This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal
Smartwatch-based Wearable EEG System for
Driver Drowsiness Detection
Gang Li, Boon Leng Lee, Wan-Young Chung, Member, IEEE
Abstract—Driver drowsiness is a major cause of mortality in
traffic accidents worldwide. Many physiological signals have been
proposed to detect driver drowsiness. Among these signals,
Electroencephalographic (EEG) signal, which reflects the brain
activities, is more directly related to drowsiness. Thus, many
EEG-based driver drowsiness detection (DDD) models gained
more and more attention in recent years. However, one limitation
of these studies is that these models merely estimate discrete labels
and thus did not allow for estimating relative severity of driver
drowsiness. This study proposes Support Vector Machine based
Posterior Probabilistic Model (SVMPPM) for DDD aimed at
transforming the drowsiness level to any value of 0~1 instead of
discrete labels. A fully wearable EEG system which consists of a
Bluetooth-enabled EEG headband and a commercial smartwatch
was used to evaluate the proposed model in real-time way. Twenty
subjects who participated in one-hour monotonous driving
simulation experiment were used to develop this model with
fifteen subjects for building model and five subjects for testing
model. According to a video-based reference, the proposed system
obtained an accuracy of 91.25% accuracy for alert group (73 out
of 80 datasets), 83.78% accuracy for early-warning group (93 out
of 111 datasets) and 91.92% accuracy for full-warning group
(91out of 99 datasets). These results indicate that the combination
of proposed SVMPPM, EEG headband and wrist-worn smart
device constitutes an effective, simple, and inexpensive wearable
solution for DDD.
Index Terms—Driver drowsiness detection, EEG, wearable
devices, smartwatch, support vector machine.
I. INTRODUCTION
D RIVER drowsiness is a major cause of mortality in traffic
accidents worldwide. The U. S. National Highway Traffic
Safety Administration (NHTSA) reports that drowsy driving is
the cause of an estimated 40,000 injuries and 1,550 deaths in
car crashes per year [1]. Many of these deaths could be avoided
if driver drowsiness was properly monitored and drivers were
given early warnings. Driver drowsiness, which is a type of
excessive sleepiness, occurs when people are driving for
extended periods of time in monotonous environments, such as
highways. The standard clinical tests for measuring sleepiness
are the Multiple Sleep Latency Test and the Maintenance of
Wakefulness Test, combined with polysomnography datasets
[2]. These measurements are very expensive and cumbersome
Manuscript received June 18, 2015. This research was supported by Basic
Science Research Program through the National Research Foundation of Korea
(NRF) funded by the Ministry of Education (2012R1A1B3004140).
Authors are with the Department of Electronic Engineering, Pukyong
National University, Busan, 608739, Korea (e-mail: ligang@pknu.ac.kr;
wychung@ pknu.ac.kr).
1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
to perform (at least 8 channels are needed: 4-EEG, 2-
Electrooculogram (EOG), 1- Electromyogram (EMG) and 1-
Electrocardiogram (ECG)); it is essentially impossible to use
these methods to detect driver drowsiness in actual driving
environments. For instance, the usage of multiple sensors can
be uncomfortable for the driver and might impede the driver’s
movements. Thus, there is a strong need for a low-cost and
wearable drowsiness detector for early recognition of drivers at
risk of drowsy driving.
In the past decade, in order to detect drowsiness simply and
effectively, there has been lot of effort to distinguish the state
between alertness and drowsiness using EEG alone. According
to previous studies, the drowsiness recognition methods using
EEG alone can be divided into three categories: FFT-based
spectral analysis [3-9], wavelet-based spectral analysis [10, 11]
and Higher Order Statistics-based analysis [2]. Among the
three methods, FFT-based methods require the least system
resources and thus are widely used in real-time driver
drowsiness detection systems [3-9]. For example, based on the
ground truth of subjects’ response time to lane departure event,
C. T. Lin et al. developed a unipolar single-channel EEG-based
wireless portable driver drowsiness detection (DDD) system,
which used θ and α-based features in occipital area (Oz) and
obtained an accuracy of 82.8% [4]. Based on the ground truth of
drivers’ self-assessment (that is, subjects press a button placing
next to them when feeling drowsy), Y. Punsawad et al.
developed a bipolar single-channel EEG-based portable device
for real-time drowsiness detection, which used weighted θ ~ β
features extracted from temporal and parietal area (T3 & P4 or
T4 & P3) and obtained 90.4% accuracy [5]. J. Park et al.
developed a dry electrode based single-channel DDD system
and obtained 83.8% accuracy [6]. Based on the ground truth of
Rechtschaffen and Kales sleep scoring standard, Hai. B. V. et
al. also developed a unipolar single-channel EEG-based DDD
system, which used dry electrodes and α and β features
extracted frontal area (Fp1) and obtained 81% accuracy [8].
According to aforementioned examples, different ground truths
were used and different detection accuracies were obtained.
Also, the feasibility of using FFT-based θ ~ β features for
real-time monitoring driver drowsiness could be verified.
Particularly, the usage of dry electrodes could significantly
increase the ability for long-term EEG measurement [6 , 9] and
make experimental results more convincing for real-life
application. However, aforementioned examples are all
threshold-based binary classification approaches, and thus are
not capable of estimating the drowsiness level. As a result, it is
unable to address driver drowsiness at its early stage, when
feedback might be the most effective.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

(a)

(b) (c)
Fig. 1. (a) The system diagram with its specification. (b) The developed EEG bio-potentials conditioning circuit and the open-source hardware platform. (c)
Prototype of the proposed wearable DDD system.

