The handheld devices or POC are popular to measure glucose values. However, their
accuracy is affected by some factors, in which the hematocrit is the most highly affecting one.
This paper presents an approach for reducing the effects of HCT, in which HCT is estimated
from the transduced current curve by employing the single hidden layer feedforward neural
network. This network is trained by non-iterative learning algorithm. Experimental results on the
handheld devices (GlucoDr) have shown the performance of correction.
7 trang |
Chia sẻ: honghp95 | Lượt xem: 459 | Lượt tải: 0
Bạn đang xem nội dung tài liệu Glucose correction in handheld devices by reducing the effect of hematocrit - Huynh Trung Hieu, để tải tài liệu về máy bạn click vào nút DOWNLOAD ở trên
Journal of Science and Technology 54 (3A) (2016) 91-97
GLUCOSE CORRECTION IN HANDHELD DEVICES BY
REDUCING THE EFFECT OF HEMATOCRIT
Huynh Trung Hieu
1, *
, Bui Dinh Tien
1
, Yonggwan Won
2
1
Industrial University of Ho Chi Minh City, 12 Nguyen Van Bao, Go Vap, Ho Chi Minh City,
Vietnam
2
Chonnam National University, Gwangju 500-757, Korea
*
Email: hthieu@ieee.org
Received: 15 June 2015; Accepted for publication: 27 July 2016
ABSTRACT
This study presents an approach for glucose correction in handheld devices by reducing the
effects of hematocrit. The hematocrit values are estimated from the transduced current curves
which are produced during the chemical reactions of glucose measurement process in the
handheld devices. The hematocrit estimation is performed by applying the single-hidden layer
feedforward neural network which is trained by the non-iterative learning algorithm. The
experimental results show that the proposed approach can improve the accuracy of glucose
measurement by using the handheld devices.
Keywords: glucose measurement, handheld device, neural network, hematocrit, glucose
correction.
1. INTRODUCTION
Diabetes mellitus is characterized by the inadequate ability of the pancreas to handle the
blood glucose concentration. This is one of the leading diseases worldwide with the long-term
complications including hypoglycemia diabetic ketoacidosis, hyperosmolar, retinopathy,
cardiovascular, nephropathy (kidneys), and neuropathy (nerves and feet). The current treatment
methods for insulin dependent diabetes such as continuous infusion of insulin or subcutaneous
insulin injection require frequently evaluating the variation of glucose concentration.
The major tools for managing the glucose concentration are the point-of-care (POC) or
handheld blood glucose meters. These meters are easy to use and relatively cheap, however they
are inaccuracy in various clinical abnormalities. Many studies reported that the hematocrit is one
of the most highly affecting factors for POC or handheld glucose measurements [1, 2]. A low
hematocrit is associated with overestimation, while a high hematocrit is associated with
underestimation of glucose results [3 - 5]. The new born may have the hematocrit levels as high
as 63 % [6], and there may be 2 – 5 % of new born infants with erythrocytosis [7]. The normal
glucose concentration of neonates are often lower that of adults [8], and there may be up to 20 %
of neonates with hypoglycemia [9]. The hypoglycemia also occurs frequently in perioperative
critically ill patients. The adult patients may have low hematocrit value, and management of
Huynh Trung Hieu, Bui Dinh Tien, Yonggwan Won
92
their glucose concentrations is considered as very important factor for positive clinical
outcomes.
Hence, improving the accuracy of glucose measurement plays an important role to help the
healthcare professionals to act intermediately certain conditions. One of the approaches for this
issue is to reduce the effects of hematocrit level. Estimation of hematocrit level can be
performed by employing commercial impedance analyzers with traditional centrifugation
measurements. It also can be performed by dielectric spectroscopy [10]. However, the above
approaches are in vitro, quite complicated or require individual devices. Recently, we developed
approaches for hematocrit estimation from the transduced current curve which is generates
during the chemical reaction in glucose measurement. These approaches are adequate for
glucose correction in the handheld glucose meters [11, 12].
