Noninvasive and fast measurement of blood glucose in vivo by near infrared (NIR) spectroscopy

Accepted Manuscript Noninvasive and fast measurement of blood glucose in vivo by near infrared (NIR) spectroscopy

Xue Jintao, Ye Liming, Liu Yufei, Li Chunyan, Chen Han PII: DOI: Reference:

S1386-1425(17)30128-2 doi: 10.1016/j.saa.2017.02.032 SAA 14954

To appear in:

Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy

Received date: Revised date: Accepted date:

25 November 2016 15 January 2017 16 February 2017

Please cite this article as: Xue Jintao, Ye Liming, Liu Yufei, Li Chunyan, Chen Han , Noninvasive and fast measurement of blood glucose in vivo by near infrared (NIR) spectroscopy. The address for the corresponding author was captured as affiliation for all authors. Please check if appropriate. Saa(2017), doi: 10.1016/j.saa.2017.02.032

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

Noninvasive and Fast Measurement of Blood Glucose in vivo by Near Infrared (NIR) Spectroscopy Xue Jintao1,2, Ye Liming*2, Liu Yufei1, Li Chunyan1,3, Chen Han2 1. School of Pharmacy, Xinxiang Medical University, Xinxiang 453002, Henan Province, PR China 2. West China School of Pharmacy, Sichuan University, Chengdu 610041, Sichuan Province, PR China

RI

PT

3. Sanquan Medical College, Xinxiang 453002, Henan Province, PR China.

SC

Corresponding author ： Ye Liming, e-mail: [email protected], tel: 8615090093659, postal address: School of Pharmacy, Xinxiang Medical University,

AC

CE

PT E

D

MA

NU

Xinxiang 453002, Henan Province, PR China.

ACCEPTED MANUSCRIPT Abstract This research was to develop a method for noninvasive and fast blood glucose assay in vivo. Near-Infrared (NIR) spectroscopy, a more promising technique compared to other methods, was investigated in rats with diabetes and normal rats. Calibration models are generated by two different multivariate strategies: partial least

PT

squares (PLS) as linear regression method and artificial neural networks (ANN) as non-linear regression method. The PLS model was optimized individually by

RI

considering spectral range, spectral pretreatment methods and number of model

SC

factors, while the ANN model was studied individually by selecting spectral pretreatment methods, parameters of network topology, number of hidden neurons,

NU

and times of epoch. The results of the validation showed the two models were robust, accurate and repeatable. Compared to the ANN model, the performance of the PLS

MA

model was much better, with lower root mean square error of validation (RMSEP) of 0.419 and higher correlation coefficients (R) of 96.22%.

D

Keywords: near-infrared spectroscopy; neural networks; partial least squares;

Abbreviations

PT E

mathematical models; bioanalysis.

CE

2h-PG (2h post- OGTT plasma glucose), ANN (artificial neural networks), COE (Constant Offset Elimination), HFD (high fat diet), MMN (Min Max Normalization),

AC

MSC (Multiplicative Scatter Correction), MSE (mean square rrror), NIR (Near-Infrared Spectroscopy), OGTT (oral glucose tolerance test), PLS (partial least-squares), R (correlation coefficient), Rcal (the correlation coefficient of the calibration set), Rpre (correlation coefficient of prediction set), RMSECV (the root mean square error of cross-validation), RMSEP (the root mean square error of validation), RPD (the residual predictive deviation), SLS (Straight Line Subtraction), STZ (streptozotocin), VN (Vector Normalization).

ACCEPTED MANUSCRIPT

1. Introduction Nowadays, noncommunicable chronic diseases have become the leading causes of mortality and disease burden worldwide as the result of changes in human behaviour and lifestyle in recent decades [1, 2]. The past two decades have seen a dramatic

PT

increase in the incidence of diabetes worldwide, and the mortality from diabetes doubled and increased to 1.3million deaths worldwide [1, 3, 4]. In China, the

RI

prevalence of diabetes also has increased significantly. In a subsequent national

SC

surveys which was conducted in 2000-2001, the prevalence of diabetes was 5.5%, and the most recent national survey in 2007 was 9.7%, representing an estimated 92.4

NU

million adults in China with diabetes. Diabetes is likely to remain a huge threat to public health in the years to come, which has been declared a global epidemic by

MA

World Health Organization [1, 2].

