Biomedical Signal Processing and Control 30 (2016) 98105

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

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

A human-machine interface based on single channel EOG and

patchable sensor
Xuhong Guo a , Weihua Pei a, , Yijun Wang a , Yuanfang Chen a , He Zhang a , Xian Wu a ,
Xiaowei Yang a , Hongda Chen a , Yuanyuan Liu b , Ruicong Liu b
a
Institute of Semiconductors, CAS Center for Excellence in Brain Science and Intelligence Technology, Chinese Academy of Sciences, Beijing, China
b
Engineering Research Center for Semiconductor Integrated Technology, Chinese Academy of Sciences, Beijing, China

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

Article history: Goal: This study demonstrates that single channel electrooculography (EOG) recorded by ultrathin and
Received 1 December 2015 exible electrodes can be practically used to control computer or machine, with the aid of proper recog-
Received in revised form 29 May 2016 nition algorithms and efcient controlling methods.
Accepted 27 June 2016
Methods: First, micro-fabrication process and transfer technology were used to develop a patchable sen-
Available online 14 July 2016
sor including three electrodes (a measurement electrode, a ground electrode, and a reference electrode).
Each electrode was composed of golden ribbon in the form of a lamentary serpentine mesh to provide
keywords:
conformal contact of skin and stretchability. Second, EOG was recorded by the proposed sensor installed
Flexible patchable sensor
Electrooculography
above the eyebrow. The peak and trough of eye movement signals were extracted as features to recognize
Human-machine interface three types of eye movements (blink, upward and downward) using a threshold-based recognition algo-
rithm. Finally, a human-machine interface (HMI) system was realized by converting eye movements to
computer commands including scroll up, scroll down, and close. To verify the effectiveness of the system,
eight subjects were trained to use their eye movements to navigate a document on the screen.
Results: The sensor was approximately triangular with a 5 cm side-length and a 70 um thickness. The
electrode can be stretched to 10% longer without any damage. The weight of the sensor was 180 mg. The
demonstration system was capable of making continuous controls with an average accuracy of 84%.
Conclusion: Single channel EOG recorded with a patchable sensor is feasible for developing a wearable
HMI system. The proposed system provides a comfortable user experience, a stable control method and
a simple systematic framework for developing practical HMI systems.
2016 Elsevier Ltd. All rights reserved.

1. Introduction
In recent years, due to the great advances in biomedical instru-
mentation, applications of bioelectrical signals are now widely
used in monitoring human activities rather than medical diagno-
sis [1]. A number of human-machine interfaces (HMIs) based on
bioelectrical signals have been proposed. For example, electroen-
cephalogram (EEG) has been used to control external devices, such
as a computer cursor [2] or a virtual helicopter [3]. Electromyo-
graphy (EMG) is another approach for controlling output devices
such as a prosthetic limb [4]. In addition, electrooculography (EOG)
has been shown to be efcient for controlling a keyboard [5].
These bioelectrical signals-based HMIs provide new communica-
tion methods for both disabled and healthy people.
Corresponding author.
E-mail address: peiwh@semi.ac.cn (W. Pei).
Traditional HMI systems are only suitable for subjects who stay
still or have few motions. Wet electrodes, multi-channel biopoten-
tial signals and a series of hardware, which were used to induce
or synchronize bioptentials, must be employed to ensure the feasi-
bility and reliability of collecting bipotentials in the HMI. Although
excellent performance can be achieved, these bulky HMI systems do
not t the requirements of a wearable system due to their uncom-
fortable user experience and unwieldy equipment.
There is a clear trend that the HMI system is getting smaller,
lighter, and more comfortable to wear [6]. In order to solve the
related problems in HMI systems, an extensive amount of work
has been invested in recent years.
To facilitate preparation and removal of traditional wet
electrodes, novel electrodes like dry-contact and non-contact
electrodes have been developed. Forvi et al. fabricated a micro-
needle array based dry electrode [7]. It can achieve lower contact
impedance without the use of conductive gel. However, this kind
of sensor is slightly invasive as the microneedles penetrate the

