Available online at www.sciencedirect.

com

ScienceDirect
Cognitive Systems Research 52 (2018) 715–728
www.elsevier.com/locate/cogsys

Electroencephalography based fatigue detection using a novel

feature fusion and extreme learning machine
Action editor: Sebastien Helie
Jichi Chen, Hong Wang ⇑, Chengcheng Hua
Department of Mechanical Engineering and Automation, Northeastern University, 110819 Shenyang, Liaoning, China

Received 12 July 2018; received in revised form 5 August 2018; accepted 16 August 2018
Available online 23 August 2018

Abstract

The unsafe behaviors of operators in fatigue state not only lead to declines of work efficiency but also higher error rates and more
injuries and even deaths. Automated fatigue detection using Electroencephalogram (EEG) due to helping us decline the occurrence prob-
ability of related accidents has gained more and more attention in recent years. Hence, designing a suitable feature extraction approach
and choosing an efficient classification methodology are considered as the key to successful implementation. We first propose a new min-
imum spanning tree (MST) feature extraction approach on the basis of the phase coherence (PC) and a power spectrum density (PSD)
method, respectively. Then to further improve the detection performance, we perform feature fusion (FBN-PSD-FF), where the func-
tional brain network (FBN)-based feature characterizing the relationship between brain network organization and fatigue and PSD-
based feature characterizing the relationship between power variation and fatigue. Furthermore, an automated fatigue detection system
has been developed, which is integrated between the novel fusion feature (FBN-PSD-FF) and extreme learning machine (ELM). Finally,
a driving simulation experiment is designed and conducted to demonstrate the proposed detection system, and the Karolinska Sleepiness
Scale (KSS) and the Stanford Sleepiness Scale (SSS) are employed to validate the results. Experimental results indicate that the proposed
method gives a sensitivity of 95.71%, a specificity of 94.29%, an accuracy of 95.00%, and a highest value of area under the receiver oper-
ating curve (AUC-ROC = 0.98). The ELM is employed for fatigue detection, reducing the time consuming greatly. Our proposed system
can be a viable solution for detecting operators’ fatigue and has great potential to reduce fatigue-related crashes in many circumstances
such as navigation, driving, aviation, construction industry, etc.
Ó 2018 Elsevier B.V. All rights reserved.

Keywords: Electroencephalography; Fatigue detection; Functional brain network; Power spectrum density; Extreme learning machine

1. Introduction construction industry, induced by fatigue was 10.6 per

Fatigue is one of the crucial factors resulting in lower
productivity, poor quality of work, increases accidents in
the construction industry and causes crashes in the trans-
portation industry (Aryal, Ghahramani, & Becerik-
Gerber, 2017; Chen, Taylor, & Comu, 2017). Based on a
report released in 2015, the non-fatal injury rate in
⇑ Correspondence author.
E-mail address: hongwang@mail.neu.edu.cn (H. Wang).
10,000 workers (Aryal et al., 2017). Usually, construction
workers perform a large number of heavy tasks under
harsh environmental conditions, which are likely to result
in fatigue (Cheng, Teizer, Migliaccio, & Gatti, 2013). Espe-
cially, in cold and hot environmental temperatures, physi-
cal fatigue not only leads to higher error rates but also
more work-related injuries (Rowlinson, Yunyanjia, Li, &
Chuanjingju, 2014). For the transportation industry, fati-
gue driving is a leading factor in road accidents which
results in serious injuries and fatalities (Chai et al., 2017;

