Journal of Neuroscience Methods 274 (2016) 1–12

Contents lists available at ScienceDirect

Journal of Neuroscience Methods

journal homepage: www.elsevier.com/locate/jneumeth

Optimizing detection and analysis of slow waves in sleep EEG

Armand Mensen ∗ , Brady Riedner, Giulio Tononi
Center for Sleep and Consciousness, Department of Psychiatry, University of Wisconsin, United States

h i g h l i g h t s

• We introduce an open-source toolbox for individual detection and analysis of slow waves in sleep electroencephalography.
• Novel and previously applied automatic detection algorithms are introduced and explored.
• Individual slow waves are detected in sleep recordings from participants along a large search-space of parameter settings.
• Properties of detected slow waves are compared across parameter settings on a range of outcome measures of interest.
• Visualization options for toolbox users are introduced, including the possibility to manual score sleep.

a r t i c l e i n f o a b s t r a c t

Article history: Background: Analysis of individual slow waves in EEG recording during sleep provides both greater sen-
Received 22 March 2016 sitivity and specificity compared to spectral power measures. However, parameters for detection and
Received in revised form analysis have not been widely explored and validated.
19 September 2016
New method: We present a new, open-source, Matlab based, toolbox for the automatic detection and
Accepted 20 September 2016
analysis of slow waves; with adjustable parameter settings, as well as manual correction and exploration
Available online 20 September 2016
of the results using a multi-faceted visualization tool.
Results: We explore a large search space of parameter settings for slow wave detection and measure
Keywords:
Slow waves
their effects on a selection of outcome parameters. Every choice of parameter setting had some effect on
Automatic detection at least one outcome parameter. In general, the largest effect sizes were found when choosing the EEG
Sleep reference, type of canonical waveform, and amplitude thresholding.
EEG Comparison with existing method: Previously published methods accurately detect large, global waves
Toolbox but are conservative and miss the detection of smaller amplitude, local slow waves. The toolbox has
additional benefits in terms of speed, user-interface, and visualization options to compare and contrast
slow waves.
Conclusions: The exploration of parameter settings in the toolbox highlights the importance of careful
selection of detection
Methods: The sensitivity and specificity of the automated detection can be improved by manually adding
or deleting entire waves and or specific channels using the toolbox visualization functions. The toolbox
standardizes the detection procedure, sets the stage for reliable results and comparisons and is easy to
use without previous programming experience.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction
For much of the 20th century sleep was considered to be a global
phenomenon of the brain, and its macro-architecture was of pri-
mary interest (Jones, 2005; Saper et al., 2005; Siegel, 2009). The past
few decades has witnessed a shift in interest to the spatial domain
and local aspects of sleep (Krueger et al., 2008; Nobili et al., 2011;
∗ Corresponding author.
E-mail address: research.mensen@gmail.com (A. Mensen).
Vyazovskiy et al., 2011; Marzano et al., 2013). Much of the research
into local patterns of sleep has used the measure of ‘slow wave
activity’ (Werth et al., 1997; Huber et al., 2004; Stadelmann et al.,
2013), reflective of the changes in the power spectra in the lower
frequencies (typically around 1–4 Hz), measured across a whole
night or single cycle of sleep. This measure is, however, under-
determined since both slow wave incidence and amplitude will
affect power.
Slow waves can originate in a small region of the cortex and then
propagate to other cortical regions based on both EEG and intracra-
nial recordings (Amzica and Steriade, 1998). Therefore, there are,

http://dx.doi.org/10.1016/j.jneumeth.2016.09.006
0165-0270/© 2016 Elsevier B.V. All rights reserved.
2 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12
at least, five distinct ways in which local slow wave activity could
increase or decrease (Massimini et al., 2004; Menicucci et al., 2009;
Murphy et al., 2009): 1, origins and traveling parameters remain
constant but there are local changes in the amplitudes of the slow
waves as they travel over particular regions of the cortex; 2, there
is an increase in the incidence of local waves which originate in
that particular part of the cortex; 3, slow wave origins remain dis-
tributed but more waves travel to or through a specific area of
the cortex; 4, there are local changes in the speed of propagation
resulting in a frequency shift of power spectra; 5, some particular
combination of the above factors. Notably, if any of the above mech-
anisms have opposing effects they may cancel out in the power
spectra.
Conventional power-based methods are relatively easy to cal-
culate as the appropriate tools have already been developed and
standardized over time. Thus, results are relatively comparable
across studies, in turn leading to power spectral measures being
used repeatedly in research. Nevertheless, an examination of the
properties of individual slow waves is now also possible and can
yield an increase in specificity and sensitivity without the asso-
ciated cost of having to perform new measurements (Riedner
et al., 2007). For example, the analysis of the origin of slow waves
revealed a local increase indicative of post-sleep learning in a visual
perception task (Mascetti et al., 2013). Moreover, slopes of individ-
ual slow waves correlate with neural development (Fattinger et al.,
2014), and epileptic spike waves can impair individual slow waves
(Bölsterli Heinzle et al., 2014). The advent of parallel computing
and the lower cost of technology has made individual waveform
analysis practically feasible. However, the necessary tools for the
detection of individual slow waves, and the subsequent calcula-
tion of their traveling parameters have not yet been made freely
available, standardized, or validated. Moreover, tools have not been
made sufficiently simple or generalized so that researchers and
clinicians can easily explore their own data and have confidence
in the results while also comparing them against set standards.
Here we describe an open source toolbox with the principal
purpose of providing a reliable interface to detect and analyze
individual slow waves found in EEG sleep recordings. The tool-
box is version-controlled using git and freely available at https://
github.com/Mensen/swa-matlab. The main goal of the current arti-
cle is to introduce the main features of the toolbox for slow-wave
detection. Secondly, we use several full-night sleep recordings to
examine how a large search-space of different parameter settings
influence the detection and properties of slow waves. While this is
not intended to be an extensive overview of how to most accurately
detect slow waves, it is meant to showcase the toolbox’s functional-
ity, typical work-flow, and visualization capabilities while making
the user aware of the strengths and weaknesses of various settings
and what aspects to consider when analyzing recordings of their
own.
2. Materials and methods
2.1. Toolbox overview
2.1.1. Sleep scoring
Given that different types of waves in sleep occur during differ-
ent periods, detection can be improved by parsing the night into its
various sleep stages. To this end a user-friendly interface was cre-
ated for the manual visual scoring of sleep stages and arousal events
of high-density EEG channels. This has previously been available
primarily through proprietary software in certified sleep centers
and only once specified channels from the high-density caps had
been exported. This set of toolbox functions allows for the on-line
adjustment of displayed channels and specified references, as well
as individual channel filtering options for each channel displayed.
Navigation and sleep scoring can be performed using the mouse
or keyboard keys (e.g. left and right arrows to navigate, number
keys to indicate sleep stage). Importantly, the data used for scor-
ing can be loaded directly from the original recording and is saved
in the same file. This is a useful feature, as the file sizes for long
recordings of high-density channels are typically in the 5–20 Gb
range. This feature also reduces the number of copies necessary
while simultaneously maintaining a controlled history of process-
ing. Similarly, the scoring toolbox uses dynamic memory mapping
to load into random-access-memory (RAM) only the part of the
file that is currently displayed (e.g. 8 channels of 30 s duration).
This feature allows for the scoring of files on a typical computer
where the amount of RAM memory may be smaller than the dataset
one needs to score. Since filtering is only applied to the displayed
portion of data, and not the raw data itself, this graphical user inter-
face (GUI) is also a good way to visualize raw recordings and mark
artifacts manually, prior to any preprocessing methods. Channel
montages and filtering options are saved within the file and can
be used with other datasets to keep the settings consistent across
participants or recording nights.
2.1.2. Wave detection
The detection of slow waves, regardless of specific settings, fol-
lows 4 key stages.
1 Calculation of the canonical wave(s).
2 Detection of individual slow waves within the canonical wave(s).
3 Detection of corresponding waves within the actual channels.
4 Examination of each wave for its traveling streams and its prop-
erties.
There are two conceptually distinct approaches available for the
calculation of the canonical wave: either the mean activity over
a specified region; or the negative envelope of the channels. The
simplest canonical wave can be created by calculating the mean
activity over a single circular region of the electrode array, defin-
ing its center and radius. Alternatively, multiple canonical waves
can be computed by taking the mean activity of distinct regions.
Regions can be specified over the mid-line, leading to a frontal,
central and posterior canonical time series. Another option is to
calculate the mean activity of four regions equidistant around the
center, as was done in the original slow wave detection algorithm
(Massimini et al., 2004). These four regions can be arranged as a
square (left/right; frontal/posterior), or as a diamond (single frontal
and posterior regions; left and right central region). Users may also
directly specify their own canonical wave.
One issue with regional methods is that local waves outside the
specified region will not be represented in the canonical wave and
therefore have no chance of being detected. Thus, only waves that
pass through a substantial portion of at least one of the regions can
be detected – a source of bias that may underestimate the amount
of locality in sleep. Conversely, if a single region is too large, then
the mean activity within the region may no longer be represen-
tative and waves could be missed (see Supplementary Fig. 1 for
comparison of regional methods). Secondly, such a method does not
scale well to more sparsely recorded arrays where those predefined
regions may only contain a few channels.
A solution is to calculate the negative envelope of all chan-
nels. This is akin to examining the butterfly plot (overlaying all
channels) and tracing the negative contour. Here, we calculate
the mean activity of the most negative 2.5% of channels at each
sample independently in the time series. The advantage is that
the most negative portion can potentially come from any channel
in the dataset and is not restricted to a particular region. Fur-
thermore, a single canonical wave is representative of the entire
 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12 3