http://dx.doi.org/10.1016/j.bspc.2016.06.018
1746-8094/ 2016 Elsevier Ltd. All rights reserved.
 X. Guo et al. / Biomedical Signal Processing and Control 30 (2016) 98105
Fig. 1. A schematic diagram of the system.
stratum corneum and can cause minor pain or infection. Non-
contact sensors fabricated by Matthews et al. can acquire signals
without directly contacting the skin [8,9]; this kind of capacitive
sensor is very sensitive to motion and relies on a complex amplier
to acquire signals due to high source impedance.
Many groups try to simplify the biopential-based HMI systems
by reducing the number of electrodes, developing more efcient
control paradigms and signal processing algorithms. Abdelkader
et al. proposed a classication algorithm based on two channels of
EOG, which can identify six classes of eyes movements [10]. How-
ever, there is still room for improvement in the accuracy (85.2%)
and sensitivity (77.6%). Abo-Zahhad et al. proposed a HMI system
based on single channel EOG recorded with a commercial headset
(NeuroSky, CA, US) [11]. High accuracy (97.3%) was achieved, but
the information transfer rate (ITR) was limited because the system
can only identify one type of eye activity. Ma et al. designed a hybrid
HMI system that can send 28 commands to four robots based on
EOG and EEG [12]. The system required 11 channels and the oper-
ation method was relatively complex. Overall, these HMI systems
are still not convenient to use.
From the above analysis we can see the traditional wet electrode
is not suitable for long-term recording. In order to ensure good
signal quality, users need to replace electrode frequently. In addi-
tion, traditional HMI systems use numerous recording channels,
therefore, users need to spend a lot of time to prepare electrodes
and ensure the stable wear. Those troubles make users reluctant to
use the HMI system. As far as we know, these two problems were
always studied separately. Very few studies combined the advan-
tages of dry electrode and a small number of recording channels to
solve these problems. For example, in Ref. [13], Tejz et al. designed a
four-channel EOG-based assistive system using thin copper plates
as dry electrodes. The proposed system could reach 100% accuracy
in the classication of three kinds of eye movements (blink, left
and right). To further reduce the amount of the channels, systems
were developed with a single-channel commercial headset (Neu-
roSky, CA, US) [11,14,15]. The sensor was a dry electrode made of
stainless steel. The system could only detect one type of eye activity
(i.e., eye blink). In Ref. [16], with an optimized algorithm, the sys-
tem designed by Shen et al. could classify two different eye-blink
signals with the accuracy of 95% using the same headset. Only a
small number of types of eye activity can be classied in these
studies. In addition, the stainless steel electrode is sensitive to head
movements.
This study pays more attention to the user experience and
ergonomic factors for developing a wearable HMI system for practi-
cal applications. On the one hand, we design a exible dry electrode
99
to replace the traditional wet electrode or rigid dry electrodes. It can
acquire good signal quality comparable to the wet electrode. On the
other hand, we manufacture a exible sensor, which integrates a
measurement electrode, a reference electrode and a ground elec-
trode, to realize a single-channel EOG-based control paradigm. It
can signicantly reduce system preparation time and improve user
comfort. The proposed HMI system used a exible patchable sensor
attached above the eyebrow to record single-channel EOG. A new
control paradigm and an efcient classication algorithm are com-
bined to implement the HMI system, which can recognize three
kinds of eye movements (upward, downward, and blink) with high
accuracy and ITR.
2. Materials and methods
2.1. Control paradigm
EOG is the difference of the corneal-retinal potential in the pro-
cess of eye-movements. The researcher takes the fundus as the
negative pole and the cornea as the positive pole. The amplitude
of EOG ranges between 50 uV and 3.5 mV with a frequency range
of dc100 Hz [17].
Traditional EOG recording needs four or more electrodes pasted
around the eyes, which is uncomfortable. And if traditional elec-
trodes are used, it is inconvenience for the subjects wearing these
devices. In this study, we present a new control paradigm based on
a novel patchable sensor. The sensor, which can be stuck on the
forehead above the eyebrow, includes three electrodes (namely
measurement (MEA), ground (GND), and reference (REF) elec-
trodes), each in the form of a lamentary serpentine mesh with
exposed gold that contacts the skin directly. Three types of eye
movements including upward, downward, and blink can be rec-
ognized from the single-channel EOG recorded by the sensor and
converted into computer commands (i.e., scroll up, scroll down, and
close) to navigate a document on screen.
2.2. System architecture
As illustrated in Fig. 1, the system can be divided into three
modules. The rst module is the sensor for EOG acquisition. Fig. 2
shows the sensors structure, which consists of a base elastic tape
layer and three electrodes made with lamentary serpentine gold
mesh. The size of the patchable sensor is approximately 5 cm,
and the thickness is about 70 um.The size of each electrode is
2 cm 2 cm. Three electrodes are distributed in an equicrural tri-
angle manner with a center-to-center distance of 3 cm. The sensor
100 X. Guo et al. / Biomedical Signal Processing and Control 30 (2016) 98105
Fig. 2. Structure of the sensor.
Fig. 3. Example EOG signals recorded by the system.
is mounted on the forehead 2 cm above the left eyebrow. The elec-
trode closed to eyebrow acts as the measurement electrode, and
the other two electrodes act as reference and ground respectively
(see Fig. 1).The second module is the acquisition instrument. The
signals are recorded by a commercial amplier system (Blackrock
Cerebus). The sampling rate is 500 Hz and ADC is 16 bits. The band-
pass lter is set to 0.350 Hz. The data can be read or displayed on
a PC and processed online using MATLAB (Mathworks, Inc.). The
third module is the host computer control system. It includes data
processing and a graphic user interface. Data preprocessing, feature
extraction, classication, and display are developed under MATLAB.
Fig. 3 shows example EOG signals of three types of eye movements
(i.e., blink, upward, and downward) recorded using the proposed
system.
2.3. The patchable sensor
2.3.1. Fabrication process
Fig. 4 shows the fabrication process of the patchable sensor.
A four-inch silicon wafer is rinsed by acetone, ethanol and DI
water sequentially. A gold made metal layer (1um in thickness) is
deposited (Fig. 4(a)) and then a photoresist (BP 212) is spin coated
onto it (Fig. 4(b)). The photoresist is patterned via photolithogra-
phy (Fig. 4(c)) and the metal layer is patterned via wet etching (I2 :
KI: H2 O = 1:4:4) (Fig. 4(d)). Photoresist on the surface of the metal
layer is removed by acetone to get the mesh structure on the sil-
icon substrate (Fig. 4(e)). The metal electrode is transferred from
Si wafer to an ordinary elastic tape (Fig. 4(f)) to get the patchable
sensor (Fig. 4(g)).
2.3.2. Mechanical test
The mechanical property is veried by wrapping the electrode
around a pencil (0.7 cm in diameter) and straightening 10 times
repeatedly. Then the microscope is used to observe the surface of
the electrode morphology and to check the intactness of the gold
strips.
2.3.3. Contact impedance test
Contact impedance is measured by a potentiostat/galvanostat
(CHI 660D). Two exible electrodes are placed 2 cm apart on one
volunteers forearm. Sinusoidal AC signals with voltage amplitude
of 10 mV, frequencies from 0.1 Hz to 100 kHz, are applied to mea-
sure the contact impedance. For comparison purposes, commercial
wet electrodes (Ag/Agcl, Red Dot 5620, 3 M) are also measured at
the same skin positions using the same method.
2.3.4. Biopotential recording ability test
As a proof-of-concept experiment, recordings of biopotential
signals including ECG and EOG are performed to compare the per-
formance of commercial wet electrodes and the proposed exible
electrodes.
The ECG and EOG signals of eight subjects were recorded by the
amplier system (Blackrock cerebus). These signals were collected
with wet electrode and exible electrode at the same time. Fig. 5
shows the positions of electrode placement for ECG (Fig. 5(a)) and
EOG (Fig. 5(b)). REF and GND indicate the placement positions of
the reference electrode and the ground electrode, respectively. A
and B indicate the measurement electrodes of a exible electrode
and a wet electrode, respectively. For the comparison purpose, they
were 2 cm apart.
For quantitatively comparing the quality of the signals, correla-
tion coefcient and signal-to-noise ratio (SNR) are used to evaluate
the performance. The calculations are described as follows:

R = cov (x1, x2) / cov (x1, x1) cov (x2, x2) (1)
where R symbolizes the correlation coefcient, cov(.) represents
covariance, and x1 and x2 represent signals recorded by the exible
electrode and the wet electrode respectively.
Vs
SNR = 20 log10 (2)
Vn
where Vs and Vn represent the amplitude of signal and noise,
respectively.
2.4. Experiment setup
Eight subjects participated in the EOG signal measurement
experiment. The experiment was divided into two phases: training
and testing. Subjects learned the control method in the ofine train-
ing phase. For each subject, EOG data recorded from the training
phase were used to determine optimal thresholds to improve the
accuracy, specicity and sensitivity of the system. The performance
of the system was further evaluated with testing phase.
There were 30 trials in the training phase (5 for each of the six
classes: stay at the center, eyes leftward, rightward, upward, down-
ward or blink). At the beginning of a trial (t = 0 s), a xation cross
appeared on the screen. After one second (t = 1 s), an arrow as a cue
pointing either to the center, left, right, up, down or four corners
(corresponding to one of the six classes) appeared. The subjects
were instructed to make the indicated eye movement before the
cue disappeared (t = 2 s). The next trial began after another 1s. The
 X. Guo et al. / Biomedical Signal Processing and Control 30 (2016) 98105 101

