B3184 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019)

JES FOCUS ISSUE ON 4D MATERIALS AND SYSTEMS

Graphene Smart Textile-Based Wearable Eye Movement Sensor
for Electro-Ocular Control and Interaction with Objects
1 1,2,3,z
Ata Jedari Golparvar and Murat Kaya Yapici
1 Faculty of Engineering and Natural Sciences, Sabanci University, TR 34956 Istanbul, Turkey
2 Sabanci University SUNUM Nanotechnology Research Center, TR 34956 Istanbul, Turkey
3 Department of Electrical Engineering, University of Washington, Seattle, Washington 98195, USA

Wearable graphene textile embedded smart headband and its feasibility in electrooculography (EOG) applications is demonstrated
by benchmarking against clinical Ag/AgCl wet electrodes; where the recorded biopotentials displayed excellent correlation of 91.3%
over durations up to hundred seconds. Automatic eye movement (EM) detection is implemented and performance of the graphene-
embedded “all-textile” eye movement sensor and its application as a control element toward human-computer interaction (HCI)
and human-machine interfaces (HMI) is experimentally demonstrated by: 1) generating digital clock transitions directly from eye
blinks for facilitating switching requirements in HCI/HMI applications, 2) controlling and sequentially lighting up a single LED in
a 5 × 5 LED array in four directions to draw a pattern of “8”, 3) evaluating the limits of the entire system in an hour-long EOG
recording session which includes several activities like checking a phone, watching a video, reading, and performing several EMs
including blinks, saccades, and fixations. The excellent success rate ranging from 85% up to 100% for eleven different EM patterns
demonstrates the applicability of the proposed algorithm in wearable EOG-based sensing and HCI/HMI applications with graphene
textiles. The system-level integration and the holistic design approach presented herein which starts from fundamental materials level
up to the architecture and algorithm stage is highlighted and will be instrumental to advance the state-of-the-art in wearable electronic
devices.
© The Author(s) 2019. Published by ECS. This is an open access article distributed under the terms of the Creative Commons
Attribution 4.0 License (CC BY, http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse of the work in any
medium, provided the original work is properly cited. [DOI: 10.1149/2.0241907jes]

Manuscript submitted December 28, 2018; revised manuscript received March 29, 2019. Published April 30, 2019. This paper is
part of the JES Focus Issue on 4D Materials and Systems.

Eye stores a tremendous source of potential for the rise of new ap- of the sensing electrode. Typically, acquisition units for electrophysi-
plications in human-computer/machine interfaces (HCI/HMI).1 How- ological responses rely on the direct contact of disposable, pre-gelled,
ever, economic challenges and long-term performance of the earlier “wet” silver/silver chloride (Ag/AgCl) electrodes fixed on the subject’s
designs obligates successful realization of casual, consumer-driven, skin with adhesive backings. Although standard Ag/AgCl electrodes
and wearable products. Therefore, afford has been placed to fully in- are low-cost, widely available, and provide accurate signal acquisition
vestigate different methods to take possession of eye movements (EM) capabilities,18 the need for skin preparation severely limits their usabil-
in HCI/HMI interfaces. For instance, coil-based eye tracking systems ity in wearable electronic applications.19 For instance, the conductive
are invasive and are not meeting the non-clinical application needs.2 gel dehydrates in time and degrades the electrode performance, thus,
On the other hand, camera-based eye tracking setups fulfill the inva- once in a few hours, electrodes must be changed or the gel must be re-
sivity issue and display long-term functionality but they are hardly applied, which is inefficient and time-consuming and not acceptable
affordable due to their hardware (e.g. camera) and image processing for everyday and easy to use applications. Moreover, the gel can cause
requirements. Additionally, the camera has to be positioned at a loca- an itching sensation;20 as well as, red and swollen skin which devel-
tion suitable to capture the eye movements, which limits portability ops immediately upon removal by mechanical peeling of the electrode.
of such systems. Alternatively, electrooculography (EOG) is an eco- Such irritations and allergic reactions may only last for several hours21
nomical, non-invasive, and reliable method for acquiring biopotential or may even lead to dermatitis.22 Due to the above concerns, studies
signals around the eyes, and addresses the limitations of both coil and have been proposing the elimination of the gel by developing “dry”
camera-based systems.3 EOG is essentially based on the simple model electrodes, which are more suitable for continuous, autonomous and
of the human eye, which is a dipole where the potential between its unsupervised electrophysiological monitoring, and meet the desired
forward and backward facing spots generate an electric field.4 These comfort level for integration with wearable devices.23
voltage fluctuations can be detected if a pair of electrodes is attached One promising approach that emerged in recent years is based
around the eyes, and by feeding the acquired ocular biopotential signal on the use of conductive, smart textiles. Owing to their inherent ad-
into a proper processing unit, EMs can be tracked. vantages like smooth texture and the ability to be directly weaved
So far, several EOG-based rehabilitation systems were developed into garments, several methods have been suggested to develop textile
as an assistive technology for people with lock-in syndromes.5,6 These electrodes for electrophysiological signal monitoring. The main chal-
human-computer or human-machine interfaces are able to emulate lenge here is to synthesize conductive textiles from ordinary fabrics
a PC mouse,7 run a virtual keyboard,8 drive a wheelchair,9 con- like cotton, nylon, and polyester. To address this issue, different meth-
trol robots,10 and improve user experience on gaming11 and smart- ods have been investigated which include functionalization of ordinary
phone operations.12 Additionally, visual fatigue estimation using EOG fabrics with conductive inks or pastes via screen printing,24 nanowire-
was proposed to be used in 2D/3D display auto-adjustment switch coated threads,25 electroplating,26 and embroidering conductive mate-
systems.13 Moreover, EOG signals are also used in cognitive stud- rials into fabrics.27 However, earlier methods either require dedicated
ies and neurosciences including human activity recognition,14 sleep equipment or fabrication processes that are complex, expensive, and
studies and drowsiness detection,15,16 and monitoring and diagnosis incompatible for large-scale production, lack uniformity or sacrifice
of mental and sleep disorders.17 from the natural comfort of the fabric. In order to use textiles as biopo-
Despite the various demonstrators of wearable EOG devices in the tential electrodes, they need to be flexible, durable, comfortable, and,
literature, their full potential has not been realized due to limitations biocompatible and have suitable electrical characteristics for signal
acquisition. Several advantages of graphene—a single layer of carbon
atoms arranged in a hexagonal lattice, having excellent electrical con-
z
E-mail: [email protected] ductivity and elasticity combined with high ultimate strength while

