Psychophysiology, 54 (2017), 608–619. Wiley Periodicals, Inc. Printed in the USA.

C 2017 Society for Psychophysiological Research

V
DOI: 10.1111/psyp.12810

Localized component filtering for electroencephalogram artifact

rejection

~ a,b
MARCOS DELPOZO-BANOS AND CHRISTOPH T. WEIDEMANN a,c
a
Department of Psychology, Swansea University, Swansea, Wales, UK
b
Swansea University Medical School, Swansea, Wales, UK
c
Department of Psychology, University of Pennsylvania, Philadelphia, PA, USA

Blind source separation (BSS) based artifact rejection systems have been extensively studied in the
electroencephalogram (EEG) literature. Although there have been advances in the development of techniques capable
of dissociating neural and artifactual activity, these are still not perfect. As a result, a compromise between reduction
of noise and leakage of neural activity has to be found. Here, we propose a new methodology to enhance the
performance of existing BSS systems: Localized component filtering (LCF). In essence, LCF identifies the artifactual
time segments within each component extracted by BSS and restricts the processing of components to these segments,
therefore reducing neural leakage. We show that LCF can substantially reduce the neural leakage, increasing the true
acceptance rate by 22 percentage points while worsening the false acceptance rate by less than 2 percentage points in a
dataset consisting of simulated EEG data (4% improvement of the correlation between original and cleaned signals).
Evaluated on real EEG data, we observed a significant increase of the signal-to-noise ratio of up to 9%.
Descriptors: EEG, Young adults

Measurements of natural processes are inevitably contaminated by techniques, and in particular independent component analysis
extraneous signals (henceforth “noise”) from various sources. Such (ICA; Hyv€arinen et al., 2004). These outperform other methods in
noise can pose serious problems for the interpretability of the sig- rejecting high amplitude noise, such as contamination from eye
nal, especially when its magnitude rivals or exceeds that of the sig- movements (Daly, Nicolaou, Nasuto, & Warwick, 2013).
nal. For almost 100 years, researchers and clinicians have been Activity recorded at each EEG sensor represents a combination
able to noninvasively record brain activity through EEG. These of multiple sources, some of which are based on brain activity (sig-
recordings are especially vulnerable to contamination by noise, nals) and some of which are not (noise). BSS algorithms transform
because the neural signals recorded at the scalp are considerably the EEG recordings with the aim to have each dimension (compo-
smaller than other electrical activity that is regularly picked up by nent) of the data correspond to an individual source. To the extent
the sensors (e.g., due to muscle activity or interference from elec- that this separation of sources is successful and that artifactual sour-
tric activity in the vicinity of the recordings). Advances in record- ces can be identified, eliminating the corresponding dimensions
ing technology have increased the signal-to-noise ratio (SNR) of and projecting the remaining components back into EEG-sensor
EEG recordings, but contamination by noise is still a major concern space will produce a clean signal (Makeig, Bell, Jung, & Sejnow-
(Muthukumaraswamy, 2013; Fatourechi, Bashashati, Ward, & ski, 1996).
Birch, 2007). The steps involved in this approach are summarized in the left
Because noise in EEG recordings is typically considerably larg- panel of Figure 1. The data are initially pre-processed, for example
er than the neural signal, it can be quite obvious when a particular by applying filters and by rejecting sensors and/or epochs with
epoch is contaminated by noise. It is often desirable to remove the exceptionally high levels of noise that could interfere with the fol-
noise from the signal rather than simply to discard contaminated lowing BSS step. Often the next step is to classify each extracted
epochs. Common approaches to separate signal from noise in EEG component as either neural or artifactual, and to process artifactual
recordings involve the application of blind source separation (BSS) components (usually by completely rejecting them). Finally a
cleaned EEG signal is reconstructed by inverting the projection
This work was supported by a BIAL Foundation grant (#48/12) to from the BSS step using the clean(ed) components, followed by
CTW. We would like to thank Sivatharshini Sangaralingham for her any processing procedures that work better on the cleaned signal.
assistance with data collection. We would like to thank Dr. Hugh Nolan The literature contains a great variety of architectures following
for providing the full synthetic data set used in this research.
Python implementations of LCF, ADJUST, and FASTER are avail- the described procedure. For example, the classification of compo-
able online at https://github.com/mdelpozobanos/eeglcf, https://github. nents into clean and artifactual can be manual or automatic. Auto-
com/mdelpozobanos/eegadjust, and https://github.com/mdelpozobanos/ matic classifiers can be trained on (usually manually) labeled sub-
eegfaster, respectively. sets of the data (supervised techniques), or set up to achieve the
Address correspondence to: Christoph T. Weidemann, Swansea Uni-
versity, Department of Psychology, Singleton Park, Swansea, SA2 8PP, classification without training examples of contaminated and
Wales, UK. E-mail: [email protected]. uncontaminated data (unsupervised techniques). Due to the
608
LCF for EEG artifact rejection 609

Figure 1. Diagrams of feature rejection systems based on blind source separation (BSS). Left: the common architecture of existing systems. The actu-
al BSS component can be seamlessly interchanged and therefore it has been left out of the “Artifact Rejection Method” box, which defines how the
output of BSS is processed. Center: A combination of the common architecture and the proposed LCF method. Right: How LCF can be used without
any artifact rejection method. The labels of the inputs and output of the LCF step are explained in the caption of Figure 2.

