Blood Glucose Determination by Fourier Transform near Infrared

Spectroscopy

F. S. Rondonuwu

1,2

, A. Setiawan

2,4

and F. F. Karwur

Research Center for NIR Applications, Universitas Kristen Satya Wacana, Salatiga, Indonesia

Department of Physics, Faculty of Mathematics and Natural Sciences, Universitas Kristen Satya Wacana,

Salatiga, Indonesia

Faculty of Health Science, Universitas Kristen Satya Wacana, Salatiga, Indonesia

Study Center for Multidicplinary Applied Research and Technology (SeMARTy), Universitas Kristen Satya Wacana,

Salatiga, Indonesia

Keywords: Near-infrared spectroscopy, PCA, blood glucose, non-invasive

Abstract: Diabetes is a metabolic disorder that is caused by unregulated blood glucose and therefore requires regular

and intensive monitoring. Currently, blood sugar monitoring is mostly done invasively by withdrawing

blood through a needle or piercing of the fingertips. This method can cause trauma and an infection.

However, there is the potential for using a non-invasive measurement of blood glucose levels with near-

infrared spectroscopy (NIRS) combined with partial least-square regression. As a pathway to actualize it,

the spectrum of whole blood was measured with different glucose levels. A total of 72 NIR spectrum from 8

whole blood samples with different types of glucose levels were measured. A principal component analysis

(PCA) and partial least square regression (PLSR) were applied to the spectral data matrix. The results

showed that PCA is successfully classified as spectral data based on the glucose content and PLSR model

within the clinically accurate region of the Clarke error grid. These results indicate that NIRS has an

immense potential to be applied in measuring blood glucose non-invasively.

1 INTRODUCTION

Diabetes is a disease caused by a deficiency of

insulin in the body (American Diabetes Association

(2004). This disease can increase or decrease blood

glucose levels. Under normal conditions, blood

sugar levels vary from 80-130 mg/dL. Insulin is

created by the pancreas to mediate metabolic

reactions with blood and maintain glucose levels in a

normal range (Torpy et al. 2014; Center for Disease

Control and Prevention, 2016. Uncontrolled diabetes

may result in various medical conditions such as a

stroke, kidney failure, heart disease, and blindness

(Center for Disease Control and Prevention, 2016).

Recently, the number of diabetics in the world

continues to increase. This increase makes it

necessary to be able to detect blood glucose levels.

This detection is important not only for those with

diabetes but also for non-diabetes people as a part of

their routine clinical monitoring. This monitoring

often requires fast, painless, non-invasive, and self-

measurement methods (Ferrante et al. 2008;

Kurasawa et al. 2017)

Various studies and developments of detection

models have been done, such as a glucometer design

to non-invasively check blood glucose by applying

NIR at a single wavelength (Saleh et al. 2018). This

is a promising method. Nevertheless, an assortment

of other organic compounds present on the tissue

can have an effect on the accuracy. Therefore,

spectrum-based measurements are significant to

boost the absorbance dynamics for a more thorough

analysis. Yano et al. (2001) discussed the possibility

of using NIR spectroscopy to simultaneously

estimate glucose and citric acid in an aqueous

solution of a blood anticoagulant. Zhang et al.

(2014) utilized two-dimensional correlation

spectroscopy (2DCS) to make the data analysis

better. Essential investigations into the NIR

spectrum of different organic samples have been

carried out since the 1970s. The NIR glucose

spectrum was also studied by Simeone et al. (2017)

and Yano et al. (2001). In addition, a number of

302

Rondonuwu, F., Setiawan, A. and Karwur, F.

Blood Glucose Determination by Fourier Transform near Infrared Spectroscopy.

DOI: 10.5220/0010163200002775

In Proceedings of the 1st International MIPAnet Conference on Science and Mathematics (IMC-SciMath 2019), pages 302-307

ISBN: 978-989-758-556-2

measurement techniques have been devised like NIR

Raman spectroscopy Zhang et al., 2013; Lam et al.,

2010 , direct diagnostics that utilize an NIR detector

chip implanted under the skin Saleh et al., 2018;

Uwadaira et al, 2016) and wireless long-term

constant observations (Dingari et al. 2011). The

NIRS approach with an exploration into different

spectral ranges and additional measurement methods

has also been reviewed (Pandey et al. 2017) and

studied using chemometrics.