Fig. 1. Typical steps in the detection and analysis of slow waves.

A, the current default canonical wave is the negative envelope of all channels. This wave is the mean of the bottom 2.5% of channels at each sample independently such that
a slow wave could come from anywhere in the dataset. B, examining the canonical wave can take several parameters to explore the properties of either each downward
zero-crossing or local minima, to check that the potential wave matches set criteria for amplitude, wavelength, or slope. C, once slow waves have been detected in the
canonical wave, we can examine the cross-correlation to not only find which waves also show a slow wave (high positive correlation with the canonical), but at what
sampling delay they best fit giving an estimate of the channel origins of the wave and its propagation pattern. D, using a delay map created by interpolating the delay values
onto the scalp, we can find optimal trajectory paths through this gradient in terms of maximum displacement and distance traveled as well as any stream with a notably
distinct angle of travel.

dataset. Thus, the full detection criteria only has to be run on
a single time series. Despite these advantages over the regional
approach, there are also 3 limitations: the waveform may be dis-
torted because different waves contribute to its shape such that it
may not closely correspond to any particular channel; the positive
portion of the wave is eliminated from the canonical wave, leading
to meaningless peak-to-peak information; and the negative enve-
lope has rare zero-crossings for further calculation and so must
be high-pass filtered or detrended to bring the mean activity to
zero, thereby consistently, yet somewhat unpredictably, reducing
the amplitudes of the waves it intends to represent.
Once the canonical wave channel has been created, the goal
is to use a set of parameters to automatically detect potential
slow waves. The toolbox has a wide range of parameters that
can be adjusted, including those used in previous research (Mölle
et al., 2002; Massimini et al., 2004). For example, the points
between downward zero-crossings and the subsequent positive
zero-crossing can be seen as a potential slow wave. However, slow
drifts in the EEG as well as the use of the negative envelope method
may lead to waves not having reliable zero-crossings. An alterna-
tive is to examine the local minima of the canonical wave and use
these as initial points to inspect further wave properties. This elim-
inates the importance of the zero-crossing portion of the waves
altogether as the nearest local maxima are then seen as the start
and end points of that particular wave. The choice between zero-
crossings or local minima/maxima is referred to as the inspection
point.
Individual waves must also pass a crucial minimum ampli-
tude threshold. One may specify the absolute minimal amplitude
directly in microvolts (default: 80 ␮V). However, several factors
will influence the overall amplitudes of the EEG, such as the chan-
nel montage, filtering parameters, as well as physiological variables
such as scalp thickness and age (Buchmann et al., 2011; Wilhelm
et al., 2014). Therefore, the toolbox also implements a data-driven,
dynamic threshold based on the calculation of the median ampli-
tude and standard deviation of all the local minima in the canonical
wave. With this option, the user specifies the number of stan-
dard deviations from the median (default: 5 standard deviations).
The minimum and maximum wavelength should be specified for
the canonical wave (default: 250 ms–1250 ms). Depending on the
inspection point, wavelength is either defined by the point of down-
ward to subsequent upward zero crossing, or as the time between
consecutive local maxima. Finally, the potential wave must pass
a peak-to-peak criterion from the negative to positive peak of the
wave (default: 140 ␮V; not applied to negative envelope canoni-
cal waves). For multiple regions, each wave is initially examined
independently and potential waves are examined for whether a
slow wave was already found within the wavelength of another
canonical wave.
By this stage the user has determined the number of unique slow
waves in the dataset which meet specific criteria, and the goal is
now to find the corresponding channels and their properties. There
are two available channel detection methods. The same set of initial
criteria can be applied to individual channels as was applied to the
canonical wave; with the option to adjust all the criteria by a certain
ratio (default: 90%). In this way, one could use a strict set of crite-
ria on the canonical wave (to minimize false positives), but then a
more relaxed set of criteria at the channel level (to maximize sen-
sitivity). Alternatively, the user has the option to cross-correlate
the canonical wave with the individual channels over a specified
time window, looking for channels that correlate positively with
just the negative portion of the canonical wave. The correlation
method has important advantages. It is amplitude independent, so
any channel that shows the same pattern as the detected slow wave
in the canonical channel will be accepted as a match. Secondly,
the temporal delay of the wave at each channel is already calcu-
lated while searching for the maximum correlation across different
delays (illustrated in Fig. 1). The original method calculates the
delay separately according to the time point of the minimum peak
for each channel. Using correlation, the entire waveform is consid-
ered for the delay, not just the peak, and is thus robust to noise.
The limitation with the correlation approach is that if the nega-
tive envelope was used as the canonical wave, this wave may not
correlate well with any particular channel. This limitation is over-
come by also cross-correlating channels with the channel showing
the largest negative peak. Finally, a cluster test can be performed
on the active channel topography to ensure that all the channels
are connected to one another through some neighboring channels;
this eliminates island channels which are likely to be artifacts. For
either the correlation or threshold channel detection method, if no
corresponding channels are found, this individual wave is removed
from the final output.
To calculate the traveling properties of individual waves, an
interpolated delay map over the scalp is created using the individ-
ual delays across ‘active’ channels. Each channels’ coordinates are
used as a seed to examine the optimal stream lines to and from the
channel; thus calculating all potential streamlines for the traveling
wave. The streamline calculation algorithm works much like plac-
ing a leaf into a wind tunnel and recording its trajectory. At each
consecutive point of the interpolated delay map, the next point in
the stream is determined by following the gradient of the map until
a stopping condition is met. There are three conditions for stopping
4 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12

Fig. 2. Example of the sleep-wave-analysis (swa) explorer window.