inconvenience of labeling the data, unsupervised algorithms are Makarov, 2006; Joyce, Gorodnitsky, & Kutas, 2004). Increasing the
more common in the EEG literature, but instances of systems inte- threshold for identifying a component as artifactual would reduce
grating supervised classifiers can also be found (Shao, Shen, Ong, neural leakage at the cost of increasing the level of remaining noise.
Wilder-Smith, & Li, 2009). Some systems lack a classification We propose a novel methodology to improve the balance
stage entirely, treating each of the components equally (Castellanos between artifact rejection and retention of neural activity by focus-
& Makarov, 2006) and hybrid approaches differentially process ing the processing of BSS components: localized component filter-
both clean and artifactual components (Vorobyov & Cichocki, ing (LCF). The presented algorithm localizes time segments within
2002). components contaminated by artifacts, and directs the processing to
Unsupervised systems use a variety of rules to classify individu- these segments, keeping the remaining parts of the component in
al components as noisy. To this end, topological templates of arti- their original forms. This removes the need for a conservative
facts (Li, Ma, Lu, & Li, 2006, Viola et al., 2009) and statistical threshold on the identification of artifactual components, because
properties of their temporal and frequency representations components identified as containing noise undergo further scrutiny
(Delorme, Sejnowski, & Makeig, 2007; Greco, Mammone, Mora- and are not generally removed entirely, reducing the probability of
bito, & Versaci, 2007) have been extensively used, and the most neural leakage (i.e., the removal of neural signal).
successful approaches use combinations of some or all of such fea- Furthermore, we have designed LCF to be easily integrated
tures (Nolan et al. 2010; Mognon et al., 2010; Winkler, Haufe, & within existing BSS-based artifact rejection systems. The LCF
Tangermann, 2011). component can be directly embedded before the BSS21 step, with-
The BSS step is particularly crucial for the system’s success, and out any modification to the other steps (Figure 1, central panel).
a comparative study by Romero, Ma~nanas, & Barbanoj (2008) Alternatively, LCF can also be used by itself, without a separate
shows that, for all surveyed systems, some neural activity is rejected artifact rejection method (Figure 1, right panel). LCF is a general
along with artifacts (i.e., “neural leakage”; Castellanos and approach to the problem of neural leakage. In this paper, we
610 M. DelPozo-Ba~
nos and C.T. Weidemann

additional step after BSS to optimize both detection of noise and

reduction of leakage of neural activity. However, LCF is an inde-
pendent processing step that can be integrated with a wide range of
artifact rejection systems as illustrated below. LCF consists of the
following steps (Figure 2):

Feature Extraction
Instantaneous measurements (i.e., measures that are defined for
each time instant n), which are characteristic of noisy activity, are
extracted from the original component. Because blink, muscle, and
pop-off artifacts are localized in time, components carrying them
are characterized by bursts of activity. In this case, the component’s
voltage can be used to locate these artifacts. Similarly, high fre-
quency noise, when localized in time, translates in a sudden
increase of the amplitude of the component’s time derivative. Non-
instantaneous measurements could also be used by windowing the
components. Windows with abnormal values of variance, Hurst
exponent, and voltage range could indicate the presence of noise.
When noise is localized in frequency or has a well-defined power
distribution, it may be easier to detect within the frequency domain.
The short-time Fourier transform or the discrete wavelet transform
are two examples of tools that could be used to obtain instantaneous-
like frequency representations of the BSS components. The result of
the feature extraction step is a list of features, with Fi ½c; n represent-
ing the i-th feature obtained from component c at time instant n.

Integrator

Because instantaneous features are noisier than those based on the

whole signal, the application of an integration window around each
Figure 2. Diagram of the LCF step. The inputs and output are (C) the time instant is useful to stabilize the extracted features (Figure 3).
original BSS components, (L) the control signal pointing to the compo-
nents that will be mixed, (P) the processed or alternative components, The list of integrated features Fi ½c; n is defined as follows:
and (R) the resulting mixed components. Fi ½c; n  WI
Fi ½c; n5 X ; (1)
W ðkÞ
k2K I
develop a deliberately simple implementation in an effort to assess
the method, but the same framework can be used with more sophis-
where WI is a window of length NI, * is the convolution operator,
ticated approaches for identifying and eliminating artifacts.
and K is the integration range defined as

Localized Component Filtering K5½maxð0; NI =22nÞ; minðNI 21; NI =21NC 2nÞ; (2)

Even though the different components identified by BSS techni- with NC the length of the component. This range covers the entire
ques generally do not perfectly separate signal and noise, separat- integrating window except when it reaches the beginning or end of
ing the signal into different components does facilitate the component. The denominator of Equation 1 is a normalizing
identification of artifacts. We therefore propose to apply LCF as an factor that effectively transforms the integration to a weighted