https://doi.org/10.1016/j.cogsys.2018.08.018
1389-0417/Ó 2018 Elsevier B.V. All rights reserved.
716 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728
Michalaki, Quddus, Pitfield, & Huetson, 2015; Saxby,
Matthews, & Neubauer, 2017). Driver-related fatigue has
been defined as feeling tired and less alert while driving
(Chen and Xie, 2014). With the ever-growing car owners
on the road this issue will be even more severe (Wang
and Pei, 2014). As a result, an automated fatigue detection
while preserving the efficiency and simplicity is urgently
needed for reducing fatigue-related risks in the construc-
tion industry and transportation industry (Chai et al.,
2017; Fang, Jiang, Zhang, & Wang, 2015).
Currently, researchers have explored to develop various
algorithms for the automatic fatigue detection based on
physiological measures such as electrooculography
(EOG), electrocardiography (ECG), EEG, etc (Craig,
Tran, Wijesuriya, & Boord, 2006; Craig, Tran,
Wijesuriya, & Hung, 2012; Hu and Zheng, 2009; Patel,
Lal, Kavanagh, & Rossiter, 2011; Tran, Wijesuriya,
Tarvainen, Karjalainen, & Craig, 2009). Physiological mea-
sures can provide a precise and reliable assessment of the
true state of an operator and, in particular, may help to
determine the early stages of fatigue (Zhao et al., 2017).
Among these physiological measures, EEG is considered
to be the physiological ‘‘gold standard”, and it has been
verified that EEG is the most significant signal to indicate
fatigue (Zhao et al., 2017). However, the traditional visual
inspection of fatigue EEG signals for an experienced neu-
rologist is not only a challenging job but also a time con-
suming process (Nakamura, Chen, Sugi, Ikeda, &
Shibasaki, 2005). Hence, researches on automatic fatigue
detection using EEG have become more and more popular.
More specifically, designing a suitable feature extraction
approach and choosing an efficient classification methodol-
ogy are considered as the key to successful detection.
Different feature extraction methods based on EEG sig-
nals have been explored, which are mainly divided into
three categories. One type is based on time-domain analy-
sis; (Altunay, Telatar, & Erogul, 2010) used a linear predic-
tion filter to observe the presence of spikes and sharp waves
on EEG signals. Another type is based on frequency-
domain analysis; (Chen et al., 2017) introduced a EEG
approach to estimate task mental workload based on the
power spectral density (PSD) of major frequency bands.
However, the above two methods only reflect the single
characteristic of the EEG signals (Rabal, Cap, Criado, &
Alamo, 2012; Song, Hu, & Zhang, 2016). Besides, EEG
may hide more accurate information about neuronal activ-
ity, and unobtrusive features in the original full-spectrum
EEG may be obvious in the decomposed sub-bands
(Chen, Zhang, Zou, Zhao, & Wang, 2014). Hence, time-
frequency analysis represented by wavelet transform
(WT) has been investigated (Guo, Rivero, Dorado,
Rabunal, & Pazos, 2010; Kumar, Dewal, & Anand,
2014). In the current study, we first decompose and recon-
struct the collected EEG signals into multiple frequency
bands by employing wavelet packet transform (WPT),
which is an extension of the WT and leads to a full decom-
position tree (Wang, Chang, & Zhang, 2016). Considering
that phase synchronization (PS) analysis is an important
and effective means to infer neural connections (Kong,
Zhou, Jiang, Babiloni, & Borghini, 2017), and then, the
PS and the phase coherence (PC), characterizing the syn-
chronization of EEG signals, have been used in the feature
extraction. Furthermore, graph analysis not only enables
us to clearly recognize the relationship between the brain
network topology and fatigue (Zhao et al., 2017) but also
helps to extract some of the graphic characteristics as the
final features and is hence used in this paper. However,
conventional functional brain network (FBN)-based graph
analysis exists complicated problems in setting the thresh-
old procedures as well as comparing networks constructed
for a range of thresholds (Boersma et al., 2013; Tewarie,
Van Dellen, Hillebrand, & Stam, 2015). Hence, a novel
minimum spanning tree (MST) approach, which focuses
on the key structural network backbones and overcomes
the threshold problem (Stam et al., 2014; Tewarie et al.,
2015; van Wijk, Stam, & Daffertshofer, 2010), is used in
this paper and to our best knowledge has not been applied
to fatigue EEG signals.
Another main task is to develop an efficient classifier
with a low false rate to complete the fatigue detection suc-
cessfully. The commonly used classifiers applied to fatigue
detection are summarized as follow. Artificial neural net-
work (ANN), imitating the biological nervous system to
process the information, has been developed to detect driv-
ing status (Garces Correa, Orosco, & Laciar, 2014; Singh,
Conjeti, & Banerjee, 2013). Support vector machine
(SVM), which is established by construct a separating
hyper-plane in high dimensional space, has been employed
for classification (Zhao, Zhao, Liu, & Zheng, 2012; Zhao,
Zheng, Zhao, Tu, & Liu, 2011). Logistic Regression
(LR), a regression analysis method, has been successfully
applied to disease classification (Liao and Chin, 2007). K
Nearest Neighbors classifier (KNN), a supervised learning
technique (Li, Hu, Sun, & Cai, 2016), has been employed
for driver fatigue detection (Hu and Wang, 2017). How-
ever, the aforementioned classification method may result
in a decrease in learning speed due to the large sample size,
which cannot meet the requirements of real-time applica-
tions. Extreme learning machine (ELM) (Ding, Zhao,
Zhang, Xu, & Nie, 2015; Huang, Zhu, & Siew, 2006) is a
novel learning algorithm which develops on the basis of
the single-hidden layer feed-forward neural networks
(SLFN). ELM not only can avoid local optimal solution,
but also greatly improve the learning speed, thus attracting
a large number of researchers (Chen, Zhao, Zhang, & Zou,
2015; Song and Zhang, 2016; Yuan, Zhou, Li, & Cai,
2011).
In this paper, we first propose a novel minimum span-
ning tree (MST) feature extraction approach on the basis
of the phase coherence (PC) and a power spectrum density
(PSD) method, respectively. Then to further improve the
detection performance, we perform feature fusion (FBN-
PSD-FF), where the functional brain network (FBN)-
based feature characterizing the relationship between brain
 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728
network organization and fatigue and PSD-based feature
characterizing the relationship between power variation
and fatigue. Furthermore, an automated fatigue detection
system has been developed, which is integrated between
the novel fusion feature (FBN-PSD-FF) and ELM. The
main contributions of this paper lie in the following 3
aspects. (1) The correlation between brain network organi-
zation and fatigue is studied without setting the threshold.
(2) The novel fusion feature (FBN-PSD-FF) which has not
been explored previously improves the classification accu-
racy. (3) The ELM is employed for fatigue detection,
reducing the time consuming greatly.
The rest of this paper is described as below. Section 2
describes the details of the materials and methods, brief
information about the functional connectivity and power
spectrum density that we have used as features and the
ELM classifiers which we have adopted in this work. The
experimental results obtained in this work are presented
in Section 3. Section 4 reviews some similar studies and
compares our results with other methods and results in
the literature. Finally, the conclusion is presented in
Section 5.
2. Materials and methods
2.1. Participants
Twelve healthy male postgraduate students, from the
School of Mechanical Engineering and Automation,
Northeastern University (NEU), take part in this experi-
ment, ranging in age from 24 to 33 (M = 27.33,
SD = 2.81) years old. According to their self-reports, all
the participants have normal or corrected-to-normal
vision, normal hearing and are right-hand dominant. None
of them are reported of any neurological diseases in the
past nor have taken any drugs known to affect EEG
(Craig, Hancock, & Craig, 1996). All the participants
own driving licenses and have at least two years of driving
experience. They are asked to sleep adequately (at least 8 h)
at night before the experiments. Moreover, coffee, alcohol,
tea and tobacco are prohibited prior to participating in the
experiments. At the beginning of the experiments, the
experimental procedures are explained to ensure that each
participant understands the instructions. After signing the
informed consent agreements, each participant receives
10-yuan in compensation for his effort and time. All the
procedures in the experiments are approved by the local
Institute Review Board (IRB) of NEU.
2.2. Apparatus and experimental protocol
Because working in dangerous situations or driving in
fatigue cannot be investigated using real environment
(Liu et al., 2016), simulated driving is hence needed for
exploring safety issues about driving.
The driving simulator, developed by Beijing jing-
gongkeye Scientific Education Co., Ltd, is used to record
717
and measure participants’ driving behaviors. The simulator
consists of a steering wheel, a horn, a clutch, a brake pedal,
an accelerator, a shift, a chair and turn signals. Among
them, the clutch is not used because of the selection of
the automatic shift mode in the study. A 32-inch LCD
monitor shows other cars, the current speed, other road
stimuli, which plays at a rate of 24 frames per second.
The simulator is modeled after a Volkswagen (Santana)
coupe and clearly visible from the front windshield and
rearview mirrors. In addition, the simulator also provides
engine noise as well as nearby traffic noises, which has
the same performance as real vehicle.
The experiments are carried out in a controlled (sound
and temperature) laboratory environment (Chen, Wang,
& Hua, 2018). In order to reduce the differences in partic-
ipants’ familiarity with the driving simulator, each partici-
pant is asked to perform a 15-min familiarization driving
(FD) phase. In the FD phase, all the participants are famil-
iar with the simulated driving environment while wearing
EEG cap. When the participants become comfortable with
the FD phase, they are asked to perform self-evaluation of
their drowsiness level. Subsequently, a monotonous driving
(MD) begins at 2:00 PM, which is the normal working hour
in China. During the MD phase, all the participants are
required to drive and follow the road rules for 60 min with-
out any visual distraction. More specifically, in order to
induce drowsiness naturally, a bi directional and two-lane
road is selected in the MD phase, eliminating the need to
change lanes. Participants are asked to drive at a constant
speed of about 60 km/h without playing music or radio. In
addition, participants are instructed to avoid unnecessary
movements to reduce artifacts in the physiological data
recording (Fu, Wang, & Zhao, 2016). After the MD phase,
participants are again asked to perform self-evaluation of
their drowsiness level. During the whole experiment, par-
ticipants can leave the experiment with the help of an
observer at any phase if they feel uncomfortable. In addi-
tion, drowsy signs such as yawning and nodding are
obtained manually by the observer to validate the fatigue.
An illustration of the experimental protocol is shown in
Fig. 1.
2.3. Self-evaluation of drowsiness
Two scales serve as self-evaluation of drowsiness level of
the participants. One is Karolinska Sleepiness Scale (KSS)
(Akerstedt and Gillberg, 1990) which is the most com-
monly adopted subjective measure of drowsiness (Kong,
Lin, Babiloni, Hu, & Borghini, 2015; Yang, Lin, &
Bhattacharya, 2010), rating the participants on an ordinal
scale from 1 (extremely alert) to 9 (very sleepy). And
another is Stanford Sleepiness Scale (SSS) which has been
shown to be sensitive to deficits in alertness (Hoddes,
Zarcone, Smythe, Phillips, & Dement, 1973). Unlike
KSS, SSS rates the participants on an ordinal scale from
1 (feeling active) to 7 (sleep onset soon) depending on their
state of alertness. All the participants are asked to measure
718 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728

0
15 20 23
77 80
Time (min)

255
255
255
255
255

255
255

255
255

255
255
255
255
255
255
255
255

255
255
255
255

255
255
255
255
255

255
255

255

255

255
255
255
255
255
255

255

255
255
255

255
255
HEOG HEOG
VEOG Fp1 Fp2 VEOG
FP1 FP1
FP2 FP2
F7 Fz F7
F3 F3 F4 F3
FZ
FZ
F4
F7 F8 F4
F8 F8
FT7 FC3 FCz FT7
FC3 FC4 FT8 FC3
FCZ FT7 FCZ
FC4 FC4
FT8 FT8
T3 T3
C3 T7 C3 Cz C4 T8 C3
CZ
CZ
C4 C4
T4 T4
TP7 TP7
CP3 TP7 CP3 TP8 CP3
CPZ CPz Cp4 CPZ
CP4 CP4
TP8 TP8
T5 P7 P8 T5
P3 P3 P3
PZ Pz P4 PZ
P4 Scale
P4 Scale
T6 T6
O1 O1 25
OZ 25 O1 Oz O2 OZ
O2 +- O2 +-