Researchers can load the result file after successful detection to freely explore results and make manual adjustments to either the individual waves, by adding or removing
channels (in the central left panel), or adding or deleting entire waves. Any two channels used in the analysis can be viewed in the top panel along with the butterfly plot
of all channels or the canonical wave. The bottom left two panels show summary measures of the entire recording that one can choose using the drop-down menus. On
the right is the amplitude or delay topography of the current slow wave displayed. In the figure the involvement topography is shown with yellow to blue indicative of the
smallest to largest wave amplitudes. The semi-transparent streamlines indicate the longest displacement, the longest distance traveled, and the stream of highest angular
deviation from the first stream. All plots can be exported to their own figure for editing or printing for publication purposes.

the streamline search: the first is that the gradient at the next point
does not exist, indicating that no slow wave was found at the near-
est channel to that point; second that the change in gradient angle is
greater than 90◦ , indicative of counteracting gradients; and finally
if the gradient at the next coordinate is zero, corresponding to a stop
of the traveling aspect of the wave. Although streamlines for each
active channel are calculated, only a subset of 3 are saved as output:
the longest displacement (the distance between the start and end
points of the wave); the longest distance traveled (the cumulative
sum of all coordinates of the line); and the stream of most angular
deviation from the longest displacement, assuming that stream has
a different overall trajectory of at least 45◦ .
2.1.3. Visualization
Clear and interpretable visualization of the results is of utmost
importance in the toolbox. An overview of the swa Explorer tool
is given in Fig. 2. This GUI is organized into a single window with
four sections. At the top there is a time series plot of the EEG sig-
nal as it was initially loaded. The user is able to select any two
individual channels from the dataset, a ‘butterfly’ view, or the
waves canonical channel. This panel can be used to quickly nav-
igate through the dataset and missed waves can simply be added
by clicking on the axes near the wave. Manual addition of a wave
in this way will bypass the initial criteria for wave selection, cal-
culate the necessary parameters on the canonical wave and then
pass this information onto the channel detection and traveling algo-
rithms. In the left central portion, the ‘butterfly’ plot of the active
channels of each individual wave can be seen in a narrower time
window. The canonical wave is also displayed to allow for compari-
son to the rest of the active channels. This section can be expanded
by using the ’+’ symbol at the corner of the plot which opens up
the channel editing mode. This mode shows all the channels in a
‘butterfly’ arrangement with inactive channels in dashed, lighter
gray color. Clicking on a channel allows the user to manually acti-
vate or deactivate any particular channel in case the automatic
detection has missed a channel or was subject to a false positive.
The user can also zoom into any aspect of the wave and re-calculate
its traveling parameters, in the case that active channels have been
changed. In the bottom left of the window there are two smaller
plots which can display any outcome measure currently available.
All summary measures are selected using a drop-down menu. Addi-
tional wave summary measures are consistently being added but
currently include: histogram plots of the distribution of global-
ity measures, scatter plots comparing amplitude and wavelength,
angular maps of travel direction, as well as EEG topographies of
wave density, amplitude, stream origins, and wave origins. On the
right of the explorer window is a large topography map showing
either the delay map (time to channel peak), or the intensity map
(amplitude of the negative peak) of the currently selected wave.
Below the topography are several options to change the look of
the map such as whether to display the channel locations or trav-
eling streams. Nearly all the plots can be exported to their own
figure using the ‘+’ button, making it easy to create presentation or
publication ready figures; as was done in the current article.
2.1.4. Implementation
The toolbox is currently implemented as semi-independent
scripts and functions in Matlab (The MathWorks, Inc., Natick,
Massachusetts, United States), with minor functions implemented
using C for computation speed. The GUI aspects of the toolbox also
include portions of JavaTM coding to circumvent certain limitations
inherent to Matlab when it comes to visualization of data. Each
function in the toolbox is given the prefix ‘swa’, with a keyword
to the file’s function (e.g. ‘swa SleepScoring’). In principle; all that
 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12
is required to get started with wave detection is the time series;
knowledge of the sampling rate; and the electrode locations file.
However; since several options already exist for importing and
preprocessing EEG data from various sources; we have taken the
file and variable structure of the popular analysis toolbox EEGLAB
(Delorme and Makeig, 2004) as the basis for our own scripts. It
should be noted however that EEGLAB is not required for the tool-
box to function. Several template scripts are available in the toolbox
directory to help users get started with scoring sleep stages; export-
ing stages; preprocessing; or wave detection. Currently; going
from an empty Matlab environment; loading the data; loading the
default detection parameters; running all the separate detection
steps; and saving the results can be completed automatically within
minutes. The efficient algorithms employed allow the user to exper-
iment with optimal detection parameters for the study’s particular
purpose. Moreover; a batch processing script is available such that
an entire dataset can be analyzed using the same set of parameters;
without interruption.
The output of the detection process is a single Matlab file which
includes three structure arrays with fixed names Data, Info and SW.
Depending on the options used when saving the results, Data, may
contain the original EEG data, or simply a pointer to the binary
file that was used in the calculation; so that multiple copies of the
same data need not exist given the typically large file sizes of night
recordings. Data also contains the canonical wave(s), calculated in
the first stage of the detection, and may also keep the sleep staging
information. Info has various subfields holding general information
about the data that was loaded, its electrode positions, and all the
parameters used for wave detection. Finally, SW, has information
on detected wave properties. Subfields of the structure contain:
wave timing; canonical amplitudes; channels involved; and coor-
dinates of the traveling streams. This file can then be loaded into
the swa Explorer function to visualize each wave, create summary
measures and plot them, manually edit the channels involved, or
manually add missed waves or eliminate false positives.
2.2. Parameter evaluation
2.2.1. Data and preprocessing
Whole night sleep data was taken for 11 healthy controls. All
participants were right-handed males, between the ages of 25–35
with no history of mental illness or sleep-related problems. A 257
channel electrode net was used for all recordings (EGI; Tucker,
1993), initially referenced to the central electrode (Cz) and recorded
at 1000 Hz sampling rate. Using EEGLAB, recordings were down-
sampled to 200 Hz. The data was then scored into sleep stages and
artifacts were marked using the swa SleepScoring function accord-
ing to the current standards (Iber et al., 2007). Only stages scored
as N3 were exported for further processing steps; not including
the segments manually marked as artifacts. The remaining data
was then bandpass filtered between 0.3 and 30 Hz. Channels with
excessive artifacts were then manually removed from the dataset.
At this point, two versions of the data were created: one version
that underwent independent component analysis (ICA), and sub-
sequent removal of noise-related components; and another version
without the ICA steps. Finally, the removed channels were reintro-
duced using spline interpolation (Perrin et al., 1987). Data was then
either re-referenced to the average reference (the mean activity of
all channels for each sample), or average-mastoid (mean activity
of electrodes over the left and right mastoids). The resulting data
structure was then loaded into the toolbox for analysis of individ-
ual slow waves. To examine the effect of the number of channels
used in the analysis, the individual electrode arrays were spatially
down-sampled using an iterative process of continually removing
5
the channel with the closest neighbors until the desired number of
remaining channels was achieved.
2.2.2. Analysis
In order to examine how the various parameter settings affect
the detection of slow waves, we ran fully automatic detection on a
large set of possible parameter combinations. The parameter set-
tings tested are outlined in Fig. 3, where the main line reflects the
current default value for each parameter in the toolbox, and the
sub-values tested for that particular parameter are indicated by
the curved paths below. While the EEG montage reference is not
technically a detection parameter, it has a significant impact on
the waveforms. Average mastoid reference has been the general
choice for slow-wave detection likely due to it being standard prac-
tice for sleep scoring in clinical settings (Iber et al., 2007); given its
relative electric neutrality. However, electrodes near the reference
have reduced amplitude activity because shared activity between
them is eliminated. Given the mastoid’s posterior-temporal loca-
tion, this implies that, even in randomly generated activity, frontal
electrodes (those furthest from the reference) will have increased
amplitudes. This frontal bias for average-mastoid reference may be
mitigated by using the average-reference; generally recommended
for high-density EEG recordings (Pascual-Marqui and Lehmann,
1993). However, as will be discussed later, average referencing
can invert the polarity at electrode locations where the wave is
smaller than the average, making tracking the negative envelope
more difficult in the case of large, potentially global slow waves.
It should be noted that, for statistical analysis, the number of
channels analyzed is expressed using the log2 value, such that a
unit increase implies twice as many channels analyzed. Another
important point is that the low and high cut-off for range param-
eters, filter and wavelength, were manipulated together but only
the lower bound value was entered into the model (e.g. 0.5–4 Hz
was indicated as 0.5; and 1–3 Hz indicated as 1). This has the conse-
quence that a single jump in the value of these parameters generally
implies a square law, not a linear one, and that the primary driver
of any effect, that is whether it was the low or high end of the range
that drove the effect, is left unspecified. All main effects, that is
direct deviations from the default parameters while maintaining
other settings constant, were tested along with a smaller subset
of second order combinations of parameters of interest. Lastly, the
specific configuration of parameters used in the original detection
of slow waves as traveling waves, along with some variants on the
amplitude threshold were also tested (Massimini et al., 2004). In
total, automatic detection for 48 different parameter combinations
were run.
Of the possible outcome measures, 9 representative summary
measures of interest were chosen as dependent variables: wave
density, measured as detected slow waves per minute; canoni-
cal wavelength mean and standard deviation; wave amplitude, the
median over all waves (amplitude distributions tend to be highly
skewed), and deviation; mean wave globality and its deviation, mea-
sured as the percentage of channels which were involved in the
wave; anterior-posterior traveling preference with positive values
indicating more back-to-front streams; and finally the angular devi-
ation of all the streamlines calculated. Topographic density maps
were also compared across a selection of settings. Density was
defined as the number of waves found at each channel, independent
of its amplitude.
For parameter settings found to significantly affect the wave
incidence, we further examined waves that were uniquely found in
one method and not the other. We also examined wave properties
for the commonly found waves between methods. Common waves
defined as those with negative peaks within 100 ms. This compar-
ison can be done using the toolbox function swa compareOutput,
and provides an overview of the parameter changes and outcome
6 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12

