛽 is empirical PSD in b) (purple). FOOOF estimates a similar aperiodic
measured using FOOOF for all 80 frequency ranges from 1–100 Hz exponent for the simulated and the real spectrum (β = 0.61). When
to 80–100 Hz (red). The x-axis indicates the lower fitting range bor- decreasing the power of the 2 Hz delta oscillation (blue), the esti-
der, while the y-axis shows the absolute deviation from the ground mated aperiodic exponent decreases (β = 0.50) despite a constant
truth. b) Various frequency ranges commonly used for E–I estima- exponent for the simulated spectrum. When increasing the power
tion are applied to an STN-LFP PSD of a Parkinsonian patient (pur- of the delta oscillation (orange), the estimated aperiodic exponent
ple). Since many of the chosen ranges overlap with spectral peaks, increases (β = 0.72)
the estimated exponents 𝛽 are strongly differing. FOOOF parameters:

set (max_n_peaks) to 1 to account for the power line noise. ill-chosen fitting ranges, in practice more subtle (but similar)
Further, the 1–45 Hz range (orange) is inappropriate because errors might occur. Therefore the presence of oscillations
its upper fitting range border at 45 Hz lies in the middle at fitting range borders must be carefully checked for every
of the gamma peak. While these are obvious examples of single PSD of interest.

13
Neuroinformatics (2022) 20:991–1012 999

The 1–95 Hz range (purple dotted) seems to be the only independent of the chosen fitting range. If this assumption
acceptable range for this spectrum: The upper fitting range does not hold, however, the aperiodic component must be fit-
border extends beyond the beta-to-gamma peak but ends ted in the given frequency range of interest. If this frequency
before the onset of the spectral plateau. The estimated expo- range of interest coincides with low-frequency oscillations,
nent has a value of 𝛽FOOOF = 0.61. For these two frequency this challenge cannot be avoided.
ranges (1–45 Hz and 1–95 Hz), we increased the peak width The impact of (sub-)delta oscillations should therefore be
limits from 0.5–12 Hz (default) to 1–100 Hz to account for kept in mind as a limitation. If one finds a difference of the 1/f
the chosen spectral resolution (1 Hz) and enable modeling exponent between groups of investigation, one should check
of the very broad (> 12 Hz) beta-to-gamma peak. The cor- whether the delta power of the FOOOF-fits varies across con-
responding FOOOF fits of Fig. 3b are shown in SI Fig. 2. ditions. If delta power is similar across conditions but the
While the 1–95 Hz range seems best, it appears almost slope varies, it seems likely that indeed the aperiodic compo-
impossible to avoid low-frequency oscillations crossing the nent causes these differences in the estimated slopes and not
lower fitting range border. If some delta oscillations are a distortion by delta oscillations. If delta power does change
present, they lead to a steepening of the spectrum which across conditions (without a global offset of the PSD across
impacts the estimation of the 1/f exponent. We visualize this all frequencies), the change of slopes could either be caused
effect by reproducing the empirical LFP spectrum in three by a change of oscillatory delta activity (as shown in SI Fig. 1)
simulations in Fig. 3c. We set the oscillation frequencies to or by a change in the aperiodic component itself, and these
2 Hz, 12 Hz, 18 Hz, 27 Hz, 50 Hz (gamma), 50 Hz (power two possibilities cannot be differentiated with full certainty.
line), and 360 Hz. In the panels in Fig. 3c from left to right,
we only vary the delta power at 2 Hz while keeping the Recommendation
aperiodic component and all other oscillations’ amplitudes
and widths fixed. Since the delta oscillation has a bandwidth Scenario A - the 1/f exponent needs to be estimated:
crossing the lower fitting range border of 1 Hz, FOOOF- Use a fitting range at higher frequencies (for example
estimates of the 1/f exponent diverge strongly between the 40–60 Hz) to avoid distortion by low-frequency oscillations.\
three scenarios (same FOOOF parameters as for the 1–95 Hz newline
range in Fig. 3b). While the aperiodic (white noise-free) Scenario B - the aperiodic component needs to be removed
ground truth remains unchanged at 𝛽 = 1.5 for all three from the PSD:
simulations, FOOOF estimates an 18% lower 1/f exponent If the assumption of self-similarity across a wide range
(blue, 𝛽 = 0.50 ) if the delta oscillation from the middle of frequencies holds for the aperiodic part of the spectrum,
panel (green, 𝛽 = 0.61) is removed. On the other hand, it both slope and intercept of its linear fit could theoretically
estimates an 18% larger 1/f exponent (orange, 𝛽 = 0.72 ) if be obtained from any frequency range. In reality, different
we double the power of the delta oscillations. The power of exponents could be present in different frequency ranges. In
the true delta oscillations in the purple curve is, of course, that case, the exponent should be estimated in the broadband
unknown. range starting at very low frequencies. For this lower fitting
Overall, fitting and removing delta oscillation peaks range border (starting often at around 1 Hz), the challenge
seems unfeasible since they rarely occur as a single distin- cannot be avoided and should be kept in mind as a potential
guishable peak in the double logarithmic representation. limitation of the results.
Furthermore, FOOOF requires smooth input spectra to
reduce the impact of noise which at the same time hinders Challenge 3: FOOOF Cannot Characterize Oscillation
fitting sharp peaks. Therefore, we recommend 1/f estima- Peaks that are not Clearly Distinguishable
tion for a higher lower-border of the fitting range to avoid
the impact of these low-frequency oscillations. For high As illustrated in Fig. 1, FOOOF models oscillations as
lower borders of the fitting range, oscillations can be better Gaussian functions fitted to peaks in the flattened PSD.
avoided, and if they are present, they likely have less impact While this does not impose a severe challenge for clearly iso-
on the estimation. lated peaks, the modeling becomes complicated when peaks
Estimating the power of low-frequency oscillations by overlap partially. If many different peaks overlap, the result-
removing the aperiodic part of the spectrum poses a spe- ing PSD can be caused by various combinations of oscilla-
cial challenge in this regard. Many studies (Donoghue et al., tions with different frequencies and powers that are impos-
2020; El Boustani et al., 2009; Fransson et al., 2013; Free- sible to disentangle on a single spectrum. Furthermore,
man & Zhai, 2009; Miller et al., 2009; Wen & Liu, 2016) whereas spectral leakage from oscillations at neighboring
have conceptualized the aperiodic part of the spectrum frequencies but same Fourier phase can add up in different
as self-similar, or fractal, across a wide range of frequen- combinations to yield similar power spectra, oscillations at
cies, such that the estimation of the 1/f exponent should be different phases can also subtract power from other peaks.

13
1000 Neuroinformatics (2022) 20:991–1012

In the right panel of Fig. 4a, we present a real PSD that exponents of 𝛽pre = 1.52, 𝛽seiz = 2.31, and 𝛽post = 1.52. One
might exemplify a spectrum containing many strongly over- could interpret this finding as an increase of the aperiodic
lapping oscillation peaks. The underlying time series was 1/f exponent during the seizure, indicating (quite counterin-
recorded from a subject with epilepsy during an absence sei- tuitively) stronger neural inhibition. However, even though
zure using a bipolar montage of EEG electrodes F3-C3. Pre- FOOOF subtracts a substantial part of the harmonic peaks
seizure activity is highlighted in turquoise, seizure activity by modeling them as four broad peaks with center frequen-
in red, and post-seizure activity in yellow. Absence seizures cies at 11 Hz, 22 Hz, 37 Hz, and 50 Hz (see SI Fig. 3), it is
are proposed to be related to cortico-thalamic E–I dysbal- not clear whether it can correctly estimate the peak heights.
ance (Onat et al., 2013) caused by reduced cortical inhibi- A peak height is the power of an oscillation on top of the
tion (Tan et al., 2007), hyperexcitable somatosensory neu- aperiodic component. However, there is no reference point
rons (Karpova et al., 2005), ­GABAB receptor dysfunctions for the aperiodic component from which the height could
(Inaba et al., 2009; Merlo et al., 2007), changes in NMDA be measured in the scenario of many overlapping oscilla-
(D’Arcangelo et al., 2002; Pumain et al., 1992), or mGLU2/3 tions. Hence, it might be that the aperiodic exponent does
receptors (Ngomba et al., 2005). It would be interesting to not change during the absence seizure–instead, the inaccu-
complement these molecular rodent studies by non-invasive rately removed 3 Hz-harmonics likely caused the increased
human electrophysiological recordings. Specifically, using 1/f exponent value.
the 1/f exponent as a biomarker of E–I balance before, dur- Figure 4b shows a time series of simulated 1/f noise
ing, and after the seizure might help to gain new insights with an exponent of 𝛽sim = 1.8 . During the same time
into absence seizures. However, the non-sinusoidal 3 Hz interval in which the absence seizure in Fig. 4a occurs, we
spike-wave discharges might create many harmonic peaks add a saw-tooth oscillation of 3 Hz to the signal. As in the
throughout the spectrum. Applying FOOOF (default param- example of the real seizure in Fig. 4b, FOOOF estimates
eters) in a frequency range of 1–100 Hz yields estimated 1/f a strongly increased 1/f exponent even though the ground