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019) B3185

Read-out circuitry Microcontroller Laptop

Interface & Real-time

Signal Conditioning A/D Converter Monitoring
3
2.5

Amplitude [V]
2
Digital Signal 1.5
UART
Processing 1
0.5
0
0 5 10 15 20 25 30 35 40 45 50
lead I reference lead II Time [s]

(right) (left) Feature Extraction

5x5 LED Matrix

Classification row 1 row 1

row 2 row 2

Control Signal column 1 column 2

Generator

Figure 1. Block diagram of the proposed EOG-based HCI interface consisting of front-end read-out circuitry and graphene textile embedded smart headband for
signal acquisition, microcontroller unit for signal processing, and display for real-time monitoring and visualization.

being extremely lightweight,28 leads to the direct application of it in measured. If identified features are in match with hardcoded thresh-
electronic textiles (e-textiles) or smart garments. Owing to these fea- olds (set beforehand during calibration sessions) the signal is classified
tures, the merger of graphene on a variety of textiles was recently as a unique pattern. Subsequently, according to the detected patterns,
demonstrated based on a low-cost, gel-free, washable, and scalable control signals are generated. Many application scenarios can be ad-
approach29 and the feasibility of the developed graphene textile elec- ministered by these commands such as controlling a simple virtual
trodes were experimentally verified and benchmarked successfully keyboard, eye mouse, wheelchair, or by mimicking remote control
against Ag/AgCl in multiple testing scenarios.30 protocols, home appliances like TV can be operated by EMs. Here,
This paper provides a systematic analysis on the possible locations for demonstrating the robustness of the system an LED is turned on
on the forehead to record oculo biopotentials and reports the integra- sequentially in a 5 × 5 matrix to trace a pattern of “8”.
tion of textile electrodes into an ordinary elastic sports headband with Finally, both the digitized data and detected EMs are streamed
embedded electronics. The capability of the graphene textile integrated continuously to a general purpose computer through USB port with
headband in EOG monitoring and automatic recognition of multiple UART baud rate of 128000 b/s where they are displayed on a costume-
eye movement patterns are demonstrated along with complete discus- designed graphical user interface (GUI). The GUI was developed using
sion on the detection algorithm, system architecture and system-level visual basic for application (VBA), and provides an easily configurable
integration. Experimental results obtained in several functionality test- platform to adjust system settings, a trouble-free calibration session,
ing scenarios demonstrate the excellent performance and potential and includes features like beep sounds to alert users, data storage, and
of graphene e-textiles toward the development of advanced human- real-time monitoring capabilities.
computer interaction (HCI) interfaces as well as medical studies.
The present work extends our pioneering efforts on wearable
graphene textiles toward object control and interaction. We show the Synthesis of graphene textiles and integration with clothing.—
functionality of electro-conductive wearable graphene textiles by suc- Conductive textiles were synthesized based on a low-cost and scalable,
cessful acquisition of EOG biopotentials that arise due to various eye three-step, dip-dry-reduce coating approach where graphene clads
movements and report an automatic detection algorithm to differen- around a variety of ordinary fabrics (e.g. nylon, cotton, or polyester)
tiate between different eye movements. The eye movements are then to form a conformal layer.29 The process consists of preparing of
used to sequentially light-up an LED in a 5 × 5 array, as a proof-of- graphene oxide (GO) suspension based on the modified Hummer’s
concept demonstrator toward applications in human-computer inter- method, followed by dipping of plain textiles into GO solution. Next,
action (HCI) and human-machine interfaces (HMI). the wetted textile was left to dry at moderate temperatures (∼80°C)
which allowed layering of GO around individual textile fibers. The
GO-coated textile was then chemically treated with reducing agents
Experimental
like hydrazine or hydrogen iodide and rinsed in deionized water to
System-level architecture.—Fig. 1 illustrates the system level form stable, conductive, graphene cladding on textiles.
block diagram of the proposed HCI interface. In the front-end read- In order to detect electrooculograms from different spots on the
out circuitry, heavily noise contaminated surface biopotentials are re- forehead, the prepared graphene textile piece was cut into desired di-
ceived with the fabricated graphene textile electrodes and fed into mensions (∼3 × 3 cm) and mounted on an elastic headband with
several filters and amplification stages. Specifications of the signal flexible sticky foams which were sandwiched between metallic snap
conditioning circuitry include very sharp roll-off band-pass filter in fasteners in order to establish electrical connection with the front-end
the frequency range of 0.3 to 10 Hz, adjustable gain from 600 to circuitry (Fig. 2).31 To acquire ocular biopotentials, electrodes should
4600 V/V, and variable offset of ±5 V.30 be positioned on the skin surface and have stable contact. In com-
Upon denoising, the signal is digitized by the built-in ADC of an mercial electrodes this is achieved by gels and adhesives; which, on
8-bit microcontroller unit (atmega328, ATMEL, CA, USA) with a the other hand limit their use in wearable applications. Alternatively,
sampling frequency of 100 Hz. In order to further normalize the sig- we have used elastic bands with Velcro straps and foam paddings
nal, a digital rolling average filter is applied. Later on, the normalized (polyethleyene-based) to provide pressures in the range of few mmHg
signal goes through a feature detection where its distinctive attributes (up to 5 mmHg),32 which supported the contact of graphene textile
like rise/fall times, amplitude, and local maxima/minima points are electrodes on skin surface and ensured interface stability.

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 B3186 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019)

Figure 2. EOG headband with graphene textile electrodes for HCI/HMI appli-
cations; insets show flexible graphene textiles after synthesis (bottom right) and
stand-alone version of a pair of graphene textile electrodes with foam padding
and snap fasteners prior to headband integration (top left).

