See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/339290182

Estimation of Blood Glucose from PPG Signal Using Convolutional Neural

Network

Conference Paper · November 2019

DOI: 10.1109/BECITHCON48839.2019.9063187

CITATIONS READS

16 7,206

6 authors, including:

Mst Shamima Hossain Bidya Debnath

Daffodil International University Bangladesh University of Engineering and Technology
9 PUBLICATIONS 41 CITATIONS 9 PUBLICATIONS 55 CITATIONS

SEE PROFILE SEE PROFILE

Sabyasachi Biswas Md. Junaed Al-Hossain

Bangladesh University of Engineering and Technology North Carolina State University
12 PUBLICATIONS 64 CITATIONS 19 PUBLICATIONS 68 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Md. Junaed Al-Hossain on 05 January 2021.

The user has requested enhancement of the downloaded file.

2019 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (B ECITHCON)
28-30 N ovember, 2019 , Dhaka, Bangladesh

Estimation of Blood Glucose from PPG Signal

Using Convolutional Neural Network
Mst. Shamima Hossain, Bidya Debnath*, Sabyasachi Biswas, Md. Junaed Al-Hossain, Adrita Anika
Syed Khaled Zaman Navid
Department of Electrical and Electronic Engineering
Bangladesh University of Engineering and Technology, Dhaka-l20S
*Email: [email protected]

Abstract-The purpose of this paper is to develop a system Noninvasive BG monitoring is a diabetes management
that can measure the blood glucose level in a non-invasive way. approach that allows an in-vivo continuous blood glucose
Since blood glucose is an important indicator of health issues, monitoring while also providing a convenient, and nonintrusive
frequent blood glucose measurement is required by a great
number of people. Non-invasive methods are getting popular measurements technology to patients [6]- [7]. Several noninva-
for such measurements since painful collection of blood is not sive biosensors, based on different measuring approaches such
required. In non-invasive methods, blood glucose level is recorded as reverse iontophoresis [8], combined reverse iontophoresis
by exploiting the relationship of blood glucose level with various and enzyme-based aerometric biosensor [9], Raman Spec-
biological signals and parameters. This paper addresses the troscopy [10], and Near Infrared Spectroscopy (NIRs) [11],
relationship of glucose level with PPG (photoplethysmoram)
signal which can provide information regarding the volume have been proposed. Most of the spectroscopy technique is
of blood flowing in the test region through exposure of that impractical to be used in the clinical practice for development
particular region to infrared light of certain wavelengths. A of portable and handy glucose monitoring device. With respect
dataset has been created by collecting the PPG signal of 30 people to spectroscopy, Infra-red (IR) is less affected by these matters
along with their blood glucose level, measured in invasive way as well as physical conditions. So now-a-days more emphasis
both before and 2 hours after breakfast. A CNN model has been
trained with this dataset to exploit the relationship of glucose is given on sensors based on IR spectroscopy.
level with PPG signal and thus the glucose level of blood have Based on biological parameters of human body closely
been measured. Results obtained from our system have been related with glucose level, in this paper it is tried to develop
compared with corresponding glucose level,recorded in invasive a sensor based glucose measuring device which will measure
way.
body glucose level in non-invasive way. Pulse Oximeter sensor
Index Terms-Non-invasive, Invasive, PPG signal, Deep learn-
ing, Blood glucose. is used for the purpose of recording the change in blood
volume flowing in the fingertip known as photoplethesmogram
1. INTRODUCTION
signal. For machine learning algorithm CNN model has been
used as it can extract various frequency levels which contains
Diabetes is a chronic disease in which the body fails to the necessary information related to the signal [12]. by using
regulate blood glucose concentration in the normal range (5- learn able filters. The goal is to replace the present invasive-
8milimole/litre). This instability can lead to serious problems, glucose measurement with our designed non-invasive glucose
e.g. cardiovascular diseases, kidney failure, blindness, etc. [1]. monitoring device. The device can be afforded by mass people
According to the World Health Organization, the pervasiveness due to its low cost nature. Also it is very easy to use and should
of diabetes was estimated to be 9% around the globe, 4.9 fit to most of the patients finger very easily. The training data is
million deaths were caused by diabetes in 2014 alone. It is taken from the user. The results are compared with the clinical
expected that by 2030, diabetes will use up to 11.6% of the value obtained in an invasive way. However, the estimation
total expenses made in the healthcare domain [2]. There are of blood glucose level is satisfactory for most of the user
two types of diabetes. The cause of type 1 diabetes is not with a low error rate. This work paves the way of monitoring
known and it is not preventable with current knowledge [3]- and preventing diabetes with a non-invasive wearable solution
[4]. Type 2 diabetes results from the bodys ineffective use of considering the cost and portable capability of the system.
insulin. Until recently, this type of diabetes was seen only in Most of the previous methods mentioned before are relatively
adults but it is now also occurring increasingly frequently in more complex than this one.
children [5].
Inspite of being painful, and with the risk of infection, inva- II . BACKGROUND
sive measurement method has been the most widely practiced
method across the world. For decades, various techniques have Studies demonstrating the relation between PPG signal
been attempted for pain-free blood glucose monitoring, includ- and blood glucose level [15] reveals that blood glucose is
ing minimal invasive and non-invasive methods, depending on approximately proportional to blood viscocity and inversely
their transduction mechanism. proportional to blood flow rate i.e, increased blood glucose

