G Model

ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS

Artificial Intelligence in Medicine xxx (2017) xxx–xxx

Contents lists available at ScienceDirect

Artificial Intelligence in Medicine

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

An EEG-based functional connectivity measure for automatic

detection of alcohol use disorder
Wajid Mumtaz, Mohamad Naufal b Mohamad Saad, Nidal Kamel, Syed Saad Azhar Ali,
Aamir Saeed Malik ∗
Center for Intelligent Signal and Imaging Research (CISIR), Department of Electrical and Electronic Engineering, Universiti Teknologi PETRONAS, 32610, Seri
Iskandar, Perak, Malaysia

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

Article history: Background: The abnormal alcohol consumption could cause toxicity and could alter the human brain’s
Received 30 May 2017 structure and function, termed as alcohol used disorder (AUD). Unfortunately, the conventional screen-
Received in revised form 15 August 2017 ing methods for AUD patients are subjective and manual. Hence, to perform automatic screening of
Accepted 10 November 2017
AUD patients, objective methods are needed. The electroencephalographic (EEG) data have been uti-
lized to study the differences of brain signals between alcoholics and healthy controls that could further
Keywords:
developed as an automatic screening tool for alcoholics.
Alcohol use disorder (AUD)
Method: In this work, resting-state EEG-derived features were utilized as input data to the proposed
Alcohol abuse (AA)
Alcohol dependence (AD)
feature selection and classification method. The aim was to perform automatic classification of AUD
Electroencephalography (EEG) patients and healthy controls. The validation of the proposed method involved real-EEG data acquired
Resting-state EEG (REEG) from 30 AUD patients and 30 age-matched healthy controls. The resting-state EEG-derived features such
Synchronization likelihood as synchronization likelihood (SL) were computed involving 19 scalp locations resulted into 513 features.
Furthermore, the features were rank-ordered to select the most discriminant features involving a rank-
based feature selection method according to a criterion, i.e., receiver operating characteristics (ROC).
Consequently, a reduced set of most discriminant features was identified and utilized further during
classification of AUD patients and healthy controls. In this study, three different classification models
such as Support Vector Machine (SVM), Naïve Bayesian (NB), and Logistic Regression (LR) were used.
Results: The study resulted into SVM classification accuracy = 98%, sensitivity = 99.9%, specificity = 95%,
and f-measure = 0.97; LR classification accuracy = 91.7%, sensitivity = 86.66%, specificity = 96.6%, and
f-measure = 0.90; NB classification accuracy = 93.6%, sensitivity = 100%, specificity = 87.9%, and f-
measure = 0.95.
Conclusion: The SL features could be utilized as objective markers to screen the AUD patients and healthy
controls.
© 2017 Elsevier B.V. All rights reserved.

1. Introduction report academic consequences because of drinking, including miss-

Alcohol use disorder (AUD) is usually characterized by a severe
alcohol intake (Association, 2013). Currently, about 17.6 million
adults in the U.S. suffer from alcohol abuse or dependence [1].
Unfortunately, several million more people engage in risky, binge
drinking patterns that can lead towards alcohol-related prob-
lems [1]. Every year, about 1825 students aged between 18–24,
incur alcohol-related unintentional fatal injuries including motor-
vehicle crashes [2]. In addition, about one in four college students
∗ Corresponding author.
E-mail address: aamir saeed@petronas.com.my (A.S. Malik).
ing classes, falling behind in class, doing poorly on exams or papers
and receiving lower grades overall [2]. Chronic heavy drinking
eventually leads to AUD, alcohol abuse (AA), or alcohol dependence
(AD). More specifically, the AD is a more severe form of AA. Unfor-
tunately, heavy consumption of alcohol and its cumulative toxic
effects may lead to medical, neurological, psychiatric and social
problems. According to the definition, alcohol consumption less
than 48 g per day or 144 g per week is characterized as safe [3].
According to the Diagnostic and Statistical Manual of Mental Dis-
orders V (DSMV), people with AA keep drinking despite social and
personal problems (Association, 2013). In addition, people with AD
not only fulfill the criteria of AA, but also develop increased tol-

https://doi.org/10.1016/j.artmed.2017.11.002
0933-3657/© 2017 Elsevier B.V. All rights reserved.

Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
2 W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx
erance and withdrawal symptoms once abandoned drinking, also
termed as alcoholics [4].
Conventionally, the screening and assessment of AUD patients
involves questionnaire-based techniques such as AUDIT (Alcohol
Use Disorder Identification Test) [5]. However, the subjective feed-
backs observed from AUD patients may confound the screening
process. For example, many AUD patients are less candid and
unable to measure alcohol consumption [6–8]. Hence, assessments
with questionnaire-based techniques could implicate into misjudg-
ments during screening and assessing actual quantity of alcohol
consumption unless supported with neuroimaging modalities such
as electroencephalography (EEG) [9,10]. The EEG is a standard
modality and has been utilized for various applications such as
monitoring depth of anesthesia [11], assessment of spontaneous
perceptual switching [12], assessing learning processes [13], and
modeling purposes [14,15].
In the literature, various EEG features have shown clinical rele-
vance with AUD, e.g., the inter-hemispheric coherence, phase delay
and synchronization likelihood (SL) have been proposed to explore
functional influences among different brain regions [16]. A study
investigated whether heavily drinking students, although drinking
for a shorter period than alcoholics, already show differences in the
brain functional connectivity (FC) compared to light-drinking con-
trols [17]. EEG was recorded from male student drinkers (11 light
and 11 heavy drinkers) during eyes closed (EC), and EC plus men-
tal rehearsal of pictures. The FC was assessed with SL. As a result,
only the heavy drinkers exhibited increased EEG synchronization
implicated as the changes in hippocampal–neocortical connectiv-
ity. In another study, the brain connectivity has been found different
between alcoholics and non-alcoholics [18]. These studies have
shown the significance of synchronization differences between the
alcoholics and healthy controls.
The inter-hemispheric coherence can quantify the functional
coupling between two spatially located EEG sensors represent-
ing distinct brain regions. However, contradictory findings have
been reported regarding the changes in the inter-hemispheric
coherence. For example, Tcheslavski et al. highlighted significant
reduction of EEG power, inter-hemispheric coherence and phase
synchronization in alcoholics as compared with controls [19]. On
the contrary, higher inter-hemispheric coherence in first degree
male relative (parents, full siblings, or children) of alcoholics was
found in the frontal and centro-parietal regions than in controls
without a family history of AUD [20]. Moreover, studies based on
EEG data have identified differences of neuronal activities among
different brain regions in alcoholics and healthy controls [21–23].
EEG data are of composite nature and can be decomposed into
frequency bands such as delta (0.5–4 Hz), theta (4–8 Hz), alpha
(8–12 Hz), beta (12–30 Hz) and gamma (> 30 Hz). The EEG spec-
tral power analysis has been the most popular EEG method to
discriminate alcoholics and healthy controls. For example, higher
theta power has been reported in alcoholics when compared with
healthy controls [17,18,24–26]. This abnormal increase may inhibit
the ability to encode new information [27]. Similarly, an increase
of theta power at all scalp loci, prominent at central and parietal in
males and at the parietal for females has been reported [25]. In addi-
tion, significant changes in theta power has been associated with
cortical atrophy (Coutin-Churchman et al., 2006; Saletu-Zyhlarz
et al., 2004). A higher low-voltage alpha (LVA) (< 10 ␮V) has been
reported in alcoholics than healthy controls [28,29]. However, the
observed difference could not exhibit statistical significance. In a
study, an increased beta power is reported as a primary character-
istic feature for alcoholics and high risk subjects, associated with
benzodiazepines intake, that was mainly used for alcohol detoxifi-
cation [24,30]. Therefore, beta power needs to be considered as a
marker during alcohol treatment. On the contrary, theta and delta
bands were found significantly increased in alcoholics when com-
Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
pared with healthy controls. In addition, these bands were not
affected by medication or found in people with family history of
AA. However, the findings based on the alpha and gamma bands
are not matured yet and considered as active research areas.
Recently, machine learning (ML) techniques have shown
promising results to solve issues related with screening alcoholic
subjects from healthy controls [31–34]. However, the requirements
for clinical application are tough and need strong evidences regard-
ing EEG as a modality that could be utilized to classify the AUD
patients and healthy controls [35]. Hence, the primary objective of
this work is to investigate the EEG-based measure, i.e., SL to quan-
tify the FC from EEG data. In addition, the study emphasized the
importance of SL as a feature to discriminate the alcoholics and
healthy controls. Moreover, the SL is compared with other FC mea-
sures such as the mutual information and the interhemispheric
coherence. The secondary objective is to develop a less complex ML
method than the ones presented in the literature [34] that show
high efficiencies based on the EEG data acquired from the AUD
patients and healthy controls. In brief, the proposed ML method
involved a general methodology of feature extraction, selection,
and classification validated with 10-fold cross validation (10-CV).
2. Method
2.1. Study participants
In this work, thirty (30) AUD patients (mean 55.4 ± 12.87 years)
and thirty (30) age-matched healthy controls (mean
42.67 ± 15.90 years) were recruited from clinic Bingkor, Sabah, East
Malaysia. Table 1 shows the available demographic information for
the study participants. The study protocol for experimental data
acquisition was approved by the ethics committee of Universiti
Malaya, Malaysia. All participants were volunteers and had signed
the consent forms of participation. The experimental procedure
was briefed to the study participants. The healthy participants
were assessed for any neurological disorder and were found naïve.
According to the study’s inclusion criteria, the AUD patients must
met the diagnostic criteria defined by the Alcohol Use Disorders
Identification Test (AUDIT) [36]. In this study, the participants with
AUDIT score greater than seven were categorized as AUD patients
[37].
2.2. Experimental data acquisition
In this study, the experimental data included the acquisition
of resting-state EEG data and the clinical assessment scores based
AUDIT data. The resting-state EEG data were recorded with the Dis-
covery 24E system involving EEG caps with 19 electrodes. As shown
in Fig. 1, the electrodes were located on the scalp according to the
international 10–20 system [38]. The Discovery 24E EEG system
digitized the brain signals at a sampling rate of 256 samples per sec-
ond. The EEG data were filtered at 0–70 Hz with an additional 50 Hz
notch filter to supress the line noises. The resting-state EEG record-
ings included five minutes of eyes-closed (EC) and five minutes of
eyes-open (EO) data. The participants were instructed to sit relaxed
in a semi-recumbent position during the recordings. During EO ses-
sion, the participants were instructed to sit relaxed with less eye
Table 1
The demographic information of the study participants.
Information AUD Patients Healthy Controls
Age (Mean ± SD) 55.4 ± 12.87 years 42.67 ± 15.90 years
Gender (Males/Females) 16/14 18/12
Smoking/Non-Smoking 3/27 1/29
Handedness (right/left) 20/10 23/7
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx
Fig. 1. Topomap showing EEG sensors locations on the scalp.
movements to reduce the artifacts. During the recording sessions,
it was make sure that the quality of the EEG data was maintained.
Hence, by the end of the recording, enough data were recorded so
that the clean data were achieved for all study participants.
2.3. EEG noise reduction
The EEG data were confounded with noise or artifacts such as
eye blinks, eye movements and muscular activities, e.g., heart beats.
The data confounded with artifacts may not truly represent the
underlying neural activities; therefore, reduction of noise from the
recorded EEG data was desirable and termed as EEG data prepro-
cessing. In this paper, the preprocessing was performed involving
the multiple source modeling technique [39], implemented in
the standard brain electric source analysis (BESA) software [40].
According to the technique, the noise topographies and a head
model were utilized to effectively clean the EEG noise. For example,
to remove eye-blinks, the noise vectors for eye blinks were con-
structed from the recorded EEG data. Based on these noise vectors,
eye-blink topographies were constructed. These noise topogra-
phies and a head model, already provided in BESA, were then
utilized to successfully clean the EEG data for eye-blinks. In addi-
tion, noise topographies for the head movement, heart or muscular
activities were constructed. Finally, a head model and artifacts
topographies due to muscular and heart artifacts were utilized to
clean head movement, heart or muscular activities from the data.
The construction of noise topographies including eye-blinks, mus-
cular and heart activities was based on the recorded EEG data and
then can be applied to whole recording. This method preserves the
information in the data as it allows the EEG data affected by arti-
facts to be utilized in the data analysis. In comparison, the methods
based on artifact rejection also delete the EEG data which may have
result into loss of information.
2.4. The proposed ML method
Fig. 2 shows an overview of the proposed ML method involving
EEG feature extraction, selection, classification and validation. As
detailed in previous subsection, the EEG artifact reduction impli-
cated into clean EEG data which was followed by EEG feature
extraction. For feature extraction, the artifact-free EEG data were
segmented into one-minute epochs per study subject. In this study,
the raw EEG data have been subjected to preprocessing that means
reduction of artifacts while keeping the underlying data intact.
Hence, the selection of one-minute artifact-free EEG data was pos-
sible for each study participant.
Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
3
As a result, a large number (Nc) of candidate features were
extracted and arranged column wise in a matrix, termed as the EEG
data matrix. Columns of the matrix represented features and each
column was denoted as xi, where i = 1. . . Nc. Rows of the matrix
represented the study participants (AUD patients and healthy con-
trols) with physiological conditions (both EC and EO) per study
participant; denoted by L = [(xi,yi), i = 1 . . . Nc] including both the
EEG features and the corresponding output class labels or targets,
i.e., y = [Alcoholics or Controls]. A detailed description for each sub-
process is provided in the respective subsections.
In this study, the feature selection was based on a rank-based
feature selection method [41]. In brief, each feature in the EEG data
matrix was assigned a weight and rank-ordered accordingly. The
weights were computed according to the area under curve (AUC)
of each feature, termed as z-values. Next, the features were rank-
ordered such as the feature with highest z-value was placed at the
top of the list and other features were arranged in descending order
of their respective z-values. A detailed description on the feature
selection method is provided in the respective subsection.
The classifier training and testing was performed after rank-
ordering the features. For computing the classifier test accuracies
as a function of number of features, each feature was progressively
added into a reduced set of features, e.g., starting with the high-
est z-value feature among the list of rank-ordered features. Upon
addition of each feature, the size of reduced set of feature matrix
kept increasing. For example, during first iteration, the matrix size
was 60 × 1 (where 60 corresponded to number of examples and
the 1 corresponded to the feature itself). For this matrix, classi-
fier performance was computed. During second iteration, adding
another feature into the reduced set of feature matrix implicated
into a matrix size of 60 × 2, and the classifier performance was com-
puted. The process kept repeating until all features were utilized,
i.e., 60 × 513. As a result, the classifier performance was plotted as a
function of the number of features. The feature set that showed best
accuracy among others was selected and considered as selected
features. Since the process is repeated for each fold of 10-CV, the
final values were computed by taking average across each fold. As
a result, all features set were tested accordingly. To compute the
classifier accuracy, sensitivity, specificity and f-measure, a confu-
sion matrix was constructed involving the classifier predictions and
the actual labels. In addition, the classifier accuracies were com-
pared based on the ROC plots as well. In the rest of this section, a
detailed description on feature extraction, selection, classification
and validation is provided.
2.4.1. Feature extraction
The SL is a measure to quantify synchronization of two times
series data [42]. Because of the quantification the SL value may have
values between 0 and 1: a minimum value such as zero represents
complete non-synchronization. On the other hand, a maximum
value i.e., one corresponds to the completely synchronized EEG data
at the specified brain location. For example, the EEG data recorded
during the eyes closed (EC) rest state and eye open (EO) rest state
is found to have small synchronization values. In short, the SL is a
good measure to quantify the EEG data to discriminate the mental
states.
In this study, the recorded EEG data consisted of multiple chan-
nels which were denoted as M simultaneously recorded time series
xk,i , where k denoted the channel number and (k = 1,2,3, . . ., M) and
i denoted discrete time (i = 1,2,3, . . ., N). As described in Eq. (1),
the EEG data corresponding to a channel was used to construct
embedding vectors Xk,i by using the time-delay embedding [43]:
X k,i = (xk,i ,xk,i+l ,xk,i+2l , ...,xk,i+(m−1)l ) (1)
where l is lag and m is the embedding dimension. For each channel
ε
k and each time i, a probability Pk,i was computed such that embed-
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
4 W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx

Fig. 2. Proposed ML method to classify the AUD patients and healthy controls.

ding vector are closer to each other than a distance ␧ as described where the embedding vectors Xk,i and Xk,j will be closer together
in equation than the critical distance εk,i :


M
Hi,j = (εk,i − |Xk,i − Xk,j |) (3)
1 N
ε
Pk,i = (ε − |Xk,i − Xk,j |) (2) k=1
2(ω2 − ω1 )
j=1 This number lies in a range between 0 and M, and reflects how
ω1 < |i − j| < ω2 many of the embedded signals ‘resemble’ each other. Now, for each
channel k and each discrete time pair (i,j), the synchronization like-
lihood in Eq. (4) as defined:
Here the |.| is Euclidean distance and  is the Heaviside step Hi,j − 1
function, (x) = 0 if x ≤ 0 and (x) = 1 for x > 0. Here ␻1 and ␻2 are two if |Xk,i − Xk,j | < εk,i : Sk,i,j =
M−1 (4)
windows; ␻1 is Theiler correction for autocorrelation effects and
if |Xk,i − Xk,j | ≥ εk,i : Sk,i,j = 0
should be at-least of the order of autocorrelation time [44]; ␻2 is a
window that sharpens the time resolution of the synchronization
As mentioned in Eq. (4), by averaging over all j synchronization
measure and is chosen such that ω1 < < ω2 < <N.
likelihood Sk,i is obtained as follows:
Now for each k and each i the critical distance εk,i is deter-
ε
mined for which Pk,i = pref , where pref< <1. In this study, the value 
N
1
of the critical distance εk,i was found via empirical process such as Sk,i = Sk,i,j (5)
it should fulfil the requirement Pk,i ε
= pref . 2(ω2 − ω1 )
j=1
Now for each time pair (i,j) and within the considered window
ω1 < |j − i| < ω2
(ω1 < |i-j|< ω2 ), it was easy to determine the number of channels Hi,j

Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx 5

In this study, the SL was computed for each channel pair involv- Table 2
Pseudo code for feature ranking method.
ing frontal (Fp1, Fp2, F3, F4, F7, F8, Fpz), temporal (T3, T4, T5,
T6), parietal (P3, P4, P7, P8), occipital (O1, O2), and central (C3, y = patterns(:,2);
C4). For example, the synchronization likelihood computed for Fp1 p = cumsum(y = = 1);
tp = p/sum(y = = 1);
included channel pairs such as Fp1-Fp2; Fp1-F4; Fp1-F8; Fp1-T4;
n = cumsum(y = = −1);
Fp1-T6; Fp1-P4; Fp1-P8; Fp1-O2; Fp1-C4. The extracted features fp = n/sum(y = = −1);
were arranged column-wise in a matrix, termed as the EEG data patterns = [x y];
matrix. Furthermore, the EEG data matrix was subjected to features patterns = sortrows(patterns,−1);
selection because the features could be redundant or irrelevant. n = length(tp);
Y = (tp(2:n) + tp(1:n − 1))/2;
X = fp(2:n) − fp(1:n − 1);
2.4.2. EEG data matrix and z-score standardization auc = sum(Y.*X)−0.5;
The features extraction has resulted into EEG data matrix involv-
ing the number of rows (data points = 60). The matrix might not
be centered and unequally distributed. Hence, the data stan- Table 3
dardization was performed involving z-score standardization. The Sample EEG data.
standardization was performed by computing values column-wise Sample ID i j Label
by subtracting each element value with its column-wise mean and
1 −0.2 +0.5 (−)
divided by the corresponding standard deviation. A true assessment 2 −1.4 −1.4 (−)
of any classification model requires independence between the test 3 +0.8 −0.9 (−)
and train samples. In this study, the independence was maintained 4 −0.8 +0.2 (+)
during EEG features extraction, standardization and classification. 5 +0.1 −2.5 (+)
6 +0.5 +1.4 (−)
More specifically, the z-scores were computed for each feature
7 +1.6 −0.3 (+)
involving the feature mean and standard deviation. Therefore, sep- 8 −2.1 −1.2 (−)
arate computation of z-score did not affect the performance of the 9 −0.3 +2.2 (+)
proposed ML scheme. Moreover, the separate treatment of the test 10 +3.4 −1.7 (−)
and train samples during each step implicated into test accuracy,
sensitivity, specificity and f-measure. Therefore, in this study, the
z-scores were computed separately for test and train samples. Table 4
Intermediate variables values.

2.4.3. Rank-based feature selection method Feature values (sorted in labels p n tp fp

Feature selection remains as challenging research topic and car- descending order)

ries critical importance during data analysis. The extracted features 3.4 (−) 0 1 0 0.1667
might be either redundant or irrelevant. From the classification 1.6 (+) 1 1 0.25 0.1667
0.8 (−) 1 2 0.25 0.333
point of view, high dimensional datasets may easily over-fit or
0.5 (−) 1 3 0.25 0.5
under-fit a classification model. Therefore, feature selection was 0.1 (+) 2 3 0.5 0.5
desirable to reduce the feature space, i.e., from Nc to a lower dimen- −0.2 (−) 2 4 0.5 0.6667
sion Nr . −0.3 (+) 3 4 0.75 0.6667
In this study, the feature selection involved a rank-based fea- −0.8 (+) 4 4 1 0.6667
−1.4 (−) 4 5 1 0.8333
ture selection method [41,45]. According to the method, the area −2.1 (−) 4 6 1 1
under curve (AUC) for each feature was computed termed as z-
value. The z-value represents relevance of a feature with the class
labels. A high z-value (equal or near 0.5) corresponded to the abil-
ity of a feature to discriminate within classes. The z-value could In addition, the columns ‘i’ and ‘j’ represent two unique features,
take any value between 0 and 0.5 indicating bad to good classifi- accordingly. The last column shows the corresponding class labels.
cation ability, accordingly. In this study, the features were ranked Tables 4, 5, and 6 lists the intermediate values of different vari-
in descending order according to the z-values such as the most ables during the computation of the AUC for the feature ‘i’ (as listed
noteworthy features were top-listed. To find minimum number in Table 3). The computations follow the pseudo code provided in
of features that would be sufficient to train the classifier model Table 2. As shown in Table 8, the first step is to sort the feature
without over-fitting, an empirical process was adopted. The min- values in a descending order (1 st column) and the corresponding
imum number of features were determined based on iteratively labels are also adjusted (2nd column), accordingly. Further, the val-
observing performance of the classification models for each fea- ues of intermediate variables (i.e., p, n, tp, and fp) are computed and
ture subsets selected from top-listed features such as 1st, 2nd, 3rd, listed in the respective column.
4th, 5th until complete features in the EEG data matrix. The fea- Table 5 provides the computation of the intermediate variable
ture set that showed best accuracy among others was selected and Y according to the formula Y = (tp(2:n) + tp(1:n-1))/2. The Table 4
considered as selected features. computed Y based in the values provided in Table 8.
To generate a sufficient statistical distribution of classifier per- Similarly, Table 6 provides the computation of the intermediate
formance metrics such as the accuracy, sensitivity and specificity variable X according to the formula ‘X = fp(2:n) − fp(1:n-1)’. The
for each subgroup, 100 times simulations were performed and box- Table 5 shows the computed X based on the values obtained in
plots were plotted. Table 3.
Table 2 provides pseudo code for the rank-based feature selec- Finally, the AUC is computed based on the formula
tion method. Let x be a vector that represents a feature and the ‘auc = sum(Y.*X)-0.5 . Tables 7 and 8 shows the detailed values of
vector y represents the target labels (-1, +1). In this study, both x intermediate variables Y and X and the AUC, respectively.
and y have same dimensions. As shown in Table 7, the value obtained for AUC (z-value) is
Table 3 enlisted the sample EEG data to further explain the zero which means that the feature ‘i’ would not be a good option
pseudo code for the rank-based feature selection method. The table for further classification process and could be rejected. The process
listed the sample data for 10 examples as shown in the first column. is repeated for all features in the EEG data matrix.

Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
6 W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx

Table 5
Computation of Y = (tp(2:n) + tp(1:n-1))/2.

tp(2:n) 0.25 0.25 0.25 0.5 0.5 0.75 1 1 1

tp(1:n-1) 0 0.25 0.25 0.25 0.5 0.5 0.75 1 1

(tp(2:n) + tp(1:n-1))/2 0.25 0.5 0.5 0.75 1 1.25 1.75 2 2
Y 0.125 0.25 0.25 0.375 0.5 0.625 0.875 1 1

Table 6
Computation of X = (fp(2:n)-fp(1:n-1)).

fp(2:n) 0.1667 0.333 0.5 0.5 0.6667 0.6667 0.6667 0.8333 1

fp(1:n-1) 0.1667 0.1667 0.333 0.5 0.5 0.6667 0.6667 0.667 0.8333
X 0 0.1667 0.1667 0 0.1667 0 0 0.1667 0.1667

Table 7 declared as AUD patients, and otherwise associated with control

Computation of AUC = sum(Y.*X)-0.5.
group.
AUC = sum(Y.*X)-0.5;
AUC = (0.125 × 0 + 0.25 × 0.1667 + 0.25 × 0.1667 + 0.375 × 0 + 0.5 × 0.1667 1
+ 0.625 × 0 + 0.825 × 0 + 0.1667 × 1 + 0.1667 × 1)-0.5
F(z) = E(Y/x) = (6)
1 + e−z
AUC = 0.5–0.5 = 0
where Y indicates the class labels and assigned a value either ‘MDD’
or ‘Controls’. In addition, x represents a combination of different
Table 8
EEG features, i.e., the wavelet coefficients achieved after applying
Performance of extracted EEG (Infinity Ref. Data) features based on Logistic Regres- WT technique, power computations and inter-hemispheric asym-
sion (LR) classification while discriminating AUD patients and healthy controls. metry. To obtain the LR model from the logistic function, we used
Sr. Classifier Parameters Value
Eq. (7):

1 Logistic Regression Link Function Logit

z = ˛ + ˇ1 X1 + ˇ2 X2 + ... + ˇk Xk (7)
Classification Distribution Binomial
Offset 1
Constant term A constant term is where z is a linear combination of ␣ plus ˇ1 multiplied with X1 , plus
added in the model ␤2 multiplied with X2, and plus ˇk multiplied with Xk , where the
2 Support Vector C for class 1 (N/2xN1) 0.787 Xk are the independent variables and ␣, and ˇi are constant terms
Machine (SVM) C for class 2 (N/2xN2) 1.3684 representing unknown parameters. Furthermore, by replacing the
Degree of polynomial 1 value of z from Eq. (7) to Eq. (6), the following Eq. (8) represents
No. of classes 2
the logistic function:
Kernal function Linear

3 Naïve Baysian Distribution Normal 1

Classification Prior Uniform distribution F(z) = E(Y/x) =  (8)
−(˛+ ˇi Xi )
for all classes 1+e

In terms of response and non-response, the risk of a person

2.4.4. Classification
Classification has been an important procedure during auto-
matic identification of patterns specific to a group or disease
condition. In this study, 3 different classifiers were selected named
as logistic regression (LR) classifier, support vector machine (SVM)
and Naïve Bayesian (NB) classifier. The selection of LR and NB clas-
sifier was since they could easily be trained with small sample
data. In addition, the problem at hand was to discriminate between
two conditions: AUD Vs Controls which is a dichotomous study
and suited for classifier such as LR. On the other hand, SVM has
been considered as a high-performance classifier. Since it has been
considered as a bench mark, its sole purpose was to provide com-
parison with LR and NB. SVM is also good for low sample size
data because it involves support vectors. Once the support vectors
could be identified, SVM can perform better for small data sets as
well.
In this study, the LR classifier was used to model the relation-
ship between the reduced set of features and the corresponding
treatment outcomes (AUD patients and controls) y = [AUD, Controls],
according to Eq. (6) [46]. For LR classifier, the coefficients estima-
tion was based on maximum likelihood method. The LR classifier
resulted into a likelihood value l(x), where 0 < l(x) ≤ 1, which was an
indication of subjects, associated either with AUD patients or con-
trols. If l(x) was greater than the threshold = 0.5, the subject was
to be non-responder or a responder is estimated and represented
by Y or l(x). The LR classifier resulted into a likelihood value l(x),
where 0 ≤ l(x) ≤ 1, which was an indication of subjects, associated
either with R or NR category. If l(x) was greater than the thresh-
old = 0.5, the subject was declared as R (responder), and otherwise
as a NR (non-responder). In summary, the LR classifier generated
probability values to cater MDD patients as either R or NR to the
treatment.
The second classification model employed was SVM classifier
with linear kernel [47]. It can classify the feature space based on
a ‘hyperplane’ that separated AUD patients and controls accord-
ing to the class labels. The SVM is considered as a high efficiency
classifier model and used here for a comparison purposes. Accord-
ing to SVM, a linear decision boundary can be found based on this
high dimensional space. Use of linear kernel instead of a nonlin-
ear kernel reduced the risk of over-fitting the data and improves
the performance for our data and significantly reduces the overall
model complexity. In summary, the LR classifier generated prob-
ability values to categorize AUD patients as either AUD patients
or controls and the SVM concluded a hyperplane to achieve the
maximum classification accuracy.
The third classification model was the NB classification [48],
based on generating the conditional posterior probabilities for each
sample involving the target groups, i.e., AUD patients and healthy
controls. The classifier was formed by assigning the sample to the
class for which the sample has higher posterior probabilities.

Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx 7

2.4.5. Validation of classification models

After classifier design, a fair evaluation requires assessment of
its performance over a range of selected features and classifier
design (with suitable coefficient values until convergence) that cor-
responds to large number of subjects. To address this consideration,
we evaluated classification performance based on 10-fold cross val-
idation by dividing the data sample points (Study participants) into
10 equal segments. During each round, 9 of the segments were uti-
lized as training subset and the remaining 1 as test subset. Ten-fold
cross validation provides a fair test of validation in cases where the
data points are limited while utilizing features for both test and
train classifier models.
For each feature subset, 10 times run of the simulations were
performed involving 10-fold cross validation to achieve the accu-
Fig. 3. the ROC plots for SVM, LR and NB.
racies, sensitivities and specificities. Since the individual iteration
resulted into 10 different values of performance metrics (the accu-
racy, the sensitivity and the specificity), the final confusion matrix were supposed to originate from two classes, i.e., AUD patients
was computed by averaging over 10 times. The performance met- and healthy controls. The offset value was set to 1, whereas the
rics computed from the confusion matrix were presented by Eq. mathematical model of LR classifier has included a constant term.
(9)–(11). The sensitivity of a classification model corresponds to Regarding the SVM classifier, the C values were assigned as 0.787
the percentage of true cases (TP) which are correctly classified as for AUD patient’s class and 1.3684 for the healthy control’s class,
cases defined by Eq. (9). The specificity of a classification model accordingly. The parameter value ‘C’ indicates an optimum value
refers to the percentage of true non-cases (TN) which are correctly adjusted because of automatic parameter estimation during train-
classified as non-cases as described by Eq. (10). The accuracy of a ing process. The values were computed with formula (N/2 × N1 )
classification model illustrates the percentage of correctly classi- and (N/2 × N2 ), respectively. The ‘N’ denoted total number of study
fied cases and non-case among all the example points as depicted participants; N1 indicated the number of AUD patients and N2 indi-
in Eq. (11). In the scenario of a confusion matrix, both the FP and cated the number of healthy controls. Other parameters such as
FN are considered as type-I and type-II errors. More particularly, ‘Degree of polynomial’, ‘No. of classes’, and ‘Kernal function’ were
the FP stands for false positive, e.g., an AUD patient that is wrongly assigned as 1, 2, and ‘Linear’, respectively. The parameters for Naïve
identified as a healthy control. On the contrary, a FN implies as a Bayesian were assigned with normal uniform distribution for AUD
healthy person that is wrongly identified as depressed patient. patient’s class and the healthy control class. The parameter value ‘C’
TP indicates an optimum value adjusted because of automatic param-
Sensitivity = (9)
TP + FN eter estimation during training process.
TN Fig. 3 has shown the ROC plots for SVM, LR and NB classifiers
Specificity = (10) implied in this study. According to the plots, the classification effi-
TN + FP
ciencies are comparable. However, the SVM classifier performed
TP + TN better than others.
Accuracy = (11)
TP + TN + FP + FN Table 9 describes the most significant features found during
To compare the three classification models, F-score as defined the feature selection stage. According to the selected features, the
in Eq. (12), was applied as a measure of choice for interested class. abnormal brain areas such as the frontal, temporal, and occipital
F-score (Rijsbergen, 1979) could be interpreted as a weighted har- regions have shown the features which are most significant.
monic average of precision and recall value where precision was Since the features were tested one-by-one as mentioned earlier
probability that a (randomly selected) patient analyzed to be AUD and the best accuracy achieved has been reported only.
was really AUDs and recall was the probability that an (randomly Table 10 shows results of classification performance based on
selected) AUD patient was correctly identified as AUD. F-score was the LR classification model. According to the table, the highest per-
calculated using harmonic averaging because it preferred the bal- formance (accuracy = 91.7%, sensitivity = 86.66%, specificity = 96.6%,
ance between precision and recall so it would determine better for and f-measure = 0.90) was achieved with synchronization likeli-
the optimal pair in highly unbalanced datasets. Since we didn’t have
any prior information to either precision or recall, the beta value
Table 9
was set to 1 and the F-score was also named F1 score:
Most significant functional connectivity between scalp locations.
precision × recall
F = (1 + ˇ2 ) × (12) Functional connectivity Frequency band Absolute z-values p-value
ˇ × precision + recall between Electrode pairs
The comparison between the classification models is shown Fp1-Fp2 Delta 0.5000 0.022
while constructing their Receiver operating characteristics (ROC) C3-F4 Delta 0.4990 0.008
plots [49]. The plot has shown the performance comparison in F7-T3 Delta 0.4980 0.010
F3-T7 Delta 0.4970 0.016
graphical form involving their sensitivity and 1-specificity.
F7-F8 Theta 0.4960 0.022
T4-T5 Theta 0.4949 0.022
3. Results F8-T3 Theta 0.4939 0.008
F3-T4 Delta 0.4929 0.010
Table 8 shows specific values of the parameters assigned for F4-T7 Theta 0.4919 0.002
F8-Fp1 Delta 0.4909 0.045
each classifier during training and testing. Regarding the LR clas- T3-F4 Theta 0.4899 0.0021
sifier, the link function showed the relationship between the EEG C4-T3 Delta 0.4879 0.016
features and clinical outcomes. The value was set as ‘logit’ since the F8-T3 Theta 0.4869 0.022
classifier used was logistic regression. Binary classification assumes T4-T3 Delta 0.4859 0.013
P3-Fp1 Theta 0.4848 0.022
binomial distribution; this was set as binomial because the data

Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
8 W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx
Table 10
LR classification while discriminating AUD patients and healthy controls.
EEG Features Accuracy
Synchronization Likelihood 91.7%
Interhemispheric Coherence 77.85%
Mutual information (MI) 79.1%
Integration of SL, Coherence and MI 92%
hood. The second highest accuracy was achieved with mutual
information measure. The integration of these features provided
the performance such as (accuracy = 92%, sensitivity = 88.3%, speci-
ficity = 94.5%, and f-measure = 0.90).
Table 11 shows results of classification performance based on
the NB classification model. According to the table, the highest
performance (accuracy = 93.6%, sensitivity = 100%, specificity = 87.9%,
and f-measure = 0.95) was achieved with synchronization likeli-
hood. The second highest accuracy was achieved with mutual
information measure. The integration of these features provided
the performance such as (accuracy = 92.6%, sensitivity = 100%, speci-
ficity = 84.5%, and f-measure = 0.93).
Table 12 shows results of classification performance based on
the SVM classification model. According to the table, the high-
est performance (accuracy = 98%, sensitivity = 99.9%, specificity = 95%,
and f-measure = 0.97) was achieved with synchronization likeli-
hood. The second highest accuracy was achieved with mutual
information measure. The integration of these features provided
the performance such as (accuracy = 94.7%, sensitivity = 98.3%, speci-
ficity = 91.4%, and f-measure = 0.95).
Fig. 4 shows the diagnosis accuracy as a function of number of
features (while showing for 100 features). The figure reveals the fact
that by increasing the number of features the classification accu-
racy increases and at a certain point show constant behavior. The
figure depicts a constant accuracy as the number of features keeps
increasing because the addition of any new features could not add
useful information to the classifier model.
In this study, all features were tested one-by-one. It has been
observed that the SVM classifier accuracy remained constant after
certain number of features. In the results section, the classifica-
tion tables have reported the best accuracies only. According to
this scheme, the features that show maximum classification accu-
racy would be considered as a reduced set of features sufficient for
classification.
In this study, a highest performance accuracies (98%) has been
reported which is not shown previously, e.g., a recent study has
reported 97% accuracy of classifying the alcoholism subjects from
the healthy controls [50].
Fig. 4. SVM accuracy as a function of number of features.
Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
Sensitivity Specificity F-measure
86.6% 96.6% 0.90
83.3% 71.6% 0.77
83.3% 82.0% 0.78
88.3% 94.5% 0.90
To remove the likelihood that the resulting classifier models are
concluded due to noise present in the EEG data, we have adopted
the following precautions. First, during preprocessing, the artifacts
are carefully removed and tested by plotting their histograms plots
to check the presence of any kind of outliers and found the data
to be suitable for classification purposes. Second, we have selected
equal sample sizes in both the groups. In addition to this the gender
distribution is equal among the groups as well. This is eliminating
the gender biasness from the conclusive results. Third, the logis-
tic regression and Naive Bayesian classifier models are relatively
simpler than the support vector machine classifier. Therefore, the
SVM takes more number of features to produce the classification
results. The incorporation of classifier with 3 different structures
has proved the validity of our data also. Fourth, the over-fiting may
happen; therefore, we have incorporated 100-time permutation
test with 10-fold cross validation to improve the robustness of the
underlying models.
4. Discussion
In this paper, a ML method is proposed that utilizes EEG-
based features as input data to discriminate the AUD patients from
healthy controls. In this paper, the primary finding is that the EEG
features such as the EEG powers and interhemispheric coherences
computed from theta, delta, and high gamma bands can be used as
physiological markers for the screening of AUD patients. In addi-
tion, these features are used as input data for the proposed ML
models to classify the AUD patients and healthy controls. In con-
trary, the conventional methods for screening require subjective
feedbacks from the AUD patients that may confound the screening
process due to the human errors because quantification of the AUD
intake is a tedious task.
The SL has been considered as a marker for diagnosing metal
illness such as unipolar depression and schizophrenia. For exam-
ple, the MDD patients have demonstrated abnormal functional
connectivity among different scalp locations in the frontal and
temporal regions when compared with healthy controls [51]. The
literature has shown that the EEG-based assessment of functional
connectivity between different brain regions could be promising
marker for depression [52–55]. For example, the aberrant neuronal
interactions between different brain regions have been associated
with MDD patients when compared with healthy controls. In addi-
tion, the synchronization likelihood (SL) was used to diagnose
Alzheimer’s disease [56].
In Table 1, the theta and high gamma bands show a significant
difference between the groups which is in accordance with litera-
tures [25,30,57]. These findings implicate that the theta, beta and
high gamma bands are most significant while analyzing the AUD
patients and healthy controls. In Table 1, spectral power shows
their ability in analyzing and discriminating AUD patients from
healthy controls. Spectral power shows overall differences between
AUD patients and controls but show insignificance at every elec-
trode location. Using spectral power would help to explain the
alteration of brain activities of AUD patients. In comparison, AP
and RP show more discrimination between the two groups. Espe-
cially, theta power has proved its potential by outperforming other
frequency bands with a remarkable power increase exhibited in
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx 9

Table 11
NB classification while discriminating AUD patients and healthy controls.