The handheld glucose meters often employ the biosensors. These biosensors use an enzyme
to break the blood glucose down. It has been reported that an optical sensor could be applied to
the determination of glucose content in beverage samples with good results [13] or it could be
applied to the pulse hematometry [14]. However, most optical biosensors developed so far are
not sensitive as the electrochemical biosensors. They also affect by the interference from some
species in biological samples which result in the biosensor device very complicated in design to
reduce the effects of interferences. In this paper, we investigate an approach for improving the
accuracy of handheld devices in glucose measurement by reducing effects of hematocrit. The
biosensors mentioned in this study is electrochemical glucose ones. The remaining parts of this
paper are organized as follows. Section 2 describes the materials and methods, the experimental
results and discussions are presented in the section 3, and the last section is conclusions.
2. MATERIALS AND METHODS
2.1. Effects of hematocrit (HCT) on glucose measurement
Although the handheld devices can provide quick measurements, their performance is
affected by critical care variables, in which the hematocrit is the most highly effecting factor.
This problem can be illustrated in Fig. 1. In Fig. 1a, three current curves from time point 11.5s to
14s provides the same value (the measured values of three curves at time point 14s are the same
as 17.3439) even though glucose values corresponding to these current curves are different those
are 147 mg/dL, 161 mg/dL and 157 mg/dL for the hematocrit of 27 %, 45.6 % and 39.4 %,
respectively. Otherwise, with the same glucose value of 262 mg/dL, the measured values on
three current curves corresponding to different hematocrit levels are different as shown in Fig.
1b.
Figure 1. Effects of Hematocrit on Glucose Measurement: (a) same measured value on current curve but
different glucose values, (b) different measured values on current curves but same glucose value.
Glucose correction in handheld devices by reducing the effect of hematocrit
93
Furthermore, relationship between hematocrit levels and errors of glucose measurement was also
reported by researchers [14 - 16]. The results from Louie et al. [16] has shown that the difference
of glucose measurements by portable device minus the primary reference glucose measurements
is a function of hematocrit. Thus, in order to improve measurement performance of portable
devices, the effects of hematocrit must be reduced.
2.2. Error correction for glucose measurement by reducing the effects of hematocrit
Let us denote t
ref
be the measured glucose values from the primary glucose measurement, t
m
be the measured glucose values from portable device, and t
c
be the corrected glucose values of
t
m
. The main goal of correction process is to find mapping
ψ: tm tc (1)
so that dependence of residuals on hematocrit is reduced as much as possible. This is illustrated
in Fig. 2.
Figure 2. Glucose correction.
Denoting R
m
be the residual which is defined as
R
m
= t
m
- t
ref
, (2)
the proposed mapping for ψ is given by:
t
c
= t
m
-R
m
. (3)
From the previous results, it showed that R
m
is a variable which depends on the hematocrit
density. Thus, there is a function g mapping from hematocrit to residuals R
m
as follows:
g:HCT
m
R
m
R
m
= g(HCT
m
), (4)
where HCT
m
is hematocrit estimated from portable devices.
In this study, the mapping function g is determined by using the linear model and the
hematocrit is estimated from the transduced current curves. These curves are produced by
chemical reaction between the blood and the enzyme coated on the biosensor. The enzyme
commonly used in biosensors to measure the glucose levels is the glucose oxidase (GOD); it is
used to catalyze the oxidation of glucose by oxygen to produce gluconic acid and hydrogen
peroxide.
Glucose+O2+GO/FAD→Gluconic acid+H2O2+GO/FADH2
The reduced form of the enzyme (GO/FADH2) is oxidized to its original state by an
electron mediator. The resulting reduced mediator is then oxidized by the active electrode to
produce a current. Figure 3 illustrates a current curve in the first 14 seconds. Normally, the first
eight seconds do not contain much information for HCT estimation. In the next six seconds, the
Huynh Trung Hieu, Bui Dinh Tien, Yonggwan Won
94
current curves are sampled at frequency of 10Hz to produce current patterns. Let x=[x1, x2, ,
xd] be a current vector of d sampled patterns from the current curve.