978-1-7281-5389-6/19/$31.00 ©20 19 IEEE 53

level results in increased blood viscosity and decreased blood A. Data Acquisition
flow rate [16]. Thirty subjects willingly took part in the experimentation
PPG is an optical technique based on the blood volume of which fifteen were male and the rest of them were female.
changes in the micro vascular bed of tissue measured at the Each participant gave their consent in the prescribed consent
skin surface. Depending on the energy of incident photons, paper before data acquisation . Students of department of EEE ,
bond deformation or vibration at different energy level of BUET and both diabetic patients and staffs of BUET Medical
different bonds occurs. So, only the photon with energy that Center were asked to appear for blood glucose check up at
corresponds to the difference between two of its energy levels early morning in empty stomach and later do the same at two
can be absorbed. The absorptivity decreases rapidly as one hours after having food and in some cases with drinking 75g
goes from fundamental to the overtones in sequence [13]- [14]. glucose water in case of the patients were ignorant about their
For non-invasive glucose level measurement, the concept of diabetic condition. We have managed to take This two phase
transmission PPG is used. Table I shows the relative absorbp- data of thirty patients were managed to taken, one third of
tivities of C-H bonds in a glucose molecule. The absorbance which are diabetic patient with blood glucose level higher than
of the glucose bonds(C-H, O-H) is strong in the region of normal. The data were recored in milimole per litre. A sample
fundamental frequencies (2-2.5 urn) and first overtone region of the recorder data is given in Table II and the distribution
(1.53-1.82 urn) but suffers from drawbacks like high cost of of data over various range is shown in Table III :
the sensors, strong absorption due to water and scattering of
fatty tissue. For this reason, second overtone region (0.8-1.6 TABLE II
urn) has been considered for the selection of the IR sensor SAMPLE OF RECORDED DATA
where the glucose has good absorption compared to the other
Patient ID I Age I Weight I Gender I PBS I Not-fasting I
chromophores of the blood.
01 28 75 Male 5.1 5.8
02 53 63 Male 4.4 5
TABLE I 03 40 64 Female 9.9 13.4
RELATIVE ABSORPVITIES OF C-H BONDS 04 42 75 Female 5.8 6
Overtone I Wave number I Relative absorpvity
Fundamental 3019 I The PPG signals were recorded by sampling at 115.2 Hz
First overtone 5912 0.088 and the invasive data was taken as the average of three times
Second overtone 8677 0.0032 recording in a fixed interval.

For getting glucose level in blood, the absorbance which is TABLE III
DISTRIBUTION OF DATA
mainly due to blood glucose is calculated from AC component
of PPG wave as follows: Range(mmol elL) I Number of data segments
0-5 807
5-10 689
f10D).. = log(l + ~?(ti~) 10-15 435
x ti+l 15++ 276
Where f10D).. is difference between optical densities at time
ti and ti+1, f1!>-. (ti) is the pulsatile component at time ti and
It can be seen from Table III that the data is not uniformly
I(ti+d is the intensity of light at time ti+}.
distributed. The number of data segment for higher blood
III. METHODOLOGY glucose level is relatively low.
The glucose estimation system can be divided into three E. Preprocessing
steps. The process flow is as follows: Before preprocessing data is divided into lOs segment
• Data Acquisition containing 1152 samples and segments with missing data were
• Preprocessing eliminated.Total 360 segments were obtained from the data set
• Blood glucose estimation which was used as training data. Preprocessing step involves
The flow chart of the working mechanism of the overall two steps. They are described in the following:
system is given in Fig 1. • Filtering
This step involves elimination of very low frequency
which was done using a high pass filter with -3dB cutoff
frequency of 4 Hz.
• Eliminating corrupted signals
Corrupted signals were eliminated using different param-
eters extracted from the filtered PPG signal. Different
Fig. 1. Flow Chart of the working mechanism of the overall system measures were used to differentiated between good and
corrupted signal.

