Home Search Collections Journals About Contact us My IOPscience

An unsupervised eye blink artifact detection method for real-time electroencephalogram

processing

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2016 Physiol. Meas. 37 401

(http://iopscience.iop.org/0967-3334/37/3/401)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 193.140.216.7
This content was downloaded on 27/06/2016 at 22:13

Please note that terms and conditions apply.

 Institute of Physics and Engineering in Medicine Physiological Measurement

Physiol. Meas. 37 (2016) 401–417 doi:10.1088/0967-3334/37/3/401

An unsupervised eye blink artifact

detection method for real-time
electroencephalogram processing
Won-Du Chang, Jeong-Hwan Lim and Chang-Hwan Im
Department of Biomedical Engineering, Hanyang University, Seoul, Korea

E-mail: [email protected]

Received 13 August 2015, revised 27 December 2015

Accepted for publication 14 January 2016
Published 19 February 2016

Abstract
Electroencephalogram (EEG) is easily contaminated by unwanted
physiological artifacts, among which electrooculogram (EOG) artifacts due
to eye blinking are known to be most dominant. The eye blink artifacts are
reported to affect theta and alpha rhythms of frontal EEG signals, and hard
to be accurately detected in an unsupervised way due to large individual
variability. In this study, we propose a new method for detecting eye blink
artifacts automatically in real time without using any labeled training data.
The proposed method combined our previous method for detecting eye blink
artifacts based on digital filters with an automatic thresholding algorithm. The
proposed method was evaluated using EEG data acquired from 24 participants.
Two conventional algorithms were implemented and their performances
were compared with that of the proposed method. The main contributions
of this study are (1) confirming that individual thresholding is necessary for
artifact detection, (2) proposing a novel algorithm structure to detect blink
artifacts in a real-time environment without any a priori knowledge, and (3)
demonstrating that the length of training data can be minimized through the
use of a real-time adaption procedure.

Keywords: electroencephalogram (EEG), electrooculogram (EOG), ocular

artifact, eye blink
S Online supplementary data available from stacks.iop.org/PM/37/401/
mmedia

(Some figures may appear in colour only in the online journal)

0967-3334/16/030401+17$33.00 © 2016 Institute of Physics and Engineering in Medicine Printed in the UK 401
Physiol. Meas. 37 (2016) 401 W-D Chang et al

1. Introduction

Scalp electroencephalogram (EEG) is easily contaminated by other physiological signal

sources (Hagemann and Naumann 2001, Durka et al 2003, Lawhern et al 2013) because there
are many signal sources near the brain, such as heart, teeth, facial muscles, and eyes (Croft
and Barry 2000, Jiang et al 2007, Yong et al 2008). The signals from other non-brain sources,
known as artifacts, are mixed with the brain signal and interfere with the accurate analysis of
EEG. For this reason, most EEG studies adopt a preprocessing procedure for obtaining arti-
fact-free EEG signals (Jung et al 2000, Hagemann and Naumann 2001, Mognon et al 2010).
Eye blinks are one of the most influential sources of artifacts contaminating electroen-
cephalogram (EEG) signals (Hagemann and Naumann 2001). In most EEG paradigms using
visual stimuli or visual feedbacks, EEGs are inevitably contaminated by eye blink artifacts
because spontaneous blinking is an autonomic process of the human body (Pult et al 2013).
Although theta band is distorted most severely when an EEG signal is contaminated by eye
blink artifacts, eye blink artifacts have also been reported to significantly affect alpha and
beta band EEG signals (Hagemann and Naumann 2001). Therefore, cleaning EEG signals
contaminated by eye blink artifacts is an important preprocessing procedure to obtain accurate
EEG analysis results.
The general approach for the EEG data cleaning is either rejecting the artifact-contaminated
time windows (contaminated epochs) or recovering the artifact-free data (Croft and Barry
2000), where both approaches commonly require artifact detection process. Although detect-
ing contaminated epochs is obviously indispensable for the rejection approach (Krishnaveni
et al 2006, Kook et al 2008), the recovery approach also utilizes the artifact detection pro-
cedure in order to minimize distortion of source data by applying the recovery procedure
only to the detected regions (Krishnaveni et al 2006) or applying different correction factors
for eye blink artifacts and other muscle movement artifacts (Gratton et al 1983). Detection
of artifacts is also helpful for independent component analysis (ICA) of EEG signals as the
artifact-contaminated independent components need to be detected for the automated artifact
removal (Shao et al 2009, Mognon et al 2010).
There have been many previous studies on the automatic detection of eye blink artifacts,
which can be roughly classified into two main categories: feature-based approaches and dis-
tance-based approaches. Feature-based approaches determine the presence of artifacts in an
epoch using features extracted after certain preprocessing procedures such as band-pass filter-
ing (Hoffmann and Falkenstein 2008, Klein and Skrandies 2013) and component separation
(Mognon et al 2010). Various features have been introduced, including maximum absolute
value (Nolan et al 2010), kurtosis (Barbati et al 2004, Mognon et al 2010), entropy (Barbati
et al 2004), second-order difference (Klein and Skrandies 2013), and the Teager-Kaiser energy
operator (Breuer et al 2014).
Distance-based approaches use a template representing the generic shape of the target arti-
fact. After the template is built by an expert (Aarabi et al 2009) or generated by averaging
sample artifact signals (Kim and McNames 2007), the distance (or similarity, which is the
inverse of distance) between the template and a part of the signal is calculated. A variety of
methods such as support vector machine (SVM; (Shao et al 2009), dynamic time warping
(DTW; (Chang and Im 2014), cross-correlation (Li et al 2006, Kim and McNames 2007), and
correlation between EEGs in different channels (Durka et al 2003) have been used to calculate
the distance.
In general, both approaches to detect artifacts need to set a certain threshold value to make
binary decisions on whether a specific epoch is contaminated by artifacts or not. The sim-
plest way is to use a constant threshold value determined empirically from previous analyses.