In order to detect driver drowsiness as early as possible,
several decision-making models have been proposed to classify
driver drowsiness into several levels. For example, M.
Murugappan et al. classified drowsiness into 4 levels with 84.4%
accuracy by using subtractive fuzzy classifier [12]. Khushaba.
R. N. et al. classified drowsiness to 5 levels with about 95%
accuracy by using kernel spectral regression based linear
discrimination analysis and the ground truth of Wierwille scale
[13]. However, most of this kind of studies used conventional
wet Ag/AgCI electrodes. In addition, their methods were not
evaluated in real-time way. In a more recent work [9], C. T. Lin
et al. developed a dry electrode based driver vigilance
monitoring system for real-time monitoring driver drowsiness.
Based on the ground truth of subjects’ response time to lane
departure event, this system could classify drowsiness to
maximum 8 levels by using the combined approach of
threshold-based method and a support vector regression model.
In this paper, we propose to apply Support Vector Machine
(SVM)-based posterior probabilistic model (SVMPPM) for
automated drowsiness detection and estimation of the severity
of drowsy driving based on dry EEG sensor signals obtained
from simulated driving experiment. To evaluate the real-time
performance of the proposed model, a wireless EEG headband
was developed and combined with a commercial smartwatch to
extract FFT features and implement the probabilistic model.
Compared with previous studies, this system has several
advantages. First, this system uses probabilistic model which
can estimate the probability of drowsy driving with a value
range from 0 to 1 and thus can effectively increase the detection
resolution. Second, an early-warning threshold probability was
developed to detect drowsiness at its early stage. Third, the
ground truth of percentage of eyelid closure over pupil over
time (PERCLOS) was used in all the experiments. Unlike other
ground truth mentioned above, PERCLOS techniques has been
commercialized and intensively tested for more than ten years
by U. S. NHTSA for field application (actual truck drivers
driving commercial trucks) [14-16]. Therefore, it can ensure
more convincing experimental results. Fourth, the proposed
model was implemented on wearable smartwatch which could
effectively prevent the distracted driving caused by the usage of
smartphones or tablets in cars [6, 9, 17-19].
The rest of the paper is organized as follows. Section II
presents the whole system architecture includes hardware parts
and proposed SVMPPM. Section III presents the experiment
design for simulated driving. Section IV presents the validation

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

of the system with further discussion. Section V concludes the
paper with future work.
II. SYSTEM ARCHITECTURE
A. System Overview
As shown in Fig. 1, the wearable EEG system is powered by
a 3.6 V 100-mAh lithium-ion rechargeable battery and consists
of a wireless EEG acquisition device and a commercial
smartwatch. The wireless EEG acquisition device consists of an
EEG sensory input unit (SIU) as well as a sensory processing
unit (SPU). The analog EEG data (i.e., voltage output) from the
SIU are converted to digital data and wirelessly transmitted to
the smartwatch via a Bluetooth Low-Energy (BLE) or so-called
Bluetooth Smart module by SPU [20]. The smartwatch further
extracts EEG features and estimates the probability of drowsy
driving by a well-trained SVMPPM. If the estimated
probability exceeds a particular threshold, a drowsy warning
can be activated. For real-life application, the wireless EEG
acquisition device is sewed to an elastic headband. The EEG
headband and the smartwatch can continuously work 5.2 and
4.5 hours respectively. The following sections introduce in
detail the major components of this system.
B. SIU
SIU consists of three EEG dry electrodes and an EEG
bio-potentials conditioning circuit. Among the three dry
electrodes, one ear-clip dry electrode, from Laxtha Co., Ltd
[21], serves as a ground electrode and the remaining two dry
electrodes, adapted from [22], are used as EEG signal
electrodes. The main difference between the current system and
that presented in [22] is that we sewed conductive fabric on the
surface of the copper fill plate instead of using copper fill
directly to sense EEG bio-potentials . This alteration effectively
reduces motion artifacts as well as the skin-electrode interface
impedance, as illustrated by Chin in [23]. Its comparison with
conventional wet Ag/AgCl electrodes can be found in Section
IV. According to the International 10–20 system, they are
placed at occipital regions O1 and O2, where is highly
order to acquire the most useful EEG bands for DDD
application (θ (4-7 Hz), α (8-12 Hz) and β (13-30 Hz)), the
output signal of IA is transferred to High Pass Filter (HPF) and
then filtered by the Low Pass Filter (LPF). Finally, the signal is
amplified by amplifier to meet the ADC’s input requirement
(0~3.3V).
C. SPU
The SPU consists of an inexpensive open-source hardware
(OSHW) platform [25] and a BLE transmitter. The OSHW
platform was chosen because it is inexpensive, lightweight, and
specifically designed for wearable devices with sewing holes. It
includes a low-power 8-bit MCU with a clock speed of 8 MHz.
It has 6 ADC input pins. The present application uses a single
ADC pin. The digitalized EEG data is saved to SRAM and then
directly transmitted to the smartwatch via BLE interface. BLE
is used because the smartwatch has embedded BLE
microprocessors with capabilities for connecting wirelessly to
external biomedical sensors by lower power consumption.
D. EEG Signal Analysis and Feature Extraction
A digital band-pass filter (4-30Hz) is implemented in
smartwatch to further filter EEG signals, particularly the power
line noise (60Hz). In order to reduce individual differences and
enhance the generalization of the SVM model, relative EEG
power ratio (power percentages) are used instead of the
absolute EEG power values as input features for building the
SVM model. EEG power is calculated as the sum of the squared
FFT magnitude of the EEG signal using 2 s Hamming windows;
the algorithm is implemented in real-time way on the
smartwatch. Then, the power percentage (Per) is calculated by
dividing the FFT power of one EEG band by the sum of the
FFT power of all three EEG bands, as shown in Equation 1,
where, zi={ θ, α, β }.
Power ( z i )
Per ( z i ) = 3
´ 100% (1)
å Power ( z i )
i =1

correlated with the driver’s vigilance level [24]. The ground

electrode is placed on the earlobe and given 1.65v bias to
satisfy the required condition of the SPU’s ADC which is
powered by a single positive reference voltage (+3.3V).
EEG signals are usually weak and easy to be interfered by
undesired noises (EEG signals typically have amplitude in the
range of 10u-100uV). Therefore, both amplifying and filtering
are required for further signal processing. In addition, human
skin typically provides source impedance on the order of
1M-5Mohm. Thus, amplifiers must match the source
impedance or have greater input impedance than the source
skin impedance to acquire EEG signals. For these reasons, the
output of the bipolar electrode is transferred to an EEG
bio-potentials conditioning circuit and firstly amplified by an
instrument amplifier (IA). The IA features a high differential
input-impedance of 10Gohm || 2pF and a high common mode
rejection ratio (CMRR) when the gains are 10X. Therefore, IA
can balance the electrode-skin impedance well and reject Fig. 2. EEG feature extraction procedures for one participant. A single feature
common mode noise as much as possible. Additionally, in extraction procedure marked by dotted box will repeat 60 times during
one-hour driving experiment.