Diabetes mellitus is a chronic disease in which the blood glucose level is higher

D

than normal [4]. According to the Diabetes Control and Complications Trial Research

PT E

Group, most of the long-term complications was associated with diabetes, such as heart disease, stroke, kidney damage, blindness and nerve damage, result from sustained hyperglycemia (blood glucose exceeding 120 mg·dL-1). In addition,

CE

Hypoglycemia (blood glucose concentrations less than 60 mg·dL-1) can lead to insulin shock as well as death [3, 5, 6]. It is widely agreed that diabetics need to supervise

AC

their glucose levels closely and measure them several times a day for the maintenance of blood glucose level within the physiological range [4, 7]. Current blood glucose monitoring for the self-monitoring is accomplished through invasive methods, such as a finger prick for withdrawing a drop of blood. It requires that a diabetic patient suffer pain in several times a day and risk infection, and it is also expensive due to the number of test strips. For many years it has been assumed that the above reasons were why blood glucose tests were not carried out as often as recommended [7, 8]. Therefore, a method for non-invasive and fast blood glucose

ACCEPTED MANUSCRIPT assay which can not only detect it painlessly, safely and duly by patients themselves, but also be less expensive is highly desired [9]. Notably NIR and Raman spectroscopy has shown substantial promise in this regard [6, 7]. NIR with chemometrics holds great promise for clinical chemistry measurements on the basis of the light in tissues in the range of 1–100 mm of depths for noninvasive

PT

assay, nondestructive, and reagent-less [10]. NIR is based on focusing on the body a beam of light in the 12000~4000cm-1 spectrum. The most prominent absorption bands

RI

of NIR are related to the overtones and combinations of fundamental vibrations

SC

exhibited by –CH, –NH, –OH and –SH functional groups [8, 11].

In our previous study, NIR or Raman spectroscopy were investigated in the

NU

artificial plasma [12], skin tissue phantom and whole blood in vitro to develop a method for noninvasive and fast blood glucose assay. In the study of artificial plasma,

MA

the NIR and Raman spectroscopy models were generated by performing PLS and validated for the determination of glucose, and the results show the models

D

established can be used for noninvasive measurement of glucose, and the performance

PT E

of NIR spectroscopy was better than Raman spectroscopy [12]. In the research of skin tissue phantom, the NIR model was generated successfully by PLS regression, while Raman spectroscopy is inappropriate for glucose noninvasive measurement as the

CE

reason of the noise and matrix background interference [13-15]. In whole blood, The NIR model established are robust, accurate and repeatable for fast and noninvasive

AC

blood glucose assay.

On the basis of the above previous study, this research was to develop a method for noninvasive and fast blood glucose assay in vivo. Rats with diabetes was induced by high fat diet (HFD) for 4 weeks and streptozotocin (STZ). The calibration models were constructed by using PLS as linear models and ANN as non-linear calibration models, respectively. The two models were was optimized individually and validated for the determination of glucose.

ACCEPTED MANUSCRIPT 2. Materials and Methods 2.1 Reagents Streptozotocin (STZ) (Sigma, USA, batch NO.: B64927); glucose injection (Wuhan Fuxing Bio-Pharmceutical Co., Ltd., Wuhan, China, specification: 20ml: 10g); water was purified by an ultrapure water instrument. All other reagents were of

PT

analytical grade. 2.2 Animals and Experimental Design

RI

Adult male Sprague-Dawley rats weighing 180–200g were obtained from

SC

Laboratory Animal Center, Sichuan University (Sichuan, China). Each rat was kept under controlled temperature (20±2ºC) and lighting (08:00–20:00) conditions with

NU

food and water available ad libitum. After being fed a standard diet for 1 weeks while acclimating to the facility, the animals were randomly divided into 2 groups: the

MA

normal group (n=12) and the hyperglycemia group (the diabetes model group, n=18). The hyperglycemia group: Rats were then fed HFD, added with 12% W/W

D

coconut oil and 9% W/W glucose, for 4 weeks. After 12 h starvation period with

PT E

enough water, 1% STZ in citric acid - sodium citrate buffer (pH4.2-4.5) intraperitoneally injected at a dose of 40 mg·kg-1 body weight [16-18]. The normal group: Fed a standard diet, and 4 weeks later, citric acid - sodium

CE

citrate buffer (pH4.2-4.5) was injected in intraperitoneal for 40mg·kg-1 according to body weight after 12 h starvation period with enough water.