Fig. 4. Fabrication process of the patchable sensor.

Fig. 5. Positions of electrode placement for (a) ECG and (b) EOG.

training phase, the system generated the thresholds for online test-
ing phase.
The purpose of the experiment in the testing phase was to sim-
ulate a scenario that a subject controls scroll up, scroll down and
close of a document on a computer screen, which can be realized
by three classes of eye movements (eyes upward, downward and
blink). After the training phase, the system generated the thresh-
olds for testing. There were 30 trials in the testing phase (10 for
each of the three classes). The experiment procedure (see Fig. 6(b))
was similar to the training phase except that online feedbacks were
provided in the testing phase. At t = 2 s, the classication result
appeared and stayed on the screen for 1s. Note that, to ensure
higher sensitivity of the system and less fatigue of subjects, we use
trimmed thresholds (see details below) in the testing phase.

2.5. Online classication algorithm

Fig. 6. The experiment procedure in (a) training phase and (b) testing phase.
experimental procedure was detailed in Fig. 6(a). After the ofine
2.5.1. Processing
A 1-s time window (1 s2 s in Fig. 6.) was used to extract the
data. Because the raw signals contained the low frequency noise
like baseline drift and high frequency noise like the power line
interference, the data were smoothed using a Savitzky-Golay l-
ter to get the trend clearly. Two parameters of the lter (i.e., order
102 X. Guo et al. / Biomedical Signal Processing and Control 30 (2016) 98105

Fig. 7. EOG signals for the three types of eye movements before and after ltering.

and point) determined the cutoff frequency. To balance the calcu-

lation time and denoising effect, we chose the parameters of order
7 and 31 points. The resultant cutoff frequency was 33.7 Hz.
Fig. 7 shows the signals before and after ltering. The insets in
the left column are raw signals of eyes upward, downward and
blink from top to bottom, and the insets in the right column are
the corresponding signals after ltering. The signal of blink has a
peak and a trough, but the signals in the case of eye upward and
downward only has a peak or a trough. Moreover, different types of
eye movements have different speed and time duration. The char-
acteristic waveform shapes indicate that these eye movements can
be discriminated by extracting appropriate features from EOG.