Fig. 4  FOOOF cannot characterize oscillation peaks that are not FOOOF fits. Note the increase of the 1/f exponent during the seizure.
clearly distinguishable. a) Left: Time series of an absence seizure b) Left: Simulated 1/f noise and temporarily (red) added 3 Hz saw-
measured using EEG. Turquoise: Pre-seizure, red: seizure, yellow: tooth signal. Right: Aperiodic FOOOF fits. Note the increase of the
post-seizure activity. Right: Corresponding PSDs and aperiodic 1/f exponent despite constant ground truth of 𝛽truth = 1.8

13
Neuroinformatics (2022) 20:991–1012 1001

truth exponent remains constant. The corresponding model For example, if we chose only two resampling fac-
(default parameters) is shown in SI Fig. 3. tors hset = {2, 3}, the spectrum would be up-sampled by
Note that it is possible to enable FOOOF fitting of the hup1 = 2 and hup2 = 3 and down-sampled by hdown1 = 1∕2
many harmonious peaks by reducing the maximum peak and hdown2 = 1∕3. As a result, a fitting range of 10–100 Hz
width limits to 1 Hz. While it is not feasible to tune the would correspond to four evaluated frequency ranges of
parameters across conditions (there is an alpha peak with a 20–200 Hz, 30–300 Hz, 5–50 Hz, and 3.3–33 Hz. Of these
peak width larger than 1 Hz in the pre-and post-condition), four resampled spectra, IRASA takes the median. The lower
even with the specifically tuned parameters, FOOOF returns border of the evaluated frequency range feval. min can be cal-
increased 1/f exponents 𝛽 = 2.28 and 𝛽sim (tuned) = 1.91 for culated from the minimum fitting range border ffit min divided
real and simulated data, respectively SI Fig. 3. by the maximum resampling factor hmax according to Eq. (1).
The upper evaluated frequency border feval. max corresponds
Recommendation to the upper fitting range border ffit max multiplied by hmax
according to Eq. (2).
Scenario A - the PSD appears as a straight line with well-
distinguishable peaks on top of this line:
feval. min = ffit min ∕hmax (1)
Challenge 3 does not apply.
Scenario B - the PSD might contain overlapping peaks: feval. max = ffit max ⋅ hmax (2)
The more peaks overlap, the less accurate the model
While evaluating a larger frequency range than the actual
results will be. The lower and upper fitting range borders
fitting range can be advantageous, as shown in Fig. 5a, it can
must vastly extend the overlapping oscillation peaks (chal-
also lead to severe challenges, as shown in Fig. 5b. Here, we
lenge 2) to enable peak removal. Estimating the power of
simulate an aperiodic PSD with 𝛽 = 2 , which is highpass
overlapping peaks will be difficult.
filtered at 1 Hz. We then apply IRASA in a fitting range of
Scenario C - almost the full PSD seems to consist of over-
2–30 Hz for three different h-sets with maximum resampling
lapping peaks (as in Fig. 4):
factors hmax = 2, hmax = 8, and hmax = 15, respectively. The
Avoid fitting the aperiodic component.
fitting ranges, indicated in green, orange, and red, are the
same, but the evaluated frequency ranges, shown in the cor-
responding transparent colors, increase with increasing hmax.
IRASA
Note that the highpass filter disrupts the 1/f power law
for low frequencies and violates IRASA’s assumption of a
Challenge 1: The Evaluated Frequency Range
resampling-invariant aperiodic component. With increasing
is Larger than the Fitting Range
hmax , IRASA evaluates substantially larger parts of the low-
frequency stopband which increasingly biases its 1/f esti-
While FOOOF tries to iteratively fit all oscillatory peaks to
mates towards smaller values. A good agreement with the
obtain a periodic model, IRASA takes the median of spec-
ground truth of 𝛽 = 2 is only obtained for hmax = 2, which
tra after up- and down-sampling to eliminate the peaks, as
corresponds to an evaluated frequency range of 1–60 Hz,
shown in Fig. 1. As a result, it aims to obtain the pure ape-
avoiding the stopband of the highpass-filtered spectrum.
riodic component that is assumed to be invariant to resam-
Apart from low-frequency fitting artifacts due to highpass
pling. As an advantage over FOOOF, IRASA can overcome
filtering, care must also be taken to avoid fitting artifacts at
challenge 2: Even if a peak crosses the fitting range border
high frequencies. For example, in Fig. 5c, the high-frequency
(at the original sampling rate), it can be removed success-
spectral plateau disrupts the 1/f power law. Even though the
fully due to the resampling procedure.
upper fitting range border of IRASA is set well below the
In Fig. 5a, we replot the spectrum of Fig. 3a and estimate
plateau onset to ffit max = 30 Hz, IRASA does nevertheless
the 1/f exponent for all frequency ranges between 1–100 Hz
evaluate the plateau due to the upsampling step. Therefore,
and 80–100 Hz. In contrast to FOOOF, IRASA has minimal
with growing hmax , IRASA biases the 1/f estimates towards
errors for all frequency ranges. The reason is that the fit-
smaller values again.
ting range of FOOOF has well-defined borders: If the lower
Even in the absence of a spectral plateau, care must
border is set to 5 Hz (the center frequency of the first peak),
be taken to avoid the resampled Nyquist frequency. For
it cannot identify and model the 5 Hz peak correctly. On
example, for a sampling rate of fsample = 2400 Hz and
the other hand, for IRASA, the fitting range is blurry: By
hmax = 10 , the resampled Nyquist frequency reduces from
up- and down-sampling the spectrum, the peaks are shifted
fNyquist = 1200 Hz to fNyquist resampled = 120 Hz. Accord-
towards lower and higher frequencies. Therefore, the evalu-
ingly, the upper fitting range border must not exceed this
ated frequency range of IRASA is much more extensive than
value. The same holds true for a potentially applied lowpass
the actual fitting range.

13
1002 Neuroinformatics (2022) 20:991–1012

Fig. 5  IRASA’s evaluated frequency range is larger than the fitting ing transparent colors. IRASA’s error of the 1/f estimation increases
range. a) Upper panel: Same simulation as in Fig. 3a. Lower panel: with larger resampling rates hmax (and lower resampling rates 1/hmax,
The lower fitting range border is shown on the x-axis, the absolute respectively). c) Same as b) with a spectral plateau disrupting the 1/f
deviation from the ground truth on the y-axis. IRASA correctly power law. d) FOOOF 1/f estimate within 1–30 Hz for a spectrum
estimates the 1/f exponent for all used fitting ranges. b) Simulated obtained from voxel data after MEG source reconstruction. e) IRASA
aperiodic PSD with a ground truth of 𝛽 = 2. A 1 Hz highpass filter 1/f estimates for an evaluated frequency range of 1–30 Hz (green)
disrupts the 1/f power law. IRASA’s fitting range for the maximum and an evaluated frequency range of 0.3–90 Hz (green-dashed, corre-
resampling factor hmax ∈ {2, 8, 15} is indicated as bright-colored sponding to a fitting range of 1–30 Hz at h­ max = 3). f) FOOOF (blue)
lines upon the fitted aperiodic components, with the evaluated fre- and IRASA (green) estimates of the 1/f exponent for the same fitting
quency ranges after up- and down-sampling indicated in correspond- range of 1–30 Hz

filter. In general, to avoid accidentally fitting spectra above algorithm. However, since only the minimum and maximum
Nyquist frequency or in the stopbands of lowpass or high- resampled spectra contain the 1 Hz and 30 Hz borders of
pass filters, it is advisable to choose hmax as small as pos- interest, IRASA emphasizes the estimation of the 1/f expo-
sible. Furthermore, the evaluated frequency range should nent from intermediate frequency values above 1 Hz and
always be checked by calculation from the fitting range and below 30 Hz. Therefore, 1/f exponents estimated by IRASA
hmax. cannot be directly compared to 1/f exponents estimated by
Given that IRASA evaluates a more extensive frequency FOOOF.
range than the fitting range, the meaning of the fitting range We visualize this effect for a spectrum of voxel data
becomes imprecise. For example, if we are interested in fit- obtained by MEG source reconstruction in the lower pan-
ting the 1/f exponent from 1–30 Hz and use hmax = 2 , we els d) – f) of Fig. 5. In d), FOOOF estimates an 1/f expo-
should choose 2–15 Hz as a fitting range for the IRASA nent of 𝛽FOOOF = 1.41 in the fitting range of 1–30 Hz. Due