Figure 3. Examples of F1 and F2 defined in the implemented (LCF) block. Their integrated version Fi is smoother, allowing for a more robust detec-
tion of Ai (in the figure we expanded the amplitude of Ai to the length of the ordinate). For illustrative purposes, we set b 5 0 for the calculation of Ai
(see Equation 8) in this figure.
LCF for EEG artifact rejection 611

average around each time instant. It also counteracts the boundary The resulting components can then be back-projected to the
effect of the convolution, so that artifacts at the beginning and end original signal space. For each component, only those samples that
of components can be correctly detected. have been identified to either contain artifacts or to lie near artifacts
are affected by these processing steps. This should reduce the leak-
age of neural activity and thus produce superior results to methods
Decision Logic that eliminate entire components whenever components do not sep-
Once features are extracted and integrated, we need to create a vec- arate signal and noise sufficiently well.
tor that signals the presence of noise for each component c at each Although we suggest here that LCF could be used without a
separate artifact rejection method, its integration within a tool
time instant n: A½c; n5f ðfFi ½c; nj8igÞ, where f is the decision log-
that classifies each component with respect to the noise it con-
ic function and Fi ½c; n is the integrated feature vector described tains has some important benefits. Specifically, cleaning only
above. Similar to the classification block of a traditional artifact those components that have been identified as containing arti-
rejection tool, the decision logic function can take many forms. facts allows for more aggressive LCF detection of artifacts by
The simplest implementation would be a fixed threshold. More reducing the risk of false positives. Moreover, if the classifica-
advanced approaches include supervised and unsupervised learning tion tool differentiates between types of artifacts (blinks, pop-
algorithms, such as support vector machines or expectation maxi- off, white noise, electrocardiogram, etc.), LCF could be adapted
mization. Overall, these are generally more accurate options at the to exploit this information to more accurately detect the differ-
expense of higher computational costs and complexity. In the deci- ent kinds of artifacts.
sion logic, we default to a simple binary logical vector that indi-
cates the presence or absence of artifacts at each time instance. Limitations of LCF

When applied to components identified as containing artifactual

Mixer activity, LCF serves to focus the artifact removal on specific
The final step requires the output of the artifact rejection system, instances of artifactual activity within these components. Thus,
the processed components P½c; n, to be mixed with the original LCF can only decrease the amount of reduced noise relative to
unprocessed components C½c; n (in the absence of a separate arti- the complete removal of the affected component. To the extent
that LCF is successful, any reduction in the amount of removed
fact rejection system, P½c; n can be simply set to zeros). The mix-
noise should be small relative to the preserved neural activity
ing is governed by a mixing signal M defined as follows:
that would otherwise have been removed along with the noise.
M½c; n5A½c; n  WM ; (3) It is important to note that LCF is aimed at improving the clean-
P
where WM is a window of length NM and WM ½n51, used to ing of components with localized artifacts, such as blinks, muscle,
round the edges of detected areas, which, in turn, smooths the tran- or pop-off artifacts. Components with continuous contamination,
sitions in the mix. A½c; n refers to the decision vector defined such as continuous white noise, or line noise cannot be effectively
above, and * is the convolution operator. The resulting (cleaned) tackled by LCF. Such components should be either removed entire-
component is defined as ly or appropriately filtered.

Q½c; n5P½c; n  M½c; n1C½c; n  ð12M½c; nÞ: (4) Example Implementation of LCF

Q thus is a mixed vector that consists of either C (where the deci- We now present applications of LCF to both simulated and real
sion logic indicates an absence of artifacts), P (where the decision EEG data sets. In an effort to illustrate the principles of LCF, we
logic indicates the presence of artifacts), or a mix of both (around kept the implementation deliberately simple, but note that the
transitions between samples with and without artifacts) that choices at each processing step can and should be adapted to best
fit the aims of the particular application.
smooths the transition to avoid discontinuities (Figure 4).
Feature extraction. Because, within (mainly) artifactual compo-
nents, absolute voltage of artifactual EEG activity often exceeds
the amplitude of activity from neural sources, a threshold on abso-
lute voltage can provide a simple instantaneous indicator for the
presence of noise. Likewise, because artifactual activity is often
associated with sudden changes, large absolute values of the first
time derivative of the voltage can also be a simple and instanta-
neous indicator for the presence of noise.
We normalized both of these features as follows:

X2mX^
HðXÞ5 ; (5)
sX^
Figure 4. Mixing of the component C whose features are depicted in
Figure 3 when its processed version P is all zeros. Transitions between
where m and s respectively refer to the mean and standard devia-
C and P are smoothed to avoid discontinuities (to facilitate the figure’s tion of the variable denoted in the subscript. X^ is a trimmed version
representation, we modified the amplitude of M to match that of the of the vector X that excludes those samples that deviate more than
ordinate such that the shaded proportion of the ordinate below M three sX from mX.
denotes the proportion of P in Q). Thus, features are formally defined as
612 M. DelPozo-Ba~
nos and C.T. Weidemann

Figure 5. Examples of the simulated data set created by Nolan et al. (2010). Contaminated channels and blink artifacts are particularly prominent.

F1 5HðjCjÞ (6) regions and we used this value throughout unless otherwise noted.
and For the sake of simplicity, and because features were normalized to
z-scores, we used a single threshold (manually set to s 5 1) for all
F2 5HðjdC=dtjÞ (7) features.
for each component C determined in the BSS step1. To produce an overall decision vector that combines the
information from all features, we simply combined the individ-
Integrator. We used a Hamming window of 0.2 s as the integra- ual decision vectors for each feature with the logical “or” opera-
tion window WI (Figure 3). tor Ú:
_
Decision Logic. We used a simple threshold (s) to determine
A½c; n5 Ai ½c; n:
i
whether a given feature indicated the presence of noise (which cor-
responds to a value of 1 in the decision vector A), so that Furthermore, whenever more than 75% of a trial was labeled as
(
1 if maxð½Fi ½c; n2b; . . . ; Fi ½c; n1bÞ > s noise, we rejected the entire trial. Likewise, any component where
Ai ½c; n5 (8) more than 75% of the samples were classified as noise was rejected
0 otherwise; entirely.
where b corresponds to a buffer around each time point to allow
samples directly adjacent to samples identified as noise to also be Mixer. We used a Hamming window of 0.1 s as the mixer window
classified as noise. We found that a buffer of 0.1 s produced contin- WM (Figure 4).
uous noise regions that were well separated from adjacent noise
Materials
1. If components are classified as clean or noisy by the artifact rejec-
tion method, only artifactual components need to be processed by the We evaluated (LCF) using an artificial data set created by Nolan
LCF block. et al. (2010), as well as an unpublished data set recorded by us.
LCF for EEG artifact rejection 613