Figure 3. Anodic current curve.
In this study, the current vector is used as the input vector or neural network to determine
the hematocrit. Several neural network architectures have been proposed. However, it was
shown that the single hidden layer feedforward neural networks (SLFNs) can approximate any
function if the activation function and the number of hidden units are chosen properly. One of
the most important problems in neural networks is training. Assuming that there are n training
patterns (xj, tj), j=1, 2, , n, where xj=[xj1 xj2 xjd]
T
and tj=[tj1 tj2 tjC]
T
are the j-th input
pattern and its target, respectively, the main goal of training process is to determine the network
weights to minimize the error function given by
2
j j
1
n
E o t
j
, (5)
where oj is the output vector corresponding to the j-th input pattern. An efficient algorithm for
training SLFNs is extreme learning machine (ELM) which was proposed by Huang et al. [17]. In
ELM, the minimization process of error function is based on the linear model:
HA = T, (6)
where H is called as the hidden-layer-output matrix of SLFN and defined as [17]:
H=
1 1 1 N 1 N
1 n 1 N n N
( ) ( )
( ) ( )
f f
f f
b b
b b
w x w x
w x w x
, (7)
T = [t1 t2 tn]
T
, (8)
and
A = [ a1 a2 aC]. (9)
Note that wm = [wm1 wm2 wmd]
T
is the weight vector connecting from the input units to the m-th
hidden unit, bm is its bias, ai = [ai1 ai2 aiN]
T
is the weight vector connecting from the hidden
units to the i-th output unit, and f(·) is the activation function of hidden nodes. The input weights
(wm’s) and biases (bm’s) of hidden units in ELM are randomly assigned, and the output weights
(ai’s) are determined by
Glucose correction in handheld devices by reducing the effect of hematocrit
95
 = H
†
T, (10)
where H
†
is the pseudo-inverse of H. Thus, the network weights are determined by the non-
iterative learning algorithm it consists of two steps: (1) randomly choose the input weights and
hidden biases and (2) determine the output weights by using the pseudo-inverse operation. This
algorithm can offer good performance with high learning speed in many applications.
3. RESULTS AND DISCUSSION
The data set used in our experiments was obtained from 191 blood samples from randomly
selected volunteers. Each sample was applied to measure the accurate hematocrit using
centrifugation method, accurate glucose using YSI2700, glucose values using handheld device
(GlucoDr) and the anodic current curves. From the second part of curve, which is after the
incubation period, we obtained 59 current points by sampling the current curve at the frequency
of 10 Hz.
Figure 4. Plot of paired-differences of GlucoDr glucose measurements minus the YSI2007
glucose measurements.
The plot of paired-differences of GlucoDr glucose measurements minus the YSI 2007
glucose measurements against changes of collected hematocrit is shown in Fig. 4. From this
figure, we can see that there is a relationship between hematocrit and residuals which are defined
as differences of GlucoDr glucose measurements minus the YSI2007 glucose measurements. In
addition, using the test statistic for the slope given by
0
slope
slope
slope
t (11)
and using the P-test we see that the slope value is significantly different than 0 (p < 0.01).
Therefore, we can conclude that effect of hematocrit on glucoDr measurements is significant
which is consistent with the previous reports.
Thirty percent of data set is used for training in order to find the network parameters and
the mapping function g from hematocrit to residuals, and the remaining 70 % is used for
evaluation. The RMSE for GlucoDr on the test set without error correction is 16.4149 while that
with error correction is 13.7418. The t-test for slope without error correction is -3.846 (p-value <
Huynh Trung Hieu, Bui Dinh Tien, Yonggwan Won
96
0.001) which shows dependence of residuals on hematocrit levels, while the t-test for slope with
the error correction is 0.26, these results show that the effects of hematocrit are reduced after
error correction. On the error tolerance of 15 %, without error correction give 91.6 %, while with
error correction can give 94.6 %.
4. CONCLUSIONS
The handheld devices or POC are popular to measure glucose values. However, their
accuracy is affected by some factors, in which the hematocrit is the most highly affecting one.