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

Therefore, a set of three features (Per (θ), Per (α) and Per (β)) (e.g., radial basis function (RBF), polynomial kernel function
were extracted from each 2-s EEG epoch. In order to (Poly)), sigmoid kernel function (Sig) and regulation parameter
synchronize the extracted EEG features with the 1-min ground C, SVM is able to do this mapping work and accomplish
truth measurements (PERCLOS), every 30 sets of feature were classification without the need to know explicitly what the
successively averaged before applying SVM operation. The mapping function is.
feature extraction procedure for one participant is demonstrated
in Fig. 2. 2) SVMPPM
E. Drowsiness Detection Model In addition to classifying the labels into the categories of
“drowsy” or “alert,” the posterior class probabilities, i.e.,
We propose to apply SVMPPM for automated drowsiness r
detection and estimation of the severity of drowsy driving P(class=+1| x ) can be calculated based on the outputs of binary
based on EEG signals. As shown in Fig. 3, applying this model SVM classification model using the method initially described
includes three stages - Stage I: Data collection, Stage II: Model by Platt [26]. Briefly, its main idea is given by Equation 2,
r r r
building and Stage III: Model testing. The Stage I and Stage III where, w × F ( x ) + b is just the SVM output for input dataset x .
involve one-hour simulated driving experiment (see Section A and B are parameters that are determined by an improved
III), during which both EEG signals and features were Platt algorithm [27].
recorded. The Stage II was implemented on MATLAB in PC in
off-line way, which aims to build an optimized model and r 1
P (class = +1 | x ) = r r (2)
develop the threshold probabilities for recognizing early 1 + exp( A( w × F ( x ) + b) + B )
drowsy warning group and full drowsy warning group. The
well-trained model was then transferred to smartwatch for Therefore, compared to previous multi-class classification
real-time final testing. The following subsections introduce in based models, the SVMPPM can further increase the resolution
detail the major components of proposed approach. of detected drowsiness level (any value from 0 to 1). For
example, the probability “0” indicates a full alert state, whereas
a value close to “1” indicates a severely drowsy driving state. In
addition, through the combination of true labels and estimated
posterior probabilities, a threshold probability could be found
to provide early warning before drivers actually become
severely drowsy.
F. Wearable Application
Wrist-worn smart devices, particularly
Android-Wear-OS-based smartwatches, have emerged as
miniature but powerful devices for data acquisition and
processing. In this study, a commercial smartwatch is
combined with external EEG sensors, in order to design a fully
wearable DDD system. For real-life application, a simple user
interface was designed (as shown in Fig.4), where the traffic
light principle was used to indicate driving status. The green
light and yellow light indicate safe driving and early warning
respectively. For the full warning, the red light is ON, and
additionally a long-lasting vibration, which is produced by the
built-in vibration sensor of the smartwatch, is activated
simultaneously. Also, a pie chart is used to imitate a sort of
watch face and display the extracted EEG features in real-time
way. The complete working flowchart of this wearable
Fig. 3. Flowchart of estimation of driver drowsiness using SVM-based application is described in Fig. 5. It is very important to note
posterior probabilistic model. that besides the usage of BLE, two Services that incorporate
with two Activities are carefully designed for this smartwatch
1) Binary SVM Classification application aiming to meet the low power requirement. In
A conventional binary SVM classifier outputs a class label Android OS, each Activity corresponds to each user interface.
(e.g., +1: drowsy driving or -1: alert driving) for each input Each user interface has a short life time due to the power save
r
dataset x comprising several features (e.g., Per (θ), Per (α) and mode of smartwatch screen. So, the simplest and Activity-based
Per (β) in this study), according to the sign function way to collect data continuously is to keep the screen always
r r r r
f ( x ) = sign( w × F ( x ) + b), where the w is a vector perpendicular ON by using the WindowManager component; while Service is
to the decision surface and b is a scalar (decision surface bias). independent of Activity and specifically designed to run the
r
F (x ) is the mapping function which is used to map each input repeating task in background. So, two Services are employed to
r
dataset x from the linearly non-separable input space  n to collect data continuously in background instead of keeping
linearly separable feature space H. By using kernel functions screen ON. In this case, much power can be saved.

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal
Fig. 4. Smartwatch screen showing the designed user interface for this
wearable application (termed “Brainwatch”). The three circles on the top side
display the driving status. Green circle (left side) indicates safe driving.
Yellow circle (middle) indicates early warning. Red circle (right side)
indicates full warning. The pie chart shows the EEG features.
Fig. 5. The working flowchart of the proposed wearable application.
III. EXPERIMENT DESIGN AND MATERIALS
According to our previous experiences [28, 29], one-hour
monotonous driving after lunch (typically around 1:00-2:30 pm)
results in drowsiness in a majority of subjects. Therefore,
twenty subjects (twelve males and eight females, aged
30.8±9.37 years), possessing valid driver’s licenses,
participated in the one-hour monotonous driving experiment.
Experimental data from fifteen subjects (subjects No.1-15) was
used for building drowsiness detection model. Final testing
phase was performed for the remaining subjects (subjects
No.16-20).
The driving simulation environment consisted of a
commercial truck driving simulator (Euro Truck Simulator 2), a
Logitech® steering wheel, and acceleration and brake pedals
1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
(Fig. 6). During the monotonous driving experiment, the
participants drove with a few road simulators on a highway and
experienced various realistic cognitive loads, such as keeping
or changing lanes to avoid collisions, turning on windshield
wiper when driving in rainy conditions, and early deceleration
before red lights. On the day of the experiment, subjects were
prohibited from drinking tea or consuming anything containing
caffeine. Moreover, consumption of soporific medicines, such
as standard cold medications, was prohibited. Before the
monotonous driving experiment, each subject was given ten
minutes to familiarize himself/herself with the operation of the
driving simulator.
Fig. 6. Example of experimental setup for evaluating drowsiness detection.
The true alert and drowsy data are labeled based on the
combination of percentage of eyelid closure over pupil over
time (PERCLOS) and the number of adjustment (NOA) on
steering wheel. PERCLOS is a video based real-time driver
drowsiness monitoring technology. It assesses drowsiness by
measuring slow eyelid closure and estimating the proportion of
time for which the eyes are closed over 1-min interval (high
sensitivity) or 3-min interval (medium sensitivity) or 5-min
interval (low sensitivity) [16]. The approach of NOA on
steering wheel is developed by our group by using the built-in
accelerometer and gyroscope of smartwatch [30]. The basic
assumption of this approach is that alert drivers who are driving
on a highway (not city driving or countryside driving) would
have more NOA on steering wheel if compared to drowsy
drivers. We found that some drivers’ faces are occasionally
outside of the camera field of view while their NOA data can
remain stable. Thus, in order to provide a robust ground truth,
we combined the two methods together in this study. According
to [16] and [30], PERCLOS ≥ 12% and NOA ≤ 9 was chosen as
the threshold to identify the presence of true drowsy driving
and PERCLOS < 8% and NOA > 26 for true alert driving.
PERCLOS measures between 8% and 12% and NOA measures
between 9 and 26 can be used as the early-warning reference. In
our previous study [31, 32], we already developed a
smartphone application to measure PERCLOS. Thus, in this
study, the measures of PERCLOS were automatically
implemented on the smartphone which was placed in front of
the steering wheel.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