Fig. 3. Typical work-flow and parameters.

Several choices in the preprocessing of EEG recording during sleep have significant impacts on the automatic detection algorithm. Here we explored the effects of whether
the data had been subjected to an ICA artefact removal, the choice of EEG reference with which to compare all channels, the number of channels actually recorded and the
data filtering bandpass window. Detection of individual slow waves is heavily dependent on the range of parameter settings used in the canonical waveform. Once slow
waves have been detected in the canonical wave there are more significant parameter choices to move from waves in the canonical to individual channel detection. After
completed detection across the night we can extract certain summary statistics of interest such as the mean globality (percentage of channels involved in each wave), or the
topographic density of all slow waves in the recording (number of waves detected at each channel).

measures for the unique and shared waves between any two result
files. These are crucial comparisons because if the method affects
both the number of waves and some other outcome parameter, one
cannot be sure that the differences can be attributed to waves that
are uniquely detected by the method, or whether properties of the
same waves are altered by the parameter settings. This concern
does not apply to all parameter changes, such as amplitude thresh-
old, where differences are directly due to the unique waves found
for lower thresholds, since all common waves are guaranteed to
have identical properties.
Data were analyzed using a linear mixed model approach using
SPSS version 22 (IBM Corp., Version 22.0. Armonk, NY, USA). All
parameter settings and their interactions were treated as fixed
factors (given that they were directly specified) and the interac-
tion of all parameters was given as the repeated measure. The
amplitude threshold parameter was nested within the amplitude
criteria since the meaning of each is directly dependent on whether
amplitude was chosen according to the deviation from the mean
or an absolute value. In the same way, the channel correlation
threshold was nested within detection method as the thresholding
value could either correspond to a minimum Pearson’s correla-
tion value between the channel and the canonical wave, or the
adjustment ratio of the original thresholds by which to measure the
individual channels. The default covariance structure of ’variance
components’ was used, although others were tested and compared
using the Bayesian Information Criteria (BIC) but were not found to
be significant improvements in modeling the data. Each outcome
measure was modeled with a main effect factor for each parame-
ter change, along with 4 selected 2-way interactions of the binary
parameter settings: EEG reference with canonical method; canoni-
cal method with inspection point; canonical method with detection
method; and EEG reference with amplitude criteria. Search window
was left out of the model in the case of wave density, amplitude and
wavelength outcome measures, since it had essentially no effect
on the outcome measures, had therefore zero variance, causing
convergence errors or non-positive Hessian matrices.
Topographies were statistically compared using mass-
univariate statistics for each channel independently. Multiple
comparisons were handled using threshold-free cluster-
enhancement (TFCE), followed by non-parametric maximum
permutation statistics (Mensen and Khatami, 2013, 2015). Briefly,
each channel statistic is individually examined for support from
 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12
Fig. 4. Parameter changes and their significant impact on outcome measures.
Both the left and right column indicate parameter settings which were adjusted to explore their effect on several outcome measures of interest (the center column). Dashed
lines indicate that the factor had a significant effect only in interaction with a second parameter of interest. Grey scale of the lines themselves are only to add in visualization
and have no significance.
its neighboring channels, such that the statistic is enhanced if
neighboring channels show a similar pattern of differences. These
enhanced statistics are then subjected to a single permutation test
whereby the allocation of each dataset to each parameter setting
is randomized 5000 times. This creates a single empirical null
distribution to which each channel is compared to find its indi-
vidual p-value (corresponding to the percentage of randomized
permutations which show a larger maximum statistic).
3. Results
Fig. 4 depicts which parameter settings had a significant impact
on which outcome measure. Model parameter estimates and their
standard error are described in supplementary Tables 1 and 2.
Parameter estimates indicate how large an effect a change to the
parameter value has on a given outcome measure. For example, if
the relative amplitude criterion is raised by 1, the wave density of
the subsequent output is expected to decrease by 2.98 waves/min.
It is also crucial to consider the scale of each parameter when exam-
ining its estimate. For example, amplitude thresholds for relative
criteria reflect the number of standard deviations from the mean,
hence a unit of change may correspond to several tens of micro-
volts, yet for absolute threshold criteria the units are in microvolts
directly. Therefore the estimates between these parameters will
vary drastically even when their standardized effects would be sim-
ilar. While each outcome measure is examined in detail individually
below, the supplementary tables can be used a quick reference
when wishing to explore parameter settings for given outcomes.
For example, if the primary research question involved the ampli-
tude of detected slow waves, it is irrelevant whether the user selects
to detect waves based on the zero-crossing or negative peak meth-
ods; by contrast, this distinction is crucial when comparing wave
density.
3.1. Wave density
Wave density is measured by waves per minute and was calcu-
lated by dividing total number of waves detected by the number
of minutes spent in N3. Overall, recordings had a mean of 12.1
waves/min (se = 0.62). As Fig. 4 indicates, nearly all the parame-
ters had a significant influence on the number of waves detected
in each participant; all but the specific channel detection thresh-
old and the number of channels. The most influential parameters
affecting the total number of waves found were the amplitude
thresholds for both relative and absolute values, wavelength and
the EEG reference. Changing the amplitude threshold by a single
deviation results in a mean increase in wave density of approxi-
mately 3 waves per minute, which when we consider a mean time
of 120 min spent in N3 during the night indicates we would have
an additional 360 waves per recording. Similarly, if we lowered our
7
absolute threshold by 10 ␮V, we would expect to find another 1.7
waves per minute, or a mean of about 200 slow waves per night.
Adjusting the wavelength range by only 100 ms, for example from
0.25–1 s to 0.35–0.9 s, we expect to find similar differences as those
of amplitude changes of over 3 fewer waves per minute.
Crucially, whether one chooses to use average reference or aver-
age mastoids had significant interactions with both the canonical
wave method measure and also whether one chooses to use abso-
lute or relative amplitude thresholds. If the user chooses to use
the negative envelope canonical wave method, there is essentially
no difference in wave density between average and average mas-
toid EEG references (mean difference = 0.2, sd = 0.7), however if
four regions are used to create the canonical waves, the choice
of EEG reference is crucial (mean difference = 6.9, se = 1.1), with
average reference leading to a significant reduction in the wave
density (F1,81 = 40.90, p <0.001). Additionally, only when using
the four regions reference does the choice of channel detec-
tion method, either correlation or threshold, make a significant
difference (F1,87 = 12.66, p = 0.001). Here, finding corresponding
channels using a thresholding approach leads to more waves being
rejected at this stage compared to the correlation method (mean
difference = 3.2, se = 0.9). In contrast, it is only the negative enve-
lope method that is significantly affected by the choice of wave
inspection point (F1,81 = 40.60, p < 0.001). Using peak detection
as opposed to zero crossing measures yielded significantly more
waves (mean difference = 2.8, se = 0.6). Lastly, wave density was
significantly affected by the choice of EEG reference when abso-
lute as opposed to relative criteria where used to define the cut-off
amplitude thresholds (F1,97 = 15.69, p <0.001). When using aver-
age mastoid reference, absolute thresholding detected, on average,
6.335 (se = 1.6) more waves per minute, whereas the value for the
relative thresholding was similar to the average reference values.