13
Neuroinformatics (2022) 20:991–1012 1003

to the highpass filter, IRASA obtains a lower value of the range of resampling factors is too small or the peaks
𝛽IRASA = 1.09 for the same fitting range which, however, too broad, the resampled peaks overlap. In that case, peak
actually corresponds to an evaluated frequency range of removal by taking the median will not be successful.
0.33–90 Hz at hmax = 3 (Fig. 5e). Hence, setting the evalu- In Fig. 6a, we replot Fig. 5a. However, by increasing the
ated frequency range to 1–30 Hz (by setting the fitting range peak widths from the left to the right panels, the 1/f estima-
to 3–10 Hz) yields 𝛽IRASA = 1.38 which is similar to the tion error of IRASA increases strongly. This is because the
FOOOF estimate. peaks cannot be fully separated. As a result, IRASA’s calcu-
Matching the evaluated frequency range of IRASA to lated aperiodic component, shown in grey, still contains the
the fitting range of FOOOF is not always possible, though. up and downsampled peaks after taking the median. Note
Consider, for example, the fitting range of 30–45 Hz shown that not the peak width Δf itself must be sufficiently small
in Fig. 5e. At hmax = 3, the evaluated frequency range of to get separated, but instead, Δflog , as it appears in the loga-
IRASA is 10–135 Hz. Due to the spectral plateau, IRASA rithmic frequency scale, the logarithmic peak width needs to
estimates a much smaller exponent of 𝛽IRASA = 1.22 com- be sufficiently small. For this reason, a peak width of 4 Hz
pared to 𝛽FOOOF = 2.11. This time, we cannot shrink IRA- at a center frequency of 5 Hz has a similar effect as a peak
SA’s fitting range to match its evaluated frequency range width of 12 Hz at a center frequency of 35 Hz.
with FOOOF’s fitting range. At hmax = 3, the lower fitting We visualize this effect in panel Fig. 6b by simulating a
range border of IRASA must be 3 ⋅ 30 Hz = 90 Hz to match PSD with two oscillations at f1 = 30 Hz and f2 = 300 Hz.
the lower fitting range border of FOOOF at 30 Hz. However, The peak width of the second peak is 70 Hz and therefore
the upper fitting range border needs to be 45 Hz ∕3 = 15 10 times as large as the peak width of the first peak. How-
Hz to match the upper fitting range border of FOOOF. This ever, on the logarithmic frequency axis, they appear with
would lead to an inverse fitting range of 90–15 Hz. Here, it the same width. A maximum resampling factor of hmax = 2
cannot be avoided that IRASA evaluates a much more exten- is sufficient to remove the peaks correctly. Thus, they are
sive frequency range than 30–45 Hz. As a result, FOOOF fully eliminated from the aperiodic component shown in
and IRASA cannot yield comparable 1/f estimates for this turquoise. However, when the peak widths are increased to
frequency range. a logarithmic value of 0.2 log(Hz), hmax = 2 is not sufficient
anymore: The up- and down-sampled peaks remain visible
in the estimate of the aperiodic component. If we increase
Recommendations
hmax to a value of 8, however, peak removal works well.
For a further increase of the logarithmic peak width to 0.3
Always calculate the evaluated frequency range from the
log(Hz), however, hmax = 35 is necessary. We visualize this
fitting range and hmax according to Eqs. (1) and (2). Choose
challenge on empirical data of MEG and LFP data of dataset
the maximum resampling factor hmax as small as possible in
3 in SI Fig. 4.
order to 1) avoid fitting artifacts, 2) to improve comparability
We calculated the logarithmic peak width as Eq. (1)
with other methods, and 3) to improve the interpretability of
Δflog = log10 (f 2 ∕f1 ) where f1 corresponds to the lower
the investigated frequency range.
bound of the peak and f2 to the upper bound of the peak. The
Set the evaluated frequency range–and not the fitting
bounds were found by calculating the first bin of the PSD,
range–to the frequency range of interest.
which deviates above a threshold of 0.001 from the aperiodic
ground truth. Note that there is no exact equation/heuristic
Challenge 2: Broad Peak Widths Require Large to calculate the minimum hmax as a function of peak width
Resampling Factors because always many resampling factors h are calculated,
which will lead to a gradual peak removal depending on the
In challenge 1, we recommend choosing the maximum degree of peak separation.
resampling factor hmax as small as possible. However, for
IRASA to work correctly, the resampling factors must be Recommendations
sufficiently large. This is because IRASA shifts the peaks
in the frequency scale up and down through up- and down- Choose hmax as small as possible (challenge 1) while keeping
sampling. Therefore, a single peak appears multiple times it large enough to obtain peak-free estimates of the aperiodic
on the frequency scale (Fig. 1b). For a range of sufficiently component (challenge 2). If the peaks are very broad and
large (and small) resampling factors, the resampled peaks hmax cannot be chosen sufficiently large without avoiding
are completely separated and, by taking the median of their challenge 2, IRASA cannot be applied.
geometric mean, subsequently eliminated. However, if

13
1004 Neuroinformatics (2022) 20:991–1012

Fig. 6  Broad peak widths require large resampling factors. a) Upper error increases with larger peak widths. b) Simulation of a 30 Hz and
panel: Similar as in Fig. 5a) but with increasing peak widths from 300 Hz peak with increasing peak widths from left to right. Larger
left to right. Note that removal of peaks from the aperiodic compo- peak widths require larger resampling factors. Note that not the abso-
nent (grey) worsens with broader peak widths. Lower panel: The lute peak width but rather the logarithmic peak width Δflog deter-
lower fitting range border is on the x-axis, the absolute deviation mines the minimum resampling factors
from the ground truth on the y-axis. The 1/f exponent estimation

Challenge 3: IRASA Cannot Characterize Oscillation If, however, we now add two strongly overlapping oscil-
Peaks that are not Clearly Distinguishable lations at 10 Hz and 25 Hz, IRASA is no longer capable of
successfully removing the peaks. As a result, it now esti-
Similar to FOOOF, IRASA cannot separate strongly overlap- mates an exponent of 𝛽seiz = 3.05 – much larger than the
ping peaks. However, as shown in Fig. 7b, IRASA performs ground truth at 𝛽truth = 1.8.
quite well for dataset 2 because the harmonic peaks do not
strongly overlap above 10 Hz. Instead, many local power Discussion
minima in between the harmonic peaks are very close to
the power of the aperiodic ground truth. As a consequence, Both periodic and aperiodic components of power spectra
adding the 3 Hz sawtooth signal only slightly increases the are frequent targets of investigation in electrophysiologi-
estimated 1/f exponent from 𝛽pre/post = 2.24 to 𝛽seiz = 2.46. cal studies. The separation of both components before
In the middle panel of SI Fig. 5b, the extracted oscillatory analysis helps to disentangle their relative contribution
component of IRASA is shown in orange, indicating a good to the spectrum. FOOOF and IRASA are commonly used
extraction of harmonic peaks at multiple integers of 3 Hz. for this purpose. It should be highlighted, though, that

13
Neuroinformatics (2022) 20:991–1012 1005

Fig. 7  IRASA cannot characterize oscillation peaks that are not c) IRASA’s performance on the simulation drops significantly if
clearly distinguishable. a) and b) left panel: Same as Fig. 4 a) and two strongly overlapping peaks in the alpha (10 Hz) and beta range
b) Right panel: Same as Fig. 4 but showing the 1/f fits by IRASA. (25 Hz) are added. Ground truth: 𝛽truth = 1.8

the methods follow different concepts: Whereas FOOOF PSDs such as 1) spectral plateau onsets at relevant fre-
models periodic components and a single aperiodic quencies, 2) hidden low-frequency oscillations, and 3)
component and outputs the corresponding parameters, overlapping peaks.
IRASA only separates them, allowing further independ- Here, we evaluated common challenges of the separation
ent processing. Other methods to separate periodic and procedure based on two popular methods and summarized both
aperiodic PSD components exist too, for example, eBOSC general and method-specific challenges. These challenges apply
(Kosciessa et al., 2020). However, they cannot overcome to EEG, MEG, and LFP data obtained by independent research
the method-unspecific challenges in electrophysiological groups, indicating the general applicability of the results.

