Regression Analyses between Physiological Indexes and Level of
Understanding with VAS of a Listening Task
Masaki Omata
1
and Kazuya Nakazawa
2
1
Interdisciplinary Graduate School, University of Yamanashi, Kofu, Japan
2
Department of Computer Science and Engineering, University of Yamanashi, Kofu, Japan
Keywords: Level of Understanding, Physiological Index, Regression Analysis and Visual Analog Scale.
Abstract: This paper describes logistic regression analyses and multiple regression analyses to explain relationships
between physiological signals and subjective self-reported levels of understanding of a second language
listening task by using a visual analog scale (VAS) of 999 degrees. Mean contribution ratio of the logistic
regression expressions was 0.72, and mean contribution ratio of the multiple regression expressions was 0.70.
Power value of theta band of brain waves has a certain tendency to change according to the level of
understanding. Accuracy of the regression expressions using VAS was the same or more than that of the four-
level scale as our previous work.
1 INTRODUCTION
It is crucial to measure level of understanding as a
metric of learning achievement, and this metric is
used for self-reflective learning and also to develop
heuristic guidelines to improve instruction methods
and select appropriate materials. Although
conventional quizzes and questionnaires are common
methods for measuring the level of understanding,
there are two major problems with such methods.
First, these methods can impose the burden of
answering a question even when the learner does not
require this process because they already possess an
effective level of understanding. Second, the methods
interrupt learner’s activity even when the learner
wants to continue concentrating on that activity.
Several studies have reported that changes in
physiological signals reflect changes in the levels of
understanding of a learning task. For a verbal task,
some results indicate that the power values of alpha
and beta brain waves change relative to differences of
in the difficulty level of the text in a reading task, e.g.,
brain blood flow increases when reading text in a
secondary language compared to reading in the
primary language, skin conductance response
increases with greater English proficiency, and the
power value of alpha waves during reading English
content words is reduced compared to reading
functional words.
We have previously proposed using physiological
signals to validate the relationships between the
signals and a learner’s level of understanding, and we
have developed multiple regression expressions to
estimate understanding level (Omata, 2018). Using
such signals allows us to observe the state of a learner
without imposing a burden and interrupting activities;
thus, the estimation can represent a solution to the
aforementioned problems. Our multiple regression
expressions could estimate the level of individual
understanding while reading each second-language
sentence as at least 55% and at maximum 81% on a
four-level ordinal scale.
In this paper, we propose using a logistic
regression model to achieve a higher-accuracy
regression method. In addition, we propose the use of
a visual analog scale (VAS) as a rating scale for high-
resolution estimation. Previously, we reported that
the contribution rate of a logistic regression model
was greater than that of a multiple regression model
when estimating emotion from physiological signals
in a subject drawing a picture (Omata, 2014).
Watanabe et al. and Yamada et al. have reported that
a VAS method was more effective than Likert-scale
evaluations (Watanabe 2015, Yamada 2014).
This paper describes an experiment conducted to
examine the relationships between physiological
signals and self-reported levels of understanding of a
second-language listening task using a VAS. In
addition, this paper compares a logistic regression
Omata, M. and Nakazawa, K.
Regression Analyses between Physiological Indexes and Level of Understanding with VAS of a Listening Task.
DOI: 10.5220/0006892200210029
In Proceedings of the 5th International Conference on Physiological Computing Systems (PhyCS 2018), pages 21-29
ISBN: 978-989-758-329-2
Copyright © 2018 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
21
model to a multiple regression model relative to
contribution ratio to further explain these
relationships.
The contributions of this paper are summarized as
follows.
We demonstrate that physiological signals,
particularly the power of theta brain waves, are
related to levels of understanding in a listening
task.
VAS is more effective to record a detailed self-
reported level of understanding than a four-
level scale.
We demonstrate that the accuracy of a logistic
regression model is similar to that of a multiple
regression model relative to regression
analyses of such relationships.