Near-infrared spectroscopy (NIRS) facilitates the

workflow analysis and enables measurements of a

large number of samples in a reasonably quick

amount of time. It can gauge the concentration of

several constituents. In numerous instances, a

specific spectrum of a constituent can be connected

to its fingerprint. Samples containing functional

groups such as OH, CH, and NH are susceptible to

NIR due to the overtones of their fundamental

vibrations (R-H) in the IR region which match with

the NIR absorptions. Even though the C=C and C-C

bonds are not visible in the NIR region, their C-H

vibrational frequencies can reveal the C=C and C-C

bonds. The NIR absorption is commonly more

comprehensive compared to the IR absorption

because of the overlapping overtones and

combination bands discussed above. Consequently,

NIR analyses are complex and necessitate more

detailed processes. Fortunately, the development of

chemometrics enables NIR data to be utilized in

appropriate processes.

This research inspected the application of NIRS

to decide the glucose content in whole blood from a

healthy volunteer between a range of 80 and 130

mg/dL. This study strove to elaborate on the

previous measurements of glucose in an aqueous

solution and examine the likelihood of using NIRS

and PLSR as a substitute method to devise a non-

invasive blood glucose apparatus.

2 METHODS

2.1 Sample Preparations

In this experiment, lifeblood samples were retrieved

from a healthy volunteer with the intention of only

focusing on the effects of glucose. All the blood

drawings were completed within 120 minutes after

the volunteer had finished eating and drinking

sugary drinks. Blood drawings were taken in 15-

minute intervals using a lancet that punctured the

individual’s fingertips. The drawing procedures

followed standard measures using a portable

glucometer. The amount of blood drawn each time

was about one drop. A small amount of blood was

used to measure the blood sugar levels with a

glucometer while the rest was used for scanning by

NIRS. There was a total of 8 blood drawings with

the glucose levels indicated by a glucometer at 84,

86, 98, 100, 111, 116, 119, and 121 mg/dL. Within 2

hours, the blood sugar levels then returned to their

initial levels.

2.2 Data Acquisition

Each blood sample was put on a metal reflector

covered by optical glass. The space between the

glass and metal reflector was 0.2 mm (thus a 0.4 mm

path length with a double pass). A Fourier transform

near-infrared spectrometer (Buchi NIRFLEX 500

solid) with a spectral region of 4000-10000 cm

-1

was

applied in a reflectance mode using fiber optics to

test each of the sample spectra. Each spectrum had 4

-1

intervals (thus, each spectrum consisted of 1250

data points) and averaged over 32 measurements.

The sample temperatures were sustained at 26°C

during the spectral acquisitions. A total of 73 spectra

was collected with 9, 8, 9, 9, 11, 9, 8, and 10 spectra

for the blood samples with 84, 86, 98, 100, 111, 116,

119, and 121 mg/dL of glucose, respectively.

2.3 Data Analysis

A PCA analysis was applied for the 73 spectra after

the smoothing, normalizations, and derivatives. The

smoothing procedure was implemented using the

Savitzky-Golay method employing a third order

polynomial at a frame size of 21. Spectral

normalizations were applied to eliminate

multiplicative scattering and baseline variations. The

details for PLSR have been clarified elsewhere. A

total of 73 spectra were divided into two groups, 37

spectra for the calibration set, and 36 spectra for the

validation set. The calibration spectra were utilized

to devise a prediction model using the partial least

square regression (PLSR) method. PLSR attempts to

show the relationships between groups of observed

variables and latent variables. Validation spectra

were applied to cross-validate them by using the

PLSR parameters to estimate the concentrations of

the validation samples. Both PCA and PLSR

procedures were coded using Matlab version 2017b.

Blood Glucose Determination by Fourier Transform near Infrared Spectroscopy

303

3 RESULTS AND DISCUSSION

Red blood cells (erythrocytes), white blood cells

(leukocytes), platelets (thrombocytes), and plasma

are the main constituents of whole blood (Basu and

Kulkami 2014). Nearly half of whole blood is

plasma, and about 90% of it is actually water, while

the remaining 10% is protein. Therefore, the NIR

spectrum of whole blood resembles the NIR

spectrum of water. Figure 1 shows the NIR spectra

of 27 samples, the spectrum of glucose (red) and

water (blue). Water and glucose spectra are

presented to indicate their bands' positions and

widths to the whole blood spectra. The whole blood

spectra are characterized by the well-recognized two

strong water absorptions that appear at around 5200

-1

and 7000 cm

-1

. The absorption of glucose is

much weaker when compared to that of water.

However, a small skirt at around 4800 cm

-1

strongly

suggests that glucose is present in the whole blood.