13
1006 Neuroinformatics (2022) 20:991–1012

Neurophysiological Interpretation et al., 2019), 1–10 Hz (Schaworonkow & Voytek, 2021),

1–20 Hz (Bédard et al., 2006), 1–30 Hz (Wen & Liu,
Spectral Plateau 2016), 1–40 Hz (Colombo et al., 2019), 1–20 and 20–40 Hz
(Colombo et al., 2019), 1-15 and15-125 Hz (Chaoul &
The spectral plateau can hinder a correct separation of the PSD Siegel, 2021), 1–100 Hz (He et al., 2010), 2–24 Hz (Voytek
components. If the fitting range of FOOOF or the evaluated et al., 2015), 3–30 Hz (Pereda et al., 1998), 3–55 Hz
frequency range of IRASA includes a spectral plateau, the 1/f (Waschke et al., 2021), 10–100 Hz (Freeman & Zhai, 2009),
exponent will be estimated too low. In the presence of periodic 20–65 Hz (Bédard et al., 2006), 30–50 Hz (Gao et al., 2017;
components at the spectral flattening, a faulty aperiodic power Lendner et al., 2020; Stolk et al., 2019) and 40–60 Hz (Gao
estimation will lead to a faulty periodic power estimation. In et al., 2017).
addition, large periodic components could hide the onset of the If a study using FOOOF aims to generally estimate an
spectral plateau. This hinders a proper decision on the choice of 1/f exponent and is free to choose any fitting range for that
the upper fitting range border. If the 1/f exponent is estimated too purpose, we generally recommend avoiding low lower fit-
low due to the spectral flattening, this could be misinterpreted as ting range borders. It is unknown how hidden low-frequency
an increased E–I ratio since typically flatter spectra are associ- oscillations at the lower fitting range border might impact
ated with more pronounced excitability (Gao et al., 2017). the 1/f estimate. For example, if the 1/f exponent is com-
E–I balance estimation is usually applied to estimate rela- pared between two conditions and in one condition there
tive 1/f differences between conditions. If the spectral pla- are larger delta oscillations, a fitting range starting from
teau onset were to occur at exactly the same frequency for 1 Hz could have a larger y-intercept due to the presence
all spectra, a relative 1/f comparison would still be viable. A of low-frequency oscillations. This could lead to a larger
random fluctuation of the onset would introduce noise to the 1/f exponent and could be misinterpreted as stronger neural
estimates, and a systematic difference of the plateau onset inhibition. The same holds for any other lower fitting range
between conditions would lead to type 1 errors. border. Therefore, the lower border should be chosen to best
The origin of the spectral plateau at high frequency is avoid known oscillation frequencies, depending on the study.
likely rooted in Gaussian properties of amplifier noise and The same problem applies to IRASA but less severely
impedance noise (Scheer et al., 2006; Waterstraat et al., since this method is not based on a single frequency range
2015a, b). (Waterstraat et al., 2021) showed that the onset of but rather, due to up- and downsampling, to a set of differ-
the plateau starts at higher frequencies if recordings are done ent frequency ranges. This comes at the cost of only vaguely
with a low-noise MEG system. In addition to system noise, defined upper and lower fitting range borders, hindering an
biological high-frequency noise caused by electromyography easy comparison with fitting ranges used in other studies.
(EMG) from the head muscles can contribute to spectral In general, there is no one-range-fits-all fitting range
flattening, although EMG does not necessarily have a flat applicable to all kinds of PSDs. Therefore, we recommend
spectrum (Muthukumaraswamy, 2013). However, even when examining the PSDs of interest carefully and choosing the
recording LFPs from the subthalamic nucleus using a low- fitting range that best avoids the challenges discussed so far
noise amplifier, which can be considered as hardly affected in addition to further possible data-specific or goal-specific
by EMG activity, a spectral plateau could be observed one challenges.
order of magnitude above the system's noise level (unpub- Finally, if the purpose of the 1/f estimation is not to obtain
lished data). Neuronal population spiking activity probably the 1/f exponent but rather the removal of the aperiodic com-
contributes to this spectral plateau (Belluscio et al., 2012; ponent for better periodic power assessment, a broadband
Buzsáki et al., 2012; Zanos et al., 2011). Better understand- range (such as 1–100 Hz) should be chosen.
ing the origins of the spectral plateau is of major interest
and requires further research. If it is caused by noise such as Overlapping Peaks
systems noise or EMG, an identification of the origin could
help to clean the data from this high-frequency plateau. If it The stronger periodic components overlap, the more difficult
has a neurophysiological origin, a thorough analysis of the estimating their power becomes. In the shown exemplary
plateau might yield novel neurophysiological insights. data, overlapping peaks occurred mainly in STN data and
in dataset 2. MEG and EEG cortical data of healthy par-
Fitting Ranges ticipants typically have peaks in the ranges 8–13 Hz and
18–25 Hz which does not impose considerable challenges
The choice of the fitting range depends on the goal of the for the estimation of the aperiodic part of the spectrum.
study and the properties of the data. In the literature, the Assessing the 1/f exponent is still feasible if the overlap-
1/f exponent was investigated for different frequency ranges ping periodic components make up only a minor part of the
such as 0.01–0.1 Hz (He et al., 2010), 0.5–35 Hz (Miskovic fitting range. On the other hand, if the overlapping peaks

13
Neuroinformatics (2022) 20:991–1012 1007

make up a majority of the frequency range to investigate, a maximum logarithmic peak width as IRASA is, in theory,
as in Figs. 4, 7, and 8b, a separation of the periodic and able to fit any peaks if the h-values are chosen sufficiently
aperiodic components is not recommended and will likely large. However, in practice, the h-values must not exceed a
lead to imprecise results. In the case of the absence seizure certain range to avoid too low (highpass, Fig. 5b) or too high
shown in Figs. 4 and 7, neural inhibition during the seizure (spectral plateau, Fig. 5c) fitting ranges, calculated using
is likely overestimated due to overlapping peaks leading to Eq. (1) and (2).
a false 1/f estimation.
Estimating E–I Balance
Broad Peak Widths
(Gao et al., 2017) proposed to use the 1/f exponent as an
In contrast to FOOOF, IRASA cannot handle very broad indicator of E–I balance. Subsequent studies indicated the
peaks well. This limitation is especially severe for the analy- usefulness of this idea also for non-invasive EEG/MEG
sis of LFPs of Parkinsonian patients (datasets 1 and 3). In data (Colombo et al., 2019; Gao et al., 2017; Lendner
the original article (Wen & Liu, 2016), IRASA was only et al., 2020; Miskovic et al., 2019; Waschke et al., 2021).
evaluated on pure sine oscillations, for which the method While we outlined that 1/f exponent estimation is affected
works very well. We, therefore, do not recommend using by many possible error sources, we do not argue that it
IRASA if the peaks seem to have broad logarithmic peak should be avoided altogether. While proper 1/f estimation
widths Δflog . It is not possible to give a threshold value for seems to be beyond reach for some PSDs (for example,

Fig. 8  “Easy” and “hard” PSDs. a) Left: Voxel MEG PSD of a Par- “easy” because it avoids all the discussed challenges. b) LFP data of
kinsonian patient on a semilogarithmic scale. Right: Same PSD on a Parkinsonian patient on a semilogarithmic scale. Right: Same PSD
a double logarithmic scale. FOOOF, IRASA, and simply connecting on a double logarithmic scale. FOOOF, IRASA, and “straight” yield
the PSD value at 1 Hz to the PSD value at 95 Hz as a straight line different 1/f exponents. We regard such a spectrum as “hard” because
(“straight”) yield similar 1/f exponents. We regard such a PSD as it contains many challenges