Since the amplitude of the EOG signal is sensitive to relatively

small variations in electrode positioning,4 depending on the appli-
cation, different electrode counts and locations were investigated.33
Commonly in clinical monitoring, a signal acquisition unit with two
channels, one for horizontal EOG (hEOG) and the other for verti-
cal EOG (vEOG), is used to record raw biopotential signal referred
to as the electrooculograms. In this configuration, five electrodes are
used, where one electrode is placed at the outer canthus of the left and
right eye for detecting lateral EMs; whereas the remaining two are at-
tached above and below an eye for picking up transverse eye activity
and the last electrode is placed centrally on the forehead as a refer-
ence. Probably, the most comfortable approach on the user-end for
wearable devices is to achieve electrooculogram detection only from
locations on the forehead such that electrodes can be easily integrated Figure 3. Systematic analysis of electrode positioning in forehead EOG. The
into an elastic headband. Mostly, forehead EOG uses two-channel fabricated electrodes were cut into ∼3 × 3 cm dimension to test different place-
configuration by having four electrodes where one of the electrodes is ment configurations. Waveforms show the induced electrooculograms from (a)
shared between channels and detects 4 different saccadic movement locations 4, 6, 8; (b) locations 5, 6, 7; (c) locations 1, 2, 3; (d) locations 5, 6, 8;
patterns (left, right, up, and down EMs) to execute various control (e) locations 4, 6, 7; where the first, second, and third digit corresponds to the
commands.34,35 Here, we propose a new electrode positioning con- location of the left, reference, and the right lead, respectively. The performed
figuration to detect the same number of differing EM patterns, hence eye movements were: (I) voluntary blink, (II) slow left, (III) slow right, (IV)
swift left, and (V) swift right.
control commands, with only three electrodes and one channel. In this
electrode placement, three electrodes are to be fixed on the forehead
where two of the electrodes were placed roughly above the left and
right eye toward the temples (Fig. 3, locations 4 and 8), and a refer- the same location of the forehead, measurements were taken at dif-
ence electrode was placed halfway in between (Fig. 3, location 6). This ferent times but from the same location with two different types of
configuration is chosen through an experiment where a volunteer was electrodes.
asked to perform different blink, saccadic and fixation EMs. Induced The participant was a 23 years-old healthy male with no known
waveforms are included in Fig. 3. First configuration (I) is selected vision problems. The participant was instructed to sit and face a pre-
to be most appropriate one to be automatically identified with thresh- determined central point around which two other side points were
olding algorithms since amplitudes and patterns of EM differ a lot in located such that he had to make ∼ 300 saccadic gaze movements
comparison with other configurations and they are stronger in terms when looking at them (Fig. 4a). Prior to the experiment, the participant
of magnitude. was trained on the experimental protocol (Fig. 4b) which consisted of
several horizontal saccadic movements, fixations (i.e. the duration in
Performance evaluation.—In order to evaluate the performance which the participant held his stare at a single point), and volunteer
of the fabricated smart headband in sensing electrooculograms, it was blinks. The first stage begins with a voluntary blink while maintaining
benchmarked against the clinical standard, pre-gelled Ag/AgCl elec- eyes in the primary position, then a levoversion (left gaze) is performed
trodes (Ref 2228, 3M Red Dot, USA). As demonstrated, in EOG mea- where both eyes are moved to the left and fixed for 10 s, after which
surements, the location of the electrodes on the forehead has a direct eyes are brought back to the primary position. Continuing with the sec-
effect on the characteristics of the induced signal patterns, thus, in ond stage, a dextroversion (right gaze) is performed where eyes are
performance evaluation experiments locations of both embedded tex- now moved from the primary position to the right and fixed for 10 s, and
tile electrodes and wet electrodes should be kept the same. Since it is subsequently returned to the primary position. The third stage begins
physically not possible to place two different types of electrodes on with a blink followed by left and right movements (levoversion and

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019) B3187

Figure 4. (a) Schematic diagram showing the position of the eyeballs with respect to pre-determined points located in the center (X0 ), left (X1 ) and right (X2 ). (b)
A tabular summary of the sequence of EMs in the three-stage testing protocol. EOG signal acquired from (c) the developed smart headband showing the unique
EM patterns, and (d) Ag/AgCl electrodes.

dextroversion) without waiting on the sides (i.e. no fixations), and the embedded headband and Ag/AgCl electrodes demonstrate the func-
protocol is concluded by a blink at the primary position. Throughout tionality of the developed smart garment in EOG applications. Slight
this paper, saccadic moves with a fixation duration on sides are be- differences in amplitude are observed between the signals recorded
ing called “slow movements” and moves with no fixation on the sides with graphene textile and Ag/AgCl electrodes due to mismatches in
are called “swift movements”. To clearly distinguish the different EM measurement conditions and electrode characteristics. Additionally, as
patterns in the recorded electrooculogram, the duration between sub- graphene textile electrodes are manually-sized, this causes inevitable
sequent movements was maintained at approximately 10 s by timing size variations from textile to textile, and also between textiles to
the participant and alerting them with a beep sound for each move- Ag/AgCl electrodes, which can potentially affect the signal quality.
ment. In the occurrence of a spontaneous blink, the experiment was With automated handling, closer size match between electrodes could
not interrupted. Therefore, the recorded electrooculograms inevitably be achieved. To better assess the effect of electrode size on the recorded
contain waveforms due to involuntary blinks which are identified in electrooculograms, we have fabricated and tested electrodes of vary-
Figures 4c and 4d. The participant was also instructed not to talk, ing contact areas of 1 cm2 , 4 cm2 , and 9 cm2 . Fig. 5 shows the acquired
avoid any body movement, and stare only at pre-determined points as waveforms from different electrode sizes as a result of saccadic eye
defined in the protocol. movements and blinks. It has been suggested that textile electrodes
Signals from the forehead were first collected using the headband with larger contact areas could achieve better signal quality (i.e. less
without any prior skin preparation. After mounting, 5-minute wait pe- noise contamination) due to the smaller skin-electrode impedance.37 In
riod was allocated to allow electrode stabilization.36 Approximately our measurements this phenomena was not apparent due to sharp on-
10 minutes after completion of the first measurement, textile elec- board filtering, and minimal to virtually no respiration-related change
trodes were removed and Ag/AgCl electrodes were placed around in contact conditions in EOG, unlike typical ECG applications. More-
the same spots and the second part of the experiment was performed. over, while there are slight differences in the waveforms recorded by
Figs. 4c and 4d display the recorded electrooculogram signals from the different electrode sizes, this is attributed primarily due to mismatches
smart headband and Ag/AgCl electrodes respectively, along with in- in experimental conditions and amplitude variations due to electrode
terpretation of the performed EMs. For both electrode types, the same size were observed to be insignificant, which is also in alignment with
acquisition unit was used and its gain and offset values were kept con- earlier studies on ECG.38
stant for better comparison of the recorded signals (i.e. ∼4400 V/V
and 1.5 V). To quantify the overlap between obtained signals, the Algorithm development.—A quick visual analysis of the recorded
built-in linear correlation function of MATLAB (Mathworks, Nat- EOG waveforms show that there is an exclusive signal pattern for each
ick, MA, USA) was used which revealed a correlation of 91.3% for and every defined EM. These patterns mainly alter in shape, mag-
the entire measurement period of 100 seconds. The high correlation nitude, and duration. By hard-coding, the unique signature of each
or overlap of signals obtained with the wearable graphene textile- EM pattern into the software, automatic detection of EMs can be

