Noninvasive Blood Glucose Measurement Based on NIR Spectrums and Double ANN Analysis

This paper presents a new noninvasive blood glucose monitoring method based on four near infrared spectrums and double artificial neural network analysis. We choose four near infrared wavelengths, 820 nm, 875 nm, 945 nm, 1050 nm, as transmission spectrums, and capture four fingers transmission PPG signals simultaneously. The wavelet transform algorithm is used to remove baseline drift, smooth signals and extract eight eigenvalues of each PPG signal. The eigenvalues are the input parameters of double artificial neural network analysis model. Double artificial neural network regression combines the classification recognition algorithm with prediction algorithm to improve the accuracy of measurement. Experiments show that the root mean square error of the prediction is between 0.97 mg/dL 6.69 mg/dL, the average of root mean square error is 3.80 mg/dL.


Introduction
Diabetes has become a common disease in modern society.The blood glucose is too high or too low both will cause a significant impact on health, and the complications of diabetes are very serious, such as liver cirrhosis and neuropathy (Garcia-Compean et al. 2009, Mitrović et al. 2014) [1] [2].Diabetes is a disease difficult to cure, only through diet or insulin injections to control blood glucose.It is necessary to know their blood glucose realtime and accurately for the patients.However, the current method to measure blood glucose is mainly direct drawing blood from patients, which is a electrochemical way.Using invasive way to measure blood glucose for a long-term, the patients suffer from great physical pains, and the risk of infection increases.All kinds of testing strips of invasive blood glucose are expensive disposable consumables.These factors are not conducive to patients to facilitate and timely understand their blood glucose condition (Ramachandran et al. 2012) [3].
In order to overcome the disadvantages of invasive blood glucose, the non-invasive blood glucose measure-  9].This study proposes using four-channel 820 nm, 875 nm, 945 nm and 1050 nm infrared light as the transmission wavelengths.After extracting the eigenvalues of each PPG signal with the wavelet transform algorithm, the estimation model is developed with double artificial neural network.

PPG Signals
Though analyzing the fingertip photoplethysmography (PPG) signals, the physiological information of human blood and tissues can be extracted (Elgendi 2012, Ubeyli et al. 2007) [10] [11].When a bunch of near-infrared light transmit through the human fingertip and reach the corresponding photoelectric sensor, the absorption of skin tissue and blood will cause the beam weaken, thus the photoelectric signal of photoelectric sensors will be a certain of attenuated.
The near-infrared light absorption of skin tissue, muscle, bone, and venous blood of human fingertip in a short time remains constant.Only the changes of arterial blood will cause beam attenuation changes.As the heart pumps blood, blood will periodically circulation, the human fingertip artery blood volume will follow cyclical changes.When the body is systolic, the heart blood is supplied to the whole body, and the finger arterial blood will become filled.At this time, the absorption of the near-infrared light is strongest and electrical signals of photoelectric sensors are weakest.On the contrary, the absorption of the near-infrared light is strongest.In this way, the system can get cyclical PPG signals.
As shown in   [12], which eliminates the background noise:

The Selection of Wavelengths
The molecular formula of glucose contains a multiple of O-H, C-H chemical bond.In the wave bands of 935 nm -960 nm and 1040 nm -1080 nm exist the near infrared spectrums peak absorption of glucose (Ramasahayam et al. 2013) [13].
The wavelengths of 945 nm and 1050 nm are selected as key wavelengths, which are the peak wavelengths.It can reflect near-infrared light absorption of blood glucose.
Other selected wavelengths are 820 nm and 875 nm, whose absorption is relatively low.These wavelengths are taken as reference wavelengths.PPG signals caused by key wavelengths not only contain the information of blood glucose, but also other substances information.Combining key reference wavelength with wavelength to develop model can effectively eliminate the influence of other substances.
According to Lambert Beer's law In which, k is the molar absorption coefficient, d is optical path that means the thickness of the fingertip, c is liquid concentration that means blood glucose concentration here.For the four PPG signals of one person at the same moment, the blood glucose concentration c and the thickness of the fingertip d is the same.