2.5.2. Feature extraction

The goal of feature extraction is to nd a transformation that
converts the raw signals into a low-dimensional feature space.
Feature extraction should be able to preserve the discriminative
information in the raw signals. From Fig. 7, we have learned that
different types of eye movements have different amplitudes of pos-
itive and negative peaks. In order to reduce the time and complexity
of calculation, only the maximum peak and minimum trough values
were extracted as feature vectors for identifying the eye move-
ments. Fig. 8. Distribution of characteristic feature vectors of a subject. (For interpretation
To illustrate the difference between the different eye move- of the references to colour in the text, the reader is referred to the web version of
ments more obviously, the maximum peak (the X-axis) and the this article.)
minimum trough (the Y-axis) were displayed as a feature point in a
two-dimensional plane. Fig. 8 shows the distribution of these fea-
ture points when a subjects eyeballs move left (red asterisk), right class as the boundary. A large threshold can be determined by the
(red plus), up (black circle), down (black asterisk), blink (black plus) most lateral feature points of the nearby class as the boundary.
and stay at the center (red circle) ve times respectively. It can be A trimmed threshold between those two threshold ranges can be
clearly seen that different types of eye movements can be projected obtained. The calculation process is as follows:
into different areas. Through the appropriate thresholds of peak and
trough (the dotted lines), the system can identify these actions. 1 Extract the feature vectors of each training sample: maximum
peak (Vp) and minimum trough (Vt).
2.5.3. Determine the threshold and classication 2 Locate the border of each class: C[Vp+ Vp Vt+ Vt]. C = (center,
The threshold is determined by the distance of adjacent two fea- upward, downward, leftward, rightward, blink). max peak
ture points in two different classes. Therefore, each eye movement (Vp+), min peak (Vp-), max trough (Vt+) and min trough (Vt).
can be identied by three thresholds. For each class, a small thresh- 3 Calculate the classication threshold: C[Thp Thp+ Thv Thv+].
old can be determined by the most lateral feature points within this 4 Choose the threshold below according to the actual application:
 X. Guo et al. / Biomedical Signal Processing and Control 30 (2016) 98105
Fig. 9. Threshold values for eyes move up.
For the small threshold:
Thp = CVp.hp + = CVp + .Tht = CVt.Tht+ = CVt + .
For the large threshold:
Thp = NVp + .Thp + = NVp.Tht = NVt + .Tht+ = NVt.
For the trimmed threshold:
Thp = 0.5 (CVp + NVp+ ).Thp+ = 0.5 (CVp+ + NVp).Tht
= 0.5 (CVt + NVt+ ).Tht+ = 0.5 (CVt+ + NVt)
Here C is the class used for calculating the threshold, and N is the
nearest class.
Classify the testing sample X (Vp, Vt) according to the threshold:
If CThp < Vp < CThp+ and CTht < Vt < CTht+, sample X belongs
to the class C. Center, upward, downward, leftward, rightward, and
blink are tested in order.
If Vp and Vt do not satisfy the two conditions above, X does not
belong to any classes.
Fig. 9 shows an example of three thresholds for eyeballs when
they move up. Five black circles represent the feature points of eyes
upward, and the red plus and black plus correspond to its two near-
est classes: eyeballs move right and down. From inside to outside,
three dotted lines represent the small, trimmed and large threshold
respectively.
Obviously, different thresholds will inuence the sensitivity and
accuracy of the system. When the threshold increases, the false
negatives will be reduced, but the false positives may be increased.
At this moment, the sensitivity increases and the accuracy declines.
On the contrary, the sensitivity declines and the accuracy increases.
Therefore, according to different applications and practical envi-
ronments, the system requires different settings of the threshold.
For example, if we want the system to be sensitive to the users
command, a large threshold is more appropriate. A small threshold
is more preferable towards higher accuracy. If there is no special
requirement, the trimmed threshold is better because it can balance
the sensitivity and accuracy of the system.
In this study, subjects were asked to read a document on the
computer screen. The system should have sensitive and correct
responses to subjects actions simultaneously in order to avoid
fatigue. Therefore, the system employed the trimmed threshold.
103
Fig. 10. The micrographs of soft electrode (a) before the test and (b) after the test.
Because of the location of the sensor (see Fig. 1), the vertical eye
movements (up, down and blink) are easier to discriminate than
the horizontal movements (left, right and no move). In order to
guarantee the accuracy and feasibility of practical application, we
only used the three types of vertical eye movements for HMI.
2.6. Performance evaluation
The performance of the system is evaluated using the accuracy,
sensitivity, specicity and ITR, which are calculated as follows:
Acc = (TP + TN) / (TP + TN + FP + FN) (3)
Sen = TP/ (TP + FN) (4)
Spe = TN/ (TN + FP) (5)
 1 Acc