AC

Two weeks after induced to diabetes, 50% W/V glucose injection was administered a dose of 2.5g•kg-1 body weight by gastric perfusion in the morning after a 9 h starvation period, then the rats’ plasma samples and spectra were collected simultaneously at certain time points as follows: 0min, 15min, 30min, 45min, 60min, 90min, 120min, 180min and 360min after glucose injection. The whole experiment is completely followed by the Animal Ethics protocol and was approved by the Animal Ethics Committee of Sichuan University. 2.3 Data Collection

ACCEPTED MANUSCRIPT The NIR spectra at a resolution of 8 cm−1 over a wavelength range of 12000–4000cm-1 were recorded with 32 scans per spectrum using a Bruker Matrix-F FT-NIR spectrometer (Bruker Optik, Ettlingen, Germany) equipped with a PbS detector and a fiber optic probe. The system was operated by OPUS software (Bruker Optik, Ettlingen, Germany). According to the time points in 2.2 Animals and

PT

Experimental Design and the process shown in Fig.1, the NIR spectra was collected with 2 times measured for each sample, and Fig. 1D shows the original NIR spectra of

RI

all the rats’ hind leg. The abnormal spectra was recommended by the OPUS software

SC

and removed manually during the modeling process. The analysis process was at room temperature (25℃) with the humidity at ambient level in the laboratory.

NU

The blood glucose were measured by Hitachi 7020 automatic biochemical analyzer (Hitachi, Tokyo, Japan) with the corresponding kit (Maccura Biotechnology

MA

Co., Ltd., Sichuan, China). 2.4 Data processing

D

The most prominent absorption bands of NIR are related to the overtones and

PT E

combinations of fundamental vibrations exhibited by –CH, –NH, –OH and –SH functional groups [11]. The intensity of the measurements at different wavenumbers can be correlated to the concentrations of the relevant components in the sample

CE

through a series of mathematical procedures such as multivariate statistical calculations such as multiple linear regression, principal component regression, the

AC

partial least squares (PLS). These calibration approaches assume a linear relationship between the measured parameters for the sample and the intensity of its absorption bands. PLS was the most frequently used in these methods [11, 19]. At the same time, the presence of substantial non-linearity. e.g., those that arise from scattered light or intrinsic non-linearity in the absorption bands, call for the use of alternative calibration procedures to correct non-linearity, and ANN are also among the most widely used mathematical algorithms for overcoming non-linearity [20-21].

ACCEPTED MANUSCRIPT Therefore the NIR calibration models were constructed respectively by PLS with the OPUS software and ANN with NeuroSolutions for Excel 6.30 (Neurodimension Inc., Gainesville, FL). In each algorithm, the calculation strategy was as following: the content range of the calibration set must be wide enough to cover the range of the validation set, so the

PT

samples with maximum and minimum concentrations were selected in the calibration set, then the samples were selected randomly for each set based on concentration. In

RI

the PLS algorithm, 60 samples were selected randomly for the validation set and the

SC

remaining 131 samples were for the calibration set. In the ANN algorithm, the total data was split into three set: 134 samples were selected randomly as the training set

NU

for generating and training the networks, and 29 samples were selected randomly as test set for the validation of the networks. The remaining 28 samples as the validation

D

3. Results and discussion

MA

set were for the evaluation and validation of the optimal ANN model.

PT E

3.1 Development of the PLS model

To develop a robust model, different spectral range, spectral pretreatment methods and number of model factor are often selected to eliminate noise and matrix

CE

background interference, and enhance the spectral features to extract the relevant information before PLS modeling [6, 12]. In this process, the PLS model is validated

AC

by the leave-one-out cross-validation (LOOCV) algorithm where each sample of the calibration set is removed, predicted, and replaced in a sequential manner. To assess the predictive ability, every PLS methods were computed and selected basing on the correlation coefficients of the calibration set (Rcal), the root mean square errors of cross-validation (RMSECV) and the residual predictive deviation (RPD). The model with the best prediction ability was chosen according to highest R (both in calibration and validation) and RPD as well as lowest RMSECV and RMSEP were considered optimal [11, 22].