Figure 6. Examples of the real data set recorded from a memory task. Artifacts are substantially more complex than in the synthesized (EEG).

The artificial data set consists of 100 epochs of simulated EEG We first high-pass filtered the EEG data at 0.5 Hz using a finite
data to which artifacts were randomly added following the proce- impulse response (FIR) filter of order 99 and Hamming window.
dure described in Delorme et al. (2007). In addition, a random We chose a FIR filter to avoid distorting the phase of the signal or
number of channels were contaminated with white noise, and some introducing any ripple in the pass-band. We then removed the base-
of the epochs were also corrupted by a high-amplitude (30–150 line from the signal and rejected highly artifactual channels and
lV) low frequency (1–3 Hz) signal. A total of 47 files were created epochs based on the z-scores of several statistics (variance, correla-
following the described procedure. Figure 5 shows two example tion, and Hurst exponent) following the same pre-processing proce-
epochs from this simulated data set. dures described in Mognon et al. (2010). Next, we used the
We also evaluated LCF using a real data set recorded from 18 INFOMAX ICA algorithm as the BSS step (Bell & Sejnowski,
participants while they were engaged in a recognition memory 1995) and separately applied each of the following artifact rejection
task. We recorded a total of 133 channels (128 scalp channels, 2 methods:
mastoid channels, and 3 EOG channels) using a BioSemi Active ADJUST: The automatic EEG artifact detector based on the
Two system and a sampling rate of 500 Hz. We partitioned the data joint use of spatial and temporal features (ADJUST; Mognon et al.,
from each participant into 576 epochs starting 0.5 s before the onset 2010) characterizes artifactual independent components (ICs) by
of a test item and extending to 1 s after stimulus onset. Figure 6 both temporal and spatial features, specifically: kurtosis, variance,
shows two example epochs from this data set. and the spatial distribution of IC activation. These are then automat-
ically classified as clean or artifactual ICs by an expectation maxi-
Method mization algorithm (Dempster, Laird, & Rubin, 1977). The ICs
identified as artifactual are then set to zero.
We assessed the proposed LCF methodology, using three different FASTER: The fully automated statistical thresholding for
BSS-based algorithms. All systems shared the structure shown in EEG artifact rejection (FASTER; Nolan et al., 2010) uses main-
Figure 1 and only differed in the artifact rejection step. ly temporal measures to detect artifactual ICs. In particular, this
614 M. DelPozo-Ba~
nos and C.T. Weidemann