A direct comparison between the default values of the toolbox
and the criteria used by Massimini et al. (2004) revealed highly
significant differences between the two methods (F1,12 = 64.29,
p <0.001), whereby the default settings tended to detect a mean
of 15.9 waves/min (se = 1.7), and the original criteria yielded only
2.3 waves/min (se = 0.6).
3.2. Wave amplitude
Wave amplitude indicates the amplitude value in microvolts
at the strongest negative peak among all the individual chan-
nels found to be part of the wave. Overall, the mean maximum
amplitude was found to be 103.3 ␮V (se = 9.1) across all partic-
ipants. Significant main effects were found for artifact removal,
canonical method, amplitude threshold, and the number of chan-
nels. For the canonical method parameter, there was an expected
increase in median amplitude of 15.6 ␮V (se = 6.5), when the four
region method is compared to the negative envelope. Only a
8 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12
single interaction was found between EEG reference and the type
of amplitude criteria used (F1,65 = 6.68, p = 0.012). Not surprisingly,
using an absolute threshold resulted in non-significant differences
in amplitude between average and mastoid references (F1,80 = 0.54,
p = 0.464), however, if a relative threshold was used the mastoid ref-
erences waves were significantly larger by a mean value of 24.5 ␮V
(se = 11.0, F1,79 = 4.92, p = 0.029).
Does changing the EEG reference result in different waves being
detected (ones with different amplitude), or are the same waves
found but the change in reference changes the maximum negative
channel? To explore this issue we examined the unique and shared
waves detected between both methods in a single, representative
participant. Using a relative threshold of 5 standard deviations for
both methods resulted in effective thresholds of 45 ␮V for average
reference compared to 85 ␮V for average mastoids. The compar-
ison of unique and shared waves found that of the fewer waves
found using the mastoids reference (1491–953 respectively), 14%
were unique to the method, as opposed to 40% unique waves for
the average reference. Despite the different waves found for each
method, both unique and shared waves showed a consistent larger
negative amplitude for mastoid reference (amplitude difference of
41.4 ␮V and 32.1 ␮V respectively). Thus, the differences in ampli-
tudes found for the interaction can largely be attributed to the
generally higher EEG amplitudes found using the mastoid refer-
ence.
The direct comparison between amplitudes for the default
values used in the toolbox and the original criteria was found
to be significant (F1,20 = 43.24, p <0.001). Default settings had a
mean maximum channel amplitude of −86 ␮V (se = 10), opposed
to −177 ␮V (se = 9), for the original detection criteria. Given the
large discrepancy found for wave density between the methods,
we explored the differences in unique and shared waves for a sin-
gle participant where for the default settings 1319 waves were
detected but only 672 waves for the original criteria. Of the 672
waves, 193 of them were unique to the original detection criteria
(28.7%), yet the difference in amplitudes between the two meth-
ods was identical for unique waves (43.6 ␮V) and shared waves
(43.3 ␮V). The individual standard deviation of amplitudes was
fairly stable across parameter settings yet had a relatively large
grand mean of 75.8 ␮V (se = 12.9) across all participants. Thus the
original criteria was both selective for larger waves in general, but
the common waves were also of higher amplitude (mostly due to
the average mastoid reference).
Only two parameters showed significant effects on ampli-
tude deviation; canonical method (F1,79 = 3.99, p = 0.049) and the
number of channels (F1,27 = 4.39, p = 0.046). The negative enve-
lope method resulted in more consistent amplitudes by a mean
of 20.2 ␮V (se = 10.1), compared to the regional reference. It is
worth noting that the mean amplitude deviation was also sub-
stantially lower when the data had been preprocessed using ICA
(mean = 54.4 ␮V, se = 6.3) compared to without (mean = 97.2 ␮V,
se = 23.4); and that the difference was likely to only be at trend
levels due to the high variability between subjects using standard
preprocessing (F1,15 = 3.72, p = 0.078).
3.3. Wave globality
Globality had a mean of 32.6% (se = 0.7) across all parameter set-
tings, indicating that about a third of channels were involved in
each wave. Most parameter changes resulted in significant differ-
ences to the globality measure. Key parameters were the choice
of EEG reference, the canonical method, as well as the channel
detection method and its corresponding threshold setting. Using
the cross-correlation method to detect slow waves in individ-
ual channels generally resulted in significantly higher globality
(mean difference = 18.7%, se = 0.5). Yet the correlation approach is
more sensitive to changes in correlation threshold compared to the
adjustment in absolute thresholding, where changes of 0.1 in the
threshold had expected changes of 8.5% and only 2.8% respectively.
Of the two significant interactions; EEG reference with canon-
ical method (F1,69 = 342.89, p < 0.001), and canonical method with
channel detection method (F1,39 = 24.59, p < 0.001), the former is
the most telling. When using the regional method there is a sig-
nificant but relatively small effect for higher globality when using
average referenced data compared to average mastoids (mean dif-
ference = 3.2%, se = 1.4). The opposite is the case for the default
envelope method where mastoid referenced data shows a signifi-
cantly higher proportion of channels in each wave in comparison
to average reference (mean difference = 27.4%, se = 0.7). This differ-
ence was further explored by comparing the unique and shared
waves in a single recording. 1317 waves were found in the average
referenced data, compared to just 870 waves using the mastoid ref-
erence with 744 waves common to each method (85.5% of mastoid
waves). In this example, both unique and shared waves showed
similar differences in the globality measure (31.4% and 32.2% dif-
ferences respectively). Thus, the large difference in globality found
is a general property of different wave forms created by the choice
of reference (illustrated in Fig. 5 for shared comparison, and Fig. 6
in the topographic density).
Individual participant’s level of globality deviated by a mean
of 11.7% (se = 0.4), yet this variability was significantly affected
by a host of parameter settings. Notably, this was the only out-
come measure where the search window parameter has a significant
impact. Changing the amount of time allowed for corresponding
channels to reach criteria values of 100 ms changes deviation by
2.6% (se = 0.4). All the specified second-order interactions also had a
significant effect. The most informative interaction was the interac-
tion between the EEG reference and type of canonical wave chosen
(F1,90 = 31.72, p < 0.001). For both EEG references, there was a sig-
nificant different between canonical methods, however for average
reference the variability in globality increases using the negative
envelope method compared to the regional method (mean differ-
ence = 1.5%, se = 0.2), whereas when using mastoid referenced data
the opposite was true (mean difference = −2.7%, se = 0.7).
3.4. Travel angle
Directional statistics such as the primary angle of slow wave
travel cannot be compared in the same way as linear statistics. This
is clear when taking the example that the angle 355 ◦ is much closer
to 0 than 40 ◦ for example, and that the mean of angle 355 and 5 is
0, not 180. To resolve this problem we decomposed the angle into
its directional components, and examined whether the anterior to
posterior preference in direction previously reported was affected
by any of the parameters (Massimini et al., 2004; Menicucci et al.,
2009; Murphy et al., 2009). Only the choice of EEG reference had
a significant effect on the anterior-posterior direction of travel for
the longest streamlines (F1,79 = 5.377, p = 0.023). For the mastoid
reference, the mean slow wave travels from the anterior to the
posterior on the scalp (mean y-component = −0.153, se = 0.179).
On the other hand, if the data is average referenced, we found
that the mean direction is reversed and waves tended to travel
in the posterior-to-anterior direction (mean y-component = 0.166,
se = 0.131). However, neither of these two directional preferences
for mastoid or average reference was significantly different from
zero (t10 = −0.856, p = 0.393; t10 = 1.269, p = 0.207 respectively).
While angular deviation is derived using angular statistics, it
is a linear statistic and as such is subject to the same model
used to assess previous outcome measures. Overall, there was a
high amount of deviation in the primary directions of the trav-
 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12 9