Fig 7. EEG signals (line chart) and EEG band power (pie chart) comparison between wet electrodes (the top) and dry electrodes (the bottom). For the line chart,
X-axis indicates the time. The unit is second. Y-axis indicates the amplitude of the digitalized EEG samples which are already filtered by the digital band-pass filter
(4-30Hz) in smartwatch. For the pie chart, the X-axis indicates the frequency ranged from 0 to 64Hz (the half of the sampling rate 128Hz). Y-axis indicates the
magnitude of FFT power.

Fig 8. EEG signals comparison of alert driving status (the top) and drowsy driving status (the middle). The driving status was confirmed by PERCLOS
measurement as shown on the left side of the EEG signals. The EEG signal on the bottom is the zoomed in signal that is marked by red circle aiming to clearly show
the abundant alpha wave occurred in drowsy driving status.

IV. EXPERIMENTAL RESULTS AND DISCUSSION
The LibSVM® (ver 3.17) [33] implementation for SVM was
used in this study. In model building stage, an optimized
SVMPPM were trained on a PC using leave-one-subject-out
(LOSO) cross-validation tests with different kernel functions,
such as linear kernel, RBF and Poly, and different regulation
parameters “C”. LOSO approach is a cross-validation approach,
where each subject serves as a test set. The specific steps are as
follows: (1) Omit one subject from the training set; (2) Train the
binary classification model using the remaining subjects; (3)
Test the omitted subject using the trained model in step (2) and
calculate the binary classification accuracy and posterior
probabilities; (4) Repeat the steps that are listed above until
each subject has been omitted and tested once; (5) Calculate the
overall binary classification accuracy and set the threshold
probabilities for final test. In model testing stage, the
parameters of the well-trained SVM model were hardcoded in
text-file format and stored on the Secure Digital (SD) card
memory of the smartwatch. Meanwhile, the LibSVM library
was embedded into the smartwatch. Finally, smartwatch can
call library functions such as svm.svm_load_model and
svm.svm_predict_probability to run SVM in real time way.
A. Data Collection Stage
Firstly, the dry electrodes were compared with conventional
wet Ag/AgCl electrodes. We recorded EEG signals using the
two types of electrodes in the same positions (O1 & O2) but in
separate sessions (dry electrodes were tested first) when one
subject was instructed to close his eyes for 5 seconds. What we
expected to be common between these two sessions was the
appearance of obvious rhythmicity of α rhythm in time domain
and the largest power percentage for α band in frequency
domain. What exactly we observed is shown in Fig. 7, which is
just consistent with our expectation. Similarly, every time
before formal experiment we instructed each subject to close
eyes in order to check the contact quality of dry electrodes.
Totally 210 labeled alert datasets and 227 labeled drowsy
datasets were collected from 15 subjects. Each dataset consists
of three EEG features: Per(θ), Per(α) and Per(β). Fig. 8 shows

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

the examples of the recorded EEG signals that paired with one
alert dataset and one drowsy dataset. With the obvious
rhythmicity of α rhythm shown in the time-domain signals,
Per(α) is increased from 38.8% in the alert dataset to 47.4% in
the drowsy dataset. The Box-Whiskers plots (Fig. 9) show that
the median values for all 15 subjects are 16.07%, 29.37% and
54.56% for Per(θ), Per(α) and Per(β) for alert driving group,
and 12.51%, 41.72% and 45.76% for drowsy driving group.

Fig. 10 ROC curve showing sensitivity (possibility of true drowsy event) and
1-specificity (possibility of false drowsy event) for extracted EEG features.

B. Model Building Stage

The LOSO binary classification result is summarized in
Table I, where the parameter “C”, RBF parameter “g” and Poly
parameter “d” are optimized using a simple search procedure
with C, g={0.01~10} in steps of 0.01 and d={1, 2} in steps of 1.
As expected, the best accuracy was found to be 93.59% (in RBF
kernel of g=2 and C=5) when single Per(α) were used as the
Fig. 9. Box-Whiskers plots of features extracted from labeled drowsy and
SVM inputs.
alert EEG recordings. The boxes have three lines to present the values for
first quartile (the bottom), median, and third quartile (the top) for column TABLE I
data. The length between the first quartile (Q1) and the third quartile (Q3) is LEAVE-ONE-SUBJECT-OUT CLASSIFICATION ACCURACY OF SVM CLASSIFIER
called interquartile range (IQR). Two addition lines at both ends of the WITH DIFFERENT INPUT FEATURE SETS AND DIFFERENT KERNELS
whisker indicate the Q1-1.5×IQR and Q3+1.5×IQR value of a column data. Feature sets
Kernel
Any data not included between the whiskers is plotted as outliers represented θ α β θ, α θ, β α, β θ, α, β
by “o” for mild outliers and “*” for extreme outliers. The number next to the Accuracy %
outlier is the number of the data in that column, called case number. Linear 76.66 93.14 83.75 92.91 92.91 92.91 92.91
C 0.03 0.02 1.56 0.09 0.3 0.01 0.12
Statistical analysis results show that all three features are RBF 77.57 93.59 84.67 92.91 93.16 93.14 93.14
significantly different between groups as determined by C 2 5 0.03 0.29 1.57 1.45 0.62
one-way ANOVA (F(1,435)=209.677, P<.001 for Per(θ); g 3 2 0.01 0.02 0.01 0.05 0.01
Poly 76.66 93.14 83.75 92.91 92.91 92.91 92.91
F(1,435)=735.397, P<.001 for Per(α); F(1,435)=377.589 for C 0.03 0.02 1.56 0.17 0.59 0.02 0.16
Per(β)). The higher F-statistic value for Per(α) indicates that d 1 1 1 1 1 1 1
Per(α) has better class-discrimination capability than Per(θ) and
Per(β). The class-discrimination capability of Per(α) can also In addition to classifying all datasets into the categories of
be verified by receiver operating characteristic (ROC) curve “drowsy” or “alert,” the posterior probability of being drowsy
analysis. The area between ROC curve and the reference line, driving was then calculated by using the best feature Per(α) and
denoted by ROCarea, is an effective measure of the RBF kernel. The histogram of estimated posterior probability
class-discrimination capability of a specific feature [34]. Its (Fig. 11) shows that there is a clear separation between the two
value can be any value from 0.5 to 1. A value of 0.50 implies groups. Most of alert datasets were classified into the bins of
that the features are completely overlapped and thus 0~0.1; while most of drowsy datasets were classified into the
non-separable; while a value of 1 implies that the features are bins of 0.9~1. It is also important to note that unlike the
completely separable. Fig. 10 shows that Per(α) has the largest informative class labels, the probabilities in the same group are
ROCarea (=0.979) compared to feature Per(θ) (=0.842) and varied, from about 0.55 to 0.95 for drowsy group, and 0.05 to
Per(β) (=0.916). 0.4 for alert group. This result confirmed the feasibility of
estimating the relative severity of drowsiness by using
SVMPPM.