13
1008 Neuroinformatics (2022) 20:991–1012

the one shown in Fig. 8b), it seems to be a promising and E–I balance remains a hypothesis to be further
measure for others (Fig. 8a). Thus we suggest that exist- validated.
ing methods could be enhanced by more elaborate data
cleaning, such as spatio-spectral decomposition (SSD) Computational Cost and Parameter Tuning
(Nikulin et al., 2011), independent component analysis
(ICA), or inverse modeling. Moreover, it might be possi- From a computational perspective, FOOOF is much faster
ble to develop new methods that measure the 1/f exponent than IRASA. When applied to 9 time series of dataset 1 (ca.
more reliably than the ones discussed in the present study. 180 s at fsample = 2400 Hz corresponding to ≈ 9 × 440, 000
For example, if the periodic and aperiodic components data points), parameterization with FOOOF was about 50
are assumed to vary over time independently, it could times faster than separation with IRASA when the PSD cal-
be possible to disentangle them using machine learning culation time was included. FOOOF was 100 times faster
algorithms such as non-negative matrix factorization (Lee if the PSDs were precalculated. For the comparison, we
& Seung, 1999). And it might become possible to measure used 7 runs and fitted a frequency range from 1–30 Hz.
E–I balance through other electrophysiological measures For FOOOF, the default parameters were chosen, and for
thus further validating the 1/f exponent of the PSD. We IRASA a window length of 4 s and a set of 17 resampling
elaborate on this below. factors hset = {1.1, 1.15, ..., 1.9} was used. FOOOF computa-
(Bruining et al., 2020), for example, proposed to meas- tion slows down when the PSDs have a very high resolution
ure E–I balance based on the alpha band amplitude enve- leading to many iterations of fitting noise peaks. IRASA
lope and its detrended fluctuation analysis (DFA) exponent computation slows down when the number of resampling
(Peng et al., 1995). (Stephani et al., 2020) related the N20 factors is increased and when their values are increased.
of somatosensory evoked potentials to cortical excitabil- Increasing IRASA’s resampling values can help with very
ity. Other researchers related spontaneous fluctuations of broad peak widths (challenge 2) but simultaneously enlarges
alpha-band power to E–I balance (Romei et al., 2008) and the evaluated frequency range (challenge 1). Increasing the
(Iemi et al., 2019) found alpha- and beta-band power to number of resampling factors beyond 17 or changing the
predict suppression of ERP-components, which was inter- window length does not help with the challenges presented
preted as increased inhibition. This relationship held true in this article. FOOOF requires extensive parameter tun-
even after controlling for fluctuations in the 1/f exponent, ing for optimum results, but the posed challenges cannot be
which correlates with alpha power (Muthukumaraswamy resolved by parameter selection. In general, the fitting range
& Liley, 2018). It might be possible to estimate E–I bal- of FOOOF and the evaluated frequency range of IRASA are
ance by measuring transcranial magnetic stimulation the most critical parameters for each method.
(TMS) evoked potentials using EEG. By combining these
two methods, (Massimini et al., 2005) showed a break- Conclusion
down of effective cortical connectivity during non-rapid
eye movement (REM) sleep. Effective connectivity was To study either periodic or aperiodic PSD components, it is
also related to the 1/f exponent by (El Boustani et al., useful to disentangle both components. As there are theoreti-
2009). The perturbational complexity index (Casali et al., cally infinite solutions to this inverse problem, it is probably
2013) follows these lines to separate unconscious states neither possible to perfectly separate them nor to evaluate
of low excitability (non-REM sleep, anesthesia) from and verify a performed separation since the ground truth
conscious states of high excitability (wakefulness, REM remains unknown. Some PSDs seem to be particularly easy
sleep). Indeed, (Colombo et al., 2019) could link this to separate because they avoid most of the discussed chal-
index to the 1/f exponent during wakefulness and anesthe- lenges. For those PSDs, we generally recommend perform-
sia yielding similar results with both methods. However, ing a separation to study the periodic or aperiodic compo-
it should be noted that REM sleep (a conscious state of nents in a more isolated manner. We give an example of such
mind) is associated with a larger 1/f exponent compared to an “easy” PSD in Fig. 8a.
NREM sleep (unconscious) while NREM sleep is associ- These “easy” PSDs appear as an almost straight line in
ated with a larger 1/f exponent compared to wakefulness double logarithmic space with some well-distinguishable,
(conscious) (Lendner et al., 2020). These findings agree narrow periodic peaks on top of it. There is no spectral
with in vivo calcium imaging measurements of E–I bal- plateau disrupting the 1/f power law and the y-axis of the
ance in mice during wakefulness, NREM sleep, and REM PSD extends over 4 orders of magnitude. When applying
sleep (Niethard et al., 2016). FOOOF and IRASA from 1–95 Hz or simply connecting
In the best scenario, different methods used for E–I esti- values at 1 Hz and 95 Hz to a straight line in double loga-
mation will lead to similar results and might be used in rithmic space, similar values (𝛽FOOOF = 1.65, 𝛽IRASA = 1.71,
conjunction. So far, the relationship between 1/f exponent 𝛽straight =1.53) are obtained for the 1/f exponent.

13
Neuroinformatics (2022) 20:991–1012 1009

Other PSDs seem to be very difficult to separate. For Open Access This article is licensed under a Creative Commons Attri-
such, we recommend avoiding the separation since the bution 4.0 International License, which permits use, sharing, adapta-
tion, distribution and reproduction in any medium or format, as long
results will be arbitrary and might lead to ill-informed as you give appropriate credit to the original author(s) and the source,
interpretations. An example of such a “hard” PSD is shown provide a link to the Creative Commons licence, and indicate if changes
in Fig. 8b. These spectra do not appear as a straight line. were made. The images or other third party material in this article are
They have very broad and overlapping peaks and a spectral included in the article's Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in
plateau onset at lower frequencies. As a result of this pla- the article's Creative Commons licence and your intended use is not
teau, the y-axis spans only one order of magnitude. When permitted by statutory regulation or exceeds the permitted use, you will
applying FOOOF, IRASA, or a straight-line connection need to obtain permission directly from the copyright holder. To view a
between 1 and 95 Hz, strongly diverging 1/f exponent copy of this licence, visit http://​creat​iveco​mmons.​org/​licen​ses/​by/4.​0/.
values ( 𝛽FOOOF = 0.82 , 𝛽IRASA = 1.10 , 𝛽straight = 0.62 ) are
obtained. References
Checking PSDs for the challenges discussed in this work
will help to decide whether a technique to separate neural Bédard, C., Kröger, H., & Destexhe, A. (2006). Does the 1/f frequency
oscillations from aperiodic 1/f activity should be applied, scaling of brain signals reflect self-organized critical states?.
Physical Review Letters, 97(11), 118102. https://d​ oi.​org/​10.1​ 103/​
which algorithm to use, and which parameters to choose. PhysR​evLett.​97.​118102
Belluscio, M. A., Mizuseki, K., Schmidt, R., Kempter, R., & Buzsáki,
Information Sharing Statement G. (2012). Cross-frequency phase-phase coupling between θ and
γ oscillations in the hippocampus. The Journal of Neuroscience:
The Official Journal of the Society for Neuroscience, 32(2), 423–
The datasets analyzed in the present study are not publicly 435. https://​doi.​org/​10.​1523/​JNEUR​OSCI.​4122-​11.​2012
available due to privacy regulations of patient health infor- Bódizs, R., Szalárdy, O., Horváth, C., Ujma, P. P., Gombos, F., Simor, P.,
mation but are available from the corresponding author upon Pótári, A., Zeising, M., Steiger, A., & Dresler, M. (2021). A set of
reasonable request. The code of the entire study is available composite, non-redundant EEG measures of NREM sleep based on the
power law scaling of the Fourier spectrum. Scientific Reports, 11(1),
at https://​github.​com/​moritz-​gerst​er/​separ​ating_​perio​dic_​ 2041. https://​doi.​org/​10.​1038/​s41598-​021-​81230-7
from_​aperi​odic_​PSDs. Bruining, H., Hardstone, R., Juarez-Martinez, E. L., Sprengers, J.,
Avramiea, A.-E., Simpraga, S., Houtman, S. J., Poil, S.-S., Dallares,
Supplementary Information The online version contains supplemen- E., Palva, S., Oranje, B., Matias Palva, J., Mansvelder, H. D.,
tary material available at https://d​ oi.o​ rg/1​ 0.1​ 007/s​ 12021-0​ 22-0​ 9581-8. & Linkenkaer-Hansen, K. (2020). Measurement of excitation-
inhibition ratio in autism spectrum disorder using critical brain
Acknowledgements This work was supported by Deutsche Forschungsge- dynamics. Scientific Reports, 10(1), 9195. https://d​ oi.o​ rg/1​ 0.1​ 038/​
meinschaft (German Research Foundation) Project ID 424778381 TRR s41598-​020-​65500-4
295. The Wellcome Centre for Human Neuroimaging is supported by core Buzsáki, G., Anastassiou, C. A., & Koch, C. (2012). The origin of
funding from Wellcome [203147/Z/16/Z]. VL is grateful to the National extracellular fields and currents–EEG, ECoG LFP and Spikes.
Hospital for Neurology and Neurosurgery clinical team for their help with Nature Reviews Neuroscience, 13(6), 407–420. https://​doi.​org/​
data collection. 10.​1038/​nrn32​41
Buzsáki, G., & Draguhn, A. (2004). Neuronal oscillations in cortical
networks. Science, 304(5679), 1926–1929. https://​doi.​org/​10.​1126/​
Author’s Contributions Conceptualization: MG, GW, VN, GC; Meth-
scien​ce.​10997​45
odology: MG, GW, VN, GC; Data curation: VL, KL, EF, AS; Formal
Casali, A. G., Gosseries, O., Rosanova, M., Boly, M., Sarasso, S.,
analysis and investigation: MG; Visualization: MG; Writing–original
Casali, K. R., Casarotto, S., Bruno, M. -A., Laureys, S., Tononi,
draft preparation: MG; Writing–review and editing: MG, GW, VL,
G., & Massimini, M. (2013). A theoretically based index of con-
KL, AS, EF, GC, VN; Funding acquisition: VN, GC; Resources: Not
sciousness independent of sensory processing and behavior. Sci-
applicable; Supervision: VN, GW, GC.
ence Translational Medicine, 5(198), 198ra105. https://​doi.​org/​
10.​1126/​scitr​anslm​ed.​30062​94
Funding Open Access funding enabled and organized by Projekt Cellier, D., Riddle, J., Petersen, I., & Hwang, K. (2021). The development
DEAL. of theta and alpha neural oscillations from ages 3 to 24 years. Develop-
mental Cognitive Neuroscience, 50, 100969. https://​doi.​org/​10.​1016/j.​
Declarations dcn.​2021.​100969
Chaoul, A. I., & Siegel, M. (2021). Cortical correlation structure of
aperiodic neuronal population activity. NeuroImage, 118672.
Ethics Approval Please refer to the methods section.
https://​doi.​org/​10.​1016/j.​neuro​image.​2021.​118672
Colombo, M. A., Napolitani, M., Boly, M., Gosseries, O., Casarotto, S.,
Consent to Participate Please refer to the methods section.
Rosanova, M., Brichant, J.-F., Boveroux, P., Rex, S., Laureys, S.,
Massimini, M., Chieregato, A., & Sarasso, S. (2019). The spectral
Consent for Publication Please refer to the methods section.
exponent of the resting EEG indexes the presence of conscious-
ness during unresponsiveness induced by propofol, xenon, and
Conflict of Interest The authors have no conflicts of interest to declare ketamine. NeuroImage, 189, 631–644. https://​doi.​org/​10.​1016/j.​
that are relevant to the content of this article. neuro​image.​2019.​01.​024