ITR = 60 log2 N + Acclog2 Acc + (1 Acc) log2 )]/T (6)
N1
where Acc is the accuracy, Sen is the sensitivity, Spe is the speci-
city, TP is the number of true positives, TN is the number of true
negatives, FP is the number of false positives, FN is the number of
false negatives, T is the average time for a command, and N is the
number of types of commands. In the proposed system, T = 4 and
N = 3.
3. Results and discussion
3.1. The performance of sensor
3.1.1. Mechanical property
The weight of each electrode is 60 mg and the thickness is 70um.
The electrodes are stretchable and exible. The electrode can be
stretched to 10% longer than its original length. The micrographs in
Fig. 10 shows the structure of gold strips in details before and after
the test. The gold strips keep intact after being stretched or folded,
although minor wrinkles occur after the test.
3.1.2. Contact impedance
Contact impedance is presented as the mean standard devia-
tion in Fig. 11, where the red and blue lines represent the contact
impedance of wet electrode and exible electrode at different fre-
quencies respectively. At the frequency from 0.3 to 100 Hz, the
104 X. Guo et al. / Biomedical Signal Processing and Control 30 (2016) 98105

Fig. 11. Comparison of impedance spectroscopy on skin for exible and wet elec-
trodes.

Table 1
Accuracy/sensitivity/specicity (%) and ITR (bits/min) for each subject and class in
the testing experiment.

Subject Upward Downward Blink ITR

1 90/100/85 90/70/100 90/100/85 15.2

2 56.7/40/65 60/60/60 56.7/70/50 2.7
3 83.3/60/95 83.3/90/80 83.3/100/75 11.5
4 96.7/90/100 96.7/100/95 96.7/100/95 20.1
5 90/80/95 93.3/90/95 90/100/95 15.9
6 96.7/100/95 100/100/100 93.3/90/95 20.1
7 83.3/100/75 83.3/100/75 83.3/50/100 11.5
8 73.3/90/65 73.3/70/75 73.3/60/80 7.2
Mean 83.8/82.5/84.4 85/85/85 83.3/83.8/83.1 13

average contact impedance of exible electrode and commercial

wet electrode are 359.2 159.5 K and 344.1 127.2 K respec- Fig. 12. Comparison of biopotential signals of a subject recorded with wet electrodes
and exible electrodes (a) ECG (b) EOG.
tively. The results suggest the exible electrodes have similar
impedance to commercial wet electrodes on the skin.