2 RELATED WORK
This section surveys previous work that studied
relationship between person’s responses and their
physiological signals during a verbal task or e-
learning.
Concerning signals of central nervous system
(CNS), Mellem et al. focused on the role of
oscillatory EEG dynamics in retrieving open class
vs. closed class words (Mellem, 2012). Specifically,
they investigated the robustness of the theta and
alpha effects in different contexts and in a different
population and performed time–frequency analysis
based on the context or class of the word. As the
results, they observed larger power decreases in the
alpha waves for the open class compared to closed
class words and did not observe differences in theta
power between these conditions. Safi et al. studied
to validate a protocol using near-infrared
spectroscopy (fNIRS) for the assessment of overt
reading of irregular words and nonwords with a full
coverage of the cerebral regions (Safi, 2012). The
results reported that total hemoglobin
concentrations were significantly higher than
baseline for both irregular word and nonword
reading. Oishi et al. compared brain blood flow
during a listening task for first language with that for
second language (Oishi, 2008). The results reported
that the hemoglobin concentration of the second
language significantly increased than that of the first
language. They interpreted that the second language
needed attentional capacity more than the first
language.
Concerning signals of peripheral nervous system
(PNS), Harris investigated the phenomenon
psychophysiologically, 32 Turkish–English
bilinguals rated a variety of stimuli for pleasantness
in Turkish (L1) and English (L2) by analyzing skin
conductance via fingertip electrodes (Harris, 2003).
The results show that the participants demonstrated
greater autonomic arousal to taboo words and
childhood reprimands in their L1 compared to their
L2. Nomura et al. focused on the skin temperature
changes and Electrocardiogram (ECG) changes of
students engaged in two e-learning exercises:
interactive or non-interactive (Nomura, 2012). As
the result, the skin temperature showed significant
decline when subjects were engaged in the
interactive exercise. In addition, high frequency
(0.15-0.40 Hz) values of ECG dropped just after the
start of both exercises and remained low.
We conducted multiple regression analyses to
investigate a trend of the relations between several
physiological signals and level of understanding of
second language reading of a whole text (Omata,
2016). The contribution ratio was 0.82. After that,
we developed individual adaptive estimation
expressions to estimate a learner’s level of
understanding (four-level scale) of reading each
second-language sentence by analyzing the learner’s
physiological signals (Omata, 2018). The mean
determination coefficients of the expressions was
0.69 ranged from 0.55 to 0.81.
3 PHYSIOLOGICAL SIGNALS
AND INDEXES
Tables 1 and 2 list physiological signals and indexes
calculated from each kind of signals, which were
analyzed in the experiment. This section introduces
physiological signals obtained from participants
from two parts: central nervous system (CNS) and
peripheral nervous system (PNS); and also explains
the indexes and their pre-processing before
analyzing them.
We used the ProComp Infiniti (Thought
Technology Ltd.) for the measurement which is an
eight channel, biofeedback and neurofeedback
system for real-time data acquisition. The channels
can be used with some combination of sensors
described below.
PhyCS 2018 - 5th International Conference on Physiological Computing Systems
22
Table 1: Physiological signals and indexes of CNS.
Type of
signal
Calculated index
Method of
calculating
EEG
Power of theta waves
(4 Hz - 7 Hz)
Spectral analysis
of brain waves
on a forehead
Power of alpha1 waves
(7 Hz - 8 Hz)
Power of alpha2 waves
(9 Hz - 11 Hz)
Power of alpha3 waves
(12 Hz - 13 Hz)
Power of beta waves
(14 Hz - 30 Hz)
HEG
HEG ratio based on
baseline
Absorptive
power of
oxygenated
haemoglobin on
a frontal lobe
Table 2: Physiological signals and indexes of PNS.
