Emotional-state brain network analysis revealed by
minimum spanning tree using EEG signals
Jianhai Zhang
*
College of Computer Science
Hangzhou Dianzi University
Hangzhou, Zhejiang, China
jhzhang@hdu.edu.cn
Jiajia Tang
College of Computer Science
Hangzhou Dianzi University
Hangzhou, Zhejiang, China
hdutangjiajia@163.com
Wanzeng Kong
College of Computer Science
Hangzhou Dianzi University
Hangzhou, Zhejiang, China
Kongwanzeng@hdu.edu.cn
Shaokai Zhao
College of Life Sciences
Nankai University
Naikai, Tianjin, China
lnkzsk@126.com
Tao Zhang
College of Life Sciences
Nankai University
Naikai, Tianjin, China
zhangtao@nankai.edu.cn
Guodong Yang
College of Computer Science
Hangzhou Dianzi University
Hangzhou, Zhejiang, China
yangguodong1995@gmail.com
Yong Peng
College of Computer Science
Hangzhou Dianzi University
Hangzhou, Zhejiang, China
penyong@hdu.edu.cn
Abstract—
Brain networks have been widely constructed in
affective computing. The use of it may cause the bias of
emotional brain network analysis because of the threshold
selection. In the study, we applied minimum spanning tree
(MST) to ‘spanning’ the edge of emotional brain network. The
results show that the MST effectively avoid complicated
threshold selection steps in network analysis and reduce
computational complexity on a large scale.
Keywords—emotional analysis, EEG, Minimum spanning
tree (MST), brain network analysis, phase locking value (PLV)
I. I
NTRODUCTION
At present, it has gained wide recognition by abstracting
the brain into a complex network and then introducing graph
theory to analyze the emotional patterns of the human brain
[1,2]. Several studies [3,4] based on functional connectivity
achieved significant data, which showed that the construction
of functional connectivity could capture the impact of
emotions on information flows in brain. It has shown that the
phase of the EEG signal contains more valid information
than the amplitude of its oscillation [5]. For this reason, we
extracted phase locking value (PLV) of EEG signals for
emotional-state network analysis. One of the important steps
that typically need to be experienced in traditional functional
network construction is threshold selection. It makes the
network connections sparse and ultimately forms a brain
network [6, 7]. The result of threshold selection also
determines the quality of network analysis to some extent [8,
9]. A more common threshold selection scheme is to
enumerate all possible thresholds and calculate
corresponding network metrics for different test groups.
Finally, a threshold that enables significant differences in
network metrics between test groups was adopted [10]. But
the method will undoubtedly lead to an increasing in
computational pressure. In addition, the issue of threshold
selection has also led to some of the opposite findings in
recent research literatures and the inability to compare
different research results [11].
Minimum spanning tree (MST) is a subset of the brain
connectivity that connects all nodes in the network without
forming a loop and has the smallest total weight of all
possible spanning trees [11]. The process of MST generation
ensures that the MST of a brain network with N nodes has
only N-1 edges, thus solving the problem of 1) threshold
selection, so that 2) unbiased comparison can be made
between different research results, and 3) reduce individual
differences, and finally 4) simplify the complexity of the
brain network analysis process. Although MST significantly
reduces the number of edges between nodes compared with
traditional brain networks, it can also represent the brain
network to a certain extent because it carries the main
information of the brain network. Currently, MST-based
analysis has emerged in the field of disease diagnosis: such
as epilepsy [12, 13], Parkinson [14] and so on. These
documents show that MST has enough sensitivity to capture
changes in the brain network. Since emotional stimuli cause
much less change to the brain network than that of diseases,
there is still lack of research to show whether MST is
sufficient to explore the impact of emotional stimuli on the
brain network. Therefore, we analyzed the brain network
under different emotional states based on MST. The main
process will be discussed below.
II. M
ATIRIALS AND
M
ETHODS
A. DEAP database
Th
is study is performed on the publicly available
database DEAP [15]. The EEG signals (32 channels) of 32
subjects were recorded during watching music video. The 40
one-minute one videos were carefully selected to elicit
different emotional states according to the dimensional
valence-arousal emotion model. Each video clip is rated from
1 to 9 for arousal and valence by each subject after viewing,
and the discrete rating value can be used as classification
label in emotion recognition. In the study, the preprocessed
DEAP database was used to test the performance of MST in
emotional-state analysis for four emotions (high valence:
valence ≥ 5; low valence: valence < 5; high arousal ≥ 5; low
arousal: arousal < 5). Each trial’s EEG data is filtered to the