EEG Features Accuracy Sensitivity Specificity F-measure

Synchronization Likelihood 93.6% 100% 87.9% 0.95

Interhemispheric Coherence 89% 86.6% 91.6% 0.88
Mutual Information (MI) 70.8% 60% 80.8% 0.63
Integration of SL, Coherence and MI 92.6% 100% 85.4% 0.93

Table 12
SVM classification while discriminating AUD patients and healthy controls.

EEG Features Accuracy Sensitivity Specificity F-measure

Synchronization Likelihood 98% 99.9% 95.0% 0.97

Interhemispheric Coherence 87.3% 81.6% 91.8% 0.87
Mutual Information (MI) 69.5% 67.8% 74.3% 0.62
Integration of SL, Coherence and MI 94.7% 98.3% 91.4% 0.95

Table 13
Related studies about EEG application and their limitations.

Objective EEG Brain Dynamics Algorithm Authors Results

Screening alcoholics Spectral power and coherence Locally Weight Regression [62] 66.45%
ERP’s components ANN [63] 71%
ERP’s components Learning Vector Quantization [64] 80%
Gamma Visual Evoked Potential (VEP) power Least square Support Vector Machine (SVM) [65] 82.98%
Raw EEG in F4 and P8 Hidden Markov Model [66] 90.50%
Gamma VEP MLP − BP with Elliptic filter [67] 91%
Approximate Entropy (ApEn), Sample Entropy SVM [31] 91.70%
(SampEn), Largest Lyapunov Exponent (LLE),
(high order spectra) HOS
HOS Fuzzy Sugeno Classifier [68] 92.40%
ERP’s components Random Forest [69] 94.50%
Multi gamma band VEP MLP [70] 94.55%
Yule Walker coefficient Artificial NN [71] 95.00%
Wavelet Relative Power K-nearest Neighbor [72] 95.80%
Horizontal Visibility Graph Entropy K-nearest Neighbor Zhu et al. 95.80%
Gamma VEP PCA [73] 95.83%
Gamma VEP MLP [74] 96.10%
Gamma VEP LDA [75] 97.40%
Tunable Q Wavelet Transform RBF kernal [50] 97.02%
Gamma VEP KNN [76] 98.71%
Spectral power using Haar wavelet Multilayer Perceptron Network (MLP) [77] 98.83%
Spectral Entropy Probabilistic Neural Network [78] 99.00%
VEP energy in occipital KNN OR Support Vector Data Description [79] 99.20%
Mean and variance of signals Bayes with KNN and PCA (claim to classify AA) [80] 100%

AUD patients. Theta power changes, observed in alcoholics, were
also reported in literature as a consistent indication for alcoholics
screening. In addition, power of gamma band shows significant dif-
ferences between AUD patients and controls. However, its potential
has not received adequate attention. On the other hand, the inter-
hemispheric coherence exhibits associations between different
brain regions and their variation under the effect of alcohol in AUD
patients.
The classification results presented in Table 13 reveal the sig-
nificance of the theta, and high gamma bands and their integration
have revealed classification accuracy nearly ∼87%. This implicates
the robustness of the proposed method. Since the results are based
on the logistic regression classifier which is considered as a sim-
ple classifier when compared with the SVM. Hence, the proposed
ML classifier model is simple in complexity. EEG may be utilized to
screen AUD patients, to predict relapse and to evaluate medication
effects. As summarized in Table 3, despite the rapid development in
physiological studies of alcoholic brains, few reports discuss appli-
cation of EEG for early relapse detection and medication evaluation,
because low accuracy made it impossible for EEG clinical applica-
bility for AUD patients. Besides that, in review article about clinical
and neuropsychiatric application for alcohol addiction treatment
[58,59], EEG was not mentioned in primary health care. In addition,
few studies have applied EEG for AUD treatment. For relapse pre-
diction, there are few studies using spectral power and nonlinear
features, e.g., extracting Hjorth features from REEG. Unfortunately,
their accuracy is not efficient enough for clinical practice because
of either low sensitivity [24] or low specificity [26].
Regarding AUD screening, various studies have utilized differ-
ent electrophysiological features and classification algorithms with
high accuracy (>90%) for the classification between alcoholics and
controls. These results have confirmed the difference between AUD
and controls, and provide evidence that EEG may be a potential
screening tool for AUD. However, there is no discussion or analysis
about the features and algorithms used in most of those studies.
In this study, integration of EEG features such as theta power, high
gamma power and inter-hemispheric coherence are proposed as
markers that can classify the AUD patients with an accuracy of
∼89%.
Studies based on EEG observation in patients with alcoholism
had resulted into seizures like patterns like the ones happened
during epilepsy. Ping et al. [60] has implemented an automatic
method to differentiate epileptic, controls and alcoholics using EEG
with accuracy of 98.6%. However, the validity of data was doubt-
ful because the datasets were acquired from different sources and
experiment designs such as visual oddball stimulus vs. eye closed.
The data were recorded with different equipment having 64 and
128 channels, and references systems (Cz vs. common average ref-

Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
10 W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx
erence) without any indication about the synchronization between
two datasets. Moreover, the dataset of alcoholism in the study con-
tained not only alcoholics but also control subjects [61]. There were
no standard procedures to categorize study populations during
clinical evaluation. In addition, factors such as sociodemographic
characteristics and family drinking history need to be considered
and controlled.
There is a possibility that our proposed ML models are con-
founded with some outlier’s other than the relevant patterns
extracted from the brain activities. We have ruled out this concern
by 1) properly adopting artifact removal techniques, 2) standard-
izing preprocessed data based on z-scores, 3) during classifier’s
testing and training, selecting random data points so that each data
point in the feature space can be used. Based on all these precau-
tions, we may conclude that the results shown here are un-biased
and true representation of the information from the recorded pre-
treatment EEG data. In addition, the validation of the proposed
method was not performed with adequate number of subjects. The
proper validation could be performed based on more study samples
as a future enhancement of the proposed method.
5. Conclusion
In this study, a ML method was proposed to classify the AUD
patients from healthy controls base on resting-state EEG data. The
SL features have shown significant differences between the two
groups that implicates into improved classification results. The
classification results implicate that the SL features could be utilized
as characteristic features to automatically learn the disease-specific
patterns in the resting-state EEG data acquired from the study
groups. Furthermore, the proposed ML method implicate that EEG-
based CAD tool can be developed and help in making the AUD
screening an automated and a standard procedure.
Acknowledgment
This research work is supported by the HICoE grant for CISIR
(0153CA-005), Ministry of Education (MOE), Malaysia.
References
[1] Facts About Alcohol. National Council on Alcoholism and Drug Dependence
website; 2016 (2016, July) Available: https://www.ncadd.org/about-
addiction/alcohol/facts-about-alcohol.
[2] Alcohol Facts and Statistics; 2017 (2015, July 29). Available: https://www.
niaaa.nih.gov/alcohol-health/overview-alcohol-consumption/alcohol-facts-
and-statistics.
[3] Parsons OA, Nixon SJ. Cognitive functioning in sober social drinkers: a review
of the research since 1986. J Stud Alcohol 1998;59:180–90.
[4] Moss HB, Chen CM, Yi H-y. Subtypes of alcohol dependence in a nationally
representative sample. Drug Alcohol Depend 2007;91:149–58.
[5] Maisto SA, Saitz R. Alcohol use disorders: screening and diagnosis. Am J
Addict 2003;12:s12–25.
[6] Popham RE, Schmidt W. Words and deeds: the validity of self-report data on
alcohol consumption. J Stud Alcohol 1981;42:355–8.
[7] Solomon J, Vanga N, Morgan J, Joseph P. Emergency-room physicians’:
recognition of alcohol misuse. J Stud Alcohol 1980;41:583–6.
[8] Watson CG, Tilleskjor C, Hoodecheck-Schow E, Pucel J, Jacobs L. Do alcoholics
give valid self-reports? J Stud Alcohol 1984;45:344–8.
[9] Son K, Choi J, Lee J, Park S, Lim J, Lee J, et al. Neurophysiological features of
Internet gaming disorder and alcohol use disorder: a resting-state EEG study.
Transl Psychiatry 2015;5:e628.
[10] Alhassoon OM, Sorg SF, Stern MJ, Hall MG, Wollman SC. Neuroimaging in
alcohol-use disorders: clinical implications and future directions. Future
Neurol 2015;10:345–56.
[11] Shalbaf R, Behnam H, Moghadam HJ. Monitoring depth of anesthesia using
combination of EEG measure and hemodynamic variables. Cognit
Neurodynam 2015;9:41–51.
[12] Ozaki TJ, Sato N, Kitajo K, Someya Y, Anami K, Mizuhara H, et al. Traveling EEG
slow oscillation along the dorsal attention network initiates spontaneous
perceptual switching. Cognit Neurodynam 2012;6:185–98.
Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
[13] Gutiérrez D, Ramírez-Moreno MA. Assessing a learning process with
functional ANOVA estimators of EEG power spectral densities. Cognit
Neurodynam 2016;10:175–83.
[14] Kiebel SJ, Garrido MI, Moran RJ, Friston KJ. Dynamic causal modelling for EEG
and MEG. Cognit Neurodynam 2008;2:121–36.
[15] Fründ I, Schadow J, Busch NA, Naue N, Körner U, Herrmann CS. Anticipation of
natural stimuli modulates EEG dynamics: physiology and simulation. Cognit
Neurodynam 2008;2:89–100.
[16] Herrera-Díaz A, Mendoza-Quiñones R, Melie-Garcia L, Martínez-Montes E,
Sanabria-Diaz G, Romero-Quintana Y, et al. Functional connectivity and
quantitative EEG in women with alcohol use disorders: a resting-State study.
Brain Topogr 2015:1–14.
[17] de Bruin EA, Bijl S, Stam CJ, Böcker KB, Kenemans JL, Verbaten MN. Abnormal
EEG synchronisation in heavily drinking students. Clin Neurophysiol
2004;115:2048–55.
[18] de Bruin EA, Stam CJ, Bijl S, Verbaten MN, Kenemans JL. Moderate-to-heavy
alcohol intake is associated with differences in synchronization of brain
activity during rest and mental rehearsal. Int J Psychophysiol
2006;60:304–14.
[19] Tcheslavski GV, Gonen FF. Alcoholism-related alterations in spectrum,
coherence, and phase synchrony of topical electroencephalogram. Comput
Biol Med 2012;42:394–401.
[20] Michael A, Mirza K, Mukundan C, Channabasavanna S. Interhemispheric
electroencephalographic coherence as a biological marker in alcoholism. Acta
Psychiatr Scand 1993;87:213–7.
[21] Campanella S, Petit G, Maurage P, Kornreich C, Verbanck P, Noël X. Chronic
alcoholism: insights from neurophysiology. Neurophy Clin/Clin Neurophys
2009;39:191–207.
[22] Parvaz MA, Alia-Klein N, Woicik PA, Volkow ND, Goldstein RZ. Neuroimaging
for drug addiction and related behaviors. Rev Neurosci 2011;22:609–24.
[23] Porjesz B, Rangaswamy M, Kamarajan C, Jones KA, Padmanabhapillai A,
Begleiter H. The utility of neurophysiological markers in the study of
alcoholism. Clin Neurophysiol 2005;116:993–1018.
[24] Bauer LO. Predicting relapse to alcohol and drug abuse via quantitative
electroencephalography. Neuropsychopharmacology 2001;25:332–40.
[25] Rangaswamy M, Porjesz B, Chorlian DB, Choi K, Jones KA, Wang K, et al. Theta
power in the EEG of alcoholics. Alcoho.: Clin. Exp. Res 2003;27:607–15.
[26] Winterer G, Klöppel B, Heinz A, Ziller M, Dufeu P, Schmidt LG, et al.
Quantitative EEG (QEEG) predicts relapse in patients with chronic alcoholism
and points to a frontally pronounced cerebral disturbance. Psychiatry Res
1998;78:101–13.
[27] Klimesch W. EEG alpha and theta oscillations reflect cognitive and memory
performance: a review and analysis. Brain Res Rev 1999;29:169–95.
[28] Ehlers CL, Phillips E. Association of EEG alpha variants and alpha power with
alcohol dependence in Mexican American young adults. Alcohol
2007;41:13–20.
[29] Ehlers CL, Phillips E, Schuckit MA. EEG alpha variants and alpha power in
Hispanic American and white non-Hispanic American young adults with a
family history of alcohol dependence. Alcohol 2004;33:99–106.
[30] Rangaswamy M, Porjesz B, Chorlian DB, Wang K, Jones KA, Bauer LO, et al.
Beta power in the EEG of alcoholics. Biol Psychiatry 2002;52:831–42.
[31] Acharya UR, Sree SV, Chattopadhyay S, Suri JS. Automated diagnosis of normal
and alcoholic EEG signals. Int J Neural Syst 2012;22:1250011.
[32] Bajaj V, Guo Y, Sengur A, Siuly S, Alcin OF. A hybrid method based on
time–frequency images for classification of alcohol and control EEG signals.
Neural Comput Appl 2016:1–7.
[33] Sinha R. Automated identification of chronic alcoholism from brain signals. J
Health Allied Sci 2016;14. Online.
[34] Mumtaz W, Vuong PL, Xia L, Malik AS, Rashid RBA. Automatic diagnosis of
alcohol use disorder using EEG features. Knowledge-Based Syst
2016;105:48–59.
[35] Huys QJ, Maia TV, Frank MJ. Computational psychiatry as a bridge from
neuroscience to clinical applications. Nat Neurosci 2016;19:404–13.
[36] Babor TF, Higgins-Biddle JC, Saunders JB, Monteiro MG. W.H. Organization
AUDIT: the alcohol use disorders identification test: guidelines for use in