A.
300 90
60
90
60
120 60 120 60

40 150 40

150 30 150 30
200 20 20

100
180 0 180 0

100
50
210 330 210 330

0 20 40 60 80 100 240 300 0 20 40 60 80 100 240 300

270 270

B.

80 100
60

50
40
20
0
0
−20 −50

-40
−60
-100

-80 −150

−100

-120 -200
-400 0 400 -400 0 400

150
100

100 50

50
0
−50

0
-100
−50

−150

-100
-200
−150
−250

-200 -300
-400 0 400 -400 0 400

Fig. 5. The effect of EEG reference on slow wave detection.

Exploration of parameter settings revealed that one of the most influential choices is the EEG reference used. In this analysis we examined the average reference (most often
used in general EEG research), and the average mastoid (typical in sleep research for scoring and visualization) for a single participant keeping all other parameters constant.
A, two summary measures of all detected slow waves in the example participant. Globality indicates the percentage of channels recorded where a slow wave was found.
Angle map shows the primary angle of travel for the longest displacement stream; where in the mastoid reference there is a tendency to travel from anterior to posterior,
the opposite is found for average reference. B, examples of the same wave found for each reference choice for a wave early in the recording, and another late in the recording.
While the waves have common ‘hot-spots’ of highest amplitude waves in the topography, the butterfly plot of the waves can look quite different as the average reference
finds positive going waveforms in approximately half the channels.

eling slow waves (mean = 77.1◦ , se = 0.7). Only two changes to
parameter settings had a significant, albeit fairly minor, effect on
this deviation. Angle deviation was slightly higher when using
the negative envelope method compared to the four regions
method (mean difference = 1.8, se = 0.91). Moreover, increasing the
amplitude threshold likewise increased the angular deviation by
0.6 ◦ (se = 0.23) for the relative amplitude threshold and by 0.8 ◦
(se = 0.17) for every 10 ␮V increase in absolute threshold. This
increase is likely due to decrease in the number of waves found
from which the deviation is estimated, and not actual deviations in
the angles when using higher thresholds.
3.5. Topographic density
We examined topographical differences resulting from the
interaction of changing the EEG reference and the canonical wave
method. The TFCEANOVA analysis revealed significant main effects
for both factors and a significant interaction between the fac-
tors (see Fig. 6). The percentage measure indicated how often
the channel was part of the detected slow wave. The main effect
of changing EEG reference was present at all 256 channels and
peaked around central-posterior channels and lateralfrontal chan-
nels (peak channel E223; F1 = 654.70, p < 0.001). Channels in the
10 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12

Fig. 6. Topographic density, the effect of EEG reference and canonical waveform.
Density is defined at the percentage of the total number of waves found in each channel. As expected from previous research, the frontal-central region showed the highest
density of slow waves over all parameter settings (four topographies in the top left). Topographies are also shown for the mean difference between each of the factors
and the interaction. Significant channels are marked using a white dot. Given that the incidence in the frontal region is the same across references, this implies that the
average-mastoid reference found significantly more global waves compared to the average reference (see text Section 3.3), which was particularly apparent in central and
posterior regions (top right). The choice of canonical reference, either negative envelope or four regions resulted in significantly higher proportion of waves in highly frontal
regions and central-posterior regions. The two parameters interacted to have differential effects most prominent in the central region and also in highly posterior channels.