This paper presents an approach for reducing the effects of HCT, in which HCT is estimated
from the transduced current curve by employing the single hidden layer feedforward neural
network. This network is trained by non-iterative learning algorithm. Experimental results on the
handheld devices (GlucoDr) have shown the performance of correction.
REFERENCES
1. Aynsley-Green A. – Glucose, a fuel for thought, J. Pediatr Child Health 27 (1991) 21-30.
2. Hussain K, Sharieff N. – The inaccuracy of venous and capillary blood glucose
measurement using reagent strips in the new born period and the effect of hematocrit,
Early human dev. 57 (2000) 111-121.
3. Tang Z., Lee J.H., Louie R.F., Kost G.J., Sutton D.V. – Effects of Different Hematocrit
Levels on Glucose Measurements with HandHeld Meters for Point of Care Testing, Arch.
Pathol. Lab. Med. 124 (2000) 1135-1140.
4. Kilpatrick E.S., Rumley A. G., Myin H. – The effect of variations in hematocrit, mean cell
volume and red blood count on reagent strip tests for glucose, Ann. Clin. Biochem. 30
(1993) 485-487.
5. Kaplan M., Blondheim O., Alon I. – Screening for hypoglycemia with plasma in neonatal
plod of high hematocrit value, Crit. Care Med. 17 (1989) 279-282.
6. Gatti R.A. – Hematocrit values of capillary blood in newborn infants, J. Pediatr. 70 (1967)
117.
7. Stevens K, Wirth F.H. – Incidence of neonatal hyper viscosity at sea concentration, J.
Pediatr. 70 (1980) 118-119.
8. Halamek L. P., Benaron D.A., Stevenson D.K. – Neonatal hypoglycemia: I. Background
and definition, Clin. Pediatr. 36 (1997) 675-680.
9. Brooks C. – Neonatal hypoglycemia, Neonatal Netw. 16 (1997) 15-21.
10. Treo E. F., Felice C. J., Tirado M. C., Valentinuzzi M. E., Cervantes D. O. – Hematocrit
Measurement by Dielectric Spectroscopy, IEEE Transactions on Biomedical Engineering
25 (1) (2005) 124-127.
11. Hieu Trung Huynh, Yonggwan Won, Jung-ja Kim – Neural Networks for the Estimation
of Hematocrit from Transduced Current Curves, The 2008 IEEE Int'l conference on
Networking, Sensing and Control (2008) 1517-1520.
12. Hieu Trung Huynh, Yonggwan Won, Jinsul Kim – Hematocrit estimation using online
sequential extreme learning machine, Bio-Medical Materials and Engineering 26 (2015)
S2025-S2032.
Glucose correction in handheld devices by reducing the effect of hematocrit
97
13. Nogawa M., Tanaka S., Yamakoshi K. – Development of an optical arterial hematocrit
measurement method: pulse hematometry, Proceeding of the 2005 IEEE Eng. in Medicine
and Biology (2005) 2634-2636.
14. Choi M. M. F., Pang W. S. H., Xiao D., Wu X. – An optical glucose biosensor with
eggshell embrane as an enzyme immobilization platform. Analyst 126 (2001) 1558-1563.
15. Tang Z., Louie R. F., Payes M., Chang K. C., Kost G. J. – Oxygen effects on glucose
measurements with a reference analyzer and three handheld meters, Diabetes Technol.
Ther. 2(3) (2000) 349-362.
16. Louie R. F., Tang Z. P., Sutton D. V., Lee J. H., Kost G. J. – Point-of-care glucose testing
- Effects of critical care variables, influence of reference instruments, and a modular
glucose meter design, Archives of Pathology & Laboratory Medicine 124 (2000) 257-266.
17. Huang G. B., Zhu Q. Y., Siew C. K. – Extreme learning machine: Theory and applications,
Neurocomputing 70 (2006) 489-501.
Các file đính kèm theo tài liệu này:
- 11962_103810382397_1_sm_5254_2061581.pdf