method relies on temporal correlations with electrooculogram

(EOG) channels, spectral and voltage gradients, Hurst exponent
and spatial kurtosis.
ICAW: (ICAW; Castellanos & Makarov, 2006) processes each
IC by thresholding its discrete wavelet transform (DWT) coeffi-
cients. We manually set a threshold equal to the 99.5 percentile of
the absolute wavelet coefficients.
Finally, we post-processed the cleaned reconstructed EEG by
correcting the baseline and interpolating the rejected channels
using the spherical spline technique (Perrin, Pernier, Bertrand, &
Echallier, 1989). In addition, we then re-referenced the signal to
Figure 7. Graphic representation of the true and false rejection and false
common average and interpolated those parts of the signal identi- acceptance concepts.
fied as artifactual as described in Mognon et al. (2010).
We applied each of the above artifact rejection techniques to
the data sets twice: once with and once without the LCF step. This
allowed us to assess the effectiveness of LCF under different sce- leakage). Similarly, the FAR relates to the amount of noise
narios. We also tested LCF in isolation (i.e., without using a sepa- retained, the inverse of the rejection of noise (i.e., lower FARs
rate artifact rejection method). For this case, we set the processed reflect higher levels of noise rejection). If necessary, an overall
signal that is fed to the LCF step to zeros and processed all ICs. measurement of the error of the method can be computed as
ðFRR1FARÞ=2.
To isolate the performance of LCF, we extracted the simulated
Analysis of Simulated EEG Data
signal, X, and the simulated signal contaminated with noise, Y, after
The simulated data set has the advantage of offering fully con- the pre-processing stage and calculated the rejected activity R from
trolled testing conditions. Specifically, the artifactual activity, Z, the output of the BSS21 step (cf. Figure 1).
can be extracted from the artificial EEG by subtracting the simulat-
ed signal, X, from the simulated signal that is contaminated by
noise (Y): Z5Y2X. Similarly, the rejected signal, R, is defined as ERP Analysis of EEG Data
the difference between the contaminated signal and the output of A common way to analyze EEG data is to average EEG activity
the particular artifact rejection method. across events to produce event related potentials (ERPs). To
A simple difference between original and cleaned signals is reduce the complexity of the data set, we partitioned the 128
insufficient to properly assess the performance of LCF. Because EEG channels into nine regions of interest (ROI; Figure 8): a
LCF aims specifically at reducing neural leakage, we need to be central ROI (RC) surrounded by eight ROIs labeled R0, R45,
able to differentiate between the amount of noise rejected, the R90, R135, R180, R225, R270, and R315 starting with the mid-
true rejection rate (TRR), and the amount of neural activity lost, frontal region, centered around 0 , and going clockwise in 45
the false rejection rate (FRR). In order to do so, we define the increments. We averaged the sensors within each region, com-
following measurements to be used with the artificial data set: puted the ERP for each region, and averaged those ERPs across
TRR: The TRR is the proportion of the artifactual activity that participants.
has been successfully removed. It is defined as In real data sets, without labeled artifacts, the quantities we
used to assess performance on the simulated data set are not readily
X
minðZðnÞ; RðnÞÞ available. Because the baseline period of each event in the real data
n2XX
TRR5 ; (9) set lacked external stimuli that were synchronized across events,
8n
ZðnÞ we expect the average ERP to be fairly constant and close to 0 lV
for clean EEG data. However, a small proportion of artifacts (such
where X5fn j 8sgnðZðnÞÞ5sgnðRðnÞÞg and sgn is the sign function as those produced by motion) can cause significant deviations from
(cf. Figure 7). that mean due to their large amplitude.
FRR: The FRR quantifies the amount of neural activity that is Events were locked to a stimulus presentation, which should
removed in the artifact rejection effort and is defined as produce ERP deflections reflecting its processing. Noise, however,
can obscure these synchronous effects of stimulus processing in the
X X EEG and attenuate the resulting ERP. On the other hand, removal
n2ðX\WÞ
ðRðnÞ2ZðnÞÞ1 2X
n=
RðnÞ of neural activity will also reduce the SNR and attenuate the ERP.
FRR5 X ; (10)
XðnÞ We quantify the SNR as the ratio between the mean signal
8n amplitude (l) and the standard error of the mean (SE) across events
(Gonzalez-Moreno et al., 2014). Note that this results in a normal-
with W5fn j j ZðnÞj < jRðnÞjg (cf. Figure 7).
ized mean as a function of time:
It is sometimes more convenient to reframe the performance of the
artifact rejection system in terms of the true acceptance rate (TAR), lðtÞ
SNRðtÞ5 ; (11)
which corresponds to 12FRR, and the false acceptance rate SEðtÞ
(FAR), which corresponds to 12TRR. Note that the TAR relates to
the amount of neural signal retained, in other words, the inverse of We averaged this measure over an 800 ms time window starting at
the neural leakage (i.e., higher TARs reflect lower levels of neural stimulus onset. We directly compared SNR for artifact rejection
LCF for EEG artifact rejection 615

Figure 8. Regions of interest for the ERP analysis. We refer to the central ROI as RC and to the eight surrounding regions starting with the mid-
 
frontal region (centered around 0 ) and going clockwise in 45 increments as R0, R45, R90, R135, R180, R225, R270, and R315.

pipelines with and without LCF and also calculated the difference
between these measures Table 1. Mean Percentage (SE) of Processed Independent Com-
ponents (ICs) and Pearson’s Correlation (r) Between the Origi-
Df 5SNRf1LCF 2SNRf (12) nal and Cleaned Signal for the Different Methods Applied to
the Artificial Data Set.

ICs (%) r
None 0.77 (0.07) t(46)510.530
None1LCF 100.00 (0.00) 0.95 (0.07) SE 5 0.08
Difference (%) 123.73% p < .001

ADJUST 0.97 (0.07) t(46)5-1.744

ADJUST1LCF 2.05 (0.20) 0.96 (0.07) SE 5 0.01
Difference (%) -0.14% p 5 .088

FASTER 0.93 (0.06) t(46) 5 8.639

FASTER1LCF 10.57 (0.36) 0.97 (0.06) SE 5 0.05
Difference (%) 14.39% p < .001

ICAW 0.87 (0.05) t(46)514.401

ICAW1LCF 100.00 (0.00) 0.91 (0.05) SE 5 0.01
Figure 9. Mean percentage and SE of the Pearson’s correlation (r) Difference (%) 14.26% p < .001
between the original and cleaned signal for the different methods Note. Fisher’s transform was applied before computing mean and SE
applied to the artificial data set with and without LCF. Fisher’s trans- and back-transformed to obtained the presented results. The right most
form was applied before computing mean and SE and back-transformed column lists the results of dependent t-tests between the correlations of
to obtained the presented results. the methods with and without LCF.
616 M. DelPozo-Ba~
nos and C.T. Weidemann

Figure 10. Mean percentage and SE of true acceptance rate (TAR), and false acceptance rate (FAR) for the different methods applied to the artificial
data set with and without LCF. Higher TARs reflect lower levels of neural leakage, while lower FARs reflect higher rejection of noise.

