Noninvasive Blood Glucose Analysis Based on Near-Infrared Reflectance Spectroscopy

Lü Xiao-feng, ZHANG Ting-lin, XIAO Feng, LI Guang, WANG You

State Key Laboratory of Industrial Control Technology, Institute of Cyber Systems and Control, Zhejiang University, Hangzhou 310027, China

Noninvasive Blood Glucose Analysis Based on Near-Infrared Reflectance Spectroscopy

Lü Xiao-feng, ZHANG Ting-lin, XIAO Feng, LI Guang, WANG You*

State Key Laboratory of Industrial Control Technology, Institute of Cyber Systems and Control, Zhejiang University, Hangzhou 310027, China

Noninvasive glucose detection is highly required for more convenient and less pain glycaemic monitoring. Most of currently used methods are invasive. In this paper, a near-infrared reflectance spectroscopy (NIRS) is proposed to detect blood glucose to protect patient absent of pain. NIRS is a safe, simple and efficient technology applied in many fields. Experiments, based on Oral Glucose Tolerance Test (OGTT), were conducted to collect data modeling with partial least squares (PLS) regression. 42 samples of fingertip blood and palm were measured by commercially available blood glucose meter and NIRS separately at the same time. The glucose concentration range is between 5 and 12 mmol·L-1. With leave-one-out cross-validation, we obtained a result of root mean square error of cross-validation (RMSECV) of 1.16 mmol·L-1for all the data. With the pre-processing methods of normalization and un-informative variables elimination reducing noise and eliminating some additional effects, we get a better result of 0.79 mmol·L-1. A RMSECV of 0.41 mmol·L-1for individual modeling is much less than the total modeling. It has a broad application prospect in individual customization.

Near-infrared; Blood glucose; Noninvasive

Introduction

Diabetes is caused due to genetic, metabolic disorders and other reasons leading to insulin dysfunction and causes serious metabolic disorder syndrome. Currently, the blood glucose level is the common standard for diabetes diagnosis. The disease has a high prevalence and it leads to a variety of complications, such as neuropathy, renal failure, amputations and so on.

Until now, there is no compelling way to cure diabetes. Keeping blood glucose at a low concentration by controlling diet and taking hypoglycemic medicines is the only way to treat the disease. For some diabetes patients in critical conditions, monitoring the blood glucose level daily is necessary.

There are some methods to monitor blood glucose in hospitals or at home. Most of them are invasive requiring lancing and finger bleeding. Usually these methods cost a lot of time as well as money, though they did not provide an accurate measurement. Non-invasive methods come into being[1-3].

NIRS detection technology in the last several years has been developed rapidly. It is widely used in various fields such as biochemistry, medicine, energy and so on[4-6]. This method is universally accepted due to its significant advantages. (1) High safety performance: It uses spectroscopy to characterize the properties of matter without damaging the sample and keeps the patient absent of pain. (2) Fast and efficient: Each spectral scanning only costs tens of seconds to complete. (3) Chemical Secure: During the test, no chemical reagent is needed and no any chemical contamination and biohazard material is produced. (4) Low cost and fit for continuous measurement.

In this paper, we detected blood glucose by using near-infrared diffuse reflectance measurement based on partial least squares algorithm. Fingertip blood was collected at the same time as reference value. We focus on the concentration between 5 and 12 mmol·L-1, which is the general range of subjects after the experiment of OGTT. The individual result has a significant increase for it can be used for individual customization in the future.

1 Materials and methods

1.1 Instrumentation

The spectra were all collected with BRUKER spectrometer. The wave-number range was from 12 500 to 4 000 cm-1. Each sampling was scanned 32 times with a resolution of 4 cm-1.

The blood glucose meter detecting the fingertip blood concentration as the reference value was produced by Xinli Medical Devices Co., Ltd.

Modeling and preprocessing was completed by Matlab 2007b.

1.2 Experiments and data collection

First of all, palm was chosen to be the spectral scanning part, considering the impact of the different gender or age, easy pressure control and easy sampling over the spectrometer. Blood sample was collected at fingertip as the Fig.1 showed.

The experiment was designed referring to OGTT to detect the concentration of blood glucose. OGTT was proposed by the World Health Organization in 1970s[9]. They agreed on a standard dose that a recommendation dose for adults is 75 g, and a dose for children is adjusted for weight.

According to OGTT, 75 g glucose was dissolved into 250 mL water. Three hours after dinner, every subject should have an ingestion of the beverage within 5 minutes. The spectroscopy collected five spectra each time using diffuse reflectance from the palm, followed by point sampling as the reference.

The glucose concentration increased rapidly in 30mins after eating the glucose then slowed down. The data collection was repeated after 10, 20, 30, 45, 60, 90 mins respectively.