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 B3188 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019)

with the hard-coded patterns, 6) classify the signal, and finally,

7) the algorithm should generate control signals according to spe-
cific application requirements (e.g. generating clock pulses or con-
trol comments). Since all of the mentioned tasks are “soft” real-
time and they don’t have critical deadlines, they can be scheduled
by a periodic approach with the microcontroller’s internal timer.
During algorithm development, special emphasis was placed on
avoiding the use of real-time operating system (RTOS), or compli-
cated digital signal processing (DSP) techniques, feature extraction
or classification algorithms to ensure that the developed embedded
“software” can operate on slow processing speeds (e.g. max ∼20 MHz)
and implemented on general purpose, small size, and low-cost micro-
controllers.
Timer interrupt service routine (ISR) is programmed to perform
several tasks which include, triggering of an A/D conversion according
to the desired sampling rate (e.g. 100 Hz), measuring the duration of
potential EMs, running a time window and continuously checking and
controlling the inputs and outputs (I/O).
In order to construct a pattern model, one of the primary tasks
is to regularly track the location of the real-time EOG signal which
may include various EMs (Fig. 7a). Five threshold levels were de-
Figure 5. Plot of EOG signals acquired from three different sizes of graphene
fined and named as “up margin” (UM), “baseline up-margin” (BUM),
textile electrodes; inset shows an image of the fabricated electrode samples.
“baseline”, “baseline down-margin” (BDM) and “down margin” (DM)
(Fig. 7b). These threshold lines along with the duration and peak to
peak amplitude of defined EMs are measured and hardcoded to the
accomplished. To do so, a sequential, multi-step, fixed thresholding
system in advance during the calibration session. In the literature,
algorithm was developed. Fig. 6 illustrates a summarized flowchart
most calibration methods either adjust thresholds at the software level
of the developed algorithm. The algorithm is responsible to im-
and leave hardware level parameters untouched; or, the operator ad-
plement the tasks: 1) maintain synchronize with the GUI, 2) dig-
justs signals at the hardware level according to the software threshold
itize the denoised signal, 3) normalize the data, 4) extract infor-
needs and always leave software parameters constant. Here, we do a
mation and features from the signal, 5) compare the extracted data
mixture of both where the system is calibrated during training sessions
with the addition of an offset to the signals by directing the participant
to hold their gaze at the central point and fixate eyes at the primary
Main position. The baseline value should guarantee the signal to be in the
positive domain below 5 V level; here it is fixed at 1.5 V. Once the
desired offset was ensured, several EMs of each type were performed
so that the gain level could be adjusted accordingly to prevent output
Control the saturation.
display and Is the calibration Yes Go to calibration
check GUI’s flag set in GUI? scenario Meanwhile, at the software level thresholds for UM and DM are
inputs. configured based on several constraints. First, blinks, swift moves, and
right gaze must pass through and intersect the UM but left gaze must
No not. Second, all moves must pass through and intersect DM but right
gaze must not. Here, UM and DM were found as 2.1 V and 1 V, respec-
No tively. Third, BUM and BDM levels with respect to the baseline were
Is ADC flag set by selected according to baseline fluctuation; which through measure-
timer ISR? ments was determined to be ±0.1 V. During experiments there was no
need to re-adjust BUM and BDM, and the artificial baseline set at the
beginning was hardly changed during weeks of experiments. How-
Yes ever, especially in long-term use, variation of signal amplitude due
to environmental, physiological or physical factors such as feeling of
Do the A/D tiredness or change in skin-electrode impedance could be critical and
conversion and
store the data.
require recalibration of gain and offset parameters.
In feature extraction, right after normalizing the signal using a
rolling average filter, which is implemented for minimizing the effect
Rolling of stabilization phenomena of fixation,39 if the signal appears to have
average a large value than UM, the system will label the location of it as “up”,
filter
whereas if it lies in between BUM and BDM the location will be
as designated as “center” and, if the data value is less than DM, the
Control signal system will name the location of the signal as “down”. The location
Go to feature Go to generation Stream the operator will not be changed if the signal is in between UM and BUM,
extraction classification according to the data to PC.
application needs. or DM and BDM, to avoid oscillation of location operator in critical
cases near margins. If the location operator changes, a flag will be
set to alert the algorithm to implement the necessary actions in the
I/O classification section.
Control While the algorithm detects the defined EMs, it must also avoid de-
tection of undefined EMs and response as one of the defined patterns.
Figure 6. Summarized flowchart of the developed algorithm for automatic For instance, spontaneous or reflex blinks (which can have several
detection of volunteer blink along with four different saccadic EMs in single- shapes, durations, or amplitudes depending on the context), or small
channel forehead EOG. degree saccadic eye movements (mainly resembling left/right moves