and the improvement percentage driven by LCF The different methods varied greatly with respect to the propor-
Df tion of ICs they processed, ranging from just over 2% on average
Df% 51003 %; (13) for ADJUST to 100% for ICAW. Because LCF is only applied to
SNRf
tagged ICs, the conservative classification of ADJUST limited the
with f indicating the artifact rejection method. possible impact of LCF.
These measures, however, can only provide rough estimates of While the neural leakage of FASTER and ICAW was similar
the real SNR. A visual examination of the ERPs complements the (76% and 72% of TAR, respectively), the introduction of LCF had a
assessment of the effects of LCF. To facilitate this assessment, we smaller (yet still substantial) effect on the latter. ICAW cleans tagged
examined the differences between ERPs corresponding to each of components through DWT thresholding and provides an alternative
the artifact rejection methods with and without the LCF step. To signal P, instead of fully rejecting them as ADJUST and FASTER
the extent that LCF successfully reduces only neural leakage, LCF do. The effectiveness of this cleaning process constraints the poten-
should increase the amplitude for these differential ERPs in the tial improvements on neural leakage from LCF. For both FASTER
period after stimulus onset. and ICAW, however, the addition of LCF resulted in a substantial
increase of the TAR (reduction of neural leakage) with comparative-
Results and Discussion ly small increases in FAR (noise rejection penalization). Surprising-
ly, the application of LCF in isolation resulted in relatively good
Results for Simulated EEG Data
performance with a TAR 4 percentage points lower than that of
Figure 9 and Table 1 show the correlations between the original ICAW1LCF and the lowest FAR of all techniques.
and the cleaned synthetic signal, while Figure 10 and Table 2 show In some cases, we observed that the addition of LCF increased
the percentages of TARs and FARs. In general, the addition of the proportion of removed artifactual activity (i.e., it reduced the
LCF resulted in an improved correlation between the original and FAR). This may seem counter-intuitive given that LCF limits how
cleaned signal, as well as in an increase in TAR without corre- much of an IC is removed. However, removing non-artifactual
sponding increases in FAR (differences in FAR were small relative parts of an IC, produces noise, which can result in an increase of
to the associated standard deviations). In other words, it reduced either FRR or FAR.
the neural leakage (increase in TAR) without heavily penalizing These results suggest that the proposed approach works as
the rejection of artifacts (small increase in FAR). expected: it reduces the neural leakage while preserving the ability

Table 2. Mean Percentage (SE) of True and False Acceptance Rates for the Different Methods Applied to the Artificial Data Set

TAR (%) FAR (%)

None 100.00 (0.00) t(46) 5 -26.293 100.00 (0.00) t(46) 5 -43.113
None1LCF 80.79 (0.00) SE 5 0.73 14.79 (0.00) SE 5 1.98
p < .001 p < .001
ADJUST 96.84 (0.47) t(46) 5 6.652 30.19 (4.77) t(46) 5 11.661
ADJUST1LCF 99.04 (0.47) SE 5 0.33 34.83 (4.77) SE 5 0.40
p < .001 p < .001
FASTER 75.97 (2.10) t(46) 5 10.745 22.63 (2.48) t(46)51.897
FASTER1LCF 97.63 (2.10) SE 5 2.02 24.10 (2.48) SE 5 0.77
p < .001 p 5 .064
ICAW 71.74 (0.93) t(46) 5 29.168 34.72 (1.62) t(46) 5 8.186
ICAW1LCF 84.97 (0.93) SE 5 0.45 35.85 (1.62) SE 5 0.14
p < .001 p < .001

Note. The third and last columns lists the results of dependent t-tests comparing the acceptance rates of the methods with and without LCF shown in
the respective previous column.
LCF for EEG artifact rejection 617

potential effectiveness of LCF, but at the same time uses a proc-

essed (“cleaned”) version of the ICs as alternative signal.
Figure 11 shows the SNR for each method with and without
LCF, while Figure 12 and Table 3 show the comparative results
in the form of D and D% . The addition of LCF to a given system
resulted in improvements of the SNR for all cases, with an aver-
age increase of 7% for ADJUST, 9% for FASTER, and 4% for
ICAW.
The increase in SNR should be interpreted together with two fac-
tors: First, the LCF prototype used here as a proof of concept is a rath-
er simple one. It applies the same analysis to all tagged components,
including those with continuous noise, such as white-noise. These
have no salient time instants, and therefore LCF retains the compo-
Figure 11. Mean and SE of the SNR for all channel clusters combined nent in its original form. Second, the heterogeneity of the signal quali-
obtained with the different methods applied to the real data set with and ty within the real data set introduces great variability in the results. In
without LCF. cases where the signal quality is particularly good, BSS performs bet-
ter, reducing the scope for improvement by LCF, while the opposite
to reject artifacts. When the system performs conservatively, tag- is true for poor signal quality instances. This can also be observed in
ging very few components in a bid to reduce the probability of neu- the range of the results. LCF improves SNR by a maximum of 4.00
ral leakage, the scope for improvement by LCF is reduced. With points, 2.62 points, and 0.85 points for ADJUST, FASTER, and
LCF, the system can be tuned to be more aggressive in the tagging ICAW, respectively, while in the worst cases it only decreases it by
of components, rejecting more noise, without corresponding 0.51, 1.88, and 0.50, respectively.
increases in neural leakage. When looking at the differential ERPs, comparing methods
with and without LCF (Figure 13), the differences are largely
Results for EEG Data confined within the ERP in response to stimulus onset—no big
differences in amplitude are discernible in the baseline periods
The results for the real EEG data set are comparatively noisier, but or more than about 500 ms past stimulus onset. This effectively
they tend to mirror those from the simulated data. On average, means that LCF increased the amplitude of the stimulus locked
ADJUST only tagged 5.57% ICs as artifactual, limiting the scope activity (neural activity) while maintaining the rejection of
for improvement that could be gained by LCF. In fact, this average noise. This figure also illustrates the total amount of noise and
is substantially inflated by three outliers rejecting 22, 30, and 33 neural activity removed by LCF when it is used without a sepa-
ICs. Without these outliers, the average drops to only 2%. In five rate artifact rejection method as well as the distortion intro-
out of 18 cases, ADJUST did not tag any IC for rejection. duced by ICAW’s alternative signal. Overall, LCF by itself
FASTER, on the other hand, rejected 15.45% of ICs on average, (None1LCF), performed better than we anticipated, despite the
resulting in a higher potential to reject noise, but also increased simplicity of the current implementation. The LCF-only imple-
propensity for neural leakage and hence bigger scope for improve- mentation rejected a good amount of noise while retaining most
ment by LCF. ICAW processes all ICs, which maximizes the of the ERP.

