Noninvasive Blood Glucose Monitoring System Based on Distributed Multi-sensors Information Fusion of Multi-wavelength Nir *

In this research, a near infrared multi-wavelength noninvasive blood glucose monitoring system with distributed laser multi-sensors is applied to monitor human blood glucose concentration. In order to improve the monitoring accuracy, a multi-sensors information fusion model based on Back Propagation Artificial Neural Network is proposed. The Root-Mean-Square Error of Prediction for noninvasive blood glucose measurement is 0.088mmol/L, and the correlation coefficient is 0.94. The noninvasive blood glucose monitoring system based on distributed multi-sensors information fusion of multi-wavelength NIR is proved to be of great efficient. And the new proposed idea of measurement based on distributed multi-sensors, shows better prediction accuracy.


Introduction
Diabetes has become a modern disease and more than 150 million people are suffering from it all around the world [1].In order to prevent complication, the tight blood glucose level control is very essential.Currently, patients are recommended to monitor their blood glucose level via an invasive finger-tip stick method, this way will inevitably bring patients pain and infection [2].In order to avoid the weakness of invasive method, a number of noninvasive measurements occurred, like middleinfrared emission spectroscopy [3,4], Near Infrared (NIR) spectroscopy [5,6].The main drawback of infrared measurement is its accuracy and stability.In our past work, an infrared noninvasive blood glucose measurement system based on multi-sensors and Mixture of Experts (ME) [7] was designed.ME algorithm greatly improved the precision of noninvasive blood glucose measurement.In the noninvasive blood glucose monitoring system based on distributed multi-sensors information fusion of multi-wavelength NIR, this system was designed on the purpose of continuous blood glucose monitoring for patients at home and hospital [2].Another advantage of this system was the introduction of distributed multi-sensors idea, this change has been proved to be better improvement for blood glucose prediction accuracy.In this distributed multi-sensors system, a multi-sensors information fusion model, Back Propagation Artificial Neural Network (BP-ANN), was applied.

Noninvasive Blood Glucose Monitoring Model
The experimental data for this research was obtained via NIR based on laser arrays as shown in Figure 1.The NIR light source consists of 3*3 laser diodes arrays operating at output powers of 5 mW [8].

Hardware Design
Wavelengths selection: According to the specialty of glucose molecular structure and absorption specialty, second order times frequency absorption exists between 1100 -1300 nm, and first order times frequency absorption exists between 1500 -1800 nm.Other components in blood, such as hemoglobin and water contain groups of hydrogen which can trigger NIR absorption as well.glucose measurement, the selection of an ideal measurement site is essential.Four factors should be considered: First, for the convenience of measurement, the measurement site should be exposed outside; Second, in order to decrease the influence from outside factors, individual difference like gender, age and physical state should be low; Third, the measurement site should contain rich blood and the interference from other components is low; Last, NIR light can transmit measurement site easily.In the noninvasive blood glucose monitoring system based on distributed multi-sensors, left ear lobe, right ear lobe and the right hand part between thumb and index finger are selected as distributed measurement sites.Circuit design: six channels of laser-driving and photoelectronic amplification circuit are designed in this system.Figure 2 shows the laser-driving circuit, LDdriving and the part of photodiode feedback circuit help to maintain laser operate with output power of 5 mW and forward current of 30 mA.By adjusting POT1, forward current can be adjusted between 0 -120 mA. Figure 3 shows the photoelectrical signal amplification circuit.By applying an operational amplifier POT1 and some affiliated resisters and capacities, the obtained spectral signal can be amplified by 1000 to 3V, the amplified signal is transferred to plug seat JP2-1 which connected to AD converter.

Software Design
1) Data acquisition and save: Amplified spectral signal is first conveyed to AD converter, which supports Labview drive.So the converted data can be displayed on the screen programmed by Labview language.Figure 4 shows the interface of acquired spectral data from six channels.At each channel, the average value of 10 successive acquired data is calculated as displaying information.
2) Distributed multi-sensors information fusion based on BP-ANN: BP-ANN is one kind of supervised learning network.There are two phases of positive transmitting processing and error reverse transmitting processing in the study processing of BP-ANN.
A three layers BP neural network (Figure 5) is applied to model fuse multi-sensors information and predict the blood glucose concentration in experiments.Least Square function is adopted as error function in the training of BP network, which writes as: ( ) k O is the output of note k; y k is the corresponding de- sired output; q is the number of output notes, and m is the number of training samples.For the output weight, the revise value jk w is the weight connecting the number k output note and the number j hidden note: η is the study speed, k o is output value at the number k output note; k δ is the gradient factor.For the hidden layer, revise value writes as: w is the weight connecting the number j hidden note and the number i input note; j o is the output at the number j hidden note.
Figure 6 shows the training curve in the concentration prediction of blood glucose, after training of 10 times, the goal is reached and the performance can reach 0.088 mmol/L.
3) Prediction result display based on Labview: The acquired data displaying and predicted blood glucose concentration value interface is shown in Figure 4. On the interface, 8 measurement methods options are given.Corresponding blood glucose prediction program will be launched by clicking any switch, and the predicted value will be displayed as well.

Experiments
In both glucose samples experiment and noninvasive blood glucose measurement experiment, 7 measurement methods with single, two and three wavelengths were applied, which were named A, B, C, A + B, A + C, B + C and A + B + C method respectively.

Glucose Samples Experiment
Process of glucose experiment: All samples were measured in a 5 mm quartz cell and the temperature in the sample cell was controlled at about 37˚C.Two sets of 20 samples were prepared with varying concentrations, ranging from 10 to 200 mg/dL in 10 mg/dL steps.More than 500 original data sets were obtained, after eliminating some outlier samples, 200 samples with different concentrations were left for building calibration model.In the experiment, the spectrum of empty cell was measured firstly as a background [9].
Data preparation of calibration model: It is essential to figure out the significant difference between each group of spectral data sample.Statistic method of T-test was used to prove the validity of samples