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019) B3189

Figure 7. (a) EOG trace showing the different types of auto-detected eye movements by the proposed algorithm, zoom-in images of the five exclusive signal
patterns corresponding to: (b) voluntary blink; (c) swift left-right saccadic gaze; (d) swift right-left saccadic gaze; (e) left gaze; and (f) right gaze. The labels “UM”
(up margin), “BUM” (baseline up margin), “BDM” (baseline down margin), and “DM” (down margin) represent the critical threshold levels. The notations (I) to
(IV) stand for amplitudes of blink, swift left-right, left and right movements, respectively; and the labels (1) to (8) correspond to data points between which the
duration is measured.

but with a smaller magnitude may occur during office activities like 1), then returns to center (stage 2), then falls down (stage 3), and fi-
reading or writing), and must be excluded from detection. Addition- nally returns to center (stage 4) with the following of an overshoot.
ally, the main parameter which distinguishes the swift left-right move Since the unique pattern of swift right-left gaze differs it from all other
and different types of blinks from each other is their amplitude levels.40 movements, no other threshold is required for building its model.
Therefore, measurement of the signal amplitude is critical for reliably Left gaze (Fig. 7e) first changes its location from center to down
constructing the pattern model. (stage 1) and then returns to center (stage 2) with following of an over-
In case of a flag alert for a signal location change, the system enters shoot which never reaches the UM level. The algorithm for detecting
the classification section; where the algorithm tracks the signal that left gaze relies on two timer counters, one counts the duration between
occurred to identify its pattern. The volunteer blink complex (Fig. 7b) “3” and “4” which should not pass a specific threshold, and the other
first changes its location from center to down (stage 1), then returns is a countdown timer which gives the system a short duration to check
to the center (stage 2), then rises to up (stage 3), and eventually re- and find if the signal goes to “up” location or not. The same detec-
turns to primary central position (stage 4) with the following of an tion system stands for the right gaze (Fig. 7f), which is essentially
undershoot.41 As soon as the signal enters stage 1 (marked as (1) in the reverse pattern of a left gaze. In the right gaze signal first rises up
Fig. 7b), a counter starts keeping the time and stops when the signal (stage 1) and then returns to center with following of an undershoot
reaches stage 4 (marked as (2) in Fig. 7b). The interval between time which must not intersect DM. Its timer counters control the duration
1 and 2 is measured as the signal duration and it must be lower than a threshold between its stage 1 and 2, and its down counter provides an
set threshold. interval to check if the signal will pass UM or not before detecting the
Swift left-right gaze (Fig. 7c) and the volunteer blink patterns are pattern as a valid EM.
nearly identical in terms of the locations at when a change in signal Then, the algorithm computes the amplitude of the signal and com-
pattern occurs. Therefore, the stage indicator for a swift left-right gaze pares it with its respective threshold value. For calculating the ampli-
moves like the stage variable of a volunteer blink, but with a signif- tude of the pattern, ultimate high hillock and ultimate low valley points
icantly different amplitude. Its amplitude (noted as “II” in Fig. 7c) are found out by continuously comparing the maximum and minimum
must be lower than its threshold and definitely, it is smaller than the data values with each other in a pre-defined time window. If the system
threshold introduced for the blink amplitude (noted as “I” in Fig. 7b). detects specific attribute of the EOG signal as one of the five defined
Swift right-left gaze (Fig. 7d) signal changes its pattern opposite to the EMs, it will initiate a unit pulse with different amplitude for each de-
behavior of a blink, where it first starts by rising to up position (stage tected pattern. Additionally, GUI displays the detected EM’s name,

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 B3190 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019)