0 1 2 3 4 5 0 1 2 3 4 5

Fig. 1. Experimental protocol (KSS: Karolinska Sleepiness Scale, SSS: Stanford Sleepiness Scale).

how fatigue they feel before and after MD phase using the data (post-task) for each participant are selected for further
two scales. For more detailed information about the two processing. For both pre-task and post-task, the 3-min
scales, please refer to (He et al., 2017). EEG signals from 30 channels are segmented into 2-s
epochs (2000 sample points per epoch), resulting in 90
2.4. Data acquisition and processing epochs for each subject. With the 12 participants, a total
of 1080 epochs of dataset are formed for pre-task and
EEG signals are recorded with 40 channels mounted on another 1080 epochs for post-task.
a cap (NuAmps, NeuroScan) across the experiment. 30
EEG electrodes (Ag/AgCl) are positioned at FP1 (1), 2.5. Functional connectivity
FP2 (2), F7 (3), F3 (4), FZ (5), F4 (6), F8 (7), FT7 (8),
FC3 (9), FCZ (10), FC4 (11), FT8 (12), T7 (13), C3 (14), 2.5.1. Phase synchronization and mean phase coherence
CZ (15), C4 (16), T8 (17), TP7 (18), CP3 (19), CPZ (20), The notion of phase synchronization (PS) was intro-
CP4 (21), TP8 (22), P7 (23), P3 (24), PZ (25), P4 (26), P8 duced to chaotic oscillators (chaotic signals) by Arch,
(27), O1 (28), OZ (29) and O2 (30) according to the stan- Rosenblum, Pikovsky, and Kurths (1996). In order to com-
dard 10–20 system (see Fig. 1). Additionally, 2 channels pute PS, the instantaneous phase of the two signals
EOG electrodes attached above the arch of the left eyebrow involved should be known first. This can be achieved via
and below the eye are used to monitor blink activities as the analytical signal based on the Hilbert transform. The
part of classification standard for both visual inspection analytical signal zx(t) is first defined:
and automatic detection. The impedances of all the elec- 
zx ðtÞ ¼ xðtÞ þ i xðtÞ ¼ Ax ðtÞei/x ðtÞ ð1Þ
trodes are below 5 kO and are referenced to linked mas-
toids of both sides. The EEG data is recorded with a where z(t) is complex-valued. x(t) denotes a real time series,
band-pass filter set at 0.01 –70 Hz to remove the noise and x(t) denotes its corresponding Hilbert transform,
and with a sampling rate of 1000 Hz. Furthermore, EEG which is obtained using integration as follows:
data is analyzed off-line with the NeuroScan 4.3 software Z þ1
 1 xðsÞ
program. The EEG data affected by excessive drift is first x ðtÞ ¼ PV ds ð2Þ
p 1 t s
rejected by visual inspection. Prior to EEG quantitative
analysis, eye movement correction protocols are applied where PV represents the Cauchy principal value. From Eq.
to the entire set of recorded data to remove the DC shift (1), the mathematical expressions to compute the instanta-
and eye movement artifacts. Blink artifacts in the EEG neous amplitude Ax(t) and the instantaneous phase /x ðtÞ
recordings are then removed by using a regression analysis are formulated in Eqs. (3) and (4), respectively:
in combination with artifact averaging implemented algo- qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2
rithm (Huang, Lin, & Hwang, 2017). After that, the first Ax ðtÞ ¼ ½ x ðtÞ þ ½xðtÞ2 ð3Þ
3-min of EEG data (pre-task) and the last 3-min of EEG
 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728

x ðtÞ
/x ðtÞ ¼ arctan ð4Þ
xðtÞ
Similarly, analytical signal zy(t) is also defined as follow:
zy ðtÞ ¼ Ay ðtÞei/x ðtÞ ð5Þ
Hence, the phase difference or relative phase D/xy ðtÞ is
computed as a function of time:
D/xy ðtÞ ¼ jn/x ðtÞ  m/y ðtÞj 6 const ð6Þ
where n and m are integers. Since our goal is to compute
the synchronization between EEG signals which are
obtained from the same person, we consider it most likely
to restrict the case n = m = 1. In the current paper, we
employ the phase coherence (PC) to measure phase syn-
chronization using the following Eq. (7):
 
1 XN 1 
ið/x ðtk Þ/y ðtk ÞÞ 
kðx; yÞ ¼  e  ð7Þ
N  k¼0  (BC), describing the fraction of all shortest paths which
where N denotes the number of samples and tk is discrete
time-step. The value of k ranges from 0 to 1. More specif-
ically, k = 0 implies the signal x(t) and y(t) are unsynchro-
nized, while k = 1 indicates perfect phase synchronization.
The computation of the PC method is performed using
MATLAB R2017a. Prior to the PC computation, the
EEG signal is decomposed into various sub-frequency
bands using wavelet packet transform (WPT): delta
(0–4 Hz), theta (4–8 Hz), alpha (8–12 Hz) and beta
(12–32 Hz). Then, a PC adjacency matrix (C) of N  N
(30  30)-where N is the electrode number-is obtained for
each frequency band and all epochs, and then is averaged
across all epochs. Each element Cij of C matrix indicates
the connection strength between electrodes ‘i’ and ‘j’ at
specific frequency band. The adjacency matrix C is sym-
metric, indicating that Cij = Cji, and self-connections are
excluded, meaning zeros in the diagonal of this symmetric
matrix. Hence, we obtain N  (N  1)/2 = 30  29/2 =
435 different functional PC values in matrix C.
2.5.2. Minimum spanning tree
In order to investigate how the functional brain network
topology property is associated with driving drowsiness,
the graph theoretical representation of the functional con-
nectivity matrix is established by minimum spanning tree
(MST) method (Tewarie et al., 2015). MST is a relatively
novel approach which has not been commonly used in bio-
logical signal, and to our knowledge not applied to driving
drowsiness detection of the original EEG signal.
Given a connected, undirected weight graph, a MST of
the graph is a sub-graph which connects all the nodes but
does not contain circles or loops (Stam et al., 2014;
Tewarie et al., 2015). In the current paper, the MST is con-
structed by using Kruskal’s algorithm (Kruskal, 1956).
More specifically, we start the construction of a MST with
the link from the highest PC value, and then the second
highest PC value is added as the MST connection, and so
719
on. If adding a connection leads to a circle or loop, this
connection will be skipped. Here, 30 EEG channels are
included for analyses, resulting in a MST of the graph con-
taining 30 nodes with 29 edges for each frequency band and
epoch.
For the current analysis, various MST metrics are quan-
tified based on the measures described by Stam (2014). The
degree (d) is the number of the links for a given node. The
diameter (D) represents the largest distance between any
two nodes of the tree. The leaf fraction (l) indicates the
number of nodes with d = 1 divided by the total number
of nodes in the MST. The kappa (k) is the measure of
the broadness of the degree distribution, whose mathemat-
ical formula is
hd 2 i
k¼ ð8Þ
hdi
where <> denotes the average. The between centrality
pass through a particular node, is a measure of the hubness
of the node within the MST. The tree hierarchy (Th) quan-
tifies the balance between communication path and the
overload of central nodes (Dubbelink et al., 2014). It can
be expressed in the following way:
L
Th ¼ ð9Þ
2mBC max
where L is the leaf number; m is the number of links in the
MST and BCmax indicates the maximum value of BC. In
the case of L = 2 (line-like configuration), and m ! 1,
then Th!0. In the case of L = m (star-like configuration),
then Th!0.5. For L in between these two extreme situa-
tions, Th has higher values.
2.6. Power spectrum density
Power spectrum density (PSD) analysis is considered to
be the most popular methods to analysis driver fatigue
(Craig et al., 2012; Lal, Craig, Boord, Kirkup, &
Nguyen, 2003). PSD defines how the power of the signal
or time series is distributed with the frequency. The Welch’s
method (Welch, 1967) is one of the most effective
approaches to estimate the power spectrum at any signal
or time series. The calculation of the PSD of the Welch
spectrum is as follows:
1X S
P^ w ðf Þ ¼ P^ l ðf Þ ð10Þ
S l¼1
where the Pb w(f) represents the Welch PSD estimation; S is
the number of segments, and Pb l(f) represents the peri-
odogram estimate of lth segment. In the current paper,
the PSD is computed (sliding window of 1 s, 50% overlap,
grouped into frequency bands) for each of the 2 s files per
participant (90 epochs), and then is averaged across all
epochs. All the PSD analyses are performed in Brainstorm
(Tadel, Baillet, Mosher, Pantazis, & Leahy, 2011).
720 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728