402
Physiol. Meas. 37 (2016) 401 W-D Chang et al

However, applying the same threshold value can result in significant detection errors in some
individuals’ EEGs due to the large individual variance in the shapes and amplitudes of the eye
blink artifacts (Galley et al 2003). Therefore, the threshold values need to be customized for
each individual’s EEG. A method that uses the variance in artifact distribution was proposed
by (Krishnaveni et al 2006); however, their method includes expert intervention for calculat-
ing the variance in the artifact distribution. In their method, a training procedure requires
an additional dataset, in which all the artifacts should be manually labeled by an expert
(Please note that most machine learning algorithms have the same requirements because
the most algorithms train their models in supervised ways). Mullen et al (2015) suggested a
semi-supervised method to determine a threshold from distribution of artifact-free data.
We categorized this method as a semi-supervised approach, because this method does not
require blink-labeled EEG data but does require labeled artifact-free EEG data to estimate the
distribution of normal data. This kind of supervised algorithms can hardly be applied to many
end-user applications often requiring real-time EEG processing.
To circumvent the limitation of the supervised approach, some studies proposed unsuper­
vised artifact detection methods, which adopted fully automatic training procedure and thus
do not require expert labeling. Mongon et al (2010) adjusted the threshold value using the
expectation-maximization (EM) algorithm; however, their method was not fit for the real-
time processing because of the time-consuming procedure of maximization. Geetha and
Geethalakshmi (2012) utilized Otsu’s thresholding technique (Otsu 1975) originally pro-
posed for image binarization, and Breuer et al (2014) used 80th percentile of the individual
data distribution as the threshold value, but both methods were still based on the empirical
parameterization.
The present study proposes a new approach for the real-time detection of eye blink artifacts
in an unsupervised manner. The proposed approach combines our previous method to detect
eye blink artifact from single-channel EEG using the first derivative sum in multiple sliding
windows (Chang et al 2015) with an automatic thresholding method introduced by (Kim and
McNames 2007). Both methods were modified to operate in real-time as they were originally
developed for off-line spike signal detection. The performance of the proposed approach is
evaluated in both simulated-real-time and real-time environments. In the validation process,
we tried to demonstrate the followings: (1) whether the use of individual threshold enhances
the accuracy of detecting eye blink artifacts in comparison with the common threshold; (2)
how well the real-time adaptation algorithm is working; (3) how we can balance false-posi-
tives and false-negatives in real-time thresholding.

2. Methods

2.1. Experimental data

Experiments were conducted in a simulated-real-time environment using an EEG dataset

acquired from 24 healthy participants (Chang and Im 2014). In order to simulate the real-
time data acquisition, each data sample from the pre-saved EEG dataset was read one by
one with a sampling frequency identical to that of real-time recording and then immediately
transferred to an artifact detection program, which means that the developed artifact detection
program could be directly applied to real-time artifact detection without any modification.
The original dataset was recorded using a multi-channel EEG recording system (ActiveTwo
AD-box™, BioSemi, Netherlands) at a sampling rate of 2,048 Hz (afterwards the data were
down-sampled to 64 Hz for the eye blink detection), while the study participants performed
spot-the-difference puzzles (participants were asked to find the differences between two

403
Physiol. Meas. 37 (2016) 401 W-D Chang et al

images shown on the left and right of a monitor). The ground-truth of eye blinking ranges
was marked manually by visual inspections of frontal channel EEG (Fp2 in the international
10–20 system) and vertical electrooculogram (EOG). We used the first 290 s EEG data for
the analysis. The study participants signed a consent form and received monetary reimburse-
ment for their participation; the study was approved by the institutional review board (IRB)
of Hanyang University.

2.2. Artifact detection using the first derivative sum in multiple sliding windows

This section briefly introduces an artifact detection method proposed by the authors (Chang
et al 2015), which adopted a simple digital filter to accurately extract the artifact ranges. The
filter can be written as follows:
F (t ) = S (t ) − S (t−∣W ∣),
(1)
where W is the width of a sliding window and S(t) is the t-th sample of the original signal.
The filter was named the ‘summation of first derivatives in a sliding window (SDW)’ because
the equation (1) was derived from the following equation:
t

F (t ) =
(2) ∑
S (k ) − S (k − 1).
k=t−n+1

As readily recognized from the above equation, the SDW filter emphasizes spikes in a
specific wave-width. To emphasize spikes with various wave-widths, SDWs in multiple win-
dow sizes should be considered since the durations of eye blink artifacts vary (Verleger 1993,
Fukuda and Stern 2005). Hence, an empirical procedure was introduced for selecting a value
among SDWs of different window sizes as a representative at time t. This procedure was
called MSDW (Multiple-window SDW) as it utilizes multiple sliding windows. For every t,
the following selection steps are performed:
(1) Calculate SDWs with different W ′s, considering a typical eye blink range. Let F W denote
an SDW with a window size of W .
(2) Choose the maximum F W at time t, denoted as MSDW(t ), as a representative if it satis-
fies the conditions below.
a. The numbers of local minima and maxima are the same within the range of [t , t−∣W ∣].
b. All of the first derivatives from time t to t−∣W ∣ +1 should be within S′(t−∣W ∣ +1) and
S′(t ), where S′(t ) represents the first derivatives at time t.
It is possible that there exist multiple W ′s that maximize F W . In such a case, the smallest
W is selected for MSDW(t ). When there are no SDWs that satisfy all of the conditions, the
minimum window size is used.
Let us denote the window size selected through the above procedure at time t as WMSDW(t ) .
Then, the artifact range is defined as
⎡ ⎤
R = ⎢T (Maxi − j) − WMSDW(T (Maxi− j)) , T (Mini) ⎥ ,
(3)⎣ ⎦
where Maxi − j − Mini > θ, Mini and Maxi are the ith local minimum and maximum MSDW
values, respectively, and T (Maxi) and T (Mini) are the time positions of the ith local maxi-
mum and minimum, respectively. The integer value j is chosen among integer values equal
to or larger than zero in order for Maxi − j − Mini to be maximized and in order for the time