Type of
signal
Calculated index
Method of
calculating
BVP
Power of HF waves
Spectral analysis of
BVPs on a left
thumb
LF/HF ratio
Pulse rate
(beats / minute)
HEG
SC ratio based on
baseline
Electrical
conductance
between left index
and middle fingers
RESP Breaths / minute
Expansion and
contraction motion
of abdomen
3.1 CNS Signals
3.1.1 Electroencephalogram (EEG)
EEG signals are detected on the scalp as neuronal
electric fluctuations in the brain and are classified by
their frequency: theta (4–7 Hz), alpha (8–13 Hz), and
beta (14–30 Hz). Alpha waves are frequently detected
when a person is relaxed, beta waves are commonly
detected when a person is concentrating, and theta
waves are detected when a person is feeling drowsy.
Alpha waves are further classified into alpha1 (7–8
Hz), alpha2 (9–11 Hz), alpha3 (12–13 Hz). We
recorded these data using an EEG-Z sensor (Thought
Technology Ltd.), classified the frequencies and
calculated the power of each frequency band (see
Table 1). Figure 1a shows such electrodes on a
participant’s forehead (F3 of the international 10–20
system).
(a) (b)
(c) (d)
(e)
Figure 1: Sensors and electrodes for measuring
physiological signals: (a) electrodes for EEG, (b) a HEG
headband, (c) a BVP sensor, (d) a SC sensor and (e) a RESP
sensor.
3.1.2 Hemoencephalography (HEG)
The HEG ratio is the ratio between oxygen-rich
hemoglobin and oxygen-starved hemoglobin in
cerebral blood flow (CBF), which correlates with
blood flow dynamics and cellular metabolism in
localized parts of the brain cortex. The level of
oxygenated hemoglobin increases during brain
activation due to the phenomenon of neurovascular
coupling. Therefore, we subcutaneously recorded the
ratio at 2 cm on the forehead by measuring the
absorptive power of near-infrared and red light of the
CBF because of the differences of the absorptive
power of oxygen-rich and oxygen-starved
hemoglobin. The HEG ratio is calculated as follows
 = 200


(1)
where RED is the absorptive powers of oxygenated
hemoglobin and reduced hemoglobin during visible
red light irradiation, and IR is their absorptive powers
during near-infrared light irradiation. We used a near-
infrared headband (MediTECH Electronic GmbH;
see Figure1b) on the forehead (Fp2 of the
international 10–20 system) to measure HEG ratio.
Regression Analyses between Physiological Indexes and Level of Understanding with VAS of a Listening Task
23
3.2 PNS Signals
3.2.1 Blood Volume Pulse (BVP)
BVP, also called photoplethysmography, is a relative
measure of heart rate and inter-beat interval and an
index of the coarctation and angiectasis of peripheral
blood vessels. The frequency characteristics of such
waves reflect the tone of sympathetic and
parasympathetic nerves. The high-frequency (HF,
0.15–0.4 Hz) component reflects the tone of
sympathetic nerves, while the low-frequency (LF,
0.04–0.15 Hz) component reflects the tone of both
types of nerves. We used a BVP sensor (Thought
Technology Ltd.; see Figure 1c), which bounces
infrared light against a skin surface and measures the
amount of reflected light.
3.2.2 Skin Conductance (SC)
A hand’s SC is a measure of the skin’s ability to
conduct electricity due to the eccrine sweat between
two fingers. SC represents changes in the sympathetic
nervous system. The value of the sensor increases
when the subject is in an excitatory state or a stressful
situation. We used an SC sensor (Thought
Technology Ltd.; see Figure 1d) that measures SC
between two fingers in micro-Siemens.
3.2.3 Respiration (RESP)
Respiratory movement consists of inspiration and
expiration. The RESP rate is an index of emotion with
an increase indicating tension and a decrease
indicating relaxation. We used a RESP sensor
(Thought Technology Ltd.; see Figure 1e) that is
sensitive to stretching. When strapped around a
participant’s abdomen, it converts the respiratory
movement into an electric signal. We also calculate
the RESP rate from the data.