Figure 12. Mean and SE of the difference D for each channel cluster and method. Highlighted in red and blue, cases where the nominal difference is
lower and higher than 0 respectively.
618 M. DelPozo-Ba~
nos and C.T. Weidemann

Table 3. Mean, SE, and Range of the Difference D for all Channel Clusters Combined and Each Method, Together with the Corre-
sponding LCF Improvement D%

None ADJUST FASTER ICAW

Mean (SE) -0.25 (0.04) 0.23 (0.05) 0.28 (0.06) 0.13 (0.02)
[min, max] [-2.73, 0.75] [-0.51, 4.00] [-1.88, 2.62] [-0.50, 0.85]
D% -7.06% 6.81% 8.86% 3.90%
t-test t (161) 5 -6.428 t (161) 5 4.700 t (161) 5 4.383 t (161) 5 8.474
SE 5 0.04 SE 5 0.05 SE 5 0.06 SE 5 0.02
p < .001 p < .001 p < .001 p < .001
q < .001 q < .001 q < .001 q < .001

Note. Refer to Table A1 for results for individual channel clusters. The last four rows show results for dependent t-tests comparing the SNR for each
method with and without LCF along with associated (uncorrected) p-values and False Discovery Rate q-values.

Noise Rejection versus. Preservation of Neural Activity Benefits of LCF were more pronounced in cases where the
BSS algorithm found a particularly bad solution to the problem
The right balance between the rejection of noise and preservation
of neural activity and noise dissociation (see Appendix). LCF
of neural activity depends on the particular application. Consider-
can thus function to guard against inadequate levels of noise
ations such as the prominence of the targeted signal, the quality of
rejection and/or excessive neural leakage in cases where signal
the recordings, as well as the prevalence and nature of noise, deter-
and noise are not well separated by the BSS algorithm. Indeed,
mine both the effectiveness of the BSS artifact removal and, in
we demonstrated that the application of LCF by itself (i.e., with-
turn, the benefit (or lack) of an additional LCF step. Additionally,
out a previous BSS step) can substantially reduce the noise while
the trade-offs between failing to reject artifactual activity and
limiting neural leakage.
removing signal vary. With LCF, we provide a tool that can reduce
We designed LCF to be compatible with any of the existing
neural leakage at a given level of noise rejection.
systems based on BSS. Moreover, its integration is straight for-
ward and requires no substantial changes to the system itself,
Conclusion
facilitating its adoption. LCF’s ability to retain neural activity
We propose LCF as a novel methodology to boost the performance allows for more stringent classification components (i.e., for
of existing EEG artifact rejection systems that are based on BSS more liberally labeling components as artifactual), while at
techniques. The method takes the original and processed compo- the same time limiting the danger of signal loss. The implemen-
nents as inputs, and mixes them so that the processed (cleaned) tation of LCF used here as a proof of concept is a very simplis-
components replace the original ones only when an artifact is tic one, based on the voltage amplitude and its speed of change.
detected. Quantitative and qualitative analyses on simulated and More sophisticated artifact detection methods are likely to yield
real data sets demonstrated benefits of LCF, especially in the better results and, in turn, would allow for even more sensitive
reduction of neural activity leakage. artifact rejection systems.

Figure 13. Difference between the ERPs, averaged across subjects, for multiple artifact rejection methods without and with LCF. ERPs for methods
without LCF are subtracted from ERPs for methods with LCF. Lines correspond to the ERP difference of each brain region specified in Figure 8, so
that the top line is RC and subsequent ones are R0, R45, R90, . . ., R315. The scale of the y-axes differs between panels as indicated.
LCF for EEG artifact rejection 619