404
Physiol. Meas. 37 (2016) 401 W-D Chang et al

Figure 1. The overall architecture of the proposed system.

difference between Maxi − j and Mini not to exceed the maximum window size of MSDW. We
assume that the first local maximum precedes the first local minimum in an MSDW series: any
local minimum before the first maximum is ignored when determining the order.

2.3. Proposed method for unsupervised artifact detection

This article proposes a novel method to detect eye blink artifacts in an unsupervised way.
The proposed method is composed of three main steps after EEG data acquisition. Figure 1
describes the overall architecture of the algorithm. The first main step is to calculate a feature
for assessing the presence of artifacts. The MSDW method described in the previous sec-
tion was utilized for this step because this method has several advantages over the conven-
tional eye blink artifact detection methods in terms of usability in real-time applications. First,
it does not require any expert actions such as template construction except determining the
individual threshold values. Second, it can detect artifact ranges as well as the presence of the
artifact, enhancing the overall performance of artifact detection. Third, it is computationally
light and can be applied to single frontal-channel EEG data without the need for EOG refer-
ence. Therefore, it is particularly useful in brain-computer interface applications using wear-
able EEG devices with a few frontal EEG channels.
As is illustrated in figure 1, an MSDW value is calculated using newly acquired data after
preprocessing the data with a median filter (a data buffer with the same size as the median

405
Physiol. Meas. 37 (2016) 401 W-D Chang et al

filter is necessary for this purpose). The original procedure for artifact detection is separated
into parts (the 1st step and the 3rd step) in order to insert an automatic thresholding procedure
(the 2nd step). The MSDW value is converted for automatic thresholding procedure because
the original MSDW value does not directly represent dissimilarity or similarity. Hence, we
converted the MSDW values into new values using the difference between local minima and
their adjacent maxima. If the latest MSDW value is a local minimum, the largest value among
the recent maxima was chosen to calculate the difference value. The converted MSDW feature
was defined as
f (t ) = Maxi − j − Mini,
(4)
where all of the notations are the same as in (3).
The second main step is to automatically determine the threshold from the given feature
(converted MSDW feature) series. Since storing all the feature series might be inadequate
for a long recording or might require unnecessarily large storage size, a density histogram
accumulated over time was utilized in order to investigate the feature population. The histo-
gram could be initialized with some given data and readily updated by increasing a value at a
specific bin. To calculate the histogram, the width of the histogram bin was calculated using
the following equation:
hist_width
(5) = (M − m )/(n Bin − 2),
where M is an expected maximum value of the entire distribution, m is the minimum value of
the distribution, and nBin is the number of bins. In this experiment, the number of bins was
fixed to 20, and six-standard deviation of the acquired data was assumed to M. Based on this
histogram, the threshold could be automatically determined with conventional thresholding
algorithms. We implemented three conventional algorithms and modified one of them con-
sidering our conditions. The details of the thresholding algorithms will be introduced in the
next sections.
The last main step is to detect and update the artifact region using the determined threshold.
When the current MSDW value is a local minimum at time t and the dissimilarity feature, f (t ),
is larger than the threshold, the region of the artifact is determined as
⎡ ⎤
R = ⎢T (Max) − WMax , t ⎥ ,
(6)⎣ ⎦
where Max is the local maximum to calculate f (t ), T (Max) is its time points, and WMax is
the window size of MSDW for calculating Max. Essentially, the determined region is readily
overlapped because the detection process is conducted for every local minimum. To avoid
multiple detection of the same artifact, overlapped ranges were merged into a single range.

2.4. Automatic thresholding using a fixed portion of data

Breuer et al (2014) introduced a method using a percentage of an individual data (feature)

distribution as the threshold. This can be calculated by finding m satisfying the following
equation:
m
∫0
F (k ) ⋅ dk = P ⋅ N ,
(7)

where F (k ) is a density function, P is the portion of normal data, and N is the total number of
data points in the distribution.

406
Physiol. Meas. 37 (2016) 401 W-D Chang et al

2.5. Automatic thresholding by analyzing the data distribution of offline detections

This section briefly describes an algorithm proposed by (Kim and McNames 2007), hereafter
referred to as KM. The algorithm was proposed to detect spikes in extracellular neural record-
ings in offline environment. KM utilized the kernel density function (probability density func-
tion: PDF) of local maxima to detect a spike. The PDF distribution was separated into two
modes, primary and secondary modes, where the secondary mode was assumed to contain the
spikes. The procedure of the automatic thresholding algorithm is as follows:
(1) Calculate the series of similarity (using cross-correlation in the article) and calculate PDF
using the local maxima of the series. F (k ) denotes the density at a specific similarity k.
(2) Assume that the largest mode in PDF represents normal signal and the rightmost mode
represents the spikes. If there exist only a single mode, the threshold is determined
according to the predefined signal-to-noise ratio (SNR).
(3) Find local maxima and minima in the PDF. All abscissas of the local minima between the
two modes are regarded as candidates of the threshold. The abscissas of the ith maximum
and minimum after the normal mode are denoted as Mi and mi, respectively.
(4) The largest candidate among those satisfying the following two conditions is chosen as
the threshold.
a. τmin ⩽ F (mi ) ⩽ τmax, where τmin and τmax are the minimum and maximum thresholds
known in advance, respectively.
b. F (mi ) < v ⋅ min{F (Mi ), F (Mi + 1)}, where v denotes a user-specified parameter to
eliminate the shallow valleys of the minima.