Fig.1 Schematic diagram for spectral scanning part on palm and blood sample collection point at fingertip

During the experiment, the room temperature was about 20 ℃ and the room humidity was kept under 80%. The curtain was closed to keep a low illumination. At the same time of sampling by NIRS, the position and pressure of the subject should keep immovability as far as possible.

Three subjects, named A, B, C are all have normal glucose level. Each subject conducted OGTT experiments at seven glucose concentrations, and six experiments were conducted altogether. Thereby, for all of the subjects, forty-two samples were obtained. The experimental sequence is shown in table 1.

Table 1 Experimental sequence and subjects’ information

1.3 Data processing methods

PLS regression was proposed by S. Wold and C. Albano in 1983[10]. It is a multivariate calibration method which combines factor analysis and regression analysis. It would be reasonably to select the main ingredients involved in the regression for its powerful information extraction capabilities. Projecting on the covariance direction, this method will compress spectral into lower-dimensional spatial data. It would reduce the correlation between the independent variable, and enhance the stability and accuracy of the data. This algorithm has been widely used in chemometrics and it has been proved to be extremely useful for the analysis of near-infrared measurement[11-12].

Environmental noises and the impact of the physical properties are two serious problems facing near-infrared scanning. In order to improve the signal-to-noise ratio of the spectrum, some pre-processing methods should be added to reduce noise and eliminate some additional effects. There are many pretreatment methods such as transformation, smoothing, convolution, and differential[13].

Combined with the unique nature of the near-infrared spectral, normalization and uninformative variable elimination methods were used in this study.

1.3.1 Standard normal Variable

Normalization is a simplified calculation of transforming a dimensional expression into a dimensionless expression. Normalization method used in this study is the maximum-minimum normalization algorithm. The method is a linear conversion. The data will be mapped between 0-1.

X: Original data of spectrums；XNOR: Spectral data after normalization.

1.3.2 Un-informative Variables Elimination

Un-information variable elimination (UVE) method is a commonly used band selection method, which is based on coefficientbiof partial least squares regression. It was used in comparison between experimental variables and artificial variable analyzing the stability of the model to select the band[14].

bi: Coefficient of PLS；ci: Stand for the stability of PLS model.

If theciis less than a given threshold value, the spectra of No.iis considered has no contribution to the regression modeling and can be eliminated. Otherwise, it should be kept. This way was used to reduce the number of variables and the complexity of the model.

2 Results and analysis

Fig.2 shows the reflection spectrum of a human palm in once OGTT experiment. All 35 spectra at the seven different time have been recorded for comparing with reference value by PLS regression.

The test for every rational person should have an estimated curve trend that increases first then decreases. Take the once OGTT experiment as an example, the curve is shown in figure 3.

Fig.2 Scanning spectrums of subject 2 in once experiment

Fig.3 Contrast curve of once OGTT

Reference curve refers to the concentration measured by blood glucose meter； Measurement Curve refers to the concentration predicted by spectrums

A correlation analysis between the reference concentrations and predicted ones was calculated. The correlation of reference values and measured values in figure 3 is 0.990. The deviation between reference value and measured value was caused due to two reasons. Firstly, the model calculation caused some error. Secondly, each scanning of a concentration would cost one minute approximately. At the same time, the blood glucose concentration has changed especially at about 30 min.

Differences in different people can make a significant effect to the result. We consider the modeling for all subjects and modeling for individual subject respectively.

2.1 Modeling for all subjects

Calibration models were developed through the algorithm of PLS regression. The preprocessing method used was normalization and UVE. Parameters such as the threshold of UVE and the quantity of components were selected based on the best measurement results in the glucose calibration.

All data were evaluated through cross validation by using the method of leave-one-out. It took the data of each sample as testing set and modeling with all other samples. After getting all the measured samples, RMSECV was calculated to evaluate the model performance.

Model performance was evaluated by RMSECV and RMSEP (Root Mean Square Error of Prediction) of the glucose predictions over the calibration.

Average data at each concentration with five glucose meters as the ultimate reference value.

The result was improved after processing as shown in table 2.

Table 2 RMSECV of all data with PLS

Figure 4 shows the glucose concentration of predicted ones vs reference ones. The perfect result theoretically would have all the points falling on the straight line which is 45°. Seen from the picture, the points at the ends have a worse correlation than the middles ones because they have fewer samples when modeling.

Fig.4 Measurement results of total modeling

The circles are the measurement results of each sample at different concentration using the leave-one-out method

The data were separated into two parts, which were calibration set and test set. The data in the calibration set were used to do modeling and then to verify the data in the test set. We can get a result of RMSECV=0.88 mmol·L-1and RMSEP=0.93 mmol·L-1. The results are shown in figure 5.

Fig.5 Measurement results of modeling with calibration set and test set

2.2 Modeling for individual subject

The individual modeling is introduced in this part. Because more samples can get a more stable result, we choose the data of subject B who conducted the experiment three times. Modeling with all the 21 samples in full spectra, we got a better result than total modeling with a RMSECV of 0.40 mmol·L-1.