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

Fig. 11. The histogram of estimated posterior probability for each subject (#1~#15) in training phase for alert group (left side) and drowsy group (right side). The
X-axis indicates the drowsiness level (0~1). The Y-axis indicates the number of dataset that falls into the corresponding bins of estimated drowsiness levels.

Fig. 13. The threshold probabilities for each subject in training phase for (panel A) alert group with X-axis 1~15 and (panel B) drowsy group with X-axis 16~30.
The threshold probabilities for each subject refer to the maximum probability in alert group and minimum probability in drowsy group.

The Bland-Altman plot [35] is the preferred method for

Fig. 12. Bland-Altman plot for the relationship of the difference between the
observed PERCLOS versus estimated probability versus their average values.
Mean bias, +1.96SD, and −1.96SD lines are shown. SD = standard deviation.
assessing the agreement between an established measurement
(e.g., the PERCLOS measurement) and a new measurement
technique (e.g., the SVMMP measurement). Thus, a Bland–
Altman plot (Fig. 12) for the 15 subjects was generated so that
the relationship between PERCLOS and SVMMP
measurement could be examined. For comparability among the
two measures, the PERCLOS measures were normalized to the
range of 0~1 by simply removing the unit %. For example,
PERCLOS value of 1% is equal to value 0.01.
As can be seen in Fig. 12, all data is within the 1.96SD limits
of agreement. However, there is a bias of -0.432 (95% CI
-0.467 to -0.397) indicating that systematically the SVM-based
measures have higher magnitudes than PERCLOS measures.
This is mainly caused by the big differences in measures for
drowsy driving (cluster_2) between the two methods.
Specifically, most of SVM-based probabilities for drowsy
driving are around 0.9; while the range of most of PERCLOS

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

measures for drowsy driving is from 0.12 to 0.20. Few of them which are expected as subjects started to appear some irregular
can reach to 0.30 and very few of them can reach to 0.90. In
contrast, the measures representing alert driving (cluster_1)
show better agreement because there is no big difference (near
the bias of zero). Specifically, most of SVM-based probabilities
for alert dataset are around 0.05. This corresponds to the range
of PERCLOS measures for alert dataset (min=0 and
max=0.07).
Besides the Bland-Altman plot, the threshold probabilities
for each subject were summarized in Fig. 13 and two overall
threshold probabilities which indicate early warning threshold
(TH_1:P=0.424) and fully warning threshold (TH_2: P=0.545)
were developed to validate the calculated probabilities in
real-life application. The rules for this validation are
summarized in Table II.

TABLE II
RULES (THRESHOLDS FOR DDD)
Rules Ground truth Decision
8% ≤ PERCLOS < 12% (a)
0.424 ≤ P < 0.545 Early warning
9 < NOA ≤ 26
PERCLOS ≥ 12%
P ≥ 0.545 Full drowsy warning
NOA ≤ 9

C. Model Testing Stage

Based on the detection rules in Table II, five subjects that
were not used in the model building phase were used in this
final test. The test results are summarized in Table III. The
overall classification results were 91.25% accuracy for alert
group, 83.78% accuracy for early-warning group and 91.92%
accuracy for full warning group. The calculated posterior
probabilities and classification results for subject No. 16 and 18
are shown in Fig. 14.
The computational cost of the proposed approach is also
investigated. We used the debugging tool Logcat in the
Android wear developing environment (Android Studio beta
ver 0.8.6) and let Logcat print out the system time before and
(b)
after running FFT and SVMPPM. We found that the time delay Fig.14. Posterior probability estimate P (y=+1|f(x)) calculated for each
for FFT could be ignorable at millisecond level. For SVMPPM class for (a) subject #16 and (b) subject #18.
operation, there is a 4ms time delay.
movements during that period, including yawn, rubbing the
D. Discussion eyes, face, nose and mouth, touching the hair and changing the
As can be seen in Table III, the defined thresholds are quite seating position on the chair. We examined the recorded EEG
successful at classifying the three driving groups. 91 out of 99 raw data later and found that most of these movements result in
datasets are correctly classified into drowsy group and 73 out of noise in high frequency band (β band in this study). This is the
80 datasets are correctly classified into alert group. Fig. 14 reason why most of the misclassified dataset falls into the alert
shows that many of the errors occur in the early warning group, group. For dry electrodes, the noise caused by irregular
TABLE III
RESULTS OF FINAL TEST USING DEVELOPED WEARABLE EEG DEVICE AND MACHINE-LEARNING-ENABLED SMARTWATCH
Number of detected epochs by
Number of 1-min labeled epochs Accuracy (%)
Subject No Sex posterior probabilistic thresholds
Alert Early-warning Drowsy Alert Early-warning Drowsy Alert Early-warning Drowsy
16 M 15 18 27 13 14 26 93.33 77.78 96.30
17 M 11 26 13 11 22 10 100 84.62 76.92
18 M 16 26 18 13 23 16 81.25 88.46 88.89
19 F 26 13 21 24 11 20 92.31 84.62 95.24
20 F 12 28 20 12 26 19 100 92.86 95.00
Overall 80 111 99 73 93 91 91.25 83.78 91.92

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

10

TABLE IV
COMPARISON OF EEG AND PROBABILISTIC MODEL BASED DDD STUDIES
Evaluation
Reference No. Time window Model Dry Terminal Accuracy
Signal Source Ground Truth
electrodes Device (%)
[38] 6 min DBN Two-Channel EEG × - - -

[40, 41] 1 min BBE Two-Channel EEG × - Self-assessment standard -

Present work 1 min SVMPPM Single-Channel EEG √ Smartwatch PERCLOS 88.6

The symbol “-” indicates that the corresponding content is unavailable.