2.6. Automatic thresholding for real-time artifact detection

In this study, we modified the KM’s automatic thresholding algorithm in order for the algo-
rithm to be used in real-time environments. Unlike the offline environment, a real-time system
has little a priori knowledge on the signal characteristic, and a data distribution at a certain
time may not accurately reflect the distribution of whole data. The modification was con-
ducted to remove all the factors related to a priori knowledge and to allow robust operation
with only a small amount of data. We changed the following parts in the KM’s original algo-
rithm: First, a histogram was used instead of the PDF because the calculation of a histogram is
simpler, efficient, and faster than that of the PDF. Second, a previously determined threshold
was used instead of calculating a new threshold when there is only a single mode in the distri-
bution. Because it was empirically known that the optimal threshold does not change dramati-
cally, it was better to use the previous threshold than to determine a new threshold based on
the assumed SNR. Third, the smallest candidate was chosen instead of the largest. We found
that the smallest candidate separates the signal and artifacts better, and consequently a novel
condition was introduced to identify the shallow valleys. These changes eliminated the need
for a priori knowledge on the artifact range in the determination of the threshold. The newly-
defined condition and the modified procedure for real-time automatic thresholding are given
below (hereafter the proposed algorithm will be referred to as ‘real-time KM’):
(1) Accumulate sufficient length of data for initialization. Calculate the histogram using
MSDWs. F (k ) denotes the density at a specific dissimilarity k.
(2) If there is only a single mode, the threshold will not be changed until a secondary mode
is observed.
(3) Find local maxima and minima in the histogram. The largest maximum represents the
normal mode. All of the local minima after the normal mode become candidates for the
407
Physiol. Meas. 37 (2016) 401 W-D Chang et al

Figure 2. The experimental paradigm for the evaluation of methods. Data in the
initializing phase was used to construct the initial histogram, and the histogram was
updated in the latter phase. The initializing phase must include at least one eye blink
artifact, and a different initializing time was tested in the simulations.

threshold. The abscissas of the ith maximum and minimum after the normal mode are
denoted as Mi and mi, respectively.
(4) The smallest candidate among those satisfying the condition below is chosen as the
threshold.
a. F (Mi ) < v ⋅ F (M0 ), where v denotes a user-specified parameter to eliminate the shallow
minima.
After a threshold is chosen based on the feature distribution, an adjustment process for bal-
ancing false-positives and false-negatives may be necessary. A new optional parameter (α) is
developed for this purpose. The histogram is divided into two groups, normal and blink, based
on the originally-determined threshold (θ old), which is then adjusted to include the α-standard
deviation of the blink distribution:
new
θ(8)= θ old + α ⋅ σ.
For more comprehensive understanding of our procedure, please download the library source
codes freely available at http://cone.hanyang.ac.kr/BioEST/Eng/EyeblinkMasterOnline.zip.

3. Results and discussion

A simulated real-time environment was designed for evaluation as illustrated in figure 2. The
data were read from pre-recorded EEG data at every sampling time. The data were divided
into two sub-sections. The data in the first section were used to construct an initial histogram
(let us call this the initializing section), and the histogram was updated using the data in the
succeeding section. For the successful construction of the initial histogram, the initializing
section needed to include at least one eye blink artifact. Right after the initialization of the his-
togram, the real-time eye blink detection process started. The threshold was determined using
the methods described in section 2, and the results were evaluated by counting the number of
correctly detected (true positives) and falsely detected (false positives) events. The true and
false positives were counted only in the range from 85 to 290 s in order to sufficiently separate
the initializing section and the evaluation section. Note that the length of the initializing sec-
tion varied in the simulation study. In this study, we regarded that an artifact was correctly
detected if the peak point of the artifact was included in the detected artifact range.
For the evaluation of the automatic thresholding methods described in section 2, three dif-
ferent assessment criteria of accuracy were utilized: precision, recall, and F1 score. Precision
reflects false detection, recall represents correct detection, and the F1 score is the harmonic
mean of precision and recall. The equations for the accuracy measures are as follows:

408
Physiol. Meas. 37 (2016) 401 W-D Chang et al

Table 1. Accuracy of detecting a blinking event: M1, M2, and M3 denote different
thresholding algorithms: Common threshold, fixed portion, and the proposed algorithm
(real-time KM), respectively. Unit: %.
Precision Recall F1 Score
M1 M2 M3 M1 M2 M3 M1 M2 M3
1 90.74 100.00 97.96 98.00 86.00 96.00 94.23 92.47 96.97
2 100.00 100.00 100.00 97.89 50.53 95.79 98.94 67.13 97.85
3 85.71 85.71 86.36 85.71 85.71 90.48 85.71 85.71 88.37
4 94.12 50.00 94.12 100.00 100.00 100.00 96.97 66.67 96.97
5 100.00 100.00 97.50 64.44 62.22 86.67 78.38 76.71 91.76
6 88.89 61.54 72.73 100.00 100.00 100.00 94.12 76.19 84.21
7 98.51 100.00 98.48 100.00 69.70 98.48 99.25 82.14 98.48
8 100.00 100.00 100.00 100.00 61.84 100.00 100.00 76.42 100.00
9 91.67 100.00 94.83 98.21 91.07 98.21 94.83 95.33 96.49
10 85.71 89.66 87.88 100.00 86.67 96.67 92.31 88.14 92.06
11 65.22 53.57 83.33 100.00 100.00 100.00 78.95 69.77 90.91
12 100.00 100.00 100.00 94.59 83.78 94.59 97.22 91.18 97.22
13 88.00 84.62 90.48 91.67 91.67 79.17 89.80 88.00 84.44
14 95.24 95.24 95.24 100.00 100.00 100.00 97.56 97.56 97.56
15 78.26 58.06 100.00 100.00 100.00 100.00 87.80 73.47 100.00
16 98.11 100.00 96.30 98.11 71.70 98.11 98.11 83.52 97.20
17 100.00 93.10 93.75 16.98 50.94 84.91 29.03 65.85 89.11
18 86.96 90.00 95.00 100.00 90.00 95.00 93.02 90.00 95.00
19 100.00 100.00 100.00 88.00 70.00 86.00 93.62 82.35 92.47
20 93.75 80.00 100.00 93.75 100.00 87.50 93.75 88.89 93.33
21 97.96 100.00 100.00 96.00 46.00 94.00 96.97 63.01 96.91
22 92.54 100.00 98.21 92.54 41.79 82.09 92.54 58.95 89.43
23 90.91 47.62 95.24 100.00 100.00 100.00 95.24 64.52 97.56
24 96.77 100.00 100.00 100.00 66.67 93.33 98.36 80.00 96.55
Avg 92.46 87.05 94.89 92.33 79.43 94.04 90.70 79.33 94.20
Stdev 8.27 18.33 6.66 17.84 19.46 6.39 14.34 11.19 4.55
Min 65.22 47.62 72.73 16.98 41.79 79.17 29.03 58.95 84.21
Max 100.00 100.00 100.00 100.00 100.00 100.00 100.00 97.56 100.00

Precision =∣TP∣/(∣TP∣+∣FP∣)
Recall =∣TP∣/(∣TP∣+∣FN∣)
(9)
F1 score = 2 ⋅ Precision ⋅ Recall/(Precision + Recall),
where TP , FP , and FN denote the numbers of true positives, false positives, and false nega-
tives, respectively.
Table 1 compares the accuracies of the proposed methods implemented with three
different thresholding algorithms: (1) common threshold, (2) threshold at a fixed percentile,
and (3) real-time KM. The values of the common threshold and the percentiles were chosen
experimentally to yield the highest accuracy (140 for the common threshold; 96th percentile
for the fixed percentile threshold). All procedures except for the threshold determination
were controlled to be the same for each method, i.e. the same algorithm for histogram initial-
izing and updating was employed. The initialization time was fixed to 10 s from the end of
the first eye blink, and v for the real-time KM was fixed to 0.1 empirically. The results of the
common threshold method (M1) show the necessity for individual thresholding. Although

409
Physiol. Meas. 37 (2016) 401 W-D Chang et al

the mean detection accuracies were acceptable (92.46%, 92.33%, and 90.70% for preci-
sion, recall, and F1 score, respectively), relatively low accuracies were observed for some
subjects. The F1 scores of three subjects were lower than 80% due to the imbalance between
precision and recall. Furthermore, F1 score of one subject (#17) was lower than 30%. Due
to the large individual variability, the standard deviation of the method M1 was greater than
10%. The accuracies of a fixed percentile method (M2) were rather lower than those of the
common threshold, with an average F1 score of 79.33%. This is because the number of eye
blinks varied from 24 to 141 in our dataset (the mean and standard deviation were 59.8 and
32.43, respectively). The accuracies of the real-time KM showed the best performance in all
three assessment criteria (94.89%, 94.04%, and 94.20% for precision, recall, and F1 score,
respectively). In comparison with the common thresholding algorithm (M1), the improve-
ment of the performance of real-time KM mainly resulted from the improvement of the accu-
racies in some subjects’ data that showed low accuracy in M1 (#5, 11, and 17). Accordingly,
the mean F1 score increased by 3.50 %, but the standard deviation dramatically decreased
from 14.34 to 4.55.
Tables 2 and 3 show the results of evaluating the initialization and update procedures
of the histogram. Table 2 shows the changes in the accuracy of the proposed method with
respect to the initialization time. Time zero (0) in the table indicates the point when the first
eye blink signal finishes, since the initialization must include at least one eye blink signal.
As is clearly shown, the mean F1 score was greater than 90% even when the first five-
second signal is included. The standard deviation of the score decreased dramatically when
the initialization time exceeded 10 s, after which the standard deviation kept stable. Table 3
shows the changes in the accuracy with respect to the initialization time when histogram
updating process was not included in the procedure (the same as table 2 except that there
was no histogram updating). In this case, once a threshold was determined with the initial
histogram, the threshold was not changed at all during whole processes. In comparison with
the results in table 2, the accuracy reported in table 3 generally was relatively lowered and
became unstable. These results demonstrate the importance of histogram updating in our
proposed method.
Figure 3 shows the change of FP and FN balance with respect to the new parameter
(α) introduced in (4). The default precision and recall (when α = 0) were 94.89 and 94.13%,
respectively, and they traded off as α varied. The changes in the two accuracies were gradual
when α was between  −1 and 1, whereas recall decreased dramatically when α exceeded 1.
These results show that the α value needs to be set to a value between −1 and 1 to achieve
acceptable FP and FN ratios in eye blink artifact detection.
The speed of the proposed method was measured for all subjects using the default para­
meters used in table 1. The mean processing time was 1.5 ms with a standard deviation of 0.79
when the method was coded with MATLAB and run on a Windows 7 platform on an Intel®
i5 processor with 8 GB RAM. No significant difference among subjects was found. Since the
artifact detection procedure was performed in every 15.63 ms (64 data samples per second),
the speed of the proposed method was fast enough to be used for real-time applications.
After the simulated real-time experiments, we performed a fully real-time online experi-
ment with a subject (male, 25 years old). For this online experiment, the proposed method was
implemented with C/C++ and tested with a mobile EEG device (Enobio 8, Neuroelectrics,
Spain) with a single prefrontal EEG channel (See figure 4). A movie clip demonstration
is available in the online version of this article (please find the Supplementary Movie file
attached to this article) (stacks.iop.org/PM/37/401/mmedia). Ten eye blinks were observed
for 65 s, and all the blinks were detected successfully without any false detection.