primary health care. World Health Organization; 2001.
[37] Bush K, Kivlahan DR, McDonell MB, Fihn SD, Bradley KA. The AUDIT alcohol
consumption questions (AUDIT-C): an effective brief screening test for
problem drinking. Arch Intern Med 1998;158:1789–95.
[38] Klem GH, Lüders HO, Jasper H, Elger C. The ten-twenty electrode system of
the International Federation. Electroencephalogr Clin Neurophysiol 1999;52.
[39] Berg P, Scherg M. A multiple source approach to the correction of eye
artifacts. Electroencephalogr Clin Neurophysiol 1994;90:229–41.
[40] Hoechstetter K, Berg P, Scherg M. BESA research tutorial 4: distributed source
imaging. BESA Res Tutorial 2010:1–29.
[41] Mamitsuka H. Selecting features in microarray classification using ROC
curves. Pattern Recogn 2006;39:2393–404.
[42] Stam C, Van Dijk B. Synchronization likelihood: an unbiased measure of
generalized synchronization in multivariate data sets. Phys D: Nonlinear
Phenom 2002;163:236–51.
[43] Takens F. Detecting strange attractors in turbulence. In: Dynamical systems
and turbulence, Warwick 1980. ed. Springer; 1981. p. 366–81.
[44] Theiler J. Spurious dimension from correlation algorithms applied to limited
time-series data. Phys Rev A 1986;34:2427.
[45] Liu H, Motoda H. Computational methods of feature selection. CRC Press;
2007.
G Model
ARTMED-1561; No. of Pages 11 ARTICLE IN PRESS
W. Mumtaz et al. / Artificial Intelligence in Medicine xxx (2017) xxx–xxx
[46] Hosmer Jr DW, Lemeshow S. Applied logistic regression. John Wiley & Sons;
2004.
[47] Vapnik VN, Vapnik V. Statistical learning theory, vol. 2. New York: Wiley;
1998.
[48] Mitchell TM. Machine learning. Boston, MA: McGraw-Hill; 1997. WCB, ed.
[49] Lehmann C, Koenig T, Jelic V, Prichep L, John RE, Wahlund L-O, et al.
Application and comparison of classification algorithms for recognition of
Alzheimer’s disease in electrical brain activity (EEG). J Neurosci Methods
2007;161:342–50.
[50] Patidar S, Pachori RB, Upadhyay A, Acharya UR. An integrated alcoholic index
using tunable-Q wavelet transform based features extracted from EEG signals
for diagnosis of alcoholism. Appl Soft Comput 2017;50:71–8.
[51] Olbrich S, Tränkner A, Chittka T, Hegerl U, Schönknecht P. Functional
connectivity in major depression: increased phase synchronization between
frontal cortical EEG-source estimates. Psychiatry Research: Neuroimaging
2014;222:91–9.
[52] Anand A, Li Y, Wang Y, Wu J, Gao S, Bukhari L, et al. Activity and connectivity
of brain mood regulating circuit in depression: a functional magnetic
resonance study. Biol Psychiatry 2005;57:1079–88.
[53] Bae JN, MacFall JR, Krishnan KRR, Payne ME, Steffens DC, Taylor WD.
Dorsolateral prefrontal cortex and anterior cingulate cortex white matter
alterations in late-life depression. Biol Psychiatry 2006;60:1356–63.
[54] Fingelkurts AA, Fingelkurts AA, Rytsälä H, Suominen K, Isometsä E, Kähkönen
S. Impaired functional connectivity at EEG alpha and theta frequency bands in
major depression. Hum Brain Mapp 2007;28:247–61.
[55] Hamilton JP, Gotlib IH. Neural substrates of increased memory sensitivity for
negative stimuli in major depression. Biol Psychiatry 2008;63:1155–62.
[56] Stam C, Montez T, Jones B, Rombouts S, Van Der Made Y, Pijnenburg Y, et al.
Disturbed fluctuations of resting state EEG synchronization in Alzheimer’s
disease. Clin Neurophysiol 2005;116:708–15.
[57] Porjesz B, Begleiter H. Alcoholism and human electrophysiology. Alcohol Res
Health 2003;27:153–60.
[58] Ritsner MS. The Handbook of Neuropsychiatric Biomarkers, Endophenotypes
and Genes: Volume I: Neuropsychological Endophenotypes and Biomarkers,
vol 1. 2009: Springer Science & Business Media; 2017.
[59] Tavakoli HR, Hull M, Michael Okasinski L. Review of current clinical
biomarkers for the detection of alcohol dependence. Innov Clin Neurosci
2011;8:26–33.
[60] Ng EP, Lim T-C, Chattopadhyay S, Bairy M. Automated identification of
epileptic and alcoholic EEG signals using recurrence quantification analysis. J
Mech Med Biol 2012;12:1240028.
[61] Zhang XL, Begleiter H, Porjesz B, Litke A. Electrophysiological evidence of
memory impairment in alcoholic patients. Biol Psychiatry 1997;42:1157–71.
[62] Guntaka R, Tcheslavski GV. On the EEG-based automated detection of alcohol
dependence. Int J Bioautomation 2016;17:167–76.
[63] Lopes CD, Mainardi JO, Zaro MA, Susin AA. Classification of event-related
potentials in individuals at risk for alcoholism using wavelet transform and
artificial neural network, in Computational Intelligence in Bioinformatics and
Computational Biology, 2004. CIBCB’04. Proceedings of the 2004 IEEE
Symposium on 2004 2004:123–8.
Please cite this article in press as: Mumtaz W, et al. An EEG-based functional connectivity measure for automatic detection of alcohol
use disorder. Artif Intell Med (2017), https://doi.org/10.1016/j.artmed.2017.11.002
11
[64] Lopes CD, Schuler E, Engel P, Susin AA. ERP signal identification of individuals
at risk for alcoholism using learning vector quantization network, in
computational intelligence in bioinformatics and computational biology,
2005. CIBCB’05. Proceedings of the 2005 IEEE Symposium on 2005 2005:1–5.
[65] Shooshtari MA, Setarehdan SK. Selection of optimal EEG channels for
classification of signals correlated with alcohol abusers, in Signal Processing
(ICSP). 2010 IEEE 10th International Conference on 2010 2010:1–4.
[66] Zhong S, Ghosh J. HMMs and coupled HMMs for multi-channel EEG
classification. Proceedings of the IEEE International Joint Conference on
Neural Networks 2002:1159–254.
[67] Kanna PS, Palaniappan R, Ravi K. Classification of alcohol abusers: an
intelligent approach, in Information Technology and Applications, 2005. ICITA
2005. Third International Conference on 2005 2005:470–4.
[68] Faust O, Yanti R, Yu W. Automated detection of alcohol related changes in
electroencephalograph signals. J Med Imaging Health Inf 2013;3:333–9.
[69] Kuncheva LI, Rodríguez JJ. Interval feature extraction for classification of
event-related potentials (ERP) in EEG data analysis. Prog Artif Intell
2013;2:65–72.
[70] Palaniappan R. Screening for chronic alcoholic subjects using multiple gamma
band EEG: a pilot study. J Comp Sci Technol 2007;7.
[71] Ek Z, Akg A, Bozkurt MR. The classificaton of EEG signals recorded in drunk
and non-Drunk people. Int J Comp Appl 2013;68.
[72] Faust O, Yu W, Kadri NA. Computer-based identification of normal and
alcoholic eeg signals using wavelet packets and energy measures. J Mech Med
Biol 2013;13:1350033.
[73] Ong K-M, Thung K-H, Wee C-Y, Paramesranle R. Selection of a subset of EEG
channels using PCA to classify alcoholics and non-alcoholics. Proceedings
ofthe 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference
(Shanghai China 2005) 2005.
[74] Palaniappan R, Raveendran P, Omatu S. VEP optimal channel selection using
genetic algorithm for neural network classification of alcoholics. Neural
Networks, IEEE Trans 2002;13:486–91.
[75] Palaniappan R. Discrimination of Alcoholic Subjects using Second Order
Autoregressive Modelling of Brain Signals Evoked during Visual Stimulus
Perception. Prague: IEC; 2005. p. 282–7.
[76] Palaniappan R. Improved automated classification of alcoholics and
non-alcoholics. Int J Inf Technol 2003;2.
[77] Kousarrizi MN, Ghanbari AA, Gharaviri A, Teshnehlab M, Aliyari M.
Classification of alcoholics and non-alcoholics via EEG using SVM and neural
networks, in Bioinformatics and Biomedical Engineering, 2009. ICBBE 2009.
3rd International Conference on 2009 2009:1–4.
[78] Padmanabhapillai A, Porjesz B, Ranganathan M, Jones KA, Chorlian DB, Tang Y,
et al. Suppression of early evoked gamma band response in male alcoholics
during a visual oddball task. Int J Psychophysiol 2006;60:15–26.
[79] Zúquete A, Quintela B, Cunha JPS. Biometric authentication using
electroencephalograms: a practical study using visual evoked potentials.
Electrónica e Telecomunicações 2010;5:185–94.
[80] Yazdani A, Setarehdan SK. Classification of EEG signals correlated with alcohol
abusers, in signal processing and its applications, 2007. ISSPA 2007. 9th
International Symposium on 2001 2007:1–4.