Professional Documents
Culture Documents
Paper
Brain–computer interfaces (BCIs) have been attracting attention as interfaces that connect a brain to an external device. The
implementation of a BCI requires establishing a technique that accurately recognizes the brain state, and a host of challenges
remains for the implementation. This paper uses the emotion fractal analysis method (EFAM), which quantifies emotions based
on data obtained from an electroencephalogram (EEG), to propose a BCI system that accurately discriminates and recognizes
emotions such as delight, anger, sorrow, and pleasure, and uses that information to manipulate an electric wheelchair. The
EFAM’s emotion recognition rates for the four emotions, namely delight, anger, sorrow, and pleasure, were 81.11, 79.25, 73.16,
and 85.42%, respectively, and its emotion isolation rates were 97.88, 98.08, 98.07, and 97.86%, respectively. Based on emotion
data obtained in real time by the EFAM, we developed a novel BCI circuit that manipulates an electric wheelchair. Using this
BCI circuit allows us to adjust the speed of an electric wheelchair in proportion to the intensity of the emotion. © 2018 Institute
of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.
Keywords: electroencephalogram (EEG); fractal dimension; emotion fractal analysis method; brain–computer interface (BCI)
1. Introduction Figure 1 shows a standard BCI system [1]. A BCI system refers
to an interface that measures brain functions, extracts features
The establishment of brain–computer interfaces (BCIs) that from biological information thus measured, and generates some
enable users to move equipment based solely on their biologi- kind of device control signals using these features to manipulate
cal information from electroencephalography without using their an external device. Noninvasive techniques for measuring brain
limbs by connecting a brain to an external device is desired. Using function include functional magnetic resonance imaging (fMRI),
BCIs allows us to support people with amyotrophic lateral sclerosis positron emission tomography, magnetoencephalography (MEG),
(ALS), people with paralysis due to brain lesions, or elderly peo- and electroencephalography; an EEG is used most often owing
ple who face difficulties moving on their own. Our research aims to its lower constraints and cost variations [1–6]. A BCI control
are set out below. A small-size, wearable electroencephalograph is system can be developed by extracting features from the EEG
placed on the head of a person in need of nursing care, such as one signals using some kind of technique and developing a novel
with ALS or a brain lesion, to keep monitoring her/his electroen- algorithm that converts these features. Concerning wheelchair
cephalograms (EEGs). Based on the EEGs, the emotions of the per- control through BCI, systems using motor imagery EEGs [7],
son are analyzed using an emotion analysis method later described SSVEP (steady-state visual evoked potential) [8], and P300s
in this paper. Based on the emotion thus analyzed, a wheelchair, have been proposed previously. To date, however, there are no
home electronics, or electronic equipment is automatically manipu- examples using emotions. The advantages of applying emotions
lated. This emotion analysis allows us to check through a monitor to machine controls are many; for example, where detecting
how the person in need of nursing care feels, and what she/he feelings of fear in the user, it can apply an emergency stop to
wants, even if she/he has difficulty communicating with her/his the wheelchair, and by considering that, these things contribute to
caregiver by reason of ALS or paralysis due to brain lesion. The ergonomic aspects. From a safety standpoint, there are measures
first advantage of this approach is that it helps in reducing the to prevent malfunctions in the relatively large machine controls of
workload of caregivers. In other words, reducing human costs can wheelchairs; however, because of the characteristics of BCI, even
lead to the resolution of the labor shortage in an aging society with in the methods proposed so far it has been difficult to achieve
a declining birth rate. The second advantage is that it is expected to 100% recognition rate. Therefore, it is important to construct safer
improve the quality of life (QoL) of people in need of nursing care. systems by combining multiple methods. In this paper, we show
This study aims to support an aging society by realizing this con- the utility of wheelchair controls through emotion in order to
cept. This paper proposes a novel form of a BCI system that ana- investigate the possibilities of using emotions in BCI development.