54
 1) At first, PPG signal was normalized and its first
derivative was determined.
2) From its first derivative negative zero crossing
points were calculated. The points indicates both
the systolic peak and diastolic peak. To find out the Feat.
Ext .

lB 0 6 ----m,.
g ...g
corrupted signals, either systolic peak positions or Block

diastolic peak positions are needed. For our thesis,

we used systolic peaks. IconVldl IconVldJ
3) To differentiate between systolic and diastolic
peaks, their prominence and width were measured
and the prominence and width of systolic peak is put

greater than that of diastolic one.

4) The notches were also determined as the most
Fig. 3. Outline of the proposed hierarchical CNN network . Here, the
negative points on the curve. feature extraction blocks extract the necessary features from the feature vector
5) Then depending on the no of peaks and troughs, obtained from the conv 1d layer and the dense layer extracts the relation
between the previous and output block.
peak to peak interval, presence of ectopic beats,
corrupted signals were identified and lastly they
were removed. is fed into a dense layer. Batch size of I is used as we
The proprocessed signal is shown in the following figure have less data than needed. For loss function, mean square
2. The proposed model requires a sequence of M consecutive error (MSE) is used and for gradient optimization, RMSprop
feature vectors [x[xr ...x f. ...x1 ]T. The sequence is prepared optimizer is used with an initial learning rate of 0.001.
by adding the segments one after another and forming a matrix. The primary objective of these feature extraction networks
Then the corresponding sequence vector [yiyr..·yr....y1 V is to extract the necessary features from the feature vector
consisting of Blood glucose level are taken as the output. sequence [x[xr ...xf. ...x1V, where each X i is a IXlI52
feature vector. The last dense layer then provides the output
Filtered Signal as a sequence of vectors [yiyr,..yr....y1 V
where each Yi
is a IXI vector containing the estimated Blood glucose level.
For continuous prediction, the result of YM is of interest.
IV. PROPOSED DESIGN AND IMPLEMENTATION
In this section, we will illustrate the proposed design and
the method of implementation one by one. The whole process
is divided into three parts:
Raw PPG Signal • Hardware Development
• Software Development
Fig. 2. Raw PPG signal and PPG signal after preprocessing • Invasive Data Collection
The prototype model which is used for recording the PPG
C. Blood Glucose Estimation signal samples is illustrated in Fig 4.
After preparing the input and output for the model, the data
for each patient has been split to 80% for training and 20% for
validation and data collected instantaneously from the patients
whose data were not presented in the dataset were used for
testing. Data are extracted from lOs window or 1152 samples
at 115.2 Hz.
The CNN network mainly consists of one convld layer
with two stacked feature extraction layers connected to it. The
convld layer has the filter size of 64, kernel size of 5 with
strides of 2 and tanh activation function. The output of the
convid layer is fed into a maxpooling layer and a nonnaliztion
layer. The feature extraction layer consists of three convid
layer and two maxpooling layers . The first convld layer has Fig. 4. The prototype device model
the filter size of 96 with no maxpooling layer and rest of the
convid layer consists of 208 and 64 filters with maxpool size
of 3 and tanh activation function. The output of 1st feature A. Hardware Development
extraction layer is fed into the 2nd one and finally the output • MAX 30105 Pulse Oximeter module