The System of Gathering PPG Signals
The hardware of near-infrared non-invasive blood glucose estimated system mainly includes three parts.vessels periodic filling and void, the system can get cyclical PPG signals.

Analysis Methods of Artificial Neural Networks
The stoichiometric method commonly used in the near infrared spectral analysis includes linear correlation and non-linear correlation analysis.The linear correlation analysis mainly includes: multiple regression analysis (MLR), principal component regression (PCR), partial least squares (PLS), etc.The non-linear correlation analysis mainly includes: artificial neural network (ANN) and support vector machine (SVM) (Lam et

Double ANN Analysis
To improve the accuracy of the regression analysis, the study uses double ANN analysis: one is based on artificial neural network classification algorithm, and the other is based on artificial neural network prediction algorithm (Ping et al. 2005) [16].
First, the test data are applied to the network trained by classification algorithm, which will get the result R1.The result R1 is used to estimate the range which the blood glucose concentration value belongs.
Then, the test data are applied to the network trained by prediction algorithm, which will get the result R2.If R2 belongs the interval judged by R1, the R2 is right and retained as the predicted results.On the contrary, the R2 is considered inaccurate and discarded.
Since every time will gather enough PPG signals data that obtain a plurality of predicted results.After removing the maximum and minimum values of the results, the remaining data are averaged as the final predicted result.

The Experiment Conclusion
5 glucose tolerance test normal and healthy men participate in the experiment, their age distribute in 23 to 26.In the three and a half hours from the morning.Every 20 minutes, the subjects are measured twice using the PPG signals gathering system, and it will obtain multiple sets of PPG signals data.The first set of data are recorded as A i and used to train the artificial neural network; the second set of data are recorded as B i and used to verify the system.It will measure ten times and Roche Excellent type glucose meter is used to measure the real blood glucose level C at the same time, that are recorded as C real-i , where i is the number of measurement.
It uses A i and real blood glucose concentration real i C − for training the net.Finally it uses PPG signals that are not involved in the training of the net and the trained net for estimating, and gets the estimated blood glucose C est-i .n 1 = 8, n 2 = 8, n 3 = 7, n 4 = 10, n 5 = 7, n is the number of the estimation model can output the blood glucose estimation.
As is shown in Figure 5, it is the compare of the estimated values and the real values of blood glucose.
To judge the accuracy of the estimation, a root mean square error index is introduced: where x i = C est-i , which is the estimated blood glucose, y i = C real-i , which is the real blood glucose.After calculating RMSE 1 = 4.10 mg/dL, RMSE 2 = 4.64 mg/dL, RMSE 3 = 2.58 mg/dL, RMSE 4 = 6.69 mg/dL, RMSE 5 = 0.97 mg/dL.The average of root mean square error is 3.80 mg/dL.To further verify the accuracy of the estimation, Clarke error grid analysis (CEG) is applied.As is shown in Figure 6, the final result is 100% distributed in the region A, which is considered to be completely accurate in

Figure 2 .
Figure 2. The block diagram of the noninvasive blood glucose measurement system.

Figure 3 .
Figure 3.The PPG signals after removing baseline drift and smoothing.
al. 2010, Yadav et al. 2014, Barman et al. 2010) [7] [14] [15].The ANN classification recognition-prediction double algorithms are used to develop the estimation model.As shown in Figure 4, typical PPG signals should have the main wave P, the dicrotic wave T, the main trough A and the sub-trough V.The wavelet transform algorithm is used to remove the baseline drift of PPG signals and smooth it.

Figure 4 .
Figure 4.The eigenvalues of PPG signals.

Figure 5 .
Figure 5.The estimated values compared with the real values of blood glucose.

Figure 6 .
Figure 6.The Clarke error grid analysis of predicted blood glucose and reference blood glucose (mg/dl).