lyzes emotions based on human EEGs and manipulates an electric Musha et al. proposed a typical feature extraction technique
wheelchair according to the estimated emotion, in order to confirm called the emotion spectrum analysis method (ESAM) [9,10]. The
that we are indeed on a right path toward realizing this possibility. ESAM is reported to have a recognition rate of around 50% with
regard to the isolation of four types of emotions [9]. This paper
a Correspondence to: Nobuaki Kobayashi. E-mail: proposes the emotion fractal analysis method (EFAM), which
kobayashi.nobuaki@nihon-u.ac.jp quantifies emotions. EFAM can accurately recognize affects (such
* College of Science and Technology, Nihon University, 7-24-1 Narashin- as comfortable and uncomfortable) and emotions (such as delight,
odai Funabashi-shi, Chiba 274-8501, Japan anger, sorrow, and pleasure) by quantifying emotions based on
** Faculty of Engineering, Nagaoka University of Technology, 1603-1 biological information obtained via an EEG. What distinguishes
Kamitomioka Nagaoka-shi, Niigata 940-2188, Japan the EFAM from all other feature extraction methods the most
© 2018 Institute of Electrical Engineers of Japan. Published by John Wiley & Sons, Inc.
N. KOBAYASHI AND M. NAKAGAWA
100 101
50
Voltage (μV)
log σq (τ)
100
2
10–1
–50
Rest Concentration
–100 10–2 –5
0 50 100 150 200 10 10–4 10–3 10–2 10–1 100
Time (s) log τ
(a) (b)
2.0
Fractal dimension
1.7
1.4
1.1
0.8
Rest Concentration
0.5
0 50 100 150 200
Time (s)
(c)
Fig. 2. Example of fractal dimension analysis of brain waves. (a) Brain wave signal. (b) Scaling property of variance (SPV). (c) Time-
dependent fractal dimension analysis
C = R zy (R yy ) − 1 (8)
Fp1 Fp2
F7 F8 d = z ∗ − R zy (R yy ) − 1y (9)
F3 Fz F4
yy
A1 Rj ,n = yj yn − yj yn
T3 C3 Cz C4 T4 A2 (10)
Rmn = zm ∗ yn − zm ∗ yn
yy
pleasure z = [0,0,0,1]T . If C , d , and D(t) obtained from evaluation Table I. Average recognition rate of four emotions (delight,
data are input into (11), a vector of evaluative emotion output is anger, sorrow, and pleasure) (%)
obtained.
‘Emotions’ measured in the present approach denote ‘moods Pleasure Anger Sorrow Delight
or feelings’ based on the experience of each study subject rather Pleasure 81.11 6.6 6.14 3.44
than temporary emotions such as ‘joy’, ‘sadness’, and ‘annoyance’ Anger 8.95 79.25 10.34 4.52
experienced during their daily living. Sorrow 5.36 8.27 73.16 6.62
Delight 4.58 5.88 10.36 85.42
3. Characteristics of EFAM The numbers in bold indicate that the emotions from the recollection
(column) and the assessment (row) match.
3.1. Measuring conditions EEG measurements were
performed using a Polymate (Digitex Lab. Co., Ltd., Japan), a
biosignal recorder, and the EFAM performance was assessed by Table II. Average isolation rate of four emotions (delight, anger,
implementing a real-time EFAM analytical program that estimated sorrow, and pleasure) (%)
emotions in real time using C++. The measurement conditions
were as follows: sampling frequency of 2 kHz, the digital filter was Pleasure Anger Sorrow Delight
a 50 Hz notch filter, with a low-cut filter (LCF), a high-cut filter Pleasure 97.88 0.54 0.55 0.56
(HCF), and an anti-aliasing filter that had a cutoff frequency of Anger 0.79 98.08 0.76 0.76
600 Hz. The electrode sites conformed to the International 10–20 Sorrow 0.78 0.82 98.07 0.83
system, as shown in Figs. 3, and 19 electrodes were used. The Delight 0.55 0.56 0.62 97.85
reference electrode was placed on the right earlobe. The study
subjects were five healthy adult males. Informed consent was The numbers in bold indicate that the emotions from the recollection
obtained from subjects prior to the experiment. The subjects were (column) and the assessment (row) match.