TABLE V
COMPARISON OF POWER CONSUMPTION OF BLUETOOTH-ENABLED AND EEG-BASED DDD SYSTEMS
EEG acquisition part Power Consumption /
Reference No. Terminal device
Signal Source ADC Sampling rate (Hz) Wireless module Battery life

[4] 1-ch EEG 12 bit 512 × 31mA

Bluetooth v2.0+ EDR
[9] 4-ch EEG 24 bit 256 × 33mA

Bluetooth and RF dual

[3] 4-ch EEG 8 bit 295
modules
× 356mA

19mA for EEG

headband
Present work 1-ch EEG 10 bit 128 BLE
√ 4.5 hours for
(Smartwatch) Smartwatch
(by 300mAh battery)

movements is inevitable [36] and some methods directly use
these movements to identify the early drowsy phenomenon
instead of avoiding them, such as the ground truth Wierwille
scale [37].
To compare the prediction model in current work with prior
studies, three DDD-related studies that also involve EEG
signals and probabilistic models are carefully found from the
existing literatures (as shown in Table IV). For example, in
2010, G. S. Yang et al. [38] proposed a Dynamic Bayesian
Network (DBN) based posterior probabilistic model to detect
driver drowsiness. They employed Independent Component
Analysis (ICA) to preprocess EEG signals and FFT to extract
EEG features. Then, thirty sleep-deprived subjects who
participated in a simulated driving experiment were used to
validate the proposed model. However, the EEG channel
location and the ground truth used in this study are not clear and
the detection accuracy using EEG alone was not reported. In
addition, the length of the time window was 6 minutes, which
might be a major drawback for practical application. The other
possible practical hurdles include: 1) The usage of ICA. The
ICA is a complex algorithm that is not suitable for real-time
application [39]. 2) The usage of wet electrodes. In 2011 and
2014, A. Murata. et al [40, 41] proposed a Basic Bayesian
Estimation (BBE) based posterior probabilistic model. They
used shorter time window, 1 minute, to extract FFT-based EEG
features and also conducted a simulated driving experiment to
validate the proposed model. However, the ground truth used
was subjective (subject self-assessment-based) and wet
electrodes were also used. Also, the detection accuracy using
EEG alone was not reported. Besides the practical hurdles and
technical drawbacks mentioned above, building
Bayesian-based probabilistic model needs many empirical
work including the construction of prior probability, the
conditional probabilities and transitional probability (if DBN is
used) [38]. However, the primary advantage of SVM is its
ability to minimize both structural and empirical risk [42]. In
this study, we apply SVM-based probabilistic model to detect
driver drowsiness. There are presently no ‘gold-standard’
methods for assessing the estimated probabilities thus direct
validation of the calculated posterior probability values to
another technique is not currently possible. However, based on
PERCLOS measurements, the Bland-Altman plot and a
smartwatch-based wearable EEG system were used to validate
the estimated probability values in off-line way and real-time
way respectively.
To compare the practical utility, such as power consumption,
three EEG-based DDD studies that also involve Bluetooth
communication are carefully found from the existing
literatures. For previous studies, Table V shows that the
maximum power consumption of 356mA was obtained by a
Bluetooth and RF dual-mode system [3] against 31mA and
33mA by Bluetooth v2.0+EDR based systems [4, 9]. By using
BLE module, the present work obtained the minimum power
consumption of 19mA. In addition, a major limitation behind
these previous studies is that they only reported the power
consumption of EEG acquisition device and did not include the
terminal device which is the other indispensable device that
constitutes the whole DDD system. Considering the practical
utility (e.g., the requirement of low-cost and easy-to-use), we
combined the DDD application-specific EEG acquisition
device and a commercial smartwatch together and 4.5-hour
battery life was obtained by its built-in 300 mAh battery. The
battery capacity of 300mAh is quite less than that used in
previous studies, such as 750mAh [9], 1100mAh [4] and
16000mAh [3].