2.7. Statistical analysis
Paired t-tests are used for accessing the KSS and SSS
(pre-task and post-task). Moreover, topological network
differences (PC values as well as MST values) between
pre-task and post-task are evaluated using paired t-tests
for each frequency band separately. We also apply the t-
test to examine the differences between pre-task and post-
task for the PSD for each frequency bands. The level of sig-
nificance of the above-mentioned statistical analyses is set
at p = 0.05. Before applying t-test, we check that the data
are normally distributed using the K-S test (Awais,
Badruddin, & Drieberg, 2017). Here, statistical analyses
are performed using SPSS for windows, and all the data
are presented as mean ± standard deviation.
2.8. Extreme learning machine
Extreme learning machine (ELM), first introduced by
Huang et al. (2006) is a novel learning algorithm which
develops on the basis of the single-hidden layer feed-
forward neural networks (SLFN). In general, traditional
neural networks is very time consuming while ELM can
achieve a good performance with fast learning because
ELM is not necessary to adjust the input weights and hid-
Table 1
Subjective (KSS and SSS score) across all the participants. Abbreviations:
M, mean; SD, standard deviation.
Measures Pre-task Post-task
M (SD) M (SD)
KSS 2.84 (0.75) 7.30 (1.14)
SSS 1.84 (0.55) 5.53 (0.66)
Note: Bold text represents significant results.
den layer biases only need to set the number of hidden
nodes. ELM with kernels was proposed by Huang (2014),
which does not need to tune the centers and impact widths
of kernels and simply analyze the output weights instead of
interactive adjustment. In the current research, we test
three kinds of ELM, basic ELM with sigmoid activation
function, basic ELM with sine activation function and
ELM with RBF kernels, respectively.
3. Results
3.1. Subjective evaluation
The KSS is a well-validated measure of subjective state-
related fatigue. Participants are asked to evaluate how fati-
gue they currently feel on a 9-point ordinal scale. As shown
in Table 1, the KSS score is generally increased from 2.84
± 0.75 (M ± SD) at pre-task to 7.30 ± 1.14 at post-task.
Also, KSS score varies significantly with driving
(p < 0.001). The self-reported drowsiness measured by the
SSS yields similar results (see Table 1). The SSS score is sig-
nificantly increased from 1.84 ± 0.55 at pre-task to 5.53
± 0.66 at post-task. Moreover, fatigue signs such as yawn-
ing and nodding are observed frequently in all the partici-
pants during the post-task. These illustrate that our
experiment manipulation is successful to incur states of
fatigue.
3.2. Phase coherence analysis
p value
The mean phase coherence matrices for pre-task and
post-task in the delta, theta, alpha and beta frequency
<0.001
bands are shown in Fig. 2. It can be observed that the
<0.001
phase coherence matrices for both tasks in different fre-
quency bands show complex but similar configurations.

Delta Theta Alpha Beta

0.9

0.8
Pre-task

0.7

0.6

0.5

0.4

0.3
Post-task

0.2

0.1

Fig. 2. Mean phase coherence matrices for pre-task (upper row) and post-task (lower row) in the delta, theta, alpha and beta frequency bands. The matrix
is a 30  30 square matrix, where X axis and Y axis correspond with the channels numbers, and where the entries indicate the mean strength of the PC
between specific pairs of channels. The strength of the PC is indicated with a color scale, from blue to yellow. (PC: phase coherence). (For interpretation of
the references to color in this figure legend, the reader is referred to the web version of this article.)
 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728 721

More specifically, high level and low level of coherence are
reflected by yellow and blue respectively. Visual scrutiny
already suggests a significant difference between the two
tasks, particularly in the delta, alpha and beta frequency
band. Network differences of mean phase coherence for
pre-task and post-task in different frequency bands are
tested by student statistical test (t-test), and the results
are shown in Table 2. The mean PC of the pre-task is sig-
nificantly lower than that of the post-task in the delta fre-
quency band (p = 0.031). And the mean PC of the pre-task
is also significantly lower than that of the post-task both in
alpha frequency band (p = 0.008) and beta frequency band
(p = 0.003). Moreover, there are no significant differences
in the mean PC for theta frequency band.
The graphs corresponding to the mean phase coherence
matrices of Fig. 2 are shown in Fig. 3. From Fig. 3, it can
be seen that the graphs of both the two tasks have similar
complex network configurations, which consist of frontal-
temporal-central-parietal-occipital direction of information
flux. Sketchy differences between the pre-task and post-task
can be observed via visual inspection, more specifically, the
brain network structure of the pre-task is significantly spar-
ser than that of the post-task for the delta frequency band.
Additionally, compared with the pre-task, the strength of
Table 2
T-test for mean phase coherence for each frequency band.
Frequency band Pre-task Post-task
M (SD) M (SD)
Delta band 0.708 (0.031) 0.741
Theta band 0.750 (0.029) 0.755
Alpha band 0.730 (0.042) 0.787
Beta band 0.740 (0.041) 0.812
Note: Bold text represents significant results (p < 0.05).
the functional connection for post-task is much higher. It
is the similar case for both alpha and beta frequency band.
The brain network structure is more intensive and the
strength of the functional connection is significantly higher
are found for post-task compared to pre-task.
3.3. Minimum spanning tree analysis
To overcome the problem of bias, we attempt to con-
struct the MST of the network to further explore the differ-
ences between the pre-task and post-task. The MST
matrices derived from the mean PC matrices of Fig. 2 are
shown in Fig. 4. The MST matrices are constructed by
using Kruskal’s algorithm. From Fig. 4, the differences of
the MST matrices between the pre-task and post-task can-
not be obviously observed via visual inspection. Hence, we
convert the MST matrices to MST loopless graph and then
calculate various MST metrics, as shown in Figs. 5 and 6,
respectively. The MST results suggest that the group effect
on kappa, relating to the broadness of the degree distribu-
tion, in the beta frequency band, yields significance,
p = 0.023, suggesting a higher kappa in the post-task
(M = 2.310, SD = 0.074) relative to the pre-task
(M = 2.219, SD = 0.068). The group effect on diameter
(D), reflecting the largest distance between any two nodes
of the tree, is also significant in the alpha frequency band,
p = 0.037, indicating higher D in the pre-task (M = 0.527,
SD = 0.065) compared to the post-task (M = 0.465,
p-value SD = 0.048). The group effects on leaf in delta, alpha and
beta frequency band, are significantly lower in the pre-
(0.027) 0.031* task compared to the post-task p = 0.035, p = 0.011 and
(0.026) 0.220 p = 0.008, respectively. The group effects on tree hierarchy
(0.026) 0.008* (Th) in delta, alpha and beta frequency band are also signif-
(0.068) 0.003* icant, p = 0.048, p = 0.019 and p = 0.004, respectively,
indicating lower Th in the pre-task compared to the

Delta Theta Alpha Beta

0.9
Pre-task

0.8

0.7

0.6

0.5

0.4

0.3
Post-task

0.2

0.1

Fig. 3. Mean phase coherence matrices of Fig. 2 converted to graph. Schematic image of the head seen from above, with the positions of the electrodes
indicated by small black dots and numbered according to the 10–20 system placement electrode. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)
722 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728

Delta Theta Alpha Beta

Pre-task
Post-task

Fig. 4. MST matrices derived from mean PC matrices. (MST: minimum spanning tree, PC: phase coherence).