aged 24.40 ± 4.506 years.
obtained at 0.25-s intervals. For the 30-s recall of the target emo-
3.2. Assessing recognition rates and isolation rates tion, for each attempt, a 120-element emotion output vector (four
Figure 4 shows the time table of the measurements. First, the dimensions) (maximum value of 1) was obtained. In each of the
study subjects were asked to recall four emotions (delight, anger, 120-element emotion output vectors, emotions exceeding a certain
sorrow, and pleasure) for 60 s each, and the results were used value (threshold value) were counted as emotions in the evaluation
as learning data. Second, the measured data were recorded. The data, the ratio of which was the recognition rate. On this occasion,
study subjects were asked to recall any one of the four emotions the threshold rate was 0.3. Therefore, in some cases there are
(delight, anger, sorrow, and pleasure) for 30 s, and then to recall instances where two or three emotions exceed the threshold values
the same emotion for 20 s. Subsequently, they were asked to recall and are output. Thus, among the 120-element emotion output
the same emotion for 30 s. These tasks were taken together as a vectors, by counting each case where only one emotion exceeds
set lasting 80 s, and two sets were performed for each of the four the threshold value and each case where two or three emotions
emotions. The analytical program used learning data as a reference exceed the threshold values, their ratio becomes the isolation rate.
and estimated the emotion of a study subject from his actual data. Table I shows the mean recognition rates to the evaluation
In addition, the subjects were not previously trained in recalling data (mean values of two trials) for the four emotions. The
emotions, but they were instructed to recall events from past expe- column of the table shows recollection and the row shows emotion
riences that had left a strong impression. The visual analog scale assessments. The mean recognition rates were all high, with the
(VAS) was implemented during each experiment. The VAS values figures for delight, anger, sorrow, and pleasure being 81.11, 79.25,
for each emotion were as follows: delight, 71.01 ± 4.506; anger, 73.16, and 85.42%, respectively. Furthermore, Table II shows the
82.94 ± 6.870; sorrow, 71.20 ± 8.207; and pleasure, 79.20 ± 6.853. mean isolation rates for the four emotions (mean values of two
The recognition rate was defined as the rate at which the trials). The mean isolation rates were extremely high, with the
program correctly identified the emotion of a study subject, and figures for delight, anger, sorrow, and pleasure being 97.88, 98.08,
the isolation rate was defined as the rate at which a single emotion 98.07, 97.86%, respectively, and all figures were greater than 97%.
was identified by the analytical program. Because the analysis As seen above, using the EFAM allows us to recognize and isolate
window in this analysis was W = 2000 (1 s), and the window tran- the four emotions very accurately.
sition width W s = 500 (0.25 s), the evaluation emotion output was
v
Proposed BCI system Driving sytem (back wheel)
Left Right
Real-time EFAM 0°
Forward
Steering
+6
(km/h)
(V)
θ
5
–60° +60° system
Brain-signal Feature Translation
Front (front wheel)
acquisition extraction algorithm
(degree)
0
Left –60 Stop
+60 Right
θ
turn 0 2.5 5 turn
v (V)
+ –
BCI Back
–6
controller
0
Backward
Device (a) (b)
commands
Fig. 6. Operating principle of the electric wheelchair. (a) The
system of the front and back wheels. (b) The operational voltage
Electric
wheelchair BCI circuit (hardware)
A Control
Micro
Se singnal
controller
generator
4.3. Configuration and movements of BCI circuit 4.4. Configuration and working of control signal
Figure 7 shows the BCI circuit configuration. This control circuit generator Figure 8 shows the configuration of the control-
consists of a microcontroller, which receives as input a serial signal generator (CSG). It is split into two blocks: a driving system
signal (S e ) that contains information about the emotion and its block, and a steering system block, and both share the same circuit
intensity estimated by real-time EFAM software on a PC and architecture. Each of these blocks is composed of two PhotoMOS
determines the movement parameter (A = {Af , Ab , Ar , Al }) of the relays and voltage dividers (two serially connected 1-k resistors).