References
Bell, A. J. & Sejnowski, T. J. (1995). An information-maximization Nolan, H., Whelan, R., & Reilly, R. (2010). FASTER: Fully automated
approach to blind separation and blind deconvolution. Neural Computa- statistical thresholding for EEG artifact rejection. Journal of Neuro-
tion, 7(6), 1129–1159. doi: 10.1162/neco.1995.7.6.1129 science Methods, 192, 152–162. doi: 10.1016/j.jneumeth.2010.07.
Castellanos, N. P. & Makarov, V. A. (2006). Recovering EEG brain sig- 015
nals: Artifact suppression with wavelet enhanced independent compo- Perrin, F., Pernier, J., Bertrand, O., & Echallier, J. (1989). Spherical splines
nent analysis. Journal of Neuroscience Methods, 158, 300–312. doi: for scalp potential and current density mapping. Electroencephalogra-
10.1016/j.jneumeth.2006.05.033 phy and Clinical Neurophysiology, 72, 184–187. doi: 10.1016/0013-
Daly, I., Nicolaou, N., Nasuto, S., & Warwick, K. (2013). Automated artifact 4694(89)90180-6
removal from the electroencephalogram: A comparative study. Clinical Romero, S., Ma~ nanas, M. A., & Barbanoj, M. J. (2008). A comparative
EEG and Neuroscience, 44, 291–306. doi: 10.1177/1550059413476485 study of automatic techniques for ocular artifact reduction in spontane-
Delorme, A., Sejnowski, T., & Makeig, S. (2007). Enhanced detection of ous EEG signals based on clinical target variables: A simulation case.
artifacts in EEG data using higher-order statistics and independent com- Computers in Biology and Medicine, 38, 348–360. doi: 10.1016/
ponent analysis. Neuroimage, 34(4), 1443–1449. doi: 10.1016/ j.compbiomed.2007.12.001
j.neuroimage.2006.11.004 Shao, S.-Y., Shen, K.-Q., Ong, C. J., Wilder-Smith, E., & Li, X.-P. (2009).
Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likeli- Automatic EEG artifact removal: A weighted support vector machine
hood from incomplete data via the EM algorithm. Journal of the Royal approach with error correction. IEEE Transactions on Biomedical Engi-
Statistical Society. Series B (methodological), 1–38. neering, 56, 336–344. doi: 10.1109/tbme.2008.2005969
Fatourechi, M., Bashashati, A., Ward, R. K., & Birch, G. E. (2007). EMG Tallon-Baudry, C., Bertrand, O., Delpuech, C., & Pernier, J. (1997). Oscil-
and EOG artifacts in brain computer interface systems: A survey. Clini- latory c-band 30–70 Hz) activity induced by a visual search task in
cal Neurophysiology, 118, 480–494. doi: 10.1016/j.clinph.2006.10.019 humans. The Journal of Neuroscience, 17. 722–734.
Gonzalez-Moreno, A., Aurtenetxe, S., Lopez-Garcia, M.-E., del Pozo, F., Viola, F. C., Thorne, J., Edmonds, B., Schneider, T., Eichele, T., &
Maestu, F., & Nevado, A. (2014). Signal-to-noise ratio of the MEG sig- Debener, S. (2009). Semi-automatic identification of independent com-
nal after preprocessing. Journal of Neuroscience Methods, 222, 56–61. ponents representing EEG artifact. Clinical Neurophysiology, 120,
doi: 10.1016/j.jneumeth.2013.10.019 868–877. doi: 10.1016/j.clinph.2009.01.015
Greco, A., Mammone, N., Morabito, F., & Versaci, M. (2007). Semi-auto- Vorobyov, S. & Cichocki, A. (2002). Blind noise reduction for multisenso-
matic artifact rejection procedure based on kurtosis, Renyi’s entropy ry signals using ICA and subspace filtering, with application to EEG
and independent component scalp maps. International Journal of Medi- analysis. Biological Cybernetics, 86(4), 293–303. doi: 10.1007/s00422-
cal, Health, Biomedical and Pharmaceutical Engeneering, 1, 466–470. 001-0298-6
Hyv€arinen, A., Karhunen, J., & Oja, E. (2004). Independent component Winkler, I., Haufe, S., & Tangermann, M. (2011). Automatic classifica-
analysis, volume 46. John Wiley & Sons. doi: 10.1002/0471221317 tion of artifactual ica-components for artifact removal in EEG signals.
Joyce, C. A., Gorodnitsky, I. F., & Kutas, M. (2004). Automatic removal Behavioral and Brain Functions, 7(1). doi: 10.1186/1744-9081-7-30
of eye movement and blink artifacts from EEG data using blind compo-
nent separation. Psychophysiology, 41, 313–325. doi: 10.1111/j.1469-
(RECEIVED November 6, 2015; ACCEPTED November 10, 2016)
8986.2003.00141.x
Li, Y., Ma, Z., Lu, W., & Li, Y. (2006). Automatic removal of the eye blink
artifact from EEG using an ICA-based template matching approach. Supporting Information
Physiological Measurement, 27, 425. doi: 10.1088/0967-3334/27/4/008
Makeig, S., Bell, A., Jung, T., & Sejnowski, T. (1996). Independent com- Additional supporting information may be found in the online
ponent analysis of electroencephalographic data. In Advances in Neural
Information Processing Systems, pages 145–151. Cambridge, MA: MIT version of this article:
Press. Appendix A: Table A1. Mean, SE (in Parentheses), and Range
Mognon, A., Jovicich, J., Bruzzone, L., & Buiatti, M. (2010). ADJUST:
An automatic EEG artifact detector based on the joint use of spatial and (in Square Brackets) of D for Each Channel Cluster and Meth-
temporal features. Psychophysiology, pages 229–240. doi: 10.1111/ od, Together with the Corresponding LCF Improvement D%
j.1469-8986.2010.01061.x
Muthukumaraswamy, S. D. (2013). High-frequency brain activity and mus- Appendix B: LCF Effects on the Time-Frequency Domain
cle artifacts in MEG/EEG: A review and recommendations. Frontiers Appendix C: Results of a Case with Bad BSS Solution
in Human Neuroscience, 7. doi: 10.3389/fnhum.2013.00138