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal
V. CONCLUSION
This study developed a driver drowsiness detection system
with an SVM-based posterior probabilistic model and a
smartwatch-based wearable EEG device. Based on the
proposed system, the driver drowsiness level could be
transformed to any probability value of 0~1 instead of discrete
class labels aimed at detecting drowsiness as early as possible.
Also, two threshold probabilities which indicate early warning
threshold (P=0.424) and fully warning threshold (P=0.545)
were developed to validate the calculated probabilities in
real-life application in which the proposed model was
implemented in real time for subjects when they performed a
simulated monotonous driving experiment. Further studies
considering the combination of EEG with on-head motion
sensors (e.g., integrating gyroscope into the headband) need to
be performed in order to enrich the EEG data with more context
information and increase the detection accuracy further.
REFERENCES
[1] U. S. National Highway Traffic Safety Administration: Research on
drowsy driving. [Online]. Available:
http://www.nhtsa.gov/Driving+Safety/Distracted+Driving+at+Distractio
n.gov/Research+on+Drowsy+Driving
[2] V. Swarnkar, U. Abeyratne and C. Hukins, “Objective measure of
sleepiness and sleep latency via bispectrum analysis of EEG,” Med Biol
Eng Comput. vol. 48, no. 12, pp. 1203–1213, Nov. 2010.
[3] C. T. Lin, Y. C. Chen, T. Y. Hung, T. T. Chiu, L. W. Ko, S. F. Liang, H.
Y. Hsieh, S. H. Hsu and J. R. Duann, “Development of wireless brain
computer interface with embedded multitask scheduling and its
application on real-time driver’s drowsiness detection and warning,”
IEEE Trans. Biomedical Engineering. vol. 55, no. 5, pp. 1582–1591,
May. 2008.
[4] C. T. Lin, C. J. Chang, B. S. Lin, S. H. Hung, C. F. Chao and I. J. Wang,
“A real-time wireless brain-computer interface system for drowsiness
detection,” IEEE Trans. Circuits and Systems. vol. 4, no. 4, pp. 214–222,
Aug. 2010.
[5] Y. Punsawad, S. Aempedchr, Y. Wongsawat and M. Panichkum. (2011,
April.). Weighted-frequency index for EEG-based mental fatigue alarm
system. International Journal of Applied Biomedical Engineering.
[Online]. 4(1), pp. 36-41. Available:
http://www.ijabme.org/File/vol4no1/07weighted.pdf
[6] J. Park, L. Xu, V.Sridhar, M. Chi and G. Cauwenberghs, “Wireless Dry
EEG for Drowsiness Detection, “in Proc. IEEE Int. Conf. Engineering in
Medicine and Biology Society. Boston, USA, Aug. 30 – Sep. 3, 2011, pp.
3298-3301.
[7] Y. T. Wang, C. K. Cheng, K. C. Huang, C. T. Lin, Y. J. Wang and T. P.
Jung. (2012, Nov.). Cell-phone based drowsiness monitoring and
management system. Proceedings of Biomedical Circuits and Systems
Conference (BioCAS). [Online]. ,pp. 200-203. Available:
http://sccn.ucsd.edu/~yijun/pdfs/BIOCAS12.pdf
[8] H. B. Van, S. Rhodes, B. Dunne and R. Bossemeyer, “Low-Cost
EEG-based Sleep Detection,” in Proc. IEEE Int. Conf. Engineering in
Medicine and Biology Society. Chicago, USA, Aug. 26-30, 2014, pp.
4571-4574.
[9] C. T. Lin, C. H. Chuang, C. S. Huang, S, F. Tsai, S. W. Lu, Y. H. Chen
and L. W. Ko, “Wireless and Wearable EEG System for Evaluating
Driver Vigilance,” IEEE Trans on Biomedical Circuits and Systems., vol.
8, no. 2, pp. 165-176, May. 2014.
[10] M. Akin, M. B. Kurt, N. Sezgin and M. Bayram, “Estimating vigilance
level by using EEG and EMG signals,” Neural Computing and
Applications., vol. 17, no. 3, pp. 227-236, Jun. 2008.
[11] S. Gupta, S. Kar and A. Routray, “Fatigue in Human Drivers: A Study
Using Ocular, Psychometric, Physiological Signals,” in Proc. IEEE Int.
Conf. Students’ Technology Symposium. Kharagpur, India, Apr. 03-04.
2010, pp. 234-240.
[12] M. Murugappan, M. K. Wali, R. B. Ahmmad and S. Murugappan,
“Subtractive Fuzzy Classifier based Driver Drowsiness Levels
Classification using EEG,” in Proc. Int. Conf. Communications and
Signal Processing. Melmaruvathur, India, Apr. 3-5, 2013, pp. 159-164
1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
11
[13] R. N. Khushaba, S. Kodagoda, S. Lal and G. Dissanayake, “Driver
Drowsiness Classification Using Fuzzy Wavelet-Packet-Based
Feature-Extraction Algorithm,” IEEE Trans Biomed Eng., vol. 58, no. 1,
pp. 121-131, Sep. 2011.
[14] D. F. Dinges and R. Grace. (1996, Sep). Research on vehicle-based driver
status/performance monitoring; development, validation, and refinement
of algorithms for detection of driver drowsiness. US Dept.
Transportation, National Highway Traffic Safety Admin., Washington,
DC. USA. [Online]. Available:
http://ntl.bts.gov/lib/5000/5900/5911/887.pdf
[15] D. F. Dinges and R. Grace. (1998, Oct.). PERCLOS: a valid
psychophysiological measure of alertness as assessed by psychomotor
vigilance. US Dept. Transportation, Fedaral Highway Admin.
Washington, DC. USA. [Online]. Available:
www.fmcsa.dot.gov/pdfs/tb98-006.pdf
[16] Final report. (2008, Sep.). The drowsy driver warning system field
operational test: data collection methods. US Dept. Transportation,
National Highway Traffic Safety Admin. Washington, DC. USA.
[Online]. Available:
http://www.nhtsa.gov/DOT/NHTSA/NRD/Multimedia/PDFs/Crash%20
Avoidabce/2008./810035.pdf
[17] I.S. Kim. (2012, Jun.). The risk of accidents using DMB and smartphone
when driving. Traffic.[Online]. 172, pp. 32-36. Available:
http://www.koti.re.kr/mail/news/2012-06.pdf
[18] State laws about distracted driving. [Online]. Available:
http://www.distraction.gov/content/get-the-facts/state-laws.html
[19] J. A. Healey, “Wearable and automotive systems for affect recognition
from physiology,” Thesis (PhD) in Massachusetts Institute of
Technology, 2000. [Online]. Available:
http://dspace.mit.edu/handle/1721.1/9067
[20] Bluetooth 4.0 Low Energy Mini module. [Online]. Available:
http://redbearlab.com
[21] Ear-clip electrode. [Online]. Available:
http://www.laxtha.com/ProductView.asp?Model=LXEL-EAR-01&catgr
pid=1&catidx=30
[22] Y. M. Chi, P. Ng, E. Kang, J. Kang, J. Fang and G. Cauwenberghs. (2010,
Oct.). Wireless non-contact cardiac and Neural monitoring. Proceedings
of the wireless health 2010. [Online]. pp. 15-23. Available:
http://www.isn.ucsd.edu/pubs/wh2010.pdf
[23] C. T. Lin, L. D. Liao, Y. H. Liu, I. J. Wang, B. S. Lin and J. Y. Chang,
“Novel dry polymer foam electrodes for long-term EEG measurement,”
IEEE Trans. Biomed Engineering. vol. 58, no. 5, pp. 1200–1207, May.
2011.
[24] C. T. Lin, C. H. Chuang, Y. K. Wang, S. F. Tsai, T. C. Chiu, and L. W.
Ko. (2012, Jun.). Neurocognitive characteristics of the driver: A review
on drowsiness, distraction, navigation, and motion sickness. J. Neurosci.
Neuroengin. [Online]. 1(1), pp. 61–81. Available:
http://www.ingentaconnect.com/content/asp/jnsne/2012/00000001/0000
0001/art00010
[25] LilyPad Arduino. [Online]. Available:
http://arduino.cc/en/Main/arduinoBoardLilyPad
[26] J. C. Platt, “Probabilistic outputs for support vector machines and
comparison to regularized likelihood methods,” in Advances in Large
Margin Classifiers, A. Smola, P. Bartlett, B. Scholkopf, and D.
Schuurmans, Eds. Cambridge, MA: MIT Press, 2001.
[27] H. T. Lin, C. J. Lin and R. C. Weng. (2007, Oct.). A note on Platt’s
probabilistic outputs for support vector machines. Machine Learning.
[Online]. 68(3), pp. 267-276. Available:
http://www.csie.ntu.edu.tw/~cjlin/papers/plattprob.pdf.
[28] G. Li and W. Y. Chung. (2014, Sep.). Estimation of eye closure degree
using EEG sensors and its application in driver drowsiness detection.
Sensors. [Online]. 14(9), pp. 17491–17515. Available:
http://www.mdpi.com/1424-8220/14/9/17491
[29] G. Li and W. Y. Chung. (2013, Dec.). Detection of driver drowsiness
using wavelet analysis of heart rate variability and a support vector
machine classifier. Sensors. [Online]. 13(12), pp. 16494-16511.
Available: http://www.mdpi.com/1424-8220/13/12/16494
[30] B. L. Lee, B. G. Lee, G. Li and W. Y. Chung, “Wearable driver
drowsiness detection system based on smarwatch,” in Proc. the Korea
Institute of Signal Processing and Systems Autumn Conference. Busan,
South Korea, Nov. 21-22, 2014, pp. 134-136.
[31] B. G. Lee and W .Y. Chung, “Driver alertness monitoring using fusion of
facial features and bio-signals,” IEEE Sens J. vol. 12, no. 7, pp.
2416-2422, Jul. 2012.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JSEN.2015.2473679, IEEE Sensors Journal