Figure 1. Near-infrared diffuse-reflectance spectra of whole blood (black). For comparison, the diffuse-reflectance spectra

of powder glucose (red) and pure water (blue) are also shown. In this figure, the intensity of diffuse-reflectance spectra for

powder glucose and water are unscaled.

Although the whole blood spectral patterns are

similar to each other, the baseline and intensity are

relatively different. They do not entirely overlap due

to the baseline variations and multiplicative

scattering. For PCA and PLSR analyses, each

spectrum was normalized to avoid a multiplicative

scattering effect and then the first derivative was

taken to correct the baseline variations. Figure 2

shows the first derivative spectra. The box with the

broken line shown at 4200-5000 cm

-1

indicates this

study’s target region for analyses (vide infra). The

inset at the top right corner of the figure enlarges the

spectral structure around the target region.

Figure 2. First derivative spectra of the whole blood. The target region for the analyses is indicated by a box with a broken

line and enlarged in the inset for clarity.

IMC-SciMath 2019 - The International MIPAnet Conference on Science and Mathematics (IMC-SciMath)

304

A singular value decomposition (SVD) analysis of

the data matrix was applied to determine the spectral

region that effectively contributed to the PCA and

PLSR models. It shows that the effective spectral

region is 4200 - 500 cm

-1

. The details of the PCA

and PLSR calculations used in this paper are

published elsewhere (Abdi, 2010; Tharwat (2016).

The PCA and PLSR models were then calculated

using a

target spectral region of 4200 - 5000 cm

-1

. The

specified target region was chosen based on the

minimum predicted residual error sum of square

(PRESS) at the optimum number of latent variables

in PLSR calculations. Figure 3 depicts the PCA plot

of the 73 spectral data. Each of the eight groups has

different glucose clustering at their unique pc1- pc2

-pc3 space coordinates except one spectrum of 116

mg/dL located at the 84mg/dL group coordinate.

Figure 3. PCA results of seventy-three spectral data. The spectra are grouped into eight clusters based on their glucose

content in whole blood. The value indicated in each group designates the glucose content in mg/dL.

In the PLSR analysis, the 73 spectral data were

divided into two parts; 37 spectra were used for the

calibration data set, while the remaining 36 were

used for the validation data set. It was ensured that

in each set, the spectrum from samples containing

glucose of 84, 86, 98, 100, 111, 116, 119, and 121

mg/dL were represented in almost equal numbers.

Initially, the regression parameters were calculated

using a calibration data set. Finally, these parameters

were employed to predict the glucose levels through

the corresponding spectrum in the validation data

set. The region for analysis remained the same as

that used in the PCA analysis. The number of latent

variables was N=8, by which the PRESS value

reached a minimum. The latent variable obtained in

this analysis was the same as that obtained for cases

of glucose in an aqueous solution. Figure 4 shows

the results of the NIR predictions compared to the

ones measured by electrode strips (reference). The

blue circles indicate the results of the NIR

predictions of the glucose content for samples in the

calibration data set, while the red ones represent NIR

predictions for the validation data set. The

coefficients of the determination (R

) were 0.97 and

0.75 for the calibration and validation, respectively.

For PRESS, the obtained values were 70.7 and 181.4

for the calibration and validation, respectively.

Normally, both the R

and PRESS for the validation

data will be smaller than that of the calibration data

set, as in the present case.

Blood Glucose Determination by Fourier Transform near Infrared Spectroscopy

305

Figure 4. Comparisons between the NIR predictions and the reference values of the glucose content in whole blood. The

blue and red circles represent the calibration and validation sets, respectively. The broken blue line shows the linear fitting

for the calibration data set, while the broken red line shows the linear fitting for the validation data set. The solid black line

is displayed for guidance purposes.

The results of the NIR prediction of glucose in the

present study appear relatively scattered as though

less precise when compared to the NIR predictions

obtained in previous studies Rondonuwu et al., F

(2019). In the previous studies, however, the glucose

contents were systematically prepared at a certain

level so that the reference values were highly

accurate. In this study, the reference values solely

relied on the strip electrode measurements using a

glucometer that has relatively large random

deviation values. To evaluate the NIR predictions of

the validation set in terms of clinical accuracy, we

applied the Clarke error grid analysis. All of the NIR

prediction data points in the validation data set were

then transferred into the Clarke diagram, as

indicated in Figure 5. In the Clarke diagram, region

A was estimated to be clinically accurate, while

region B was considered clinically acceptable. In

this model, all of the 36 NIR predictions fall into

region A, which means they are clinically accurate.