the long-term performance and reliability of the developed algorithm, a

multi-segment EOG, which included performing of various eye move-
ments and indoor activities was conducted for a duration exceeding
1 hour. Fig. 10 illustrates the induced electrooculograms along with the
unit pulses generated by the algorithm to identify the detected EMs. In
the first and last segments of this experiment, a similar protocol of the
one introduced in Fig. 4b which the volunteer had mastered earlier,
was carried out. In the second, fourth, fifth, and seventh segments a
single type from one of the defined EMs was performed; namely, they
are voluntary blink, swift left, swift right, slow left, and slow right
movements, respectively. In the third segment, an English text having
32 lines was read from a laptop display. In the sixth segment, the par-
ticipant was reading text on a smartphone, watching short video clips
and messaging. In the eighth segment, the participant was requested
to relax and stare at a single point and, in the tenth segment, the same
task was implemented with eyelids closed. Finally, in between some
of the segments, the participant was asked to watch a neutral video.
Table I summarizes the performed activities including their duration,
the total number of EMs (either voluntary, reflex, or spontaneous), and
the algorithm’s success rate (SR) in EM detection throughout for each
segment and for every EM. While performing this experiment the par-
Figure 8. EOG signals acquired with the smart garment (blue trace), where
ticipant’s eyes were being recorded so that later on, the recorded EMs
the recorded signal includes several voluntary blink patterns which can be
translated into a series of digital pulses (red trace) to effectively implement could be tracked visually to verify the performance of the algorithm
blink-controlled clock transitions in real-time for enabling switching require- in automatic detection of EMs. Additionally, the threshold calibration
ments of HCI devices. Detection of the spontaneous blinks that occurred in the was conducted only once at the beginning of the 1-hour EOG session.
22nd , 33rd , and 69th seconds were successfully avoided. Throughout the 1-hour-long EOG session, the algorithm correctly
detected 70 voluntary blinks out of 72 (SR = 97.2%), 60 left gazes out
of 62 (SR = 96.7%), 50 right gazes out of 61 (SR = 81.9%), 58 swift
amplitude, and duration. Moreover, a buzz sound is generated by the left gazes out of 59 (SR = 98.3%), and 64 swift right gazes out of 64
computer to alert the operator of an EM detection event. (SR = 100%), also it successfully avoided detection of 462 sponta-
neous and reflex blinks out of 507 occurrences (SR = 91.1%).
Data indicates that most of the miss detected EMs were involuntary
Results and Discussion
blinks which were interpreted by the algorithm as slow left moves.
Rather than testing the feasibility of the developed wearable EM The number of spontaneous blinks occurred while watching a video
detector on a single application, we decided to investigate the funda- (∼20 minutes) is more than double of the other segments combined and
mental needs of different applications in multiple testing scenarios. also most of the misdetection of spontaneous blinks happened while
The first experiment involves translating volunteer blinks into a trail watching the video. The higher number of involuntary blinks while
of pulses that can be used to trigger output commands for various watching the video is attributed to reflex blinks. In this context, the
control purposes or implement a switching action in an HCI interface. reflex is not due to an external signal stimulus rather we are referring
The blink command was instructed by the participant according to a to the visual stimulus in the break of senses or luminance change
prescribed protocol which involved blinking at different time inter- during the video.42 The reason for larger number of miss detected
vals; including 2.5, 5, 10, 15 s intervals such that the duty cycle of involuntary blinks could be due to the difference in the overall shape
generated pulses were kept constant at ∼ 50%, and once 15 s interval and magnitude of reflex blinks compared to spontaneous blinks,43
was reached the blink repetitions were sequentially decremented back where in our experiments the reflex blinks were observed to resemble
to 2.5 s (Fig. 8). During 5 minutes of continuous experimentation, a slow left gaze pattern and were misinterpreted by the algorithm.
40 blinks occurred, where the algorithm was able to detect all and This could be addressed simply in calibration session by increasing
achieved a perfect success rate of 100%. the threshold value for the left gaze.
In the second experimental scenario, LEDs were turned on sequen- An interesting observation was made for the recorded EOG sig-
tially in a 5 by 5 LED matrix to trace certain patterns like “S”, “5”, and nals when eyes were closed. Simple visual comparison of the induced
“8” (Fig. 9). The swift EMs cause the “on-LED” (i.e. lit-up LED) to signals in segment 8 (i.e. eyes open and staring at a single point) with
move in horizontal directions while slow horizontal saccades move the those of segment 10 (i.e. eyes closed and at primary position) reveal
on-LED in the vertical direction and voluntary blinks cause flashing that variations and magnitudes of electrooculogram when eyes are
of the on-LED. As illustrated in Fig. 7a, if right after hearing the alert closed are higher than the case where eyes are open. This also sug-
for detection of a slow left or right gaze, the user follows a natural flow gests that the same calibration parameters cannot be used for detection
and makes a reverse gaze to bring eyes back to the primary position, of slow and/or swift saccadic EMs for cases when eyes are open ver-
the induced signal complex will be a combination two gazes. In most sus closed. As for the reading activity, the recorded electrooculograms
cases, this return does not satisfy the conditions to be considered as display 30 low-amplitude saccadic left moves due to focusing on dif-
a separate left or right gaze and therefore will not be detected by the ferent lines of the text while the algorithm successfully avoids their
algorithm. However, after hearing the alert if the gaze is fixated on the detection owing to the fact that left gaze pattern for saccadic moves
sides longer than a natural detection, the algorithm will categorize the was modeled for ∼ 30° displacements, not for smaller changes like
second gaze as a valid EM. For the particular switching application, ∼10° which typically occur while reading.
this could be problematic because the second eye gaze, that is also a In the 1-hour EOG session, a decline in the SR from 97% (first
reverse of the first one, will cause the on-LED to return to its previ- segment) to 87.1% (last segment) is observed where both segments
ous position and effectively cancel the intended move. Thus, for the roughly contain the same number of EMs. The reason for the decline
demonstrated application, the extra measures were implemented at the in SR is the high number of miss detected or completely missed slow
software level to eliminate the second EM in case of its detection. right gazes (nearly half of the slow right moves in the last segment)
Another important consideration for successful development of due to the insufficiency of their amplitude thresholds, indicating that
wearable electronics is their long-term performance which is funda- the system needs recalibration session in long runs, which could be
mentally related to both hardware and algorithm design. In order to test addressed by dynamic thresholding approaches.44

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019) B3191

Figure 9. Plot of the induced EOG signal with inserted interpretations for each movement and their issued direction changes, which are used to control an array
of LED by turning them on sequentially to trace a pattern of “8”.

Conclusions correlation of 91.3%. During the period of conducting the reported

experiments which exceeded one month, no significant performance
In this study, we present a system-level approach and architec-
changes were observed in the graphene textile electrodes. As for
ture wherein we first demonstrate the functionality of graphene-based
the biocompatibility of graphene, since graphene textiles do not re-
e-textile electrodes in electrooculogram recording by benchmarking
quire prior skin preparation and effectively touch only the outer-
them against standard Ag/AgCl electrodes which revealed excellent
most layer of the skin (i.e. stratum corneum, made up of several

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 B3192 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019)

Figure 10. (a) Zoom-in samples from each performed activity (b) 1 hour-long electrooculogram (c) virtual unit pulses generated by the algorithm displaying
different amplitudes according to the detected EMs. 0.1 V and 0.2 V pulses are for slow and swift right EMs, respectively; whereas, pulses with the same amplitude
but with negative sign are indicators of slow and swift left EMs. Pulses with highest amplitude correspond to the detection of voluntary blinks.

Table I. Tabular summary of success rate of the automatic detection of eye moves in different scenarios in 1-hour long EOG.

Activity 1 2 3∗ 4 5 6 7 8 9 10 11 SR
VB C/T 10/10 47/47 - 3/4 - - 3/3 - 1/1 - 6/7 97.2%
MD 0 0 - 1 - - 0 - - - 0 1
L C/T 18/19 - - - - - 25/26 - - - 17/17 96.7%
W 0 1 - - 1 3 3 - 27 1 0 36
R C/T 19/19 - - - - - 22/26 - - - 9/16 81.9%
W 1 2 1 2 - - 4 - 1 - 0 11
MD - - - - - - - - - - 4 4
SL C/T 6/6 - - 45/46 - - - - - - 7/7 98.3%
W 1 - - 1 - - 2 - 2 - 1 7
SR C/T 6/6 - - - 51/51 - - - - - 7/7 100%
W 0 - - 1 0 - 4 - 1 - 3 9
SB MD/T 7/8 7/10 16/17 2/5 2/3 39/42 5/8 46/46 330/361 - 8/8 91.1%
Success Rate 97% 94.7% 94.1% 90.9% 98.1% 92.8% 86.6% 100% 91.4% - 87.1%