12

[32] B. G. Lee and W. Y. Chung. (2012, Dec.). A smartphone-based driver Engineering, Pukyong National University, Busan. His current
safety monitoring system using data fusion. Sensors. [Online]. 12(12), pp. research interests include wireless sensor networks, ubiquitous
17536-17552. Available: http://www.mdpi.com/1424-8220/12/12/17536
[33] C. C. Chang and C. J. Lin. (2011, May.). LIBSVM: A library for support healthcare and automobile applications, smart lighting with
vector machines. ACM Trans Intell Syst Technol. [Online]. 2(3), pp. visible light communication, and embedded systems. He is
27:1-27:27. Available: now the leader of team for next generation u-healthcare
http://www.csie.ntu.edu.tw/~cjlin/papers/libsvm.pdf technology development, which is supported by the Brain
[34] S. Theodoridis, A. Pikrakis, K. Koutroumbas and D. Cavouras, “The
receiver operating characteristic curve,” in Introduction to pattern Korea 21 (BK21) plus project. He also serves as the associate
recognition: A Matlab approach, Elsevier Press, 2010. editor in IEEE Sensors Journal.
[35] J. M. Bland and D. G. Altman, “Statistical methods for assessing
agreement between two methods of clinical measurement,” Lancet. vol. 1, Boon-Leng Lee received the B.E.
no. 8476, pp. 307–310, Feb. 1986.
[36] Y. M. Chi, T. P. Jung and G. Cauwenberghs, “Dry-contact and degree in Mechanical Engineering from
Noncontact bio-potential electrodes: Methodological review,” IEEE Multimedia University, Melaka,
Reviews in Biomedical Engineering., vol. 3, pp. 106-119, Oct. 2010. Malaysia, in 2013. Since 2014, he is
[37] W. W. Wierwille and L. A. Ellsworth, “Evaluation of driver drowsiness pursuing a Master degree in
by trained raters,” Accident Analysis & Prevention., vol. 26, no. 5, pp.
571–581, Oct.1994 Engineering at Pukyong National
[38] G. S. Yang, Y. Z. Lin and P. Bhattacharya, “A driver fatigue recognition University, Busan, South Korea.
model based on information fusion and dynamic Bayesian network,” His current research interests the
Information Sciences., vol. 180, no. 10, pp.1942-1954, May. 2010. development of wearable healthcare
[39] G. Geetha, “Scrutinizing different techniques for artifact removal from
EEG signals,” International Journal of Engineering Science and sensors, mobile programming, and
Technology., vol. 3, no. 2, pp. 1167-1172, Feb. 2011. driver drowsiness detection.
[40] A. Murata, Y. Matsuda, M. Moriwaka and T. Hayami, “An Attempt to
Predict Drowsiness by Bayesian Estimation,” in Proc. SICE Annual Conf.
Tokyo, Japan, Sep. 13-18, 2011. pp. 58-63.
[41] A. Murata, Y. Urakami and M. Moriwaka, “An attempt to prevent traffic
accidents due to drowsy driving - Prediction of drowsiness by Bayesian
estimation,” in Proc. the SICE Annual Conference. Sapporo, Japan, Sep.
9-12, 2014, pp. 1708-1715.
[42] S. Gunn. (1998, May.). Support vector machines for classification and
regression. Univ. Southampton, Southampton, U. K., ISIS Tech. Rep,
1998. [Online]. pp. 6. Available:
http://www.svms.org/tutorials/Gunn1998.pdf

Gang Li received the B.E. degree in

Communication Engineering from
Nanchang Hangkong University,
Nanchang, China, in 2007, and the M.S.
degree in Computer Engineering from
Pusan National University, Busan,
South Korea, in 2010. He was a senior
researcher in Internet offload processor
for Internet-of-Things (IoT) in WIZnet
Co., Ltd, South Korea, from 2010 to
2012. He is currently pursuing a Ph.D.
degree in Electronic Engineering at Pukyong National
University, Busan, South Korea. His current research interests
include driver drowsiness detection technologies, wearable
EEG monitoring system, brain-machine interface system and
medical IoT.

Wan-Young Chung received the

B.Eng. and Master degrees in
Electronic Engineering from
Kyungpook National University,
Daegu, South Korea, in 1987 and
1989, respectively, and a Ph.D. in
Sensor Engineering from Kyushu
University, Fukuoka, Japan, in 1998.
He was an Associate Professor with
Dongseo University, Busan, South
Korea, from 1999 to 2008. He is
currently a Professor with the Department of Electronic

1530-437X (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.