Prediction result is shown in figure 6. Although the number of data is not enough to divide into two parts, it is still easy to find out that it has a better result than the total modeling.

Fig.6 Measurement results of individual modeling

The circles are the measurement results of all the 21 samples of subject B

3 Discussion and prospect

The experimental results show that near-infrared diffuse reflectance can detect the blood glucose level invasively and effectively. Furthermore, this method could be convenient and portable for home detection. During the experiment, the experiment temperature, humidity, light intensity and the human body’s temperature, pressure should be maintained a stable status to reduce the influences. Further, the pretreatment method has a significant role on the results.

The differences between individuals due to different skin tissue and physical condition between the different subjects have an influence on the accuracy and the stability of the model. The results above have met the preliminary estimates of the patient’s need to detect blood glucose. It is obvious that the modeling for individual person has a better result and the instruments can be considered for the individual patient.

Though the instrument for individual patient has higher accuracy, the one for general patients would be more popular. People don’t need to collect blood sample for modeling at all. It is suitable for everyone. Of course, it would be more convenient for people especially old peoplewho expect simplicity especially.

More work is needed to ensure the accuracy and stability of the technology and make it easy to use and carry for a commercial product.

[1] Li C, Zhao H, Shi Z, et al. 2012, 8229: 82291H.

[2] So C F, Choi K S, Wong T K, et al. Med Devices (Auckl), 2012, 5: 45.

[3] Yadav J, Rani A, Singh V, et al. Biomedical Signal Processing and Control, 2015, 18: 214.

[4] Siripatrawan U, Makino Y, Kawagoe Y, et al. Sensors and Actuators B-Chemical, 2010, 148(2): 366.

[5] Cozzolino D. Planta Medica, 2009, 75(7): 746.

[6] Balabin R M, Safieva R Z. Energy & Fuels, 2011, 25(5): 2373.

[7] Kim Y J, Yoon G. Journal of Biomedical Optics, 2006, 11(4).

[8] Abookasis D, Workman J J. Journal of Biomedical Optics, 2011, 16(2).

[9] Alberti K G M M, Zimmet P Z, Consultation W. Diabetic Medicine, 1998, 15(7): 539.

[10] Lindgren F, Geladi P, Wold S. Journal of Chemometrics, 1993, 7(1): 45.

[11] Wold S, Sjostrom M, Eriksson L. Chemometrics and Intelligent Laboratory Systems, 2001, 58(2): 109.

[12] Bastien P, Vinzi V E, Tenenhaus M. Computational Statistics & Data Analysis, 2005, 48(1): 17.

[13] Li L N, Zhang G J, Li Q B. Modern Physics Letters B, 2009, 23(7): 925.

[14] Centner V, Massart D L, deNoord O E, et al. Analytical Chemistry, 1996, 68(21): 3851.

*通訊聯(lián)系人

O657.3

A

基于近紅外反射光譜的無損血糖分析

呂曉鳳，張婷琳，肖 鋒，李 光，王 酉*

浙江大學(xué)工業(yè)控制技術(shù)國家重點實驗室，智能系統(tǒng)與控制研究所，浙江 杭州 310027

無損血糖監(jiān)測是一種方便且無痛的血糖監(jiān)測方法。目前，大部分的血糖檢測方法都是有損的。提出了一種基于近紅外反射光譜的無損血糖檢測方法。近紅外反射光譜是一種安全、簡單并且有效的方法，被應(yīng)用于很多領(lǐng)域。采用口服葡糖糖耐量試驗來采集數(shù)據(jù)，用偏最小二乘回歸方法來建模。使用市售血糖儀采指尖血作為參考值，同時用光譜儀提取手掌光譜，共取得42組樣本。血糖濃度范圍在5～12 mmol·L-1。采用留一法交叉驗證，獲得所有數(shù)據(jù)的交叉驗證的均方根誤差為1.16 mmol·L-1。通過歸一化和無關(guān)變量消除的預(yù)處理方法來減少噪聲并消除一些額外因素，優(yōu)化的均方根誤差為0.79 mmol·L-1?；趥€人的數(shù)據(jù)進(jìn)行建模，得到了遠(yuǎn)小于整體數(shù)據(jù)的結(jié)果： 0.41 mmol·L-1。該方法在個人血糖檢測的市場化方面有廣闊的應(yīng)用前景。

近紅外； 血糖； 無損

2015-06-04，

2015-10-12)

Foundation item： the National High Technology Research and Development Program of China (2013AA041201)，Zhejiang Province Scientific and Technological Project (2015C37062)

10.3964/j.issn.1000-0593(2016)07-2312-06

Received： 2015-06-04; accepted： 2015-10-12

Biography： Lü Xiao-feng, (1988—)，female, Phd student in Zhejiang University e-mail: xxlm@zju.edu.cn *Corresponding author e-mail: king_wy@zjuem.zju.edu.cn