C: Correct detection, T: Total number of occurred moves, MD: Missing detection of occurred moves, W: Wrongly detect moves.
Note: The success rate (SR) is defined as the percentage of the correct detection of occurred eye moves.
∗ None of the small degree saccadic moves due to the text’s line change detected.
Definition of eye moves: VB: Voluntary Blink, L: Left move (both center to left side and right side to center movements), R: Right move (both center to
right side and left side to center movements), SL: Swift left-right move without waiting for fixation at left side, SR: Swift right-left move without waiting
for fixation at right side, SB: Spontaneous Blink.
Definition of activity (minute): 1: Mix of defined moves (5.46), 2: Only blink (4.34), 3: Reading (5.02), 4: Only swift left-right move (4.77), 5: Only swift
right-left move (4.62), 6: Phone checking (5.11), 7: Only the left/right move with having fixation on sides (4.94), 8: Staring at a single point (5.09), 9:
Watching video (5.08+5.03+4.69+4.91), 10: Closed eyelids without moving eyes (3.61), 11: Mix of defined moves (4.38).

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).
 Journal of The Electrochemical Society, 166 (9) B3184-B3193 (2019) B3193

tens-of-microns-thick pile of dead cells), potential concerns on toxicity 15. P. Ebrahim, W. Stolzmann, and B. Yang, “Eye movement detection for assessing
are alleviated as dermal administration of graphene has been reported driver drowsiness by electrooculography,” in Proc. IEEE Int. Conf. Sys. Man Cyber.,
Manchester, UK, 2013, pp. 4142.
to display minimal effect on the skin for moderate exposure durations 16. B. Jammes, H. Sharabty, and D. Esteve, “Automatic EOG analysis: A first step toward
and concentrations.45 automatic drowsiness scoring during wake-sleep transitions,” Somnology, 12(3), 227
Additionally, a fully-wearable, smart headband was developed (2008).
and its capability to auto-detect multiple eye movements (EMs) was 17. M. Vidal, J. Turner, A. Bulling, and H. Gellersena, “Wearable eye tracking for mental
health monitoring,” Computer Communications, 35(11), 1306 (2012).
demonstrated by system-level integration of graphene textiles with 18. S. Yao and Y. Zhu, “Nanomaterial-enabled dry electrodes for electrophysiological
read-out electronics and classifier algorithms based on sequential, sensing: A review,” JOM, 68(4), 1145 (2016).
multi-step, fixed thresholding. With the approach provided in this 19. L. Liao, I. Wang, S. Chen, J. Chang, and C. Lin, “Design, fabrication and experimental
study, a novel electrode placement for the forehead EOG was also in- validation of a novel dry-contact sensor for measuring electroencephalography signals
troduced whereby five different EM patterns could be detected only by without skin preparation,” Sensors, 11(6), 5819 (2011).
20. J. Chiou, L. Ko, C. Lin, C. Hong, T. Jung, S. Liang, and J. Jeng, “Using novel MEMS
a single channel read-out circuitry. The results presented in this work EEG sensors in detecting drowsiness application.” in Proc. IEEE Biomed. Circuits
lay down the foundation of graphene textiles toward control applica- Syst. Conf., London, UK, 2006, pp. 33.
tions specifically tailored to EOG-based human-computer interaction 21. A. J. Golparvar and M. K. Yapici, “Wearable graphene textile-enabled EOG sensing,”
(HCI) or human machine interfaces (HMI). in Proc. IEEE Sensors, Glasgow, UK, 2017.
22. W. Uter and H. Schwanitz, “Contact dermatitis from propylene glycol in ECG elec-
trode gel,” Contact Dermatitis, 34(3), 230 (1996).
Acknowledgments 23. Y. M. Chi, T. Jung, and G. Cauwenberghs, “Dry-contact and noncontact biopotential
electrodes: Methodological review,” IEEE Rev. Biomed. Eng., 3, 106 (2010).
The experimental procedures involving volunteer human subjects 24. N. Matsuhisa, M. Kaltenbrunner, T. Yokota, H. Jinno, K. Kuribara, T. Sekitani, and
T. Someya, “Printable elastic conductors with a high conductivity for electronic textile
described in this research are followed by the ethical principles out- applications,” Nature Communications, 6, (2015).
lined in the Helsinki Declaration of 1964, as revised in 2013 and 25. Y. Atwa, N. Maheshwari, and I. A. Goldthorpe, “Silver nanowire coated threads for
participants gave their informed consent for inclusion before they par- electrically conductive textiles,” J. Mat. Chemist. C, 3(16), 3908 (2015).
ticipated in the study. The authors gratefully thank the participants 26. Y. Meng, Z. B. Li, X. Chen, and J. P. Chen, “A flexible dry micro-dome electrode for
ECG monitoring,” Microsystem Technologies, 21(6), 1241 (2015).
involved in this study. 27. A. Shafti, R. B. R. Manero, A. M. Borg, and M. J. Howard, “Embroidered elec-
tromyography: A systematic design guide,” IEEE Trans. Neural. Syst. Rehabil. Eng.,
25(9), 1472 (2017).
ORCID 28. A. K. Geim and K. S. Novoselov, “The rise of graphene,” Nature Materials, 2007.
29. M. K. Yapici, T. Alkhidir, Y. A. Samad, and K. Liao, “Graphene-clad textile electrodes
Ata Jedari Golparvar https://orcid.org/0000-0002-1107-6380 for electrocardiogram monitoring,” Sens. Actuators B Chem., 221, 1469 (2015).
Murat Kaya Yapici https://orcid.org/0000-0003-4328-5985 30. A. J. Golparvar and M. K. Yapici, “Electrooculography by Wearable Graphene Tex-
tiles,” IEEE Sens J., 18, 8971 (2018).