13
1010 Neuroinformatics (2022) 20:991–1012

Dave, S., Brothers, T. A., & Swaab, T. Y. (2018). 1/f neural noise and link prestimulus neural oscillations to sensory responses. eLife,
electrophysiological indices of contextual prediction in aging. 8. https://​doi.​org/​10.​7554/​eLife.​43620
Brain Research, 1691, 34–43. https://​doi.​org/​10.​1016/j.​brain​res.​ Inaba, Y., D’Antuono, M., Bertazzoni, G., Biagini, G., & Avoli, M.
2018.​04.​007 (2009). Diminished presynaptic GABA(B) receptor function in the
D’Arcangelo, G., D’Antuono, M., Biagini, G., Warren, R., Tancredi, V., neocortex of a genetic model of absence epilepsy. Neuro-Signals,
& Avoli, M. (2002). Thalamocortical oscillations in a genetic model 17(2), 121–131. https://​doi.​org/​10.​1159/​00019​7864
of absence seizures. The European Journal of Neuroscience, 16(12), Karalunas, Sarah L., Brendan D. Ostlund, Brittany R. Alperin, Mckenzie
2383–2393. https://​doi.​org/​10.​1046/j.​1460-​9568.​2002.​02411.x Figuracion, Hanna C. Gustafsson, Erika Michiko Deming, Dan
Donoghue, T., Haller, M., Peterson, E. J., Varma, P., Sebastian, P., Gao, Foti, et al. (2022). “Electroencephalogram Aperiodic Power Spec-
R., Noto, T., Lara, A. H., Wallis, J. D., Knight, R. T., Shestyuk, tral Slope Can Be Reliably Measured and Predicts ADHD Risk
A., & Voytek, B. (2020). Parameterizing neural power spectra in Early Development.” Developmental Psychobiology, 64(3):
into periodic and aperiodic components. Nature Neuroscience, e22228
23(12), 1655–1665. https://​doi.​org/​10.​1038/​s41593-​020-​00744-x Karpova, A. V., Bikbaev, A. F., Coenen, A. M. L., & van Luijtelaar, G.
Donoghue, T., Schaworonkow, N., & Voytek, B. (2021). Methodologi- (2005). Morphometric Golgi study of cortical locations in WAG/
cal considerations for studying neural oscillations. The European Rij rats: The cortical focus theory. Neuroscience Research, 51(2),
Journal of Neuroscience. https://​doi.​org/​10.​1111/​ejn.​15361 119–128. https://​doi.​org/​10.​1016/j.​neures.​2004.​10.​004
El Boustani, S., Marre, O., Béhuret, S., Baudot, P., Yger, P., Bal, T., Kello, C. T., Brown, G. D. A., Ferrer-I-Cancho, R., Holden, J. G.,
Destexhe, A., & Frégnac, Y. (2009). Network-state modulation Linkenkaer-Hansen, K., Rhodes, T., & Van Orden, G. C. (2010).
of power-law frequency-scaling in visual cortical neurons. PLoS Scaling laws in cognitive sciences. Trends in Cognitive Sciences,
Computational Biology, 5(9), e1000519. https://​doi.​org/​10.​1371/​ 14(5), 223–232. https://​doi.​org/​10.​1016/j.​tics.​2010.​02.​005
journ​al.​pcbi.​10005​19 Kosciessa, J. Q., Grandy, T. H., Garrett, D. D., & Werkle-Bergner, M.
Engel, A. K., Fries, P., & Singer, W. (2001). Dynamic predictions: (2020). Single-trial characterization of neural rhythms: Potential
Oscillations and synchrony in top-down processing. Nature and challenges. NeuroImage, 206, 116331. https://​doi.​org/​10.​
Reviews Neuroscience, 2(10), 704–716. https://​doi.​org/​10.​1038/​ 1016/j.​neuro​image.​2019.​116331
35094​565 Lee, D. D., & Seung, H. S. (1999). Learning the parts of objects by
Fransson, P., Metsäranta, M., Blennow, M., Åden, U., Lagercrantz, H., non-negative matrix factorization. Nature, 401(6755), 788–791.
& Vanhatalo, S. (2013). Early development of spatial patterns of https://​doi.​org/​10.​1038/​44565
power-law frequency scaling in FMRI resting-state and EEG data Lendner, J. D., Helfrich, R. F., Mander, B. A., Romundstad, L., Lin,
in the newborn brain. Cerebral Cortex, 23(3), 638–646. https://​ J. J., Walker, M. P., Larsson, P. G., & Knight, R. T. (2020). An
doi.​org/​10.​1093/​cercor/​bhs047 electrophysiological marker of arousal level in humans. eLife, 9.
Freeman, W. J., & Zhai, J. (2009). Simulated power spectral den- https://​doi.​org/​10.​7554/​eLife.​55092
sity (PSD) of background electrocorticogram (ECoG). Cog- Litvak, V., Eusebio, A., Jha, A., Oostenveld, R., Barnes, G. R., Penny,
nitive Neurodynamics, 3(1), 97–103. https://​doi.​org/​10.​1007/​ W. D., Zrinzo, L., Hariz, M. I., Limousin, P., Friston, K. J., &
s11571-​008-​9064-y Brown, P. (2010). Optimized beamforming for simultaneous MEG
Gao, R., Peterson, E. J., & Voytek, B. (2017). Inferring synaptic exci- and intracranial local field potential recordings in deep brain stim-
tation/inhibition balance from field potentials. NeuroImage, 158, ulation patients. NeuroImage, 50(4), 1578–1588. https://​doi.​org/​
70–78. https://​doi.​org/​10.​1016/j.​neuro​image.​2017.​06.​078 10.​1016/j.​neuro​image.​2009.​12.​115
Gerster, M., Berner, R., Sawicki, J., Zakharova, A., Škoch, A., Hlinka, Litvak, V., Jha, A., Eusebio, A., Oostenveld, R., Foltynie, T., Limousin,
J., Lehnertz, K., & Schöll, E. (2020). FitzHugh–Nagumo oscil- P., Zrinzo, L., Hariz, M. I., Friston, K., & Brown, P. (2011). Rest-
lators on complex networks mimic epileptic-seizure-related syn- ing oscillatory cortico-subthalamic connectivity in patients with
chronization phenomena. Chaos, 30(12), 123130. https://​doi.​org/​ Parkinson’s disease. Brain: A Journal of Neurology, 134(Pt 2),
10.​1063/5.​00214​20 359–374. https://​doi.​org/​10.​1093/​brain/​awq332
Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier, Massimini, M., Ferrarelli, F., Huber, R., Esser, S. K., Singh, H., &
D., Brodbeck, C., Goj, R., Jas, M., Brooks, T., Parkkonen, L., Tononi, G. (2005). Breakdown of cortical effective connectivity
& Hämäläinen, M. (2013). MEG and EEG data analysis with during sleep. Science, 309(5744), 2228–2232. https://​doi.​org/​10.​
MNE-Python. Frontiers in Neuroscience, 7, 267. https://​doi.​org/​ 1126/​scien​ce.​11172​56
10.​3389/​fnins.​2013.​00267 Merlo, D., Mollinari, C., Inaba, Y., Cardinale, A., Rinaldi, A. M.,
Halgren, M., Kang, R., Voytek, B., Ulbert, I., Fabo, D., Eross, L., D’Antuono, M., D’Arcangelo, G., Tancredi, V., Ragsdale, D., & Avoli,
Wittner, L., Madsen, J., Doyle, W. K., Devinsky, O., Halgren, E., M. (2007). Reduced GABAB receptor subunit expression and paired-
Harnett, M., & Cash, S. S. (2021). The timescale and magnitude pulse depression in a genetic model of absence seizures. Neurobiology
of aperiodic activity decreases with cortical depth in humans, of Disease, 25(3), 631–641. https://​doi.​org/​10.​1016/j.​nbd.​2006.​11.​005
macaques and mice. bioRxiv. https://​doi.​org/​10.​1101/​2021.​07.​ Miller, K. J., Sorensen, L. B., Ojemann, J. G., & den Nijs, M. (2009).
28.​454235 Power-law scaling in the brain surface electric potential. PLoS
He, B. J. (2014). Scale-free brain activity: Past, present, and future. Computational Biology, 5(12), e1000609. https://d​ oi.o​ rg/1​ 0.1​ 371/​
Trends in Cognitive Sciences, 18(9), 480–487. https://​doi.​org/​10.​ journ​al.​pcbi.​10006​09
1016/j.​tics.​2014.​04.​003 Miskovic, V., MacDonald, K. J., Jack Rhodes, L., & Cote, K. A. (2019).
He, B. J., Zempel, J. M., Snyder, A. Z., & Raichle, M. E. (2010). The Changes in EEG multiscale entropy and power-law frequency
temporal structures and functional significance of scale-free brain scaling during the human sleep cycle. Human Brain Mapping,
activity. Neuron, 66(3), 353–369. https://d​ oi.o​ rg/1​ 0.1​ 016/j.n​ euron.​ 40(2), 538–551. https://​doi.​org/​10.​1002/​hbm.​24393
2010.​04.​020 Molina, J. L., Voytek, B., Thomas, M. L., Joshi, Y. B., Bhakta, S. G.,
He, W., Donoghue, T., Sowman, P. F., Seymour, R. A., Brock, J., Crain, Talledo, J. A., Swerdlow, N. R., & Light, G. A. (2020). Meman-
S., Voytek, B., & Hillebrand, A. (2019). Co-Increasing Neuronal tine effects on electroencephalographic measures of putative
Noise and Beta Power in the Developing Brain. In bioRxiv (No. excitatory/inhibitory balance in schizophrenia. Biological Psychi-
839258). https://​doi.​org/​10.​1101/​839258 atry: Cognitive Neuroscience and Neuroimaging, 5(6), 562–568.
Iemi, L., Busch, N. A., Laudini, A., Haegens, S., Samaha, J., https://​doi.​org/​10.​1016/j.​bpsc.​2020.​02.​004
Villringer, A., & Nikulin, V. V. (2019). Multiple mechanisms