55
 The SparkFun MAX30105 Particle Sensor is a flexible, MSE = it L:~=lIYtarget - Ypredictedl
2

powerful sensor . It has been equipped with three LEDs

where Ytarget and ypredicted are the ground truth and estimated
as well as a very sensitive photon detector. The idea
blood glucose level for the kth element of the time sequence,
is to pulse the different LEDs, then detect what shines
respectively. The convergence curve is given in Fig. 5.
back. Based on the reflected signature its possible to
detect different types of particles or materials (such as
oxygenated blood or smoke from a fire). The MAX30105 Convergence Curve
utilizes a red LED, a green LED, and an IR (Infrared) 0"9 - train
LED for presence sensing, heart beat plotting and heart 08
- test

rate monitoring among its multitude of uses, including

Pulse Oximetry. The MAX30105 is designed to operate g 0"7
UJ
0"6
at 5V and can communicate with both 3.3V and 5V
"0
~

microcontrollers.The sensor has in-built low pass filter 'g". 0.5

1Il
e
of 50Hz and 60Hz to avoid power frequency noise. The '"
v
:;:
0.4

sensor also uses both reflection and refraction of light 03

penetrating through to generate the PPG signal. In the 02
case of MAX30 105 sensor, a variation of light reflectance 0"1
due to blood flow was recorded to obtain the PPG signal. 0 20 40 60
No of epoch s
80 100 120

• Arduino
The Arduino is a microcontroller board based on the
Fig. 5. Convergence Curve of Mean Squared Error between target and
ATmega328. It has digital input/output pins (some of estimated Blood Glucose Level
which can be used as PWM outputs and some can be used
as analog inputs), a 16 MHz resonator, a USB connection, After completing the training, we found that training loss
a power jack, an in circuit system programming (ICSP) was 0.16 and validation loss was 0.19. As the validation loss
header, and a reset button. It contains everything needed is slightly greater that the training loss we can say that the
to support the microcontroller. model is a good fit. Total 104 epochs were needed.
B. Software Development The proposed model has been evaluated using mean squared
error (MSE). The prediction curve is given below:
• Tera-term & Arduino
With finger tip on the pulse oximeter or ppg sensor the
Predict ion Curve
required training set data has been recorded with the
- OigillOl
help of Arduino and Teraterm software . Teraterm later 24
- Pred icted
converts the data format in .csv format . 21
• Python
After recording the data, further analysis and trammg 18

was done by using Python. When testing the model, 15

real time data was used. For this purpose, Arduino was
incorporated with Python so that the data is directly being 12

fed into the system and the output is being processed and
9
shown on the system instantaneously. During testing tera-
term was not used as the data was directly being saved 6 L,, -_ , -_ , -- ,.-- , -- ,.-- ,.-- , -- ,-J
through python while processing. o 10 20 30 40 50 60 70 80
Sample

C. Invasive Data Collection

Fig. 6. Continuous Blood Glucose Level measurement.
In order to make training set database for our machine
learning algorithm we also have taken invasive data through The MSE error given by the proposed method is approxi-
Safe touch Blood Glucose Meter 18305805 which along with mately 0.15 and the Table IV gives the comparison between
lancet & disposable strip measures the blood glucose in the actual result and predicted result in milimole per litre.
milimole per liter. We have received invasive blood glucose
data following clinically recommended approach. TABLE IV
COMPARISON BETWEEN ACTUAL AND PREDICTED RESULT
V. RESULT ANALYSIS
Patient ID Actual Result Predicted Result
After training the data using the proposed model, we get the
convergence curve given in the following. Convergence was 31 7.1 7.9
32 6.2 6.7
measured depending upon the mean squared error given by 33 13.7 12.8
the equation,

56
 Fig. 6 shows the continuous Blood Glucose Level tracking show any distinguishable difference. Thus, it is evident that
of a patient using the proposed method. From this figure, it can the proposed model is able to estimate very high and low
be seen that the proposed model is able to track continuous Blood Glucose Level with high accuracy. As the data set was
Blood Glucose Level with better accuracy. smaller it could not predict very high Blood glucose level due
Bland-Altman analysis, Pearson's correlation coefficient to insufficient data. Also there was some corrupted data in the
analysis and box-plot analysis is also performed on the pro- dataset which caused outliers and errors.
posed model. The results are reported in the following.
Bland-Altman test, Pearson's correlation test and box plot 0
analysis have also been performed on the data. Fig. 7 shows 27 8
the Bland-Altman plot of Blood Glucose Level estimation .
For the proposed model, the limits of agreement [-1.18 1.08]
for Blood Glucose Level It means that 95% of the estimated
24

21 I I
iii
Blood Glucose Level have error less than 1.1 milimole per 18

4
litre indicates that the model is a good estimator.