wheelchair. The control signal generator (CSG) then receives A as The PhotoMOS relays used are ones available on the market.
input, and generates a movement control signal C (Section 4.2) A PhotoMOS relay produces light when a current flows in the
for the electric wheelchair. If ‘delight’ is given, it outputs ‘F’ LED in the input stage; then the light-sensitive elements receive
Se Af Ab Ar Al
Pleasure ‘F’a PWM 0 0 0
Anger ‘La’ PWM 0 0 PWM
Sorrow ‘Ra’ PWM 0 PWM 0
Delight ‘Ba’ 0 PWM 0 0
a Intensity of the emotion (0–100).
LED FET 1 kΩ
b
0V Cd
Ab
1 kΩ
c
0V
0V
Ar
Steering system block Cs
Al
20
0
80
60
40
+Duty cycle (%)
Al
20
0
80
60
40
Ar
20
0
80
60
40
Ab
20
0
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
Time (s)
(b)
2.0
1.0
0.0
5.0
4.0
3.0
Cs (V)
2.0
1.0
0.0
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
Time (s)
(c)
0.0
–3.0
–6.0
+60
+30
0
θ (°)
–30
–60
0 20 40 60 0 20 40 60 0 20 40 60 0 20 40 60
Time (s)
(d)
Fig. 10. Example of controlling electric wheelchair. (a) Strength of emotional states. (b) Operational parameter A. (c) Control voltages C .
(d) Rotational speed v and angle θ ◦
is 0, and the PWM is input into Al , C s varies from 2.5 to 0 V, respectively, for ‘pleasure’. The rotation speed v of the back
that is, the rotation angle of the front wheels changes from 0◦ wheels is determined according to the value of C d , and the
(straight movement either forward or backward) to −60◦ (left-turn) rotation angle θ of the front wheels is determined according to
in inverse proportion to changes in the +duty cycle (0–100%) of the value of C s , resulting in a movement of the wheelchair. As
the PWM. As mentioned above, appropriate input of a PWM signal shown in Fig. 10(d), with ‘delight’, v is set to around +3 km/h
corresponding to Table III into A enables the desired movements and θ to 0◦ , producing a forward movement; with ‘anger’, v is
(forward movement, backward movement, right-turn, left-turn) of set to around +3 km/h and θ to around −45◦ to -–0◦ , producing
the wheelchair. a left-turn movement; with ‘sorrow’, v is set to around +3 to
+4.5 km/h and θ to around +30◦ to +45◦ , producing a right-turn
4.5. Operation instance of proposed BCI system movement; and with ‘pleasure’, v is set to around −3 km/h and
A microcontroller within the BCI control circuit generates the θ to 0◦ , making a backward movement. As mentioned above, this
movement parameter A based on the emotion, and its intensity system realized an electric wheelchair that produces a forward
is estimated by real-time EFAM. Figure 9 shows an instance of an movement with ‘delight’, a left-turn movement with ‘anger’, a
actually observed wave pattern of a PWM signal generated as A via right-turn movement with ‘sorrow’, and a backward movement
the oscilloscope. Figure 9(a) is an instance in which the emotion with ‘pleasure’. It also demonstrated that the speed of wheelchair
estimated by EFAM is ‘pleasure’ and its intensity is 54%. The movement and the turning angle (either left or right) can be
actual emotion recalled by the study subject was also ‘pleasure’, adjusted.