3.3 Pre-processing
Note that pre-processing was required for eliminating
individual differences among participants, thus the
data of physiological indexes were standardized using
both experimental data and neutral data prior to
performing experimental tasks. The indexes of EEG,
HEG ratios and SC were calculated from the mean
and standard deviation of the same type of data by
following the equation:
=
−
(2)
where x is the raw data calculated from the signals
obtained from a sensor, and μ and σ are the mean and
standard deviation of the same kind of data,
respectively, recorded before performing an
experimental task in the neutral state as a baseline for
each participant. The RESP and BVP indexes were
standardized using the relative ratios obtained from
the experimental and neutral data.
4 EXPERIMENT
This section describes a verification experiment to
model the relationships between physiological
indexes and participants’ subjective self-reported
levels of understanding when listening to English
speech, which was a second language for the
participants.
4.1 Objective
The main objective of the experiment was to build a
regression expression to explain a learner’s level of
understanding more accurately than our previous
model by verifying the relationships between
understanding levels and physiological indexes. In
addition, a second objective was to evaluate the effect
of using a VAS as a rating scale by comparing it to
the four-level ordinal scale used in our previous work.
Therefore, our hypotheses are as follows.
H1: There is a significant relationship between the
level of understanding of a second-language
listening task and the listener’s physiological
indexes.
H2: The accuracy of a logistic regression model is
significantly greater than that of a multiple
regression model relative to a contribution ratio
comparison.
H3: A high-resolution VAS is more useful than a
low-resolution four-level scale for regression
analyses.
4.2 Experimental Setup
Figure 2 shows the experimental equipment and a
participant wearing the sensors used to measure
physiological signals during a task. A laptop with a
17.3-inch display was used to measure and record
data, and two speakers were placed approximately 30
cm from the participant. The participant’s left hand
was kept on a table and restrained from moving due
to sensors attached to the fingers.
PhyCS 2018 - 5th International Conference on Physiological Computing Systems
24
A slider (Phidgets Inc. Slider 60; Figure 3) for the
VAS was placed near the participant’s right hand. The
participant reported their subjective self-reported
level of understanding by moving the knob of the
slider along the sulcus. The resolution of the slider
was 999 degrees (0 to 998) within a 60 mm sulcus.
The degrees represent “Not at all understood” (0) side
to “fully understood” (998). The middle position
represents “neutral” degrees and was the default
position before moving. The cardboard around the
equipment was used to unify the participant’s sight by
obscuring visual information in order to allow the
participants to concentrate on the listening task.
Figure 2: Experimental equipment and a participant
wearing sensors.
Figure 3: Slider to report level of understanding.
4.3 Task
The task was to listen to English speech and report the
level of understanding of the content. In each trial, the
participant first clicked the “play” button on the
laptop’s screen. Then, they listened to English speech
played from the speakers. They then reported their
level of understanding of the content by moving the
slider knob and clicking the “record” button to fix and
record the answer. Finally, they moved the knob back
to the default start position prior to the next trial. The
participant then repeated this process with another
speech.
4.4 Stimuli
The speeches used for the listening task were
excerpted from the Grade 1, Grade Pre-1, Grade 2,
Grade Pre-2, Grade 3, and Grade 4 EIKEN tests
(Eiken Foundation of Japan, online), which is one of
the most widely used English-language testing
programs in Japan. The speeches ranged in duration
from 15 to 20 s and represented the above six
difficulty grades. Each difficulty grade included ten
speeches.
4.5 Participant
Ten male college students aged 21 to 23 participated
in this experiment. All participants were native
Japanese speakers who had studied English for over a
decade. Their TOEIC (IIBC, online) listening scores
ranged from 235 to 385 (mean score: 281.5) out of
495. Each participant was assigned a distinguishing
ID character from A to J to facilitate explaining the
results and analyses.
4.6 Procedure
A block of each difficulty grade comprised two parts:
(1) recording physiological signals during rest and (2)
performing a task with ten listening trials for each
grade.