~
15

12
Bland-Altman Plot
• •
9

6 • 6
Target Est imate d
3
Fig. 9. The box plots for Blood Glucose Level. The high amount of outliers
c;
is due to the high Blood Glucose Level of some of the subjects.
0 0
~
ur

-3 VI. FUTURE IMPLEMENTATION

Our present prototype needs to be connected to the computer
-6
• all the time for processing the signal in real time. For mobility
9 12 15 18 21 24 27
6 usage and making the device wearable, a bluetooth module
Average of Target and Esti mated
will be needed to transmit the data to the central system.
Fig. 7. Bland-Altman plot of Blood Glucose Level. The limits of agreement VII. HUMANITARIAN IMPACT
(LOA) for Blood Glucose Level are [-1.183,1.08] .
The PPG signal based CNN model has been developed to
Fig. 8 shows the regression plot for Blood Glucose Level include provision of a mobile app based system. The main
estimation. The Pearson correlation coefficient of Blood Glu- hardware of the proposed system will only be the PPG sensor
cose Level is r = 0.95 . The coefficient is close to 1.0, which for continuous real time data acquisition and Arduino based
indicates high linearity between the target and estimated Blood ancillary device for data transmission for app based system
Glucose Level. resulting in a light weighted user friendly and most impor-
tantly a cost effective solution to non-invasive blood glucose
27 measurement.A cost comparison between conventional system
24
and proposed system has been shown in table V.

21 TABLE V
COST COMPARISON OF THE PROPOSED & CONVENTIONAL SYSTEM
18
'tl
~
'" 15
E Proposed System Conventional System
til
w 12 Components Cost in USD Components Cost in USD
Arduino 14.81 Glucometer 54.23
9 MAX 30105 13.95 25 Strips 34.89
Lipo Battery 9.48 Needles(100 Pes) 17.99
6 Bluetooth Module 7.99 Antiseptic Skin Cleanser 29.99
6 9 12 15 18 21 24 27 Others 2.96 Cotton rolls 9.99
Target Bandaids(200Pcs) 9.99
Total 49.19 Total 157.08
Fig. 8. The regression plot for Blood Glucose Level. The Pearson's correlation
coefficient are r = 0.95 for Blood Glucose Level.
From the comparison, it can be seen that the initial cost of
Fig. 9 shows the box plots of Blood Glucose Level. The conventional system is much higher than the proposed method.
target and estimate box plot for Blood Glucose Level do not Moreover other than Glucometer, rest of the components need