410
 Table 2. Changes in the accuracy of the proposed method with respect to the initialization time (mean  ±  standard deviation, unit: %).
Physiol. Meas. 37 (2016) 401

Initialization
time (s) 0 5 10 20 30 40 50 60
Precision 82.18  ±  33.49 94.07  ±  14.74 96.05  ±  4.93 94.89  ±  6.66 95.02  ±  6.35 95.82  ±  4.57 95.44  ±  5 .63 95.89  ±  4.74
Recall 91.13  ±  14.05 93.76  ±  6.52 91.79  ±  9.47 94.04  ±  6.39 94.30  ±  6.15 93.97  ±  6.94 93.97  ±  8.00 93.52  ±  7.18
F1 score 79.50  ±  30.75 92.85  ±  11.38 93.54  ±  5.63 94.20  ±  4.55 94.44  ±  4.45 94.69  ±  4.23 94.50  ±  5.77 94.49  ±  4.34

411
Table 3. Changes in the accuracy with respect to the initialization time when histogram updating process was not included in the procedure
(mean  ±  standard deviation, unit: %).
Initialization
time (s) 0 5 10 20 30 40 50 60
Precision 42.89  ±  43.27 75.56  ±  35.60 81.36  ±  30.52 93.33  ±  9.79 85.25  ±  24.17 86.14  ±  22.61 95.01  ±  6.06 92.07  ±  13.21
Recall 96.71  ±  8.25 81.77  ±  33.02 91.31  ±  14.20 93.72  ±  8.60 95.40  ±  8.63 96.06  ±  3.99 93.37  ±  12.43 95.56  ±  4.70
F1 score 45.87  ±  41.03 72.26  ±  32.15 80.26  ±  26.35 92.95  ±  7.12 87.13  ±  20.56 88.41  ±  18.99 93.48  ±  8.74 93.05  ±  8.71
W-D Chang et al
Physiol. Meas. 37 (2016) 401 W-D Chang et al

Figure 3. Changes in accuracies (precision and recall) according to alpha value, which
shows the trade-off between false-positive and false-negative errors.

Figure 4. A screenshot of a real-time experiment. The proposed algorithm was tested

with a mobile EEG device with a single prefrontal EEG channel. The full movie file can
be found as a supplementary material of this manuscript.

412
Physiol. Meas. 37 (2016) 401 W-D Chang et al

The proposed method had a few optional parameters. The influences of initialization
time and the balancing factor, alpha, on the overall performance of the proposed method
were shown in the results section, where it was demonstrated that the accuracy became
stable even with a short initialization time as short as 10 ~ 20 s and the FP and FN trade-off
remained acceptable within a certain range of the balancing factor (from  −1 to 1). Table 4
shows the influence of another parameter (v) on the accuracy of the proposed method. The
parameter v was adopted to avoid misrecognition of a part of the normal (signal) distribu-
tion as an eye blink. As shown in the table 4, the influence of the parameter v was quite
smaller than expected. The accuracies increased as v increased, but became stable when
v was equal to or greater than 0.05 (5%). The proportion of the blinking signal relative to
the normal signal was about 4.17% (since the best proportion in the results of the fixed
proportion method was 96%); the proposed method exhibited stability even with twice the
original proportion.
To further confirm the general applicability of the proposed method, we applied the pro-
posed method to other EEG dataset recorded under different environment, without changing
the parameters used for the previous validation study. The additional dataset was recorded
from 20 participants using NeuroScan SynAmps2 amplifier (Compumedics USA, El Paso,
TX, USA), while the participants were performing visual oddball tasks, of which the detailed
paradigm used for the EEG recording can be found in a recent study (Kim et al 2015). EEG
signals with the length of 120 s were used for the experiment, when 20 s data from the first
blink were used for the initialization of the histogram and the data from 60 to 120 s were
used for evaluating the artifact detection accuracy. The ground-truth of eye blinking ranges
was marked manually by visual inspection of vertical EOG. Please note that any parameter
of the proposed method was not changed in this experiment, whereas the threshold for the
common threshold method was adjusted in order to yield the best detection accuracy for the
new dataset. For the statistical evaluation, all the data from 44 participants (24 original  +  20
additional experiments) were divided into two groups based on the F1 score of the common
threshold method, which were a group with low F1 (below averaged F1) and a group with high
F1 (above averaged F1). Because a group with high F1 has a very little room for improvement,
we hypothesized that the detection accuracy would be enhanced for the ‘low-F1 group’ and
unvaried for the ‘high-F1 group’ by the use of the proposed method. The detection accuracies
are summarized in table 5. In the ‘low-F1 group’, mean F1 score of the proposed method was
9.25 and 10.73 percent point higher than those of the common threshold and fixed portion
methods, respectively, when both differences turned out to be significant (p  =  0.0479 for the
common threshold, p  =  0.0085 for the fixed portion method according to Wilcoxon signed-
rank test). In the ‘high-F1 group’, there was no significant difference in F1 scores between the
proposed method and the common threshold method, while the fixed portion method showed
significantly lower F1 score than the other two methods (Bonferroni-corrected p  <  0.001 for
both methods according to paired t-test).
In spite of the significant improvement of the detection accuracies in the low-F1 group, the
accuracies are still far from the human-level accuracy in blink detection. The main reasons
for the lowered accuracy were irregular shapes of blinking artifacts as illustrated in (Chang
et al 2015) and extremely uncommon eye blinking patterns. Since the proposed method uti-
lizes a distribution of eye blinks to set a threshold, a threshold may not be correctly adjusted
when the distribution is in an extreme condition (e.g. no blinking for 10 to 20 s, or one or
more blinks in every second). We expect that this issue could be addressed by modifying
the algorithms for the construction and update of histograms, which we would like to study
further in the future.