ACCEPTED MANUSCRIPT As shown in Table 1, 10 kinds of spectral pretreatment methods including: Multiplicative Scatter Correction (MSC), Constant Offset Elimination (COE), Vector Normalization (VN), First Derivative, Second Derivative, Min Max Normalization (MMN), Straight Line Subtraction (SLS), First Derivative + SLS, First Derivative + VN and First Derivative + MSC, were compared in our study, and COE showed better

PT

performance. The selection of wavenumber range was selected individually. On the basis of the

RI

absorption peak as shown in Fig 1D, the spectra range were divided into 4 regions:

SC

11995.6-7502.0 cm-1 (A), 7502.0-6098.0 cm-1 (B), 6098.0-4601.5 cm-1 (C) and 4601.5-4246.7 cm-1 (D). According to Table 2, the results indicated that the

NU

performance of 4 spectral regions was C>B>A>D, then the combination of C and other spectral regions was compared, and the optimal spectral region was

MA

7502~4246.7 cm-1 for the PLS model.

The number of model factors (F) was studied as shown in Fig 2. RMSECV and

D

Rcal was used to choose the optimum LVs. In the PLS algorithm, much more or less F

according to Fig 2.

PT E

will lead to the overfitting or underfitting, so the optimal F of this PLS model was 10

3.2 Development of the ANN model

CE

The ANN models were trained with multilayer perceptron to generate a feed-forward network by the back-propagation which refers to a process of

AC

propagating the error information backward from the output to the hidden neurons, during which connection weights were modified by the delta learning rule. A gradient descent method is used to minimize the error. The model was chosen according to these evaluation parameters: higher R (both in calibration and validation), and lower mean square error (MSE) and RMSEP [23, 24]. To develop a robust model, the ANN model was optimized individually by selecting spectral pretreatment methods, parameters of network topology, number of hidden neurons, and times of epoch. As shown in table 3, 3 kinds of spectral

ACCEPTED MANUSCRIPT pretreatment methods (Untreated, Vector Normalization and Min max normalization) were compared; effect of the number of hidden neurons on the network performance has been studied with distinctly different architectures, transfer function, step size and momentum of hidden and output layer neurons were selected in the study of network topology; at last, times of epoch were selected and recommended by NeuroSolutions

PT

6.30 to prevent over fitting or early stopping. According to the above criteria of the optimal ANN model, the best parameters for

RI

the ANN model was the ANN-3 model as shown in Table 3.

SC

3.3 Evaluation and validation for the PLS and ANN models

The validation set were used to validate the predictive ability of the optimized

NU

PLS model as linear regression model and the optimized ANN model as non-linear regression model, respectively. As shown in Fig 3, the R and RMSEP of validation set

MA

for PLS model were 96.22 and 0.419, respectively, and the R and RMSEP for ANN model are 92.79 and 0.5602, respectively. The results show that the established 2

D

types of models give satisfactory fitting results and predictive ability, and the T and F

PT E

tests which was used to compared the predicted values of the 2 models to the reference values indicated that the accuracy was satisfactory with a significant level of 0.05, so the 2 models established are robust, accurate and repeatable.

CE

Compared to the ANN model, the performance of the PLS model was much better, with lower RMSEP of 0.419 and higher Rval of 96.22%, so the PLS model was the

AC

best model for noninvasive and fast measurement of blood glucose in vivo. 3.4 Laboratory Parameters of Animal Experiment As shown in Fig. 4, significant lesion/abnormality was observed in pancreatic tissue morphology in diabetes rats compared to normal rats. The pancreatic tissue of normal group (Fig. 4A) had round or oval cells, and no structure change in pancreatic islets; while The pancreatic tissue of diabetes rats (Fig. 4B) had pancreatic islet cells destroyed and reduced.

ACCEPTED MANUSCRIPT The oral glucose tolerance test (OGTT) is the gold standard for diagnosing diabetes. According to OGTT, diabetes were diagnosed when the fasting plasma glucose (FPG) and 2 h post- OGTT plasma glucose (2 h-PG)was exceed 7.0 mmol·L-1 (1.26 mg·ml-1) and 11.1 mmol·L-1 (2.00 mg·ml-1) , respectively [25, 26]. The OGTT of normal group and hyperglycemia group was showed in Fig. 5 with FPG of 0.620 ±

PT

0.129 mg/ml and 2.095±0.698 mg/ml, 2 h-PG of 1.333±0.080 mg/ml and 4.938±0.424 mg/ml.