city are 84.0%, 83.8% and 84.2% respectively. The average ITR is
3.1.3. Biopotential recording ability
13bits/min.
The example ECG and EOG signals recorded from one subject
One-way analysis of variance (ANOVA) (a = 0.05) was used to
are shown in Fig. 12(a) and (b) respectively.
evaluate the classication performance and compare the clas-
The results of eight subjects show high correlations of ECG from
sication results among the three eye movement classes. No
the two types of electrodes (correlation coefcient: 0.957 0.005).
signicant differences in accuracy (F(2,21) = 0.03, P = 0.97), sensitiv-
The SNRs of wet electrodes and exible electrodes are 30.6 0.2 db
ity (F(2,21) = 0.03, P = 0.97) and specicity (F(2,21) = 0.03, P = 0.97)
and 24.0 0.4 db respectively. The correlation coefcient of EOG
are observed between classes. This result indicates that the system
from two kinds of electrodes is 0.944 0.031. The SNR of wet elec-
can treat the three classes of eye movements equally.
trodes and exible electrodes are 46.1 3.8 db and 30.5 6.2 db,
Signicant differences are found between subjects for accuracy
respectively.
(F(7,16) = 224.46, P < 0.05) and specicity (F(7,16) = 7.30, P < 0.05),
Because the proposed exible dry electrode does not use con-
but no signicant differences are observed in sensitivity (F
ductive gel, the contact impedance of the dry electrode (359.2 K)
(7,16) = 1.95, P = 0.13). This result suggests that there is individual
is a little higher than the commercial wet electrode (344.1 K).
difference in classication performance obviously. However, since
So the dry electrodes have more electrical noise and relative low
the thresholds for each subject are individually optimized with
SNR than wet electrodes. But as long as using a lter, the electrical
the training data, the system reduces the individual difference of
noise can be removed and it does not affect the performance of the
sensitivity effectively.
HMI system. In terms of preparation time and comfortableness, the
To further illustrate the importance of optimizing the thresholds
exible electrode is much easier to use than the wet electrode.
for each subject, we use the thresholds from Subject 1 to clas-
sify the data for Subjects 28. On average, the accuracy, sensitivity
3.2. Experiment result and specicity are only 14.6%, 11.4% and 16.2% respectively. Fig. 13
shows the accuracy, sensitivity and specicity using subject specic
Table 1 shows the system performance in the testing experiment thresholds and the same thresholds from one subject. The accuracy,
for all subjects. On average, the accuracy, sensitivity and speci- sensitivity and specicity are signicantly improved with individ-
 X. Guo et al. / Biomedical Signal Processing and Control 30 (2016) 98105
Fig. 13. The classication results using different thresholds and the same threshold.
Table 2
Performances of different systems with EOG signals.
System 1[18] 2[19] 3[20] 4[10] 5[14] This paper
#channels 5 4 3 2 1 1
#classes 9 3 2 6 1 3
Accuracy 82100 92 95 85 7284 84
ual thresholds. This result shows that the training phase is essential
for obtaining high classication performance in the testing phase.
Table 2 shows the comparison of the proposed system to other
established systems in the literature. The proposed system only
needs one channel EOG to identify three classes of eye movements
and can achieve accuracy and ITR at an acceptable level. The pro-
posed system achieves similar classication accuracy comparable
to the other systems. The usability of the system is much higher
than other systems since the exible patchable sensor and single
channel EOG control model can enhance the comfortableness and
wearability of the system. Besides, the simple and effective clas-
sication algorithm reduces the time, complexity and memory for
computation in the system.
In order to analyze the performance of the system quantita-
tively, this study adopts a synchronous paradigm with visual cues.
In actual applications, an asynchronous paradigm is more practi-
cal for self-paced control. This problem can be solved by a sliding
window approach [21].
4. Conclusion
This study demonstrates the prototype of a novel human
machine interface that ts for a wearable system. Signal acquisition,
classication algorithms, and controlling methods are all optimized
to develop a biopotential-based HMI, which is practical for daily
applications. EOG, which is robust (compared with EEG) biopo-
tential signal, is used as the controlling source. A light and thin
patchable sensor, which integrates three exible and stretchable
electrodes, is specially designed and fabricated for EOG record-