57
to be refilled from time to time which is much expensive. [7] G. Rao, RH. Guy, P. Glikfeld, W.R LaCourse, L. Leung, J. Tamada,
The approximate weight of the proposed hardware(without RO. Potts, N. Azimi, Reverse Iontophoresis: Noninvasive Glucose
Monitoring in Vivo in Humans, Pharm. Res. An Off. J. Am. Assoc.
wrist band) is 250-300gm. This weight can be mostly reduced Pharm. Sci. 12 (1995) 18691873.
by replacing the lipo battery with suitable substitute as it [8] AJ. Bandodkar, W. Jia, C. Yardimci, X. Wang, J. Ramirez, J. Wang,
comprises most of the weight. Addition of training data will Tattoo-based noninvasive glucose monitoring: A proof-of-concept study,
Anal. Chern. 87 (2015) 394398.
further increase the accuracy of the proposed system. With this [9] R. Pandey, S.K. Paidi, T.A. Valdez, C. Zhang, N. Spegazzini, RR.
wearable device continuous monitoring of the blood glucose Dasari, I. Barman, Noninvasive Monitoring of Blood Glucose with
level of the user can be ensured with reliability and the user Raman Spectroscopy, Ace. Chern. Res. 50 (2017) 264272.
[10] K. Maruo, M. Tsurugi, Jakusei Chin, T. Ota, H. Arimoto, Y Yamada,
can be alerted immediately in the case of abnormal blood M. Tamura, M. Ishii, Y. Ozaki, Noninvasive blood glucose assay using
glucose level . a newly developed near-infrared system, IEEE J. Sel. Top. Quantum
Electron. 9 (2003) 322330.
[11] K. Yamakoshi, Y. Yamakoshi, Pulse glucometry: a new approach for
VIII. CONCLUSION noninvasive blood glucose measurement using instantaneous differential
near-infrared spectrophotometry, J. Biomed. Opt. 11 (2006) 054028.
Glucose or sugar level being one of the important health [12] A. Humayun, S. Ghaffarzadegan, Z. Feng, and T. Hasan, Learning
factors of human body, it is necessary to check glucose front-end filter-bank parameters using convolutional neural networks for
abnormal heart sound detection, vol. 2018, 07 2018, pp. 14081411.
frequently through various ways which include invasive and [13] Ramasahayam, S., Haindavi, K.S. and Chowdhury, S.R, 2015. Nonin-
non-invasive system. Invasive system requires painful blood vasive Estimation of Blood Glucose Concentration Using Near Infrared
collection while non-invasive system doesnt need any blood Optodes. InSensing Technology: Current Status and Future Trends IV
(pp. 67-82). Springer International Publishing. DOl 1O.1007/sI0916-
sample using needle. Rather some biological signal from hu- 014-0166-2
man body have showed strong relationship with glucose level. [14] Ramasahayam, S., Haindavi, K.S. and Chowdhury, S.R, 2015. Nonin-
Among those signals, PPG (photoplethysmogram) provides vasive Estimation of Blood Glucose Concentration Using Near Infrared
Optodes. InSensing Technology: Current Status and Future Trends IV
information about the volume of blood flowing in a test region (pp. 67-82).Springer International Publishing, doi:IO.1007/978-3-319-
close to the skin of body by illuminating the region of interest 12898-6-4
of body and reflected or transmitted light from that region. [15] Yamakoshi, K. and Yamakoshi, Y, 2006. Pulse glucometry: a new ap-
proach for noninvasive blood glucose measurement using instantaneous
In our paper,using this property of PPG, we measured heart differential near-infrared spectrophotometry. Journal of Biomedical Op-
rate, blood oxygen saturation using PPG sensor and studied tics, 11(5), pp.054028-054028., doi: 10.1117/1.2360919
the change of glucose level due to heart rate, blood oxygen [16] Monte-Moreno, E., 2011. Non-invasive estimate of blood glucose and
blood pressure from a photoplethysmograph by means of machine
saturation. For determining the relation, a training database learning techniques. Artificial intelligence in medicine, 53(2), pp.127-
was prepared which includes PPG sensor reading of patients 138., doi: 1O.1016/j.artrned.2011.05.001.
heart rate, blood oxygen saturation value and invasive glucose
level result. Applying machine learning, the relation of glucose
level with PPG sensors value is obtained and that relation
shows strong evidence of perfect result while comparing with
usual invasive glucose level test. Further research for more
accurate machine learning with big dataset and designing the
prototype for commercial use is future objective of our work.

R EFERENCES

[1] R Periyasamy and S. Anand, A study on non-invasive blood glucose

estimation; an approach using capacitance measurement technique, in
2016 International Conference on Signal Processing,Communication,
Power and Embedded System (SCOPES),Oct 2016, pp. 847850.
[2] S. Ramasahayam, L. Arora, S. R Chowdhury, and M. Anu-
mukonda,Fpga based system for blood glucose sensing using photo-
plethysmography and online motion artifact correction usingadaline,
in 2015 9th International Conference on Sensing Technology (ICST).
IEEE, 2015, pp. 2227.
[3] B. E. Metzger, D. R Coustan, and E. R Trimble, Hyperglycemia
and adverse pregnancy outcomes, Clinical chemistry, vol. 65, no. 7,
pp.937938, 2019.
[4] J. M. Boots, E. M. van Duijnhoven, M. H. Christiaans, B. H. Wolf-
fenbuttel, and J. P. van Hooff, Glucose metabolism in renal transplant
recipients on tacrolimus: the effect of steroid withdrawal and tacrolimus
trough level reduction, Journal of the American Society of Nephrology,
vol. 13, no. 1, pp. 221227, 2002.
[5] AJ. Bandodkar, I. Jeerapan, J. Wang, Wearable Chemical Sensors:
Present Challenges and Future Prospects, ACS Sensors. 1 (2016)
464482.
[6] H. Lee, YJ. Hong, S. Baik, T. Hyeon, D.H. Kim, Enzyme-Based Glucose
Sensor: From Invasive to Wearable Device, Adv. Healthc. Mater. 7
(2018) 114.

58

View publication stats