so the subject’s recalled emotion and the EFAM-estimated emotion
coincide. As a serial code, ‘B54’ is input to the microcontroller,
5. Conclusion
and PWM is output as Ab . Other movement parameters Af , Ar , and
Al are all 0. As the +duty cycle of the PWM signal was around This study demonstrated that an electric wheelchair can be
54%, one can tell that the PWM has been accurately output in accurately operated in real time using information only of EEG by
accordance with the intensity of the emotion. Figure 9(b) is an applying EFAM, which discriminates emotions of study subjects
instance in which the emotion estimated by the EFAM is ‘delight’ to a BCI using electrode difference signals obtained through
and its intensity is 90%. The actual emotion recalled by the study noninvasive, less constraining EEG. In addition, using a newly
subject was also ‘pleasure’, so the subject’s recalled emotion and developed BCI control circuit enables users to adjust the speed and
the EFAM-estimated emotion coincide. As a serial code, ‘F90’ angle of an electric wheelchair movement based on the intensity
is input to the microcontroller, and PWM is output as Af . Other of their emotion.
movement parameters Ab , Ar , and Al are all 0. As the +duty cycle In the future, we will seek to improve the emotion recogni-
of the PWM signal is around 90%, we can confirm that the PWM tion rate, using not only EEG information but also biological
has been accurately output in accordance with the intensity of information obtainable from near-infrared spectroscopy (NIRS).
emotion. Furthermore, we aim to develop our system into a more practical,
Figure 10 shows an instance of controlling an electric general one by building a brain–affective interface (BAI), which
wheelchair using the proposed system. Emotions recalled for 60 s can discriminate between multiple high-level emotions, to enable,
by a study subject are shown from the left in the following order: for example, the electric wheelchair to halt urgently when the user
‘delight’, ‘anger’, ‘sorrow’, and ‘pleasure’. All the horizontal feels fear.
axes represent time. Figure 10(a) is an emotion estimated by
real time EFAM, and the vertical axis represents the intensity References
of emotion. Figure 10(b) is the movement parameter A of the
wheelchair output from the microcontroller, and is ordered Af , (1) Vidal JJ. Toward direct brain-computer communication. Annual
Al , Ar , and Ab from the top. Figure 10(c) is the control voltage Review of Biophysics and Bioengineering 1973; 2:157–180.
C of the wheelchair that is output from CSG, and is ordered (2) Ramoser H, Muller-Gerking J, Pfurtscheller G. Optimal spatial
C d , C s from the top. Figure 10(d) is the rotation velocity v of filtering of single trial EEG during imagined hand movement. IEEE
the back wheels and the rotation angle θ of the front wheels, Transactions on Rehabilitation Engineering 2000; 8(4):441–446.
lined up in this order from the top. Fig. 10(a) shows that the (3) Wolpaw JR, Birbaumer N, Heetderks WJ, McFarland DJ, Peckham
intensities of each emotion matched by the emotion recalled by PH, Schalk G, Donchin E, Quatrano LA, Robinson CJ, Vaughan
the study subject are high. Thus, we can confirm that EFAM TM. Brain-computer Interface technology: A review of the first inter-
has accurately estimated the remembered emotion. The emotion national meeting. IEEE Transactions on Rehabilitation Engineering
2000; 8(2):164–173.
intensities outputs were around 50–60% for ‘delight’, 40–60%
(4) Cheng M, Gao X, Gao S, Xu D. Design and implementation of a
for ‘anger’, 50–70% for ‘sorrow’, and 50–60% for ‘pleasure’.
brain-computer Interface with high transfer rates. IEEE Transactions
Despite the fact that emotion intensities other than those of the of Biomedical Engineering 2002; 49(10):1181–1186.
remembered emotions are output, in the present system, only an (5) Naveen RS, Julian A. Brain computing interface for wheel chair
emotion whose intensity hits the highest mark and its intensity control. Computing, Communications and Networking Technologies
level are transmitted to the microcontroller in the next stage as (ICCCNT), 2013 Forth International Conference on, 2013; 1–5.