13
Neuroinformatics (2022) 20:991–1012 1011

Muthukumaraswamy, S. D. (2013). High-frequency brain activity and the first seven months of life. Developmental Cognitive Neuro-
muscle artifacts in MEG/EEG: A review and recommendations. science, 47, 100895. https://​doi.​org/​10.​1016/j.​dcn.​2020.​100895
Frontiers in Human Neuroscience, 7, 138. https://d​ oi.o​ rg/1​ 0.3​ 389/​ Scheer, H. J., Sander, T., & Trahms, L. (2006). The influence of ampli-
fnhum.​2013.​00138 fier, interface and biological noise on signal quality in high-
Muthukumaraswamy, S. D., & Liley, D. T. (2018). 1/f electrophysio- resolution EEG recordings. Physiological Measurement, 27(2),
logical spectra in resting and drug-induced states can be explained 109–117. https://​doi.​org/​10.​1088/​0967-​3334/​27/2/​002
by the dynamics of multiple oscillatory relaxation processes. Neu- Schnitzler, A., & Gross, J. (2005). Normal and pathological oscillatory
roImage, 179, 582–595. https://​doi.​org/​10.​1016/j.​neuro​image.​ communication in the brain. Nature Reviews Neuroscience, 6(4),
2018.​06.​068 285–296. https://​doi.​org/​10.​1038/​nrn16​50
Ngomba, R. T., Biagioni, F., Casciato, S., Willems-van Bree, E., Battaglia, Sharbrough, F. (1991). American Electroencephalographic Society
G., Bruno, V., Nicoletti, F., & van Luijtelaar, E. L. J. M. (2005). guidelines for standard electrode position nomenclature. Journal
The preferential mGlu2/3 receptor antagonist, LY341495, reduces of Clinical Neurophysiology: Official Publication of the American
the frequency of spike–wave discharges in the WAG/Rij rat model Electroencephalographic Society, 8(2), 200–202. https://​www.​
of absence epilepsy. Neuropharmacology, 49, 89–103. https://d​ oi.​ ncbi.​nlm.​nih.​gov/​pubmed/​20508​19
org/​10.​1016/j.​neuro​pharm.​2005.​05.​019 Singer, W. (1999). Neuronal synchrony: a versatile code for the defini-
Niethard, N., Hasegawa, M., Itokazu, T., Oyanedel, C. N., Born, J., & tion of relations?. Neuron, 24(1), 49–65, 111–125. https://d​ oi.o​ rg/​
Sato, T. R. (2016). Sleep-Stage-Specific Regulation of Cortical 10.​1016/​s0896-​6273(00)​80821-1
Excitation and Inhibition. Current Biology: CB, 26(20), 2739– Stephani, T., Waterstraat, G., Haufe, S., Curio, G., Villringer, A.,
2749. https://​doi.​org/​10.​1016/j.​cub.​2016.​08.​035 & Nikulin, V. V. (2020). Temporal Signatures of Criticality in
Nikulin, V. V., Nolte, G., & Curio, G. (2011). A novel method for reli- Human Cortical Excitability as Probed by Early Somatosensory
able and fast extraction of neuronal EEG/MEG oscillations on Responses. The Journal of Neuroscience, 40(34), 6572–6583.
the basis of spatio-spectral decomposition. NeuroImage, 55(4), https://​doi.​org/​10.​1523/​JNEUR​OSCI.​0241-​20.​2020
1528–1535. https://​doi.​org/​10.​1016/j.​neuro​image.​2011.​01.​057 Stolk, A., Brinkman, L., Vansteensel, M. J., Aarnoutse, E., Leijten,
Onat, F. Y., van Luijtelaar, G., Nehlig, A., & Snead, O. C., 3rd. F. S., Dijkerman, C. H., Knight, R. T., de Lange, F. P., & Toni,
(2013). The involvement of limbic structures in typical and I. (2019). Electrocorticographic dissociation of alpha and beta
atypical absence epilepsy. Epilepsy Research, 103(2–3), 111– rhythmic activity in the human sensorimotor system. eLife, 8.
123. https://​doi.​org/​10.​1016/j.​eplep​syres.​2012.​08.​008 https://​doi.​org/​10.​7554/​eLife.​48065
Ostlund, B. D., Alperin, B. R., Drew, T., & Karalunas, S. L. (2021). Stock, A.-K., Pertermann, M., Mückschel, M., & Beste, C. (2020).
Behavioral and cognitive correlates of the aperiodic (1/f-like) High-dose ethanol intoxication decreases 1/f neural noise or scale-
exponent of the EEG power spectrum in adolescents with and free neural activity in the resting state. Addiction Biology, 25(6),
without ADHD. Developmental Cognitive Neuroscience, 48, e12818. https://doi.org/10.1111/adb.12818
100931. https://​doi.​org/​10.​1016/j.​dcn.​2021.​100931 Tan, H. O., Reid, C. A., Single, F. N., Davies, P. J., Chiu, C., Murphy,
Ouyang, G., Hildebrandt, A., Schmitz, F., & Herrmann, C. S. (2020). S., Clarke, A. L., Dibbens, L., Krestel, H., Mulley, J. C., Jones,
Decomposing alpha and 1/f brain activities reveals their differen- M. V., Seeburg, P. H., Sakmann, B., Berkovic, S. F., Sprengel,
tial associations with cognitive processing speed. NeuroImage, R., & Petrou, S. (2007). Reduced cortical inhibition in a mouse
205, 116304. https://​doi.​org/​10.​1016/j.​neuro​image.​2019.​116304 model of familial childhood absence epilepsy. Proceedings of the
Peng, C. K., Havlin, S., Stanley, H. E., & Goldberger, A. L. (1995). National Academy of Sciences of the United States of America,
Quantification of scaling exponents and crossover phenomena in 104(44), 17536–17541. https://d​ oi.o​ rg/1​ 0.1​ 073/p​ nas.0​ 70844​ 0104
nonstationary heartbeat time series. Chaos, 5(1), 82–87. https://​ Timmermann, C., Roseman, L., Schartner, M., Milliere, R., Williams,
doi.​org/​10.​1063/1.​166141 L. T. J., Erritzoe, D., Muthukumaraswamy, S., Ashton, M.,
Pereda, E., Gamundi, A., Rial, R., & González, J. (1998). Non-linear Bendrioua, A., Kaur, O., Turton, S., Nour, M. M., Day, C. M.,
behaviour of human EEG: Fractal exponent versus correlation Leech, R., Nutt, D. J., & Carhart-Harris, R. L. (2019). Neural
dimension in awake and sleep stages. Neuroscience Letters, correlates of the DMT experience assessed with multivariate