In the rest part, all physiological signals of a
participant were measured and recorded for 60 s
while the participant rested, opened their eyes, and
thought of nothing.
Then, in the task performance part, the participant
listened to the speeches (ten voices) of each grade and
self-reported a subjective level of understanding for
each trial.
All participants conducted six blocks for all six
grades in a within-subject design. Therefore, each
participant listened to 60 speeches. Note that the order
of the blocks for each participant was
counterbalanced; thus, the order of grades differed for
each participant.
In addition, we obtained informed consent from
all participants prior to conducting the experiment.
4.7 Results
Figure 4 shows results of the self-reported level of
understanding of each participant for Grade 4, i.e., the
easiest grade, and Figure 5 shows the results of the
self-reported level of understanding of each
participant for Grade 1, i.e., the most difficult grade.
Fully understood
Start position
Not at all understood
Regression Analyses between Physiological Indexes and Level of Understanding with VAS of a Listening Task
25
Although the detailed results of other grades are
omitted in this paper, we observed that, overall, more
difficult grades resulted in lower understanding
levels. In addition, because the levels of
understanding were fully distributed, and several
values were 999 degrees on the slider, the self-
reported values and physiological indexes were
useful for analyses.
Figure 4: Distribution of self-reported levels of
understanding of each participant of Grade 4.
Figure 5: Distribution of self-reported levels of under-
standing of each participant of Grade 1.
4.8 Analyses
This section describes statistical analyses conducted
to test our hypotheses (Section 4.1).
4.8.1 Correlation Analyses
Tables 3 and 4 show the correlation coefficients
between the self-reported level of understanding
values of each participant and the physiological index
values of CNS (Table 3) and PNS (Table 4). Overall,
the sign of the coefficients of the power values of each
band of brain waves was negative. We observed that
the power values of theta band brain waves correlate
inversely with the self-reported values. The absolute
values of the coefficients of participants I and J of the
theta band were greater than 0.75.
Table 3: Correlation coefficients between level of
understanding and CNS indexes of each participant.
ID Power
of
Power
of
Power
of
Power
of
Power
of
HEG
A -0.45 -0.38 -0.25 -0.24 -0.29 0.33
B -0.52 -0.51 -0.51 -0.51 -0.55 -0.35
C -0.57 -0.31 -0.29 -0.24 -0.28 -0.27
D -0.58 -0.58 -0.59 -0.59 -0.61 -0.12
E -0.35 -0.24 -0.41 -0.26 -0.44 0.22
F -0.56 -0.33 -0.42 -0.43 -0.45 0.02
G -0.37 -0.51 -0.47 -0.49 -0.49 -0.29
H -0.44 -0.44 -0.36 -0.40 -0.45 -0.02
I -0.77 -0.52 -0.50 -0.52 -0.59 0.25
J -0.75 -0.71 -0.59 -0.51 -0.63 0.19
Table 4: Correlation coefficients between level of
understanding and PNS indexes of each participant.
ID Power
of HF
LF/HF Pulse
rate
SC ratio Breaths
/ minute
A -0.14 -0.29 -0.03 0.26 -0.25
B -0.40 0.17 -0.09 0.31 0.11
C -0.31 -0.11 -0.46 0.10 -0.21
D -0.46 0.30 0.13 -0.66 0.21
E 0.08 -0.34 -0.58 0.10 -0.10
F -0.62 0.39 0.56 0.25 -0.33
G -0.39 0.19 0.19 -0.61 0.30
H 0.02 -0.07 -0.20 -0.10 -0.27
I -0.31 -0.02 0.22 0.41 0.22
J -0.15 0.30 -0.17 -0.08 -0.25
Figure 6 illustrates the relationships between the
power values of the theta waves of participant J and
his self-reported levels of understanding as a high
correlation example. As can be seen, a higher level of
understanding, which means the level moves toward
“fully understood,” is correlated with lower theta
wave power. Figure 7 illustrates the power value of
the theta waves of each trial for each grade of
participant J. The graph shows that the power value
of theta waves varies relative to the difficulty of the
grade from easy (Grade 4) to difficult (Grade 1).