RI

Based on the results of tissue morphology and OGTT, the rats with diabetes was

SC

induced successfully in our research.

NU

4. Conclusions

In this study, NIR spectroscopy provided rapid and noninvasive analysis for blood

MA

glucose in vivo. The PLS model as linear regression model and the ANN model as non-linear regression model established was robust, accurate and repeatable. As the

D

PLS model shows good performance and extensively potential in noninvasive

PT E

measurement of blood glucose, it will be carried out for human in vivo in our future

Funding

CE

study.

This work was supported by the National Natural and Science Foundation of

AC

China for Youth Program (grant numbers 21505114); the University Key Research Projects of Henan Province (grant numbers 17A360026) and the Cultivation Fund of Xinxiang Medical University (grant numbers 505095). The authors are extremely grateful for the financial support.

References [1] P. Zimmet, K. Alberti, J. Shaw, Global and societal implications of the diabetes epidemic, Nature, 414 (2001) 782-787.

ACCEPTED MANUSCRIPT [2] Y. Xu, L. Wang, J. He, Y. Bi, M. Li, T. Wang, L. Wang, Y. Jiang, M. Dai, J. Lu, M. Xu, Y. Li, N. Hu, J. Li, S. Mi, C.S. Chen, G. Li, Y. Mu, J. Zhao, L. Kong, J. Chen, S. Lai, W. Wang, W. Zhao, G. Ning, Prevalence and control of diabetes in Chinese adults, JAMA, 310 (2013) 948-959. [3] S.K. Vashist, Non-invasive glucose monitoring technology in diabetes

PT

management: A review, Anal. Chim. Acta., 750 (2012) 16-27. [4] J. Shao, M. Lin, Y. Li, X. Li, J. Liu, J. Liang, H. Yao, In vivo blood glucose

RI

quantification using Raman spectroscopy, Plos One, 7 (2012) e48127-e48127.

SC

[5] M.W. Aslam, Z. Zhu, A.K. Nandi, Feature generation using genetic programming with comparative partner selection for diabetes classification, Expert. Syst. Appl., 40

NU

(2013) 5402-5412.

[6] R.S. Parker, F.J. Doyle, 3rd, N.A. Peppas, The intravenous route to blood glucose

MA

control, IEEE. Eng. Med. Biol. Mag., 20 (2001) 65-73. [7] Z. Chuah, R. Paramesran, K. Thambiratnam, P. Sin-Chew, A two-level partial

D

least squares system for non-invasive blood glucose concentration prediction,

PT E

Chemomtr. Intell. Lab., 104 (2010) 347-351. [8] Y. Yue, S. Zhenzhi, L. Chenxi, C. Wenliang, X. Kexin, Simulation and validation of the radial reference point in non-invasive blood glucose sensing by NIR,

CE

Nanotechno. Precis. Eng., 8 (2010) 114-119. [9] K. Shin, H. Chung, Wide area coverage Raman spectroscopy for reliable

AC

quantitative analysis and its applications, Analyst, 138 (2013) 3335-3346. [10] A. Tura, A. Maran, G. Pacini, Non-invasive glucose monitoring: Assessment of technologies and devices according to quantitative criteria, Diabetes. Res. Clin. Pr., 77 (2007) 16-40. [11] J. Xue, C. Wu, L. Wang, S. Jiang, G. Huang, J. Zhang, S. Wen, L. Ye, Dynamic prediction models for alkaloid content using NIR technology for the study and online analysis of parching in Areca Seed, Food Chem., 126 (2011) 725-730.

ACCEPTED MANUSCRIPT [12] J. Xue, H. Chen, D. Xiong, G. Huang, H. Ai, Y. Liang, X. Yan, Y. Gan, C. Chen, R. Chao, L. Ye, Noninvasive measurement of glucose in artificial plasma with near-infrared and Raman spectroscopy, Appl. Spectrosc., 68 (2014) 428-433. [13] K. Shin, H. Chung, Wide area coverage Raman spectroscopy for reliable quantitative analysis and its applications, Analyst., 138 (2013) 3335-3346.