mastoid referenced data set had a higher percentage of active chan-
nels with a mean of 26.9% (se = 3.3). The main effect of canonical
method had significant effects over two clusters of channels: one
broad frontal (107 total channels; peak channel E37; F1 = 186.56,
p <0.001); and the other central-posterior (55 total channels; peak
channel E100; F1 = 52.53, p = 0.002). In the frontal cluster, the four
regions method showed a greater proportion of active channels
(mean difference = 5.3%, se = 0.6). However, in the posterior cluster,
it was the negative envelope method that had a significantly higher
proportion of active channels (mean difference = 3.1%, se = 0.3).
For the interaction effect, 3 distinct clusters of channels were
found in the frontal, central and posterior regions. The frontal inter-
action (15 total channels; peak channel E24; F1 = 24.96, p = 0.02),
showed the same main effects described above (higher density for
mastoid and regional canonical waves), except that the differences
between canonical methods was not as pronounced when using the
mastoid reference (mean difference = 3.9%, se = 0.5), as when using
the average reference (mean difference = 7.5%, se = 0.9); likely due
to a ceiling effect of mastoid referenced data. The central interaction
cluster (35 total channels; peak channel E89; F1 = 93.35, p <0.001)
showed a dissociation of effects with the regional method showing
a higher wave proportion than negative envelope for mastoid ref-
erenced data (mean difference = 2.1%, se = 1.0), but the reverse for
average referenced data (mean difference = 4.9%, se = 0.9). Finally,
the posterior channels also showed a dissociative interaction effect
(38 total channels; peak channel E165; F1 = 59.49, p <0.001). In pos-
terior channels, the negative envelope method showed a higher
proportion of active channels for mastoid referenced data (mean
difference = 4.9%, se = 0.6), while for averaged reference data there
was no significant difference.
4. Discussion
The analysis of EEG spectral power density has proven
immensely useful in both basic and clinical sleep research. Spectral
power analysis is beginning to be complemented by the assessment
of the individual underlying waveforms, leading to further progress
in the field. An important requirement is therefore the ability
to compare methods and results between studies and research
groups. The SWA-toolbox primarily aims to provide researchers
and clinicians with the tools for the detection and analysis of
the individual waves found during sleep, and standardizes these
methods such that they can be reliably reproduced and shared.
Understanding the variability of sleep slow waves, and realizing
that detection methods and parameters have significant effects on
outcome measures is essential. The toolbox offers great flexibility of
parameters to reproduce results obtained in previous studies using
different methods, and to explore the consequences of different
parameter choices.
Exploiting the flexibility of parameter settings lead to two sur-
prising results. First, the number of channels used in the detection
and analysis did not have a significant impact on several major
outcome measures, including the total number of detected waves.
Moreover, even when the number of channels had a statistically
 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12
significant effect, the effect size was relatively low (e.g. the mean
amplitude changed by only 5 ␮V when doubling the number of
channels used). Thus, while using high-density electrode arrays is
crucial for examining topographical changes with good resolution,
there is a rich source of previously recorded data using sparse arrays
that could be usefully re-analyzed (e.g. full clinical polysomnog-
raphy recordings). A second surprising result were the marked
differences in wave detection depending on the parameters chosen.
Specifically, the parameters employed in several previous studies
(Mölle et al., 2002; Massimini et al., 2004) tended to detect fewer
waves overall, and those that were detected tended to have higher
amplitude and to travel more globally. This observation is consis-
tent with the recent proposal that there may be two distinct types
of slow waves: larger waves found in relative isolation during the
earlier epochs in the transition from sleep to wake (type I) and con-
tinuing a relatively low incidence during later stages; and smaller,
more local slow waves found throughout consolidated sleep (type
II) (Siclari et al., 2014). Therefore, the criteria used in early stud-
ies may have privileged type I waves, with decreased sensitivity
to type II waves of smaller amplitude. The toolbox default criteria
are, by design, on the opposite side of this trade-off, with increased
sensitivity to all slow waves, with the potential for an increase in
false-positives.
We found little evidence for a theoretical frontal bias when
using the mastoids reference. The average reference, which would
eliminate this potential bias, introduces its own bias due to the
global nature of many slow waves. The more channels which show
a negative deflection, the more negative the mean of these chan-
nels becomes. The result of the average reference is that essentially
half the slow waves are seen as positive-going waveforms, their
amplitude is significantly reduced, and their relationship to one
another is difficult to track. This sets an effective upper-bound of
wave globality to approximately 50%. In turn, the missed detection
of large globally traveling waves gives rise to potentially inconsis-
tent results in traveling direction. While the potential for a frontal
bias still remains when using the mastoids reference, the trade-off
when using the average reference is high. As such we recommend
persisting with the average-mastoid reference since its theoretical
flaws appear to have few practical consequences; especially given
the larger possible effects of other changes to detection parame-
ters. The mastoids reference also allows for consistent comparison
to previous work with few electrodes (i.e. <20), where the average
reference becomes impractical.
From a broader perspective, these results send a clear and
important message: slight, seemingly arbitrary changes in detec-
tion settings can have large and widespread effects on the number
and structure of slow waves found. This finding highlights the
importance of explicitly stating how waves were detected and
justifying those choices with reference to the potential effects
demonstrated here. It also points to a need for standardized param-
eter settings such that results across studies can be adequately
compared. It was not the goal of the current analysis to determine a
set of optimal parameter settings, only to show the effects settings
have on outcome measures. The development of an optimal range
of parameters requires substantial agreement on what constitutes
a ‘true’ slow wave. This could potentially be achieved in the future
using a panel of experts manually examining individual slow waves
and creating a gold standard dataset to which automatic procedures
can be compared for specificity and sensitivity. The functions in
the toolbox for dataset exploration and manually adding waves,
or changing waves at the channel level, are well-suited to such a
future study.
The search for optimal default parameters is ongoing and iter-
ative. Future studies will hopefully converge on a set of cardinal
features of EEG slow waves that can then serve to inform and
enhance the detection procedure. At present, the recommended
11
procedure would be to thoroughly evaluate which outcome mea-
sures are of primary interest prior to data analysis, and manually
examine a small subset of parameters that have significant effects
on those particular outcome measures. This could be done iter-
atively on a subset of the recordings in the study using the
toolbox explorer, and then generalized to automatically examine
the remaining recordings. Given that an entire night of recording
can be fully analyzed from preprocessed data using the toolbox
within a matter of minutes, and additional functions allow the com-
parison of results, such a processing pipeline can be highly effective.
Of course, studies may wish to report two (or more), combinations
of parameter settings, and interpret the differences according to
the effects of these parameters differences reported here.
One crucial parameter was explicitly omitted from in-depth
analysis: the topography of individual wave origins. Wave origin
has been previously defined at the scalp level as the channel to
show the earliest negative peak. From a neurophysiological per-
spective, the origin should represent the initial population that
enters the hyper-polarized, down state. However, since these small