250(2), 91–94. https://​doi.​org/​10.​1016/​s0304-​3940(98)​00435-2 EEG. Scientific Reports, 9(1), 16324. https://​doi.​org/​10.​1038/​
Podvalny, E., Noy, N., Harel, M., Bickel, S., Chechik, G., Schroeder, s41598-​019-​51974-4
C. E., Mehta, A. D., Tsodyks, M., & Malach, R. (2015). A uni- Vallat, R. (2019). YASA (yet another spindle algorithm): A fast and
fying principle underlying the extracellular field potential spec- open-source sleep spindles and slow-waves detection toolbox.
tral responses in the human cortex. Journal of Neurophysiology, Sleep Medicine, 64, S396. https://​doi.​org/​10.​1016/j.​sleep.​2019.​
114(1), 505–519. https://​doi.​org/​10.​1152/​jn.​00943.​2014 11.​1104
Pumain, R., Louvel, J., Gastard, M., Kurcewicz, I., & Vergnes, M. van Heumen, S., Moreau, J. T., Simard-Tremblay, E., Albrecht, S.,
(1992). Responses to N-methyl-D-aspartate are enhanced in Dudley, R. W., & Baillet, S. (2021). Case Report: Aperiodic Fluc-
rats with petit mal-like seizures. Journal of Neural Transmis- tuations of Neural Activity in the Ictal MEG of a Child With
sion Supplementum, 35, 97–108. https:// ​ d oi. ​ o rg/ ​ 1 0. ​ 1 007/​ Drug-Resistant Fronto-Temporal Epilepsy. Frontiers in Human
978-3-​7091-​9206-1_7. Neuroscience, 15, 646426. https://​doi.​org/​10.​3389/​fnhum.​2021.​
Robertson, M. M., Furlong, S., Voytek, B., Donoghue, T., Boettiger, C. 646426
A., & Sheridan, M. A. (2019). EEG power spectral slope differs by Van Veen, B. D., van Drongelen, W., Yuchtman, M., & Suzuki, A.
ADHD status and stimulant medication exposure in early child- (1997). Localization of brain electrical activity via linearly con-
hood. Journal of Neurophysiology, 122(6), 2427–2437. https://​ strained minimum variance spatial filtering. IEEE Transactions
doi.​org/​10.​1152/​jn.​00388.​2019 on Bio-Medical Engineering, 44(9), 867–880. https://​doi.​org/​10.​
Romei, V., Brodbeck, V., Michel, C., Amedi, A., Pascual-Leone, A., & 1109/​10.​623056
Thut, G. (2008). Spontaneous fluctuations in posterior alpha-band Veerakumar, A., Tiruvadi, V., Howell, B., Waters, A. C., Crowell,
EEG activity reflect variability in excitability of human visual A. L., Voytek, B., Riva-Posse, P., Denison, L., Rajendra, J. K.,
areas. Cerebral Cortex, 18(9), 2010–2018. https://​doi.​org/​10.​ Edwards, J. A., Bijanki, K. R., Choi, K. S., & Mayberg, H. S.
1093/​cercor/​bhm229 (2019). Field potential 1/f activity in the subcallosal cingulate
Schaworonkow, N., & Voytek, B. (2021). Longitudinal changes in ape- region as a candidate signal for monitoring deep brain stimulation
riodic and periodic activity in electrophysiological recordings in

13
1012 Neuroinformatics (2022) 20:991–1012

for treatment-resistant depression. Journal of Neurophysiology, non-invasively–An EEG tutorial. Journal of Neuroscience Meth-
122(3), 1023–1035. https://​doi.​org/​10.​1152/​jn.​00875.​2018 ods, 250, 74–84. https://​doi.​org/​10.​1016/j.​jneum​eth.​2014.​08.​013.
Voytek, B., Kramer, M. A., Case, J., Lepage, K. Q., Tempesta, Z. R., Waterstraat, G., Körber, R., Storm, J.-H., & Curio, G. (2021). Nonin-
Knight, R. T., & Gazzaley, A. (2015). Age-Related Changes in vasive neuromagnetic single-trial analysis of human neocortical
1/f Neural Electrophysiological Noise. The Journal of Neurosci- population spikes. Proceedings of the National Academy of Sci-
ence, 35(38), 13257–13265. https://​doi.​org/​10.​1523/​jneur​osci.​ ences of the United States of America, 118(11). https://d​ oi.o​ rg/1​ 0.​
2332-​14.​2015 1073/​pnas.​20174​01118
Ward, L. M. (2003). Synchronous neural oscillations and cognitive Wen, H., & Liu, Z. (2016). Separating Fractal and Oscillatory Com-
processes. Trends in Cognitive Sciences, 7(12), 553–559. https://​ ponents in the Power Spectrum of Neurophysiological Sig-
doi.​org/​10.​1016/j.​tics.​2003.​10.​012 nal. Brain Topography, 29(1), 13–26. https://​doi.​org/​10.​1007/​
Waschke, L., Donoghue, T., Fiedler, L., Smith, S., Garrett, D. D., s10548-​015-​0448-0
Voytek, B., & Obleser, J. (2021). Modality-specific tracking of Zanos, T. P., Mineault, P. J., & Pack, C. C. (2011). Removal of spuri-
attention and sensory statistics in the human electrophysiological ous correlations between spikes and local field potentials. Journal
spectral exponent. ELife, 10. https://​doi.​org/​10.​7554/​eLife.​70068 of Neurophysiology, 105(1), 474–486. https://​doi.​org/​10.​1152/​jn.​
Waschke, L., Wöstmann, M., & Obleser, J. (2017). States and traits 00642.​2010
of neural irregularity in the age-varying human brain. Scientific Zhou, Y., Sheremet, A., Kennedy, J. P., DiCola, N. M., Maciel, C. B.,
Reports, 7(1), 17381. https://d​ oi.o​ rg/1​ 0.1​ 038/s​ 41598-0​ 17-1​ 7766-4 Burke, S. N., & Maurer, A. P. (2021). Spectrum Degradation of
Waterstraat, G., Burghoff, M., Fedele, T., Nikulin, V., Scheer, H. J., Hippocampal LFP During Euthanasia. Frontiers in Systems Neu-
& Curio, G. (2015a). Non-invasive single-trial EEG detection of roscience, 15, 647011. https://d​ oi.o​ rg/1​ 0.3​ 389/f​ nsys.2​ 021.6​ 47011
evoked human neocortical population spikes. NeuroImage, 105,
13–20. https://​doi.​org/​10.​1016/j.​neuro​image.​2014.​10.​024 Publisher's Note Springer Nature remains neutral with regard to
Waterstraat, G., Fedele, T., Burghoff, M., Scheer, H. -J., & Curio, jurisdictional claims in published maps and institutional affiliations.
G. (2015b). Recording human cortical population spikes

13