Figure 6: Correlation between level of understanding and
power of theta of participant J.
0
200
400
600
800
1000
ABCDEFGH I J
Self-reprted level of
understanding
Participant's ID
0
200
400
600
800
1000
ABCDEFGH I J
Self-reprted level of
understanding
Participant's ID
-0,6
-0,4
-0,2
0,0
0,2
0,4
0,6
0,8
1,0
0 200 400 600 800 1000
Power value of theta waves
Level of understanding
PhyCS 2018 - 5th International Conference on Physiological Computing Systems
26
Figure 7: Level of power of theta waves of each grade of
participant J.
4.8.2 Regression Analyses using Two Models
This section compares the accuracy of a logistic
regression model to that of a multiple regression
model to explain the relationships between the values
of physiological indexes and the self-reported levels
of understanding using the same recorded data.
Here, the objective variable of both regression
methods is the self-reported level of understanding.
However, it was necessary to convert the values from
ratio scales to ordinal scale in the logistic regression;
therefore, the values were divided equally by five and
ranked from first to fifth because the number of
samples in each division was small in greater than
five equal divisions. On the other hand, in the
multiple regression, the values of the self-reported
levels were used as the values of the objective
variable without conversion.
To avoid multicollinearity among the indexes, the
explanatory variables of both regression methods are
the principal components obtained by principal
component analyses of the values of the physiological
indexes.
The overall contribution ratio of the logistic
regression expression obtained using all participant
data was 0.34, and that of the multiple regression
expression using the same data was 0.35. Both ratios
are less than 0.5 and less accurate to explain the
relationships (Donna, online).
Therefore, we built individual logistic regression
and multiple regression expressions to explain the
relationships of each participant using only that
participant’s data. Table 5 shows the contribution
ratios of the individual logistic regression and
multiple regression expressions of each participant.
As can be seen, all ratios are greater than 0.5, the
mean ratio of the logistic regression expressions is
0.72, and the mean ratio of the multiple regression
expressions is 0.70. These results indicate that the
regression expressions can reasonably explain the
relationships between the participant’s levels of
understanding and the physiological indexes
individually, and that there is no significant
difference between logistic and multiple regression.
Table 5: Contribution ratios of two regression models for
each participant.
ID Contribution ratio of
logistic regression
Contribution ratio of
multiple regression
A 0.68 0.68
B 0.74 0.68
C 0.79 0.75
D 0.70 0.68
E 0.58 0.53
F 0.73 0.72
G 0.78 0.70
H 0.61 0.67
I 0.85 0.86
J 0.76 0.71
4.8.3 Analyses of Accuracy of Estimation
We compared the estimated levels of understanding
using an individual multiple regression expression
with the self-reported levels of understanding for each
participant. The data for the comparison were
classified into three difficulty grades, i.e., low
(Grades 4 and 3), middle (Grades 2 and Pre-2), and
high (Grades 1 and Pre-1) for the analyses. Figure 8
shows the variance of the correlation coefficients for
all individual participants of each classified grade.
Note that values were normalized to derive the
multiple regression expressions. The graph shows
that the coefficients of the high grade converge within
a high range and that the mean coefficients of the high
grade are greater than that of the other grades.
Figure 8: Correlation coefficients of three classified grades.
-0,6
-0,4
-0,2
0,0
0,2
0,4
0,6
0,8
1,0
12345678910
Power of theta waves
Listening task trial
Grade 1 Grade Pre-1
Grade 2 Grade Pre-2
Grade 3 Grade 4
0,0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
0,9
1,0
High Middle Low
Correlation coefficients between
estimate value and reported value
Difficulty of listening task
Regression Analyses between Physiological Indexes and Level of Understanding with VAS of a Listening Task
27
Figures 9 and 10 show the estimated values by the
multiple regression expression and the actual self-
reported values for Grade 4 (Figure 9), which is the
easiest, and Grade 1 (Figure 10), the most difficult,
for participant G, whose correlation coefficient is
close to the mean of all participants. The graphs show
that the estimated values are approximately close to
the actual self-reported values.