populations of synchronized neurons are unlikely to be visible in
the EEG, the current definition of wave origin is necessarily coarse.
Possible refinements include the determination of the first chan-
nel to show a downward zero-crossing; the start of the longest
traveling stream; or the projected intersection of the starts of trav-
eling streams moving in different directions. Further studies should
explore which of these alternatives produce the most robust results
consistent with intracranial recordings.
4.1. Future directions
Functions within the toolbox are constantly being updated with
improved coding, faster implementations of algorithms, feature
requests, and bug-fixes. This refinement is expected to continue
and grow as more researchers begin to use the toolbox for their
analysis, request additional features, and ideally contribute their
own code to the toolbox. The Github repository can be used to
gain an overview of the current short-term task list by using the
search feature for the word “TODO”. Alternatively, the ‘issues’ sec-
tion of the Github page contains several open and closed issues
such as bug reports and their fixes as well as potential enhance-
ments to the toolbox and allows for discussion between users.
Mid- and long term plans include: creation of a simple GUI for
the loading and selection of parameter estimates which are cur-
rently run through a script; the implementation and validation of
automatic detection algorithms and saw tooth waves, as well as
the automatic detection of eye movements during REM to examine
their potential relationship; the implementation and optimization
of a larger variety of parameter settings and methods for spindle
detection (Ferrarelli et al., 2007) using the same methods devel-
oped for slow waves to detect epileptic events and how they travel
over the scalp; implementation of source-localization algorithms
and the development of traveling calculations through the three
dimensional source space of the brain and new corresponding visu-
alization features (Murphy et al., 2009). There will no doubt be
additional features as more researchers begin to explore the fea-
tures of the toolbox, new research made possible by the toolbox is
conducted, and novel hypotheses are generated. Through the use
of git for version control, the purposeful factorization of code, and
the emphasis on heavy commenting within the written functions,
the toolbox is designed to be both user friendly to researchers and
clinicians who wish to simply analyze their data, as well as to those
who want to contribute additional functions and expand on the
possibilities the toolbox offers.
12 A. Mensen et al. / Journal of Neuroscience Methods 274 (2016) 1–12
Acknowledgements
We thank Giulio Bernardi, Melanie Boly, Anna Castelnovo, and
Francesca Siclari for their contributions in developing and test-
ing features of the toolbox; William Marshall for his assistance
in statistical analysis; and Chiara Cirelli for her helpful com-
ments and discussion in preparing the manuscript. Armand Mensen
was supported by a Swiss National Science Foundation grant (id:
P300P3 158505).
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.jneumeth.2016.
09.006.
References
Amzica, F., Steriade, M., 1998. Electrophysiological correlates of sleep delta waves.
Electroencephalogr. Clin. Neurophysiol. 107 (August (2)), 69–83.
Bölsterli Heinzle, B.K., Fattinger, S., Kurth, S., Lebourgeois, M.K., Ringli, M., Bast, T.,
et al., 2014. Spike wave location and density disturb sleep slow waves in
patients with CSWS (continuous spike waves during sleep). Epilepsia 55 (April
(4)), 584–591.
Buchmann, A., Ringli, M., Kurth, S., Schaerer, M., Geiger, A., Jenni, O.G., et al., 2011.
EEG sleep slow-wave activity as a mirror of cortical maturation. Cereb. Cortex
21 (March (3)), 607–615.
Delorme, A., Makeig, S., 2004. EEGLAB: an open source toolbox for analysis of
single-trial EEG dynamics including independent component analysis. J.
Neurosci. Methods 134 (March (1)), 9–21.
Fattinger, S., Jenni, O.G., Schmitt, B., Achermann, P., Huber, R., 2014. Overnight
changes in the slope of sleep slow waves during infancy. Sleep 37 (February
(2)), 245–253.
Ferrarelli, F., Huber, R., Peterson, M.J., Massimini, M., Murphy, M., Riedner, B.A.,
et al., 2007. Reduced sleep spindle activity in schizophrenia patients. Am. J.
Psychiatry 164 (March (3)), 483–492.
Huber, R., Ghilardi, M.F., Massimini, M., Tononi, G., 2004. Local sleep and learning.
Nature 430 (July (6995)), 78–81.
Iber, C., Ancoli-Israel, S., Chesson, A., Quan, S., 2007. The AASM Manual for the
Scoring of Sleep and Associated Events: Rules, Terminology and Technical
Specifications. American Academy of Sleep Medicine, Westchester.
Jones, B.E., 2005. From waking to sleeping: neuronal and chemical substrates.
Trends Pharmacol. Sci. 26 (November (11)), 578–586.
Krueger, J.M., Rector, D.M., Roy, S., Van Dongen, H.P.A., Belenky, G., Panksepp, J.,
2008. Sleep as a fundamental property of neuronal assemblies. Nat. Rev.
Neurosci. 9 (December (12)), 910–919.
Mölle, M., Marshall, L., Gais, S., Born, J., 2002. Grouping of spindle activity during
slow oscillations in human non-rapid eye movement sleep. J. Neurosci. 22
(December (24)), 10941–10947.
Marzano, C., Moroni, F., Gorgoni, M., Nobili, L., Ferrara, M., De Gennaro, L., 2013.
How we fall asleep: regional and temporal differences in
electroencephalographic synchronization at sleep onset. Sleep Med. 14
(November (11)), 1112–1122.
Mascetti, L., Muto, V., Matarazzo, L., Foret, A., Ziegler, E., Albouy, G., et al., 2013. The
impact of visual perceptual learning on sleep and local slow-wave initiation. J.
Neurosci. 33 (February (8)), 3323–3331.
Massimini, M., Huber, R., Ferrarelli, F., Hill, S., Tononi, G., 2004. The sleep slow
oscillation as a traveling wave. J. Neurosci. 24 (August (31)), 6862–6870.
Menicucci, D., Piarulli, A., Debarnot, U., d’Ascanio, P., Landi, A., Gemignani, A., 2009.
Functional structure of spontaneous sleep slow oscillation activity in humans.
PLoS One 4 (10), e7601.
Mensen, A., Khatami, R., 2013. Advanced EEG analysis using threshold-free
cluster-enhancement and nonparametric statistics. Neuroimage 67 (February
(15)), 111–118.
Mensen, A., Khatami, R., 2015. EEG Permutation Testing: An Open-source Matlab
Toolbox Implementing Various Permutation Strategies for EEG Statistical
Analysis [cited 2015 Jun 9] Available from: http://dx.doi.org/10.5281/zenodo.
14787.
Murphy, M., Riedner, B.A., Huber, R., Massimini, M., Ferrarelli, F., Tononi, G., 2009.
Source modeling sleep slow waves. Proc. Natl. Acad. Sci. U. S. A. 106 (February
(5)), 1608–1613.
Nobili, L., Ferrara, M., Moroni, F., De Gennaro, L., Russo, G.L., Campus, C., et al., 2011.
Dissociated wake-like and sleep-like electro-cortical activity during sleep.
Neuroimage 58 (September (2)), 612–619.
Pascual-Marqui, R.D., Lehmann, D., 1993. Comparison of topographic maps and the
reference electrode: comments on two papers by Desmedt and collaborators.
Electroencephalogr. Clin. Neurophysiol. Evoked Potentials Sect. 88 (November
(6)), 530–531.
Perrin, F., Pernier, J., Bertrand, O., Giard, M.H., Echallier, J.F., 1987. Mapping of scalp
potentials by surface spline interpolation. Electroencephalogr. Clin.
Neurophysiol. 66 (January (1)), 75–81.
Riedner, B.A., Vyazovskiy, V.V., Huber, R., Massimini, M., Esser, S., Murphy, M., et al.,
2007. Sleep homeostasis and cortical synchronization: III. A high-density EEG
study of sleep slow waves in humans. Sleep 30 (December (12)), 1643–1657.
Saper, C.B., Scammell, T.E., Lu, J., 2005. Hypothalamic regulation of sleep and
circadian rhythms. Nature 437 (October (7063)), 1257–1263.
Siclari, F., Bernardi, G., Riedner, B.A., LaRocque, J.J., Benca, R.M., Tononi, G., 2014.
Two distinct synchronization processes in the transition to sleep: a
high-density electroencephalographic study. Sleep 37 (October (10)),
1621–1637.
Siegel, J.M., 2009. The neurobiology of sleep. Semin. Neurol. 29 (September (4)),
277–296.
Stadelmann, K., Latshang, T.D., Lo Cascio, C.M., Tesler, N., Stoewhas, A.-C., Kohler,
M., et al., 2013. Quantitative changes in the sleep EEG at moderate altitude
(1630 m and 2590 m). PLoS One 8 (10), e76945.
Tucker, D.M., 1993. Spatial sampling of head electrical fields: the geodesic sensor
net. Electroencephalogr. Clin. Neurophysiol. 87 (September (3)), 154–163.
Vyazovskiy, V.V., Olcese, U., Hanlon, E.C., Nir, Y., Cirelli, C., Tononi, G., 2011. Local
sleep in awake rats. Nature 472 (April (7344)), 443–447.
Werth, E., Achermann, P., Borbély, A.A., 1997. Fronto-occipital EEG power
gradients in human sleep. J. Sleep Res. 6 (June (2)), 102–112.
Wilhelm, I., Kurth, S., Ringli, M., Mouthon, A.-L., Buchmann, A., Geiger, A., et al.,
2014. Sleep slow-wave activity reveals developmental changes in
experience-dependent plasticity. J. Neurosci. 34 (September (37)),
12568–12575.