PT

[14] D.I. Ellis, D.P. Cowcher, L. Ashton, S. O'Hagan, R. Goodacre, Illuminating disease and enlightening biomedicine: Raman spectroscopy as a diagnostic tool, The

RI

Analyst, 138 (2013) 3871-3884.

SC

[15] E.E. Rossi, A.L. Pinheiro, O.C. Baltatu, M.T. Pacheco, L. Silveira, Jr., Differential diagnosis between experimental endophthalmitis and uveitis in vitreous

NU

with Raman spectroscopy and principal components analysis, J. Photochem. Photobiol. B., 107 (2012) 73-78.

MA

[16] S. Chengjun, Y. Zhenjun, Z. Meirong, C. Zhihong, Sericin protects against diabetes-induced injuries in sciatic nerve and related nerve cells, Neural Regen. Res.,

D

8 (2013) 506-513.

PT E

[17] W. Qu, Z. Jiang, c. Zhang, J. Zou, L. Sun, Y. Shi, Z. Liu, Regulation of C-type Natriuretic Peptides and Natriuretic Peptide Receptor-B Expression in Diabetic Rats Renal Treated by Tongluo Recipe, Chin. J. Integr. Med., 19 (2013) 524-531.

CE

[18] H. Dong, J. Wang, F. Lu, L. Xu, Y. Gong, X. Zou, Jiaotai Pill Enhances Insulin Signaling through Phosphatidylinositol 3-Kinase Pathway in Skeletal Muscle of

AC

Diabetic Rats, Chin. J. Integr. Med., 19 (2013) 668-674. [19] N. Heigl, A. Greiderer, C.H. Petter, O. Kolomiets, H.W. Siesler, M. Ulbricht, G.K. Bonn, C.W. Huck, Simultaneous determination of the micro-, meso-, and macropore size fractions of porous polymers by a combined use of Fourier transform near-infrared diffuse reflection spectroscopy and multivariate techniques, Anal. Chem., 80 (2008) 8493-8500. [20] Ş. Ercan, A. Uğur, Artificial neural network models for lot-sizing problem: a case study, Neural Comput. Appl., 6 (2013) 1039-1047.

ACCEPTED MANUSCRIPT [21] F. Olugbenga, O. Sunday, Predicting the Onset of Asphaltene Precipitation in Heavy Crude Oil Using Artificial Neural Network, Chem. Process Eng. Res., 15 (2013) 1-10. [22] X. Niu, Z. Zhao, K. Jia, X. Li, A feasibility study on quantitative analysis of glucose and fructose in lotus root powder by FT-NIR spectroscopy and chemometrics,

PT

Food. Chem., 133 (2012) 592-597. [23] N. Rizkalla, P. Hildgen, Artificial neural networks: comparison of two programs

RI

for modeling a process of nanoparticle preparation, Drug Dev. Ind. Pharm., 31 (2005)

SC

1019-1033.

[24] Ali Mirsepahi , Lei Chenb, Brian O'Neilla, A comparative approach of inverse

NU

modelling applied to an irradiative batch dryer employing several artificial neural networks, Int. Commun. Heat Mass, 53 (2014) 164-173.

MA

[25] A. Alyass, P. Almgren, M. Akerlund, J. Dushoff, B. Isomaa, P. Nilsson, T. Tuomi, V. Lyssenko, L. Groop, D. Meyre, Modelling of OGTT curve identifies 1 h

D

plasma glucose level as a strong predictor of incident type 2 diabetes: results from two

PT E

prospective cohorts, Diabetologia, 58 (2015) 87-97. [26] O. Helminen, S. Aspholm, T. Pokka, J. Ilonen, O. Simell, R. Veijola, M. Knip, OGTT and random plasma glucose in the prediction of type 1 diabetes and time to

AC

CE

diagnosis, Diabetologia, 58 (2015) 1787-1796.

ACCEPTED MANUSCRIPT

MA

NU

SC

RI

PT

Figures

D

Fig. 1 The process of collecting NIR spectra (A - rat’s hind leg shaved; B - the

AC

CE

PT E

NIR fiber-optical probe; C - collection of the NIR spectra; D - NIR spectra).

SC

RI

PT

ACCEPTED MANUSCRIPT

AC

CE

PT E

D

MA

NU

Fig. 2 Dependency of RMSECV and R on number of model factor

SC

RI

PT

ACCEPTED MANUSCRIPT

AC

CE

PT E

D

MA

NU

Fig. 3 Concentration scatter plot of the validation set

SC

RI

PT

ACCEPTED MANUSCRIPT

AC

CE

PT E

D

MA

hyperglycemia group, ×20).