ing. In an online experiment with three types of eye movements
(upward, downward, and blink), the present system achieved high
classication accuracy (84%) and ITR (13bits/min) using only one-
channel EOG.
Acknowledgements
The authors acknowledge the support from the National Natural
Science Foundation of China (61335010, 61275200, 61275145 and
61036002), National 863 plans projects (Grant No.2012AA032204),
105
the Key project of Chinese Academy of Science (KJZD-EW-L11-01)
and the Recruitment Program of Young Professionals.
References
[1] Subhas Chandra Mukhopadhyay, Wearable sensors for human activity
monitoring: a review,, IEEE Sens. J. 15 (3) (2015) 13211330.
[2] A.S. Royer, A.J. Doud, M.L. Rose, B. He, EEG control of a virtual helicopter in
3-dimensional space using intelligent control strategies, IEEE Trans. Neural
Syst. Rehabil. Eng. 18 (6) (2010) 581589.
[3] T. Liu, L. Goldberg, S. Gao, B. Hong, An online brain-computer interface using
non-ashing visual evoked potentials, J. Neural Eng. 7 (3) (2010) 036003.
[4] E. Zheng, Long Wang, Kunlin Wei, Qining Wang, A noncontact capacitive
sensing system for recognizing locomotion modes of transtibial amputees,
IEEE Trans. Biomed. Eng. 61 (12) (2014) 29112920.
[5] L.Y. Deng, C.-L. Hsu, T.-C. Lin, J.-S. Tuan, S.-M. Chang, EOG-based
human-computer interface system development, Expert Syst. Appl. 37 (4)
(2010) 33373343.
[6] L.-D. Liao, K. Lin, A.E. Wickenden, K. Gramann, T.-P. Jung, L.-W. Ko, J.-Y. Chang,
Biosensor technologies for augmented braincomputer interfaces in the next
decades, Proc. IEEE 100 (2012) 15531566.
[7] E. Forvi, M. Bedoni, R. Carabalona, M. Soncini, P. Mazzoleni, F. Rizzo, C.
OMahony, C. Morasso, D.G. Cassar, F. Gramatica, Preliminary technological
assessment of microneedles-based dry electrodes for biopotential monitoring
in clinical examinations, Sens. Actuator A: Phys. 180 (2012) 177186.
[8] R. Matthews, N.J. McDonald, P. Hervieux, The invisible electrode-Zero prep
time, ultra low capacitive sensing, in: 11th Int. Conf. Human Computer Int.,
Las Vegas, 2005, pp. 2227.
[9] W.-H. Yeo, Y.-S. Kim, J. Lee, A. Ameen, L. Shi, M. Li, S. Wang, R. Ma, S.H. Jin, Z.
Kang, Y. Huang, J.A. Rogers, Multifunctional epidermal electronics printed
directly onto the skin, Adv. Mater. 25 (20) (2013) 27732778.
[10] A.N. Belkacem, D. Shin, H. Kambara, N. Yoshimura, Y. Koike, Online
classication algorithm for eye-movement-based communication systems
using two temporal EEG sensors, Biomed. Signal Process. Control 16 (2015)
4047.
[11] M. Abo-Zahhad, S.M. Ahmed, S.N. Abbas, A novel biometric approach for
human identication and verication using eye blinking signal, IEEE Signal
Process. Lett. 22 (7) (2015) 876880.
[12] J. Ma, Y. Zhang, A. Cichocki, F. Matsuno, A novel EOG/EEG hybrid
human-machine interface adopting eye movements and ERPS: application on
robot control, IEEE Trans. Biomed. Eng. 62 (4) (2015) 876889.
[13] S.S.S. Teja, S.S. Embrandiri, N. Chandrachoodan, R.R.M, EOG based Virtual
Keyboard, in 41st Annual Northeast Biomedical Engineering Conference
(NEBEC), Troy, NY, 2015, pp. 12.
[14] A.M.S. Ang, Z.G. Zhang, Y.S. Hung, J.N.F. Mak, A user-friendly wearable
single-channel EOG-based human-computer interface for cursor control, in:
7th Annual International IEEE EMBS Conference on Neural Engineering,
Montpellier, France, 2015, pp. 865868.
[15] B. Ning, M.-j. Li, T. Liu, H.-m. Shen, L. Hu, X. Fu, Human brain control of electric
wheelchair with eye-Blink electrooculogram signal, Intell. Robot. Appl. 7506
(2012) 579588.
[16] H.M. Shen, L. Hu, K.M. Lee, X. Fu, Multi-motion robots control based on
bioelectric signals from single-channel dry electrode, Proc. Inst. Mech. Eng. H
229 (2) (2015) 124236.
[17] S. Venkataramanan, P. Prabhat, S.R. Choudhury, H.B. Nemade, Biomedical
instrumentation based on EOG signal processing and application to a hospital
alarm system, In Proceedings of IEEE ICISIP (2005) 535540.
[18] S.L. Wu, L.D. Liao, S.W. Lu, Controlling a human-computer interface system
with a novel classication method that uses electrooculography signals, IEEE
Trans. Biomed. Eng. 60 (8) (2013) 21332141.
[19] Y. Nam, B. Koo, A. Cichocki, S. Choi, GOM-Face: GKP, EOG, and EMG-based
multimodal interface with application to humanoid robot contro, IEEE Trans.
Biomed. Eng. 61 (2) (2014) 453462.
[20] Y. Punsawad, Y. Wongsawat, M. Parnichkun, Hybrid EEG-EOG brain-computer
interface system for practical machine control, In Proc. IEEE Conf. EMBS
(2010) 13601363.
[21] L.D. Lledo, A. beda, E. Inez, J.M. Azorn, Internet browsing application based
on electrooculography for disabled people, Expert Syst. Appl. 40 (2013)
26402648.