413
 Physiol. Meas. 37 (2016) 401

Table 4. Influence of the parameter v on the accuracies of the proposed method (mean  ±  standard deviation, unit: %).

v 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1

414
Precision 71.17  ±  42.31 91.96  ±  19.06 95.68  ±  4.79 95.80  ±  4.69 95.76  ±  4.71 95.78  ±  4.71 95.78  ±  4.71 94.89  ±  6.66 94.89  ±  6.66 94.89  ±  6.66
Recall 65.84  ±  39.10 75.34  ±  31.86 80.86  ±  28.71 89.12  ±  19.66 93.31  ±  7.27 94.04  ±  6.39 94.04  ±  6.39 94.04  ±  6.39 94.04  ±  6.39 94.04  ±  6.39
F1 score 49.52  ±  41.50 77.52  ±  29.50 83.54  ±  25.15 90.48  ±  17.27 94.31  ±  4.44 94.73  ±  4.05 94.73  ±  4.05 94.20  ±  4.55 94.20  ±  4.55 94.20  ±  4.55
W-D Chang et al
Physiol. Meas. 37 (2016) 401 W-D Chang et al

Table 5. The detection accuracies for two groups divided based on the F1 score
of common threshold method. M1, M2, and M3 denote different thresholding
algorithms: Common threshold, fixed portion, and the proposed algorithm, respectively
(mean  ±  standard deviation, unit: %).
Low F1 group (N  =  14) High F1 group (N  =  30)
M1 M2 M3 M1 M2 M3
Precision 81.72  ±  17.87 78.34  ±  18.96 92.38  ±  9.57 96.14  ±  4.44 89.70  ±  20.44 96.96  ±  5.71
Recall 82.95  ±  22.05 82.16  ±  18.34 84.29  ±  15.11 98.57  ±  2.86 97.52  ±  4.32 96.84  ±  4.56
F1 score 78.19  ±  16.68 76.71  ±  10.65 87.44  ±  11.12 97.26  ±  2.66 84.42  ±  16.57 96.73  ±  3.65

4. Conclusion

In this study, we introduced a novel method for automatic detection of eye blink artifacts,
which does not require any labeled training data. The novel method was based on our previ-
ous method for eye blink artifact detection called MSDW (Chang et al 2015) together with an
automatic thresholding algorithm, both of which had been designed for offline data processing.
In order to implement a real-time eye blink detection system, the architecture of the original
MSDW was disassembled and recomposed to be readily combined with the automatic thresh-
olding algorithms. Three different automatic thresholding algorithms were combined with
MSDW and their performances were compared with each other. Among the three automatic
thresholding algorithms including ‘common threshold,’ ‘fixed proportional threshold,’ and
‘real-time KM’, the proposed ‘real-time KM’ method outperformed the other two methods.
The contributions of this study are as follows: (1) This study confirmed that individual
thresholding is necessary for artifact detection. Our simulated real-time analysis showed that
individually customized threshold resulted in higher accuracy in detecting eye blink arti-
facts than common threshold methods, which implies that individual differences in eye blink
amplitude have significant influence upon the detection accuracy. (2) This study proposed a
novel algorithm structure to detect eye blink artifacts in a real-time environment without any
need for a priori knowledge on the signal properties. The results of the present study showed
satisfactory detection accuracies in comparison with the conventional thresholding methods,
and demonstrated a successful trade-off between false-positives and false–negatives. (3) The
length of training data could be minimized by using a real-time adaptation procedure. Our
experimental results showed that the minimum training time needed to stabilize detection
accuracy could be reduced by the use of adaptation procedure (initialization time of 20 s for
the proposed method; that of 50 when no adaptation procedure was applied).
It would be interesting for the proposed method to be tested with other patterns in EEG data
such as spikes, sharp waves, or a combination of these because these EEG waveforms have
similar shapes with eye blink signals. It is expected that the proposed method can be utilized
for recognizing these patterns by changing some optional parameters.

Acknowledgments

This work was supported in part by the National Research Foundation of Korea (NRF) grant
funded by the Korea government (MSIP) (No. 2014R1A2A1A11051796) and in part by the
Basic Science Research Program through the National Research Foundation of Korea (NRF),
funded by the Ministry of Education (NRF-2014R1A1A2A16052334). The authors declare
that there is no conflict of interests regarding the publication of this paper.