NU

Fig. 4 Microscopic photographs of the pancreatic tissue (A: Normal group, ×10; B:

RI

PT

ACCEPTED MANUSCRIPT

SC

Fig. 5 The changes of the rats’ blood glucose after glucose injection (G:

AC

CE

PT E

D

MA

NU

hyperglycemia group; N: normal group).

ACCEPTED MANUSCRIPT

Rcal (%)

1

Straight Line Subtraction

94.49

0.331

4.26

0.424

96.09

2

Constant Offset Elimination

94.24

0.338

4.17

0.419

96.22

3

First Derivative

93.37

0.364

3.89

0.439

95.86

4

First Derivative + SLS

92.38

0.388

3.62

0.454

95.64

5

Second Derivative

89.74

0.453

3.12

0.858

86.90

0.396

3.62

0.723

90.63

90.16

0.445

3.19

0.764

91.22

88.71

0.470

2.89

0.564

92.92

89.75

0.452

3.12

0.816

87.85

85.92

0.528

2.67

0.609

92.70

92.37

NU

6

First Derivative + VN

9

Min max normalization

D

8

PT E

Vector Normalization

MA

Correction 7

AC

CE

10 First Derivative + MSC

SC

Multiplicative Scatter

RMSECV RPD RMSEP

RI

Spectral Pretreatment Method

PT

Table 1. Influence of different pretreatment methods for PLS model Rval (%)

ACCEPTED MANUSCRIPT

Rcal (%)

RMSECV

RPD

RMSEP

Rval (%)

A

11995.6-7502.0

85.79

0.530

2.65

0.974

81.41

B

7502.0-6098.0

69.29

0.788

1.80

0.888

83.40

C

6098.0-4601.5

91.94

0.403

3.52

0.463

95.88

D

4601.5-4246.7

79.67

0.644

2.22

0.985

80.87

92.72

0.384

3.71

95.93

RI

Spectral region (cm-1)

PT

Table 2. Influence of different NIR spectral regions for PLS model

SC

11995.6-7502.0

0.443

C+B

7502.0-4601.5

94.11

0.343

4.12

0.421

96.20

C+D

6098.0-4246.7

91.09

0.425

3.35

0.467

95.82

C+B+A

11995.6-4601.5

94.23

0.343

4.16

0.425

96.16

C+B+D

7502.0-4246.7

94.24

0.338

4.17

0.419

96.22

C+A

D

PT E CE AC

MA

NU

6098.0-4601.5

ACCEPTED MANUSCRIPT

Table 3. The and information of the ANN models parameters

ANN-1

ANN-2

ANN-3

Min max

Vector

normalization

Normalization

spectral pretreatment method 18

7

transfer function

TanhAxon

TanhAxon

step size

0.9

0.9

momentum

0.7

0.7

step size

0.15

momentum

0.7

7

TanhAxon 0.9 0.7 0.15

0.7

0.7

3382

3912

4.7×10-4

3.7×10-4

4.0×10-4

5250

MA

times of epoch

0.15

NU

output layer

SC

hidden layer

RI

number of hidden neurons

PT

Untreated

MSE

set

Rcal (%)

98.86

99.12

99.14

RMSEP

0.6332

0.8392

0.5602

89.65

80.52

92.79

validation set

AC

CE

PT E

Rval (%)

D

calibration

21

ACCEPTED MANUSCRIPT

SC

RI

PT

NIR was for non-invasive and fast blood glucose assay in vivo in rats with diabetes. Two multivariate strategies, partial least squares (PLS) as linear regression method and artificial neural networks (ANN) as non-linear regression method, were studied.

AC

CE

PT E

D

MA

NU

Graphical abstract

22

ACCEPTED MANUSCRIPT Highlights 1. NIR has investigated for non-invasive and fast blood glucose assay in vivo in rats with diabetes and normal rats. 2. Two different multivariate strategies were compared: partial least squares (PLS) as linear regression method, and artificial neural networks (ANN) as non-linear

PT

regression method.

AC

CE

PT E

D

MA

NU

SC

RI

3. The results showed the two methods were robust, accurate and repeatable.

23