We also compared the estimated levels of
understanding using an individual logistic regression
expression with the self-reported levels of each
participant. Here, the mean accuracy rate of the
estimation of all individuals was 0.76, and the
standard deviation of the rates among the individuals
was 0.10.
Figure 9: Comparison of estimate values with self-reported
value of Grade 4 of participant G.
Figure 10: Comparison of estimate values with self-
reported value of Grade 1 of participant G.
4.8.4 Regression Analyses using Brain
Waves
We attempted to conduct regression analyses with
both models using only brain wave indexes to
consider whether the number of physiological signals
should be reduced relative to implementation cost.
These analyses were limited to brain waves because
the correlation coefficients of the brain wave indexes
were greater than those of the other signals, and the
signs of the coefficients of all participants were the
same (Section 4.8.1).
Table 6 shows the contribution ratios of the
individual logistic regression and multiple regression
expressions using only the values of the brain wave
indexes for each participant. As can be seen, the mean
ratio of logistic regression is 0.53, and the mean ratio
of multiple regression is 0.53. Both mean values are
less than those of the expressions using all
physiological indexes (Section 4.8.2).
Table 6: Contribution ratios of two regression models using
only brain wave indexes.
ID Contribution ratio of
logistic regression
Contribution ratio of
multiple regression
A 0.43 0.44
B 0.70 0.62
C 0.72 0.69
D 0.58 0.43
E 0.31 0.30
F 0.46 0.48
G 0.33 0.41
H 0.39 0.46
I 0.76 0.78
J 0.59 0.64
5 DISCUSSIONS
Relative to H1, we verified a significant relationship
between a listener’s self-reported level of
understanding in the listening task and the listener’s
physiological signals at an individual level. However,
the contribution ratios of the logistic and multiple
regression expressions using all participant data
collectively were low; thus, we consider that it is
difficult to develop a unified regression model to
express the relationships of all learners. We surmise
that this difficulty is due to the fact that kind or
tendency of physiological signals that change relative
to a difficulty level of understanding varies among
individuals, and that each individual has a particular
kind or tendency of the physiological changes when
he/she is finding something difficult. On the other
hand, we suggest that the power value of the theta
brain waves has a certain tendency to change
according to the level of understanding of all
listeners.
Relative to H2, we found no significant difference
between the accuracies of logistic and multiple
regressions when modelling the relationships
between the self-reported level of understanding and
the physiological indexes. Therefore, we consider it
necessary to use another statistical regression model,
to combine another type of data such as English score,
or to include other physiological signals such as an
0,0
0,2
0,4
0,6
0,8
1,0
1,2
12345678910
Normarized level of
understanding
Listening task trial
Actual self-reported value Estimate value
-4,0
-3,0
-2,0
-1,0
0,0
1,0
12345678910
Normarized level of
understanding
Listening task trial
Actual self-reported value Estimate value
PhyCS 2018 - 5th International Conference on Physiological Computing Systems
28
electromyogram to measure movement of facial
muscles, in order to increase accuracy.
Relative to H3, we have verified that the accuracy
of the 999-degree scale was greater than or equal to
that of the four-level scale in our previous work,
where the mean accuracy was 0.69 (Omata, 2018).
Therefore, we propose that it is possible to estimate
the level of understanding at sufficient resolution
using a VAS.
6 CONCLUSIONS
We conclude that our logistic regression expressions
can explain the relationships between an individual
level of understanding of listening a second language
and individual physiological indexes with 0.72
contribution ratio, and that our multiple regression
expressions can also explain the relationships with
0.70 contribution ratio. The multiple regression
expressions can explain the relationships with
convergent higher correlation coefficients when the
speech voices are difficult for the listener.