415
Physiol. Meas. 37 (2016) 401 W-D Chang et al

References

Aarabi A, Kazemi K, Grebe R, Moghaddam H A and Wallois F 2009 Detection of EEG transients in
neonates and older children using a system based on dynamic time-warping template matching and
spatial dipole clustering Neuroimage 48 50–62
Barbati G, Porcaro C, Zappasodi F, Rossini P M and Tecchio F 2004 Optimization of an independent
component analysis approach for artifact identification and removal in magnetoencephalographic
signals Clin. Neurophysiol. 115 1220–32
Breuer L, Dammers J, Roberts T P L and Shah N J 2014 Ocular and cardiac artifact rejection for real-
time analysis in MEG J. Neurosci. Methods 233 105–14
Chang W-D, Cha H-S, Kim K and Im C-H 2015 Detection of eye blink artifacts from single prefrontal
channel electroencephalogram Comput. Methods Programs Biomed. 124 19–30
Chang W-D and Im C-H 2014 Enhanced template matching using dynamic positional warping for
identification of specific patterns in electroencephalogram J. Appl. Math. 2014 528071
Croft R J and Barry R J 2000 Removal of ocular artifact from the EEG: a review Clin. Neurophysiol.
30 5–19
Durka P, Klekowicz H, Blinowska K, Szelenberger W and Niemcewicz S 2003 A simple system for
detection of EEG artifacts in polysomnographic recordings IEEE Trans. Biomed. Eng. 50 526–8
Fukuda K, Stern J A, Brown T B and Russo M B 2005 Cognition, blinks, eye-movements, and pupillary
movements during performance of a running memory task Aviat. Space Environ. Med. 76 C75–85
(PMID: 16018333)
Galley N, Schleicher R and Galley L 2003 Blink parameter as indicators of driver’s sleepiness—
possibilities and limitations Vis. Vehicles X ed A Galevol vol 82 (Amsterdam: Elsevier Science
Publishers B. V)
Geetha G and Geethalakshmi S N 2012 Artifact removal from EEG using spatially constrained
independent component analysis and wavelet denoising with Otsu’s thresholding technique
Procedia Eng. 30 1064–71
Gratton G, Coles M G and Donchin E 1983 A new method for off-line removal of ocular artifact.
Electroencephalogr. Clin. Neurophysiol. 55 468–84
Hagemann D and Naumann E 2001 The effects of ocular artifacts on (lateralized) broadband power in
the EEG Clin. Neurophysiol. 112 215–31
Hoffmann S and Falkenstein M 2008 The correction of eye blink artefacts in the EEG: a comparison of
two prominent methods PLoS One 3 e3004
Jiang J-A, Chao C-F, Chiu M-J, Lee R-G, Tseng C-L and Lin R 2007 An automatic analysis method for
detecting and eliminating ECG artifacts in EEG. Comput. Biol. Med. 37 1660–71
Jung T-P, Makeig S, Westerfield M, Townsend J, Courchesne E and Sejnowski T J 2000 Removal of
eye activity artifacts from visual event-related potentials in normal and clinical subjects Clin.
Neurophysiol. 111 1745–58
Kim S and McNames J 2007 Automatic spike detection based on adaptive template matching for
extracellular neural recordings J. Neurosci. Methods 165 165–74
Kim D-W, Shim M, Song M J, Im C-H and Lee S-H 2015 Early visual processing deficits in patients
with schizophrenia during spatial frequency-dependent facial affect processing Schizophr. Res.
161 314–21 Online: http://linkinghub.elsevier.com/retrieve/pii/S0920996414007300
Klein A and Skrandies W 2013 A reliable statistical method to detect eyeblink-artefacts from
electroencephalogram data only Brain topogr. 26 558–68
Kook H, Gupta L, Kota S, Molfese D and Lyytinen H 2008 An offline/real-time artifact rejection strategy
to improve the classification of multi-channel evoked potentials Pattern Recognit. 41 1985–96
Online: http://linkinghub.elsevier.com/retrieve/pii/S0031320307004074
Krishnaveni V, Jayaraman S, Anitha L and Ramadoss K 2006 Removal of ocular artifacts from EEG
using adaptive thresholding of wavelet coefficients J. Neural Eng. 3 338–46 Online: www.ncbi.
nlm.nih.gov/pubmed/17124338
Lawhern V, Hairston W and Robbins K 2013 DETECT: A MATLAB toolbox for event detection and
identification in time series, with applications to artifact detection in EEG signals PLoS One
8 e62944
Li Y, Ma Z, Lu W and Li Y 2006 Automatic removal of the eye blink artifact from EEG using an ICA-
based template matching approach Physiol. Meas. 27 425–36

416
Physiol. Meas. 37 (2016) 401 W-D Chang et al

Mognon A, Jovicich J, Bruzzone L and Buiatti M 2010 ADJUST: an automatic EEG artifact detector
based on the joint use of spatial and temporal features Psychophysiology 48 229–40
Mullen T, Kothe C, Chi M, Ojeda A, Kerth T, Makeig S, Jung T-P and Cauwenberghs G 2015 Real-
time neuroimaging and cognitive monitoring using wearable dry EEG IEEE Trans. Biomed. Eng.
62 2553–67 Online: http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.htm?arnumber=7274673
Nolan H, Whelan R and Reilly R B 2010 FASTER: Fully automated statistical thresholding for EEG
artifact rejection J. Neurosci. Methods 192 152–62
Otsu N 1975 A threshold selection method from gray-level histograms IEEE Trans. Syst. Man Cybern.
9 62–6
Pult H, Riede-Pult B H and Murphy P J 2013 A new perspective on spontaneous blinks Ophthalmol.
120 1086–91
Shao S-Y, Shen K-Q, Ong C J, Wilder-Smith E P V and Li X-P 2009 Automatic EEG artifact removal: a
weighted support vector machine approach with error correction IEEE Trans. Biomed. Eng. 56 336–44
Verleger R 1993 Valid identification of blink artefacts: are they larger than 50 microV in EEG records?
Electroencephalogr. Clin. Neurophysiol. 87 354–63
Yong X, Ward R K, Birch G E and Society N S 2008 Facial EMG contamination of EEG signals :
characteristics and effects of spatial filtering ISCCSP 2008 (Malta) pp 729–34

417