31. M. K. Yapici and T. Alkhidir, “Intelligent medical garments with graphene-
References functionalized smart-cloth ECG sensors,” Sensors, 17(4), 875 (2017).
32. A. Cömert, M. Honkala, and J. Hyttinen, “Effect of pressure and padding on motion
1. A. T. Duchowski, “A breadth-first survey of eye-tracking applications,” Behav. Res. artifact of textile electrodes,” Biomed. Eng. Online, 12(1), 26 (2013).
Methods, Instrum., & Comput., 34(4), 455 (2002). 33. A. Lopez, F. Ferrero, M. Valledor, J. C. Campo, and O. Postolache, “A study on
2. C. H. Morimoto and M. R. M. Mimica, “Eye gaze tracking techniques for interactive electrode placement in EOG systems for medical applications,” in Proc. IEEE Int.
applications,” Comput. Vis. Image Underst., 98(1), 4 (2005). Symp. Med. Meas. Appl., Benevento, Italy, 2016.
3. T. Eggert, “Eye movement recordings: methods,” Dev. Ophthalmol., 15 (2007). 34. Y. Zhang, X. Gao, J. Zhu, W. Zheng, and B. Lu, “A novel approach to driving fatigue
4. V. Häkkinena, K. Hirvonena, J. Hasana, M. Katajab, A. Värric, P. Loulad, and detection using forehead EOG,” in Proc. Int. IEEE EMBS NER, Montpellier, France,
H. Eskolac, “The effect of small differences in electrode position on EOG signals: 2015, pp. 707.
application to vigilance studies,” Electroencephalogr. Clin. Neurophysiol., 86(4), 294 35. J. Heo, H. Yoon, and K. S. Park, “A novel wearable forehead eog measurement system
(1993). for human computer interfaces,” Sensors, 17(7), 1485 (2017).
5. S. Wu, L. Liao, S. Lu, W. Jiang, S. Chen, and C. Lin, “Controlling a human–computer 36. M. M. Puurtinen, S. M. Komulainen, P. K. Kauppinen, J. A. V. Malmivuo, and
interface system with a novel classification method that uses electrooculography sig- J. A. K. Hyttinen, “Measurement of noise and impedance of dry and wet textile elec-
nals,” IEEE Trans. Biomed. Eng., 60(8), 2133 (2013). trodes, and textile electrodes with hydrogel.” In Proc. Conf. Proc. IEEE Eng. Med.
6. A. Larson, J. Herrera, K. George, and A. Matthews, “Electrooculography based elec- Biol. Soc., New York, NY, USA, 2016, pp. 6012.
tronic communication device for individuals with ALS.” in Proc. IEEE SAS, Glass- 37. A. A. Chlaihawi, B. B. Narakathu, S. Emamian, B. J. Bazuin, and M. Z Atashbar,
boro, NJ, USA, 2017. “Development of printed and flexible dry ECG electrodes,” Sens. Biosens. Res., 20,
7. A. López, P. Arévalo, F. Ferrero, M. Valledor, and J. C. Campo, “EOG-based system 9 (2018).
for mouse control,” in Proc. IEEE Sensors, Valencia, Spain, 2014, pp. 1264. 38. X. An and G. J. M. Stylios, “A Hybrid Textile Electrode for Electrocardiogram (ECG)
8. A. López, F. Ferrero, D. Yangüela, C. Álvarez, and O. Postolache, “Development of Measurement and Motion Tracking,” Materials, 11(10), 1887 (2018).
a computer writing system based on EOG,” Sensors, 17(7), 1505 (2017). 39. J. Otero-Millan, S. L. Macknik, and S. Martinez-Conde, “Fixational eye movements
9. R. Barea, L. Boquete, M. Mazo, and E. Lopez, “System for assisted mobility using and binocular vision,” Front. Integr. Neurosci., 8, 52 (2014).
eye movements based on electrooculography,” IEEE Trans. Neural. Syst. Rehabil. 40. K. Kaneko and K. Sakamoto, “Evaluation of three types of blinks with the use of
Eng., 10(4), 209 (2002). electro-oculogram and electromyogram,” Percept. Mot. Skills, 88(3), 1037 (1999).
10. A. Ubeda, E. Ianez, and J. M. Azorin, “Wireless and portable EOG-based inter- 41. A. J. Golparvar and M. K. Yapici, “Graphene-coated wearable textiles for EOG-based
face for assisting disabled people,” IEEE ASME Trans. Mechatron., 16(5), 870 human-computer interaction,” in Proc. IEEE 15th Int. Con. Wear. Implant. BSN, Las
(2011). Vegas, NV, USA, 2018, pp. 189.
11. A. López, M. Fernández, H. Rodríguez, F. Ferrero, and O. Postolachebc, “Develop- 42. T. Nakano, Y. Yamamoto, K. Kitajo, T. Takahashi, and S. Kitazawa, “Synchronization
ment of an EOG-Based System to Control a Serious Game,” Measurement, 127, 481 of spontaneous eyeblinks while viewing video stories,” Proc R Soc Lond [Biol], 2009.
(2018). 43. K. Kaneko and K. Sakamoto, “Evaluation of three types of blinks with the use of
12. D. Valeriani and A. Matran-Fernandez, “Towards a wearable device for controlling a electro-oculogram and electromyogram,” Percept. Mot. Skills, 88(3), 1037 (1999).
smartphone with eye winks.” in Proc. 7th CEEC, Colchester, UK, 2015, pp. 41. 44. F. Behrens, M. MacKeben, and W. Schröder-Preikschat, “An improved algorithm for
13. B. Lin, P. Wu, and C. Chen, “2D/3D-Display Auto-Adjustment Switch System,” IEEE automatic detection of saccades in eye movement data and for calculating saccade
J. Biomed. Health Inform., 22(3), 799 (2018). parameters,” Behav. Res. Methods, 42(3), 701 (2010).
14. A. Bulling, J. A. Ward, H. Gellersen, and G. Troster, “Eye movement analysis for 45. M. Pelin, L. Fusco, V. León, C. Martín, A. Criado, S. Sosa, E. Vázquez, A. Tubaro,
activity recognition using electrooculography,” IEEE Trans. Pattern Anal. Mach. and M. Prato, “Differential cytotoxic effects of graphene and graphene oxide on skin
Intell., 33(4), 741 (2011). keratinocytes, ” Nature, 7, 40572 (2017).

Downloaded on 2019-05-25 to IP 193.93.194.206 address. Redistribution subject to ECS terms of use (see ecsdl.org/site/terms_use) unless CC License in place (see abstract).