a serial code. The microcontroller will convert the serial code (6) Khare V, Santhosh S, Anand S, Bhatia M. Brain computer Interface
into the movement parameter A, which will be output as a + duty based real time control of wheelchair using electroencephalogram.
cycle PWM signal, which is in proportion to the intensity of the International Journal of Soft Computing and Engineering (IJSCE)
emotion. Figure 10(b) shows that a PWM with +duty cycle of 2011; 1(5):41–45.
around 50–60% was output to Af for ‘delight’; around 40–60% (7) Pinheiro OR, Alves LRG, de Souza JR. EEG signals classification:
Motor imagery for driving an intelligent wheelchair. IEEE Latin
was output to Af and Al for ‘anger’; around 50–70% was output
America Transaction 2018; 16(1):254–259.
to Af and Ar for ‘sorrow’; and around 50–60% was output to Ab
(8) Waytowich NR, Krusienski DJ. Development of an extensible
for ‘pleasure’. Afterwards, CSG will generate a control voltage C SSVEP-BCI software platform and application to wheelchair control.
of the wheelchair using the PWM signal. As shown in Fig. 10(c), Neural Engineering (NER) 2017; 1(1):259–264.
C d and C s were set to around 4 and 2.5 V, respectively, for (9) Musha T, Terasaki Y, Haque HA, Ivamitsky GA. Feature extraction
‘delight’; around 4 and 1 V, respectively, for ‘anger’; around from EEGs associated with emotions. Artificial Life and Robotics
4 and 4 V, respectively, for ‘sorrow’; and around 1 and 2.5 V, 1997; 1(1):15–19.
(10) Araga Y, Musha T. Digitalization and quantification of human Masahiro Nakagawa (Non-member) Masahiro Nakagawa
emotions using electroencephalogram. Bioscience and Industry 2001; obtained the Ph.D. degree from in Engineer-
59(10):693–694. ing from Faculty of Engineering, Nagaoka
(11) Kobayashi N, Nakagawa M. BCI-based control of electric wheelchair.
University of Technology (NUT), in 1988.
Proceedings of IEEE Global Conference on Consumer Electronics
(IEEE GCCE’2015), 2015; 429–430.
He was an Assistant Researcher in the
Faculty of Engineering, NUT, in 1982.
He was a foreign-based researcher from
Nobuaki Kobayashi (Member) received the B.S. degree from the March 1988 to January 1989 appointed
Faculty of Science and Engineering, Chuo by the Ministry of Education, Science,
University, in 2004, and the Ph.D. degree Sports and Culture, Japan, in the Department of Mathematics,
in engineering from the Graduate School of University of Strathclyde, UK. He is currently a Professor with
the same university in 2011. He is currently NUT. His research interests include chaos and fractal informatics,
with the College of Science and Technol- mathematical engineering, emotion information engineering,
ogy, Nihon University. His current research fractal image coding, chaos neural network, and physics of
interests include CMOS video and image liquid crystals. Prof. Nakagawa is a member of the Japan Neural
encoding processors, CMOS memory LSI, Network Society, The Physical Society of Japan, The Japan
and brain–computer interface (BCI). Dr Kobayashi is a member Society of Applied Physics, Japan Society for Fuzzy Theory and
of the Institute of Electronics, Information and Communication Intelligent Informatics, The Institute of Electronics, Information
Engineers (IEICE) and IEEE. and Communication Engineers (IEICE), and The Japanese Liquid
Crystal Society. He is the author of numerous books, including
Chaos and Fractals in Engineering (WorldScientific), Liquid
Crystal Phase Transition (Springer-Verlag Japan) (in Japanese),
and Chaos and Fractal Emotion Information Technology (Nikkan
Kogyo Shimbun, Ltd) (in Japanese).