Additionally, we suggest that VAS is useful to
answer a level of understanding and estimate it at high
resolution.
We, therefore, propose that it is possible to
develop an e-learning analysis system that
automatically estimates a detailed level of
understanding from learner’s physiological signals
and an e-learning content recommendation system
that automatically provides a flexible and adaptable
learning material based on the estimation for an
individual learner.
REFERENCES
Donna, M. R., Correlation Coefficient, MathBits.com.
Available at: <http://mathbits.com/MathBits/TISection/
Statistics2/correlation.htm> [Accessed 20 April 2018].
Eiken Foundation of Japan, Online. Test in Practical
English Proficiency in Japan. Available at:
<http://www.eiken.or.jp/eiken/en/> [Accessed 20 April
2018].
Harris, C.L., Ayçiçegi, A. and Gleason, J.B., 2003. Taboo
and emotion words elicit greater autonomic reactivity
in a first language than in a second language. In Applied
Psycholinguistics, Vol.4, pp.561-578.
Institute for International Business Communication (IIBC),
online. TOEIC Listening & Reading Test, Available at:
<http://www.iibc-global.org/english/toeic/test/lr/abou
t.html> [Accessed 20 April 2018].
Mellem, M. S., Marcel, C. M. and Rhonda, B. F., 2012.
Word class and context affect alpha-band oscillatory
dynamics in an older population. In Frontiers in
Psychology, Vol.3, No.97, pp.1-8.
Nomura, S., Hasegawa-Ohira, M., Kurosawa, Y.,
Hanasaka, Y., Yajima, K. and Fukumura, Y., 2012.
Skin Temperature as A Possible Indicator of Student’s
Involvement in E-learning Sessions. In International
Journal of Electronic Commerce Studies Vol.3, No.1,
pp.101-110.
Oishi, H. and Kinoshita, T., 2008. Daiichigengoshori to
dainigengoshori ni okeru noukasseijoutai no chigai. In
Studia Linguistica Vol. 21, pp.143-154 (in Japanese)
Omata, M. and Tanabe, S., 2018. Physiological Estimation
of Level of Understanding of Each Sentence in Second
Language Reading. In Proceedings of the 4th IEEE
Middle East Conference on Biomedical Engineering,
pp.1-6.
Omata, M., Kanuka, D. and Xiaoyang, M., 2014.
Experiments for Emotion Estimation from Biological
Signals and Its Application. In Transaction on
Computational Science XXIII. Lecture Notes in
Computer Science, Vol. 8490, pp.178-198.
Omata, M. and Tanabe, S., 2016. A regression equation to
estimate the degree of understanding in a reading using
physiological indexes. In Proceedings of the 28th
Australian Conference on Computer-Human
Interaction, pp.333-337.
Safi, D., Lassonde, M., Dang K. N., Vannasing, P.,
Tremblay, J., Florea, O., Moncet, O., Lefrancois, M.
and Béland, R., 2012. Functional near-infrared
spectroscopy for the assessment of overt reading. In
Brain and Behavior, Vol.2, No.6, pp.825-837.
Watanabe, S., Shirahama, N., Tsukamoto, H., Matsumoto,
Y., Kanagawa, M., Miyamoto, K., Nakaya, N., Tomita,
M. and Mori, Y., 2015. Distribution of the Subjective
Evaluation based on Visual Analog Scale and Likert
Scale. In Journal of Biomedical Fuzzy Systems
Association, Vol.17, No.1, pp.31-38 (in Japanese).
Yamada, K. and Erikawa, S., 2014. Evaluating the
Effectiveness of Visual Analogue Scales in a Web
Survey. In The Bulletin of Faculty of Sociology,Toyo
University. Vol.52, No.1, pp.57-70 (in Japanese).
Regression Analyses between Physiological Indexes and Level of Understanding with VAS of a Listening Task
29