Delta Theta Alpha Beta

Pre-task
Post-task

Fig. 5. MST loopless graph. Schematic image of the head seen from above, with the positions of the electrodes indicated by small black dots and
numbered according to the 10–20 system placement electrode. (MST: minimum spanning tree).

post-task. Collectively, these results suggest a more inte-
grated network organization in the post-task compared
to the pre-task.
3.4. Power spectrum density analysis
Fig. 7 shows the head plots of power spectrum density in
the delta, theta, alpha and beta frequency bands. High
activity is represented by the red-shaded areas, whereas
low activity is indicated by the blue-shaded areas. Addi-
tionally, the t-test results of all brain sites for PSD per fre-
quency band are shown in Table 3. According to Table 3
and Fig. 7, a significant increase in the delta frequency
bands at central sites (p = 0.013) in the post-task is easily
observed compared with the result in the pre-task. For
the theta frequency bands, the significant increases of
PSD are located on frontal (p = 0.025) and temporal
(p = 0.017), and significant decrease mainly focused on
occipital (p = 0.022) in the post-task compared to the
pre-task. Result for the alpha frequency band reveals sig-
nificant decreases at central (p = 0.031) and occipital
(p = 0.022) sites in the post-task relative to the pre-task.
Beta activity significantly decreases, especially in the pari-
etal (p = 0.027) and occipital sites (p = 0.013), in the
post-task compared to the pre-task.
3.5. Detection performance analysis
In this work, the functional brain network (FBN)-based
and the power spectrum density (PSD)-based features
 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728 723

(a) 2.60 (b) 0.70

2.55 Pre-task 0.65 Pre-task
2.50 Post-task Post-task
0.60 *
2.45
2.40 * 0.55

MST Diameter
MST Kappa 2.35 0.50
2.30 0.45
2.25 0.40
2.20
0.35
2.15
0.30
2.10
2.05 0.25
2.00 0.20
Delta Theta Alpha Beta Delta Theta Alpha Beta
Frequency band Frequency band

(c) 0.55 (d) 0.45

0.50
Pre-task Pre-task
Post-task 0.40 Post-task
0.45
* *
* 0.35 *

MST Tree hierarchy

0.40 * *
0.30
MST Leaf

0.35
0.30 0.25
0.25
0.20
0.20
0.15
0.15
0.10 0.10
Delta Theta Alpha Beta Delta Theta Alpha Beta
Frequency band Frequency band

Fig. 6. MST metrics between the pre-task and post-task in different frequency. (MST: minimum spanning tree).

Delta Theta Alpha 10-13 Beta

10-12 10-12 10-13
4.5 1.5
9 7
4
8
3.5 6
7
3 1 5
6
Pre-task

2.5
5 4
2 4 3
1.5 0.5 3
2
1 2
0.5 1
1
0 0 0
10-12 10-12 10-13 10-13
8 1.5 4.5
8
7 4
7
6 3.5
6
Post-task

5 1 3
5
2.5
4 4
2
3 3
0.5 1.5
2 2 1
1 1 0.5
0 0 0
10-12 10-13 10-13 10-13
3 5
6 4
4
2 3
4 3
2 2
Difference

2 1
1 1
0 0 0 0
-1 -1
-2 -1 -2
-2
-4 -3
-2 -3
-4
-6 -5 -4
-3

Fig. 7. Head plots of power spectrum density per frequency band. The top-row represents the pre-task, the middle-row indicates the post-task and the
bottom-row expresses the differences between the two tasks. (For interpretation of the references to color in this figure legend, the reader is referred to the
web version of this article.)
724 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728

Table 3 enhanced comparing with the two single features FBN-F

T-test results of all brain sites for PSD per frequency band. and PSD-F. Additionally, the proposed fusion features
Frequency band Frontal Temporal Central Parietal Occipital FBN-PSD-FF with ELM-RBF achieves a highest acc of
Delta band – – 0.013 – – 95.00%, highest sens of 95.71%, highest spec 94.29% and
Theta band 0.025 0.017 – – 0.022 much less time of only 0.038 s which shows the better per-
Alpha band – – 0.031 – 0.018 formance over other classifiers.
Beta band – – – 0.027 0.013
Note: Shows results with significant p-values (<0.05). 4. Discussion

Research has shown that the unsafe behaviors of opera-

showing significant differences are applied independently
(FBN-F, PSD-F), and together (FBN-PSD-FF), to the
classifiers (basic ELM with sigmoid activation function
and ELM with RBF kernels) using leave-one-out-cross-
validation (LOOCV). All but one participants’ data is
trained for the classifier. The remaining one’s data is used
for testing. This process is repeated 12 times. In order to
show the superior performance of the ELM, we also use
the SVM classifier for comparison. The performance of
the classifiers is accessed using the measures of sensitivity,
specificity, accuracy and consuming time, and the related
definitions are given in Table 4.
In order to selected optimal number of hidden nodes, we
attempt to gradually increase the node numbers from 5 to
30 with an interval of 5. The training accuracy of ELM
with different numbers of hidden nodes is illustrated in
Table 5. It can be seen that the training accuracy of
ELM is almost stable when the number of hidden nodes
is from 20 to 30. From the point of the avoiding the
over-fitting and consuming less time, the number of hidden
nodes is set to 20 in the current study.
We compare the detection performance between the
proposed fusion features FBN-PSD-FF and other two sin-
gle features FBN-F and PSD-F with the same classifiers
(basic ELM with sigmoid activation function, basic ELM
with sine activation function, ELM with RBF kernels
and SVM). We also calculated the consuming time in
MATLAB R2017a with Intel CoreTM i5-3210M CPU @
2.5 GHz with 6 GB RAM. The experimental results includ-
ing the sens, spec, acc, and time are shown in Table 6. It
can be seen from Table 6 that the performance of the pro-
posed fusion features FBN-PSD-FF has been greatly
tors in fatigue state not only lead to declines of work effi-
ciency but also higher error rates and more injuries and
even deaths, especially in many circumstances such as driv-
ing, construction industry, aviation, navigation, etc (Chen
et al., 2017; Zhao et al., 2017). Hence, automated fatigue
detection due to helping us decline the occurrence probabil-
ity of related accidents has gained increasing attention in
recent years. The performance of the proposed method is
verified from the following 3 aspects: (1) performance com-
parison between the proposed fusion feature FBN-PSD-FF
and other two single features PSD-F and FBN-F with the
same classifier ELM (ELM-sig and ELM-RBF); (2) perfor-
mance comparison between ELM (ELM-sig and ELM-
RBF) and SVM with the same fusion feature FBN-PSD-
FF; (3) performance comparison between the proposed
method with some similar studies that have been published
during recent years.
Usually, the receiver operating characteristic (ROC)
curve is applied as an intuitive method to demonstrate
the relationship between true positive rate (sensitivity)
and false positive rate (1-Specificity) using different thresh-
old ratios as a variable. More specifically, Fig. 8(a) is ROC
curves for SVM classifier; Fig. 8(b) is ROC curves for basic
ELM with sigmoid activation function (ELM-sig) classi-
fier; Fig. 8(c) is ROC curves for ELM with RBF kernels
(ELM-RBF) classifier; Fig. 8(d) is ROC curves for ELM
with sine activation function (ELM-sine) classifier. Fur-
thermore, in order to quantitatively evaluate the classifica-
tion performance, the areas under ROC curve (AUC) are
also computed. As shown in Fig. 8, the straight lines con-
necting (0, 0) to (1, 1) indicate random performances of
the classifiers. The curves appearing in the lower right