Note that the examination range in this study is

limited within 80 to 130 mg/dL, which is the range

of healthy subjects. A more extensive range is

necessary but requires diabetic volunteers. In this

study, diabetic volunteers were not employed since

they need specialized medical attention, and it is

relatively challenging to promote them with extra

food and beverages with high calories.

Figure 5. Clarke error grid diagram. The red circles indicate the NIR predictions of the validation data set.

IMC-SciMath 2019 - The International MIPAnet Conference on Science and Mathematics (IMC-SciMath)

306

In this study, we only employed a single volunteer

for a short time, which means the only differential

factor affecting the measurement is glucose. The rest

of the constituents, including lipids, proteins, and the

physical parameters such as viscosity and

temperature, are practically unaltered. In a more

realistic model, different volunteers having a variety

of whole blood conditions must be included in a

database of PLSR calculations to increase the

validity and accuracy of the method.

4 CONCLUSION

The glucose content in whole blood can be

determined by employing near-infrared spectroscopy

followed by the partial least squares regression

model in the 4200-5000 cm

-1

spectral region. Based

on the Clarke error grid, the NIR spectroscopy

technique followed by a PLSR analysis from a

single volunteer in the 80-130 mg/dL was

successfully predicted to be clinically accurate.

These results shed light on the NIRS technique

followed by the PLSR calculation to provide an

effective and non-invasive approach to measure

blood glucose levels.

ACKNOWLEDGMENTS

This research was supported by a grant (ID#:

c4c249c1-1f90-432b-affd-45d9544c8e89) from the

Ministry of Research, Technology, and Higher

Education, of the Republic of Indonesia, the

Directorate Research and Community Service

(DRPM) under the scheme of PDUPT 2019.

REFERENCES

Abdi H 2010 Partial least squares regression and

projection on latent structure regression (PLS

Regression) (London: Wiley Interdisciplinary

Reviews: Computational Statistics) chapter 2 pp 97–

106

American Diabetes Association 2004 Diagnosis and

classification of diabetes mellitus. Diabetes Care, pp

27, 5–10.

Basu D, Kulkami R 2014 Indian Journal of Anaesthesia 58

529.

Center for Disease Control and Prevention 2016 National

Diabetes Statistics Report: Estimates of Diabetes and

Its Burden in the United States.

Dingari N C, Barman I, Singh G P, Kang J W, Dasari R R

and Feld M S 2011 Analytical and Bioanalytical

Chemistry 400 2871

Ferrante do Amaral C E, Wolf B 2008 Med. Eng. Phys.,

30, 541

Kurasawa S, Koyama S, Ishizawa H, Fujimoto K and

Chino S 2017 Sensors 17 2702

Lam S C H, Chung J W Y, Fan K L and Wong T K S 2010

Spectroscopy 24 629

Maria L F Simone, Rafael A. C. Parrela, Robert E

Schaffert, Cynthia M B Damasceno 2017

Microchemical Journal 134 125

Oliver N S, Toumazou C, Cass A E, Johnston D G 2009

Diabet. Med., 26, 197.

Pandey R, Paidi S K, Valdez T A, Zhang C, Spegazzini N,

Dasari R R and Barman I 2017 Accounts of Chemical

Research 50 264

Rondonuwu F S, Setiawan A, Karwur F F 2019 J. Phys.:

Conf. Ser. 1307 012019

Robinson M R, Eaton R P, Haaland D M, Koepp G W,

Thomas E V, Stallard B R, Robinson, P L 1992 Clin.

Chem., 38, 1618

Saleh G, Alkaabi F, Al-Hajhouj N, Al-Towailib F, Al-

Hamza S 2018 J. Med. Eng. Technol 42(2) 140

Sekulic S, Wakemen J, Doherty P, Hailey P 1998 J.

Pharm. Biomed. Anal. 17 1285

Simeone M L F, Parrella R A C, Schaffert R E,

Damasceno C M B, Leal M C B and Pasquini C 2017

Microchemical Journal 134 125

Torpy J M, Lynm, C, Glass R M 2014 Tharwat A 2016

International Journal of Applied Pattern Recognition

197 1

Uwadaira Y, Ikehata A, Momose A and Miura M 2016

Biomedical Optics Express 7 2729

Yano T, Funatsu T, Suehara KI, and Nakano Y 2001

JNIRS 9 43

Zhang W, Liu R, Zhang W, Jia H and Xu K 2013

Biomedical Optics Express 4 789

Zhang C and Su J 2014 Acta Pharmaceutica Sinica B

4 182

Blood Glucose Determination by Fourier Transform near Infrared Spectroscopy

307