Table 4
Measures of characterizing classification performance.
Measure Definition
Sensitivity (Sens) Proportion of the correctly classified fatigue samples out of the total number of labeled fatigue samples
Specificity (Spec) Proportion of the correctly classified non-fatigue samples out of the total number of labeled non-fatigue samples
Accuracy (Acc) Proportion of the correctly classified samples out of the total number of samples
Consuming time (Time) the machine time consumed on both training and testing

Table 5
The training accuracy of ELM with different numbers of hidden nodes.
Numbers of hidden nodes 5 10 15 20 25 30
Training accuracy (%) 85.47 90.78 92.68 95.71 96.08 96.14
 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728 725

Table 6
Performance comparison among PSD-F, FBN-F and FBN-PSD-FF.
Features Classifiers Performance
Sens. (%) Spec. (%) Acc. (%) Time (s)
PSD-F ELM-sig 82.86 88.57 85.71 0.032
ELM-sine 85.71 82.86 84.29 0.030
ELM-RBF 82.86 85.71 84.29 0.047
SVM 82.29 78.57 81.43 6.115
FBN-F ELM-sig 90.00 91.43 90.71 0.025
ELM-sine 88.57 87.14 87.86 0.040
ELM-RBF 87.14 90.00 88.57 0.031
SVM 84.29 81.43 82.86 6.813
FBN-PSD-FF ELM-sig 95.71 90.00 92.86 0.040
ELM-sine 92.86 91.43 92.14 0.044
ELM-RBF 95.71 94.29 95.00 0.038
SVM 87.14 91.43 89.29 3.890
The best performance is marked in bold.

1 1 AUC=0.97
(a) AUC=0.95 (b) AUC=0.96
AUC=0.92
0.9 AUC=0.89 AUC=0.87 0.9

0.8 0.8
True positive rate (Sensitivity)
True positive rate (Sensitivity)

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3

0.2 0.2
FBN-PSD-FF FBN-PSD-FF
0.1 PSD-F 0.1 FBN-F
FBN-F PSD-F
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
False positive rate (1-Specificity) False positive rate (1-Specificity)

1
(c) AUC=0.98 (d)
AUC=0.90
0.9 AUC=0.94

0.8
True positive rate (Sensitivity)

0.7

0.6

0.5

0.4

0.3

0.2
FBN-PSD-FF
0.1 PSD-F
FBN-F
0
0 0.2 0.4 0.6 0.8 1
False positive rate (1-Specificity)

Fig. 8. ROC curves and ROC-AUCs (the areas under ROC curve) for three features PSD-F, FBN-F and FBN-PSD-FF. (a) SVM (b) ELM-sig (c) ELM-
RBF (d) ELM-sine (ROC: receiver operating characteristic, PSD-F: power spectrum density-feature, FBN-F: functional brain network-feature, FBN-
PSD-FF: functional brain network- power spectrum density – feature fusion, SVM: support vector machine, ELM-sig: extreme learning machine with
sigmoid activation function, ELM-RBF: extreme learning machine with RBF kernels, ELM-sine: extreme learning machine with sine activation function).
726 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728
triangle performs worse than random performance,
whereas the curves appearing in the upper left triangle have
superior performance classifications (Chai et al., 2017).
Obviously, a larger AUC value represents a better classifi-
cation performance (Chai et al., 2017; Chen et al., 2015;
Song et al., 2016). From Fig. 8, it is observed that the
AUC achieved by FBN-PSD-FF (AUC = 0.95 for SVM,
AUC = 0.97 for ELM-sig) is much larger than those by
other two single features PSD-F (AUC = 0.87 for SVM,
AUC = 0.92 for ELM-sig) and FBN-F (AUC = 0.89 for
SVM, AUC = 0.96 for ELM-sig). Particularly, it is note-
worthy that FBN-PSD-FF yields in largest areas
(AUC = 0.98) with ELM-RBF, which means a better per-
formance of fatigue detection can be achieved by using
ELM-RBF and FBN-PSD-FF.
We also compare the classification performance consid-
ered in this work with some similar studies that have been
published during recent years. Guo et al. (2016) collected
physiological signals including EEG from driver and
employed grey correlation analysis to select the input vari-
able to the support vector machine, and then obtained the
classification accuracy of 86% for driver fatigue detection
Yan, Coenen, Yue, Yang, and Zhang (2016) classified dri-
ver’s vigilance using a hierarchical classification system with
multiple features based video, and the overall prediction
accuracy obtained using this approach was 89%. Addition-
ally, (Tjolleng et al., 2017) developed an artificial neural net-
work (ANN) model to classify the driver fatigue based on
electrocardiography (ECG), which obtained an accuracy
of 82.0% for testing data. Li, He, Fan, and Fei (2012) estab-
lished an evaluation model for driver fatigue with the
regression equation based on the EEG signals, obtaining
an accuracy of 91.5%. In this research, an automated fati-
gue detection system has been developed, which is inte-
grated between the novel fusion feature (FBN-PSD-FF)
and extreme learning machine (ELM-RBF), which gives a
highest acc of 95.00% in comparison with the similar studies
that have been published which is 91.5% (Li et al., 2012), a
sens of 95.71%, a spec of 94.29%, and a highest value of area
under the receiver operating curve (AUC-ROC = 0.98).
The merits of our proposed method lie not only in getting
a higher acc but also in giving us a better understanding of
the correlation between brain network and fatigue. Zhao
et al. (2017) studied the relationship between brain network
reorganization and mental fatigue, and they set a series of
thresholds and choose the optimal threshold for graphical
analysis. Here, we employ a novel method-MST to con-
struct brain networks, which is a simplified representation
of a stable network core of the original network with mini-
mized connection cost, and it connects all the nodes in the
original weighted network with no loop (Stam et al., 2014;
Stam, 2014; Zare, Rezvani, & Benasich, 2016). In this
way, due to the same number of nodes and links of the
MSTs, the network properties between groups are directly
compared, which avoids the aforementioned methodologi-
cal biases and threshold problems.
Support vector machine (SVM) (Lee, Lee, & Chung,
2014; Zhao et al., 2011), which is established by construct
a separating hyper-plane in high dimensional space, and
artificial neural network (ANN) (Garces Correa et al.,
2014; Singh et al., 2013; Tjolleng et al., 2017), imitating
the biological nervous system to process the information,
have been developed to detect driving status. However,
the aforementioned classification method may result in a
decrease in learning speed due to the large sample size,
which cannot meet the requirements of real-time applica-
tions. Xu et al. (2018) encoded 32 characters with a
space-code division multiple access scheme and decoded
EEG features with a discriminative canonical pattern
matching algorithm. The experimental results demon-
strated the feasibility of using this method to achieve an
efficient BCI system. In this paper, the proposed extreme
learning machine (ELM) (Huang et al., 2006) is a novel
learning algorithm which develops on the basis of the
single-hidden layer feed-forward neural networks (SLFN).
ELM is not necessary to adjust the input weights and hid-
den layer biases only need to set the number of hidden
nodes. Additionally, ELM with kernels (Huang, 2014) does
not need to tune the centers and impact widths of kernels
and simply analyze the output weights instead of interac-
tive adjustment. Compared with other classifiers, ELM
not only can avoid local optimal solution, but also greatly
improve the learning speed greatly.
5. Conclusion
In this work, a novel minimum spanning tree (MST) fea-
ture extraction approach on the basis of the phase coher-
ence (PC) and a power spectrum density (PSD) method
are first proposed, respectively. Then to further improve
the detection performance, we perform feature fusion
(FBN-PSD-FF), where the functional brain network
(FBN)-based feature characterizing the relationship
between brain network organization and fatigue and
PSD-based feature characterizing the relationship between
power variation and fatigue. Furthermore, an automated
fatigue detection system has been developed, which is inte-
grated between the novel fusion feature (FBN-PSD-FF)
and extreme learning machine (ELM). Finally, a driving
simulation experiment is designed and conducted to
demonstrate the proposed detection system. All the exper-
imental results demonstrate that the proposed method does
a good job and can be a viable solution for detecting oper-
ators’ fatigue.
Acknowledgements
We gratefully acknowledge the financial support from
National Key R & D Program of China
(2017YFB1300300), Foundation of State Key Laboratory
of Automotive Simulation and Control (20171101), the
Fundamental Research Funds for the Central Universities
 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728
of China (N150308001) and the University Innovation
Team of Liaoning Province (LT2014006).
References
Akerstedt, T., & Gillberg, M. (1990). Subjective and objective sleepiness in
the active individual. International Journal of Neuroscience, 52(1–2), 29.
Altunay, S., Telatar, Z., & Erogul, O. (2010). Epileptic eeg detection using
the linear prediction error energy. Expert Systems with Applications, 37
(8), 5661–5665.
Arch, M., Rosenblum, M. G., Pikovsky, A. S., & Kurths, J. (1996). Phase
synchronization of chaotic oscillators. Physical Review Letters, 76,
1804–1807, 1(11), 2–5.
Aryal, A., Ghahramani, A., & Becerik-Gerber, B. (2017). Monitoring
fatigue in construction workers using physiological measurements.
Automation in Construction, 82, 154–165.
Awais, M., Badruddin, N., & Drieberg, M. (2017). A hybrid approach to
detect driver drowsiness utilizing physiological signals to improve
system performance and wearability. Sensors, 17(9).
Boersma, M., Smit, D. J., Boomsma, D. I., De Geus, E. J., Ha, D. D. W.,
& Stam, C. J. (2013). Growing trees in child brains: Graph theoretical
analysis of electroencephalography-derived minimum spanning tree in
5- and 7-year-old children reflects brain maturation. Brain Connectiv-
ity, 3(1), 50.
Chai, R., Naik, G. R., Tuan Nghia, N., Ling, S. H., Tran, Y., Craig, A., &
Nguyen, H. T. (2017). Driver fatigue classification with independent
component by entropy rate bound minimization analysis in an eeg-
based system. Ieee Journal of Biomedical and Health Informatics, 21(3),
715–724.
Chen, J., Taylor, J. E., & Comu, S. (2017). Assessing task mental workload
in construction projects: A novel electroencephalography approach.
Journal of Construction Engineering and Management, 143(8).
Chen, J., Wang, H., & Hua, C. (2018). Assessment of driver drowsiness
using electroencephalogram signals based on multiple functional brain
networks. International Journal of Psychophysiology. https://doi.org/
10.1016/j.ijpsycho.2018.07.476.
Chen, C., & Xie, Y. (2014). Modeling the safety impacts of driving hours
and rest breaks on truck drivers considering time-dependent covari-
ates. Journal of Safety Research, 51, 57–63.
Chen, L.-L., Zhang, J., Zou, J.-Z., Zhao, C.-J., & Wang, G.-S. (2014). A
framework on wavelet-based nonlinear features and extreme learning
machine for epileptic seizure detection. Biomedical Signal Processing
and Control, 10, 1–10.
Chen, L.-L., Zhao, Y., Zhang, J., & Zou, J.-Z. (2015). Automatic
detection of alertness/drowsiness from physiological signals using
wavelet-based nonlinear features and machine learning. Expert
Systems with Applications, 42(21), 7344–7355.
Cheng, T., Teizer, J., Migliaccio, G. C., & Gatti, U. C. (2013). Automated
task-level activity analysis through fusion of real time location sensors
and worker’s thoracic posture data. Automation in Construction, 29,
24–39.
Craig, A., Hancock, K., & Craig, M. (1996). The lifestyle appraisal
questionnaire: A comprehensive assessment of health and stress.
Psychology & Health, 11(3), 331–343.
Craig, A., Tran, Y., Wijesuriya, N., & Boord, P. (2006). A controlled
investigation into the psychological determinants of fatigue. Biological
Psychology, 72(1), 78–87.
Craig, A., Tran, Y., Wijesuriya, N., & Hung, N. (2012). Regional brain
wave activity changes associated with fatigue. Psychophysiology, 49(4),
574–582.
Ding, S., Zhao, H., Zhang, Y., Xu, X., & Nie, R. (2015). Extreme learning
machine: Algorithm, theory and applications. Artificial Intelligence
Review, 44(1), 103–115.
Dubbelink, K. T. E. O., Hillebrand, A., Stoffers, D., Deijen, J. B., Twisk,
J. W. R., Stam, C. J., & Berendse, H. W. (2014). Disrupted brain
network topology in parkinson’s disease: A longitudinal magnetoen-
cephalography study. Brain, 137, 197–207.
727
Fang, D., Jiang, Z., Zhang, M., & Wang, H. (2015). An experimental
method to study the effect of fatigue on construction workers’ safety
performance. Safety Science, 73, 80–91.
Fu, R., Wang, H., & Zhao, W. (2016). Dynamic driver fatigue detection
using hidden markov model in real driving condition. Expert Systems
with Applications, 63, 397–411.
Garces Correa, A., Orosco, L., & Laciar, E. (2014). Automatic detection
of drowsiness in eeg records based on multimodal analysis. Medical
Engineering & Physics, 36(2), 244–249.
Guo, M., Li, S., Wang, L., Chai, M., Chen, F., & Wei, Y. (2016). Research
on the relationship between reaction ability and mental state for online
assessment of driving fatigue. International Journal of Environmental
Research and Public Health, 13(12).
Guo, L., Rivero, D., Dorado, J., Rabunal, J. R., & Pazos, A. (2010).
Automatic epileptic seizure detection in eegs based on line length
feature and artificial neural networks. Journal of Neuroscience Meth-
ods, 191(1), 101–109.
He, J., Choi, W., Yang, Y., Lu, J., Wu, X., & Peng, K. (2017). Detection
of driver drowsiness using wearable devices: A feasibility study of the
proximity sensor. Applied Ergonomics, 65, 473–480.
Hoddes, E., Zarcone, V., Smythe, H., Phillips, R., & Dement, W. C.
(1973). Quantification of sleepiness: A new approach. Psychophysiol-
ogy, 10(4), 431–436.
Hu, J., & Wang, P. (2017). Noise robustness analysis of performance for
eeg-based driver fatigue detection using different entropy feature sets.
Entropy, 19(8).
Hu, S., & Zheng, G. (2009). Driver drowsiness detection with eyelid
related parameters by support vector machine. Expert Systems with
Applications, 36(4), 7651–7658.
Huang, G.-B. (2014). An insight into extreme learning machines: Random
neurons, random features and kernels. Cognitive Computation, 6(3),
376–390.
Huang, C. Y., Lin, L. L., & Hwang, I. S. (2017). Age-related differences in
reorganization of functional connectivity for a dual task with
increasing postural destabilization. Frontiers in Aging Neuroscience, 9.
Huang, G.-B., Zhu, Q.-Y., & Siew, C.-K. (2006). Extreme learning
machine: Theory and applications. Neurocomputing, 70(1–3), 489–501.
Kong, W., Lin, W., Babiloni, F., Hu, S., & Borghini, G. (2015).
Investigating driver fatigue versus alertness using the granger causality
network. Sensors, 15(8), 19181–19198.
Kong, W., Zhou, Z., Jiang, B., Babiloni, F., & Borghini, G. (2017).
Assessment of driving fatigue based on intra/inter-region phase
synchronization. Neurocomputing, 219, 474–482.
Kruskal, J. B. (1956). On the shortest spanning subtree of a graph and the
traveling salesman problem. Proceedings of the American Mathematical
Society, 7(1), 48–50.
Kumar, Y., Dewal, M. L., & Anand, R. S. (2014). Epileptic seizure
detection using dwt based fuzzy approximate entropy and support
vector machine. Neurocomputing, 133, 271–279.
Lal, S. K. L., Craig, A., Boord, P., Kirkup, L., & Nguyen, H. (2003).
Development of an algorithm for an eeg-based driver fatigue coun-
termeasure. Journal of Safety Research, 34(3), 321–328.
Lee, B.-G., Lee, B.-L., & Chung, W.-Y. (2014). Mobile healthcare for
automatic driving sleep-onset detection using wavelet-based eeg and
respiration signals. Sensors, 14(10), 17915–17936.
Li, W., He, Q. C., Fan, X. M., & Fei, Z. M. (2012). Evaluation of driver
fatigue on two channels of eeg data. Neuroscience Letters, 506(2),
235–239.
Li, X., Hu, B., Sun, S., & Cai, H. (2016). Eeg-based mild depressive
detection using feature selection methods and classifiers. Computer
Methods and Programs in Biomedicine, 136, 151–161.
Liao, J. G., & Chin, K.-V. (2007). Logistic regression for disease
classification using microarray data: Model selection in a large p and
small n case. Bioinformatics, 23(15), 1945–1951.
Liu, T., Liu, Y., He, W., He, W., Yu, X., Guo, S., & Zhang, G. (2016). A
passenger reduces sleepy driver’s activation in the right prefrontal
cortex: A laboratory study using, near-infrared spectroscopy. Accident
Analysis and Prevention, 95, 358–361.
728 J. Chen et al. / Cognitive Systems Research 52 (2018) 715–728
Michalaki, P., Quddus, M. A., Pitfield, D., & Huetson, A. (2015).
Exploring the factors affecting motorway accident severity in england
using the generalised ordered logistic regression model. Journal of
Safety Research, 55, 89–97.
Nakamura, M., Chen, Q., Sugi, T., Ikeda, A., & Shibasaki, H. (2005).
Technical quality evaluation of eeg recording based on electroen-
cephalographers’ knowledge. Medical Engineering & Physics, 27(1),
93–100.
Patel, M., Lal, S. K. L., Kavanagh, D., & Rossiter, P. (2011). Applying
neural network analysis on heart rate variability data to assess driver
fatigue. Expert Systems with Applications, 38(6), 7235–7242.
Rabal, H., Cap, N., Criado, C., & Alamo, N. (2012). Holodiagrams using
mahalanobis distance. Optik, 123(19), 1725–1731.
Rowlinson, S., Yunyanjia, A., Li, B., & Chuanjingju, C. (2014).
Management of climatic heat stress risk in construction: A review of
practices, methodologies, and future research. Accident Analysis and
Prevention, 66, 187–198.
Saxby, D. J., Matthews, G., & Neubauer, C. (2017). The relationship
between cell phone use and management of driver fatigue: It’s
complicated. Journal of Safety Research, 61, 129–140.
Singh, R. R., Conjeti, S., & Banerjee, R. (2013). A comparative evaluation
of neural network classifiers for stress level analysis of automotive
drivers using physiological signals. Biomedical Signal Processing and
Control, 8(6), 740–754.
Song, J.-L., Hu, W., & Zhang, R. (2016). Automated detection of epileptic
eegs using a novel fusion feature and extreme learning machine.
Neurocomputing, 175, 383–391.
Song, Y., & Zhang, J. (2016). Discriminating preictal and interictal brain
states in intracranial eeg by sample entropy and extreme learning
machine. Journal of Neuroscience Methods, 257, 45–54.
Stam, C. J. (2014). Modern network science of neurological disorders.
Nature Reviews Neuroscience, 15(10), 683–695.
Stam, C. J., Tewarie, P., Van Dellen, E., Van Straaten, E. C. W.,
Hillebrand, A., & Van Mieghem, P. (2014). The trees and the forest:
Characterization of complex brain networks with minimum spanning
trees. International Journal of Psychophysiology, 92(3), 129–138.
Tadel, F., Baillet, S., Mosher, J. C., Pantazis, D., & Leahy, R. M. (2011).
Brainstorm: A user-friendly application for meg/eeg analysis. Compu-
tational Intelligence and Neuroscience.
Tewarie, P., Van Dellen, E., Hillebrand, A., & Stam, C. J. (2015). The
minimum spanning tree: An unbiased method for brain network
analysis. Neuroimage, 104, 177–188.
Tjolleng, A., Jung, K., Hong, W., Lee, W., Lee, B., You, H., ... Park, S.
(2017). Classification of a driver’s cognitive workload levels using
artificial neural network on ecg signals. Applied Ergonomics, 59,
326–332.
Tran, Y., Wijesuriya, N., Tarvainen, M., Karjalainen, P., & Craig, A.
(2009). The relationship between spectral changes in heart rate
variability and fatigue. Journal of Psychophysiology, 23(3), 143–151.
Van Wijk, B. C. M., Stam, C. J., & Daffertshofer, A. (2010). Comparing
brain networks of different size and connectivity density using graph
theory. PLoS One, 5(10).
Wang, H., Chang, W., & Zhang, C. (2016). Functional brain network and
multichannel analysis for the p300-based brain computer interface
system of lying detection. Expert Systems with Applications, 53,
117–128.
Wang, L., & Pei, Y. (2014). The impact of continuous driving time and
rest time on commercial drivers’ driving performance and recovery.
Journal of Safety Research, 50, 11–15.
Welch, P. D. (1967). The use of fast fourier transform for the estimation of
power spectra: A method based on time averaging over short, modified
periodograms. IEEE Transactions on Audio and Electroacoustics, 15(2),
70–73.
Xu, M. P., Xiao, X. L., Wang, Y. J., Qi, H. Z., Jung, T. P., & Ming, D.
(2018). A brain-computer interface based on miniature-event-related
potentials induced by very small lateral visual stimuli. IEEE Trans-
actions on Biomedical Engineering, 65(5), 1166–1175.
Yan, C., Coenen, F., Yue, Y., Yang, X., & Zhang, B. (2016). Video-based
classification of driving behavior using a hierarchical classification
system with multiple features. International Journal of Pattern Recog-
nition and Artificial Intelligence, 30(5).
Yang, G., Lin, Y., & Bhattacharya, P. (2010). A driver fatigue recognition
model based on information fusion and dynamic bayesian network.
Information Sciences, 180(10), 1942–1954.
Yuan, Q., Zhou, W., Li, S., & Cai, D. (2011). Epileptic eeg classification
based on extreme learning machine and nonlinear features. Epilepsy
Research, 96(1–2), 29–38.
Zare, M., Rezvani, Z., & Benasich, A. A. (2016). Automatic classification
of 6-month-old infants at familial risk for language-based learning
disorder using a support vector machine. Clinical Neurophysiology, 127
(7), 2695–2703.
Zhao, C., Zhao, M., Liu, J., & Zheng, C. (2012). Electroencephalogram
and electrocardiograph assessment of mental fatigue in a driving
simulator. Accident Analysis and Prevention, 45, 83–90.
Zhao, C., Zhao, M., Yang, Y., Gao, J., Rao, N., & Lin, P. (2017). The
reorganization of human brain networks modulated by driving mental
fatigue. IEEE Journal of Biomedical and Health Informatics, 21(3),
743–755.
Zhao, C., Zheng, C., Zhao, M., Tu, Y., & Liu, J. (2011). Multivariate
autoregressive models and kernel learning algorithms for classifying
driving mental fatigue based on electroencephalographic. Expert
Systems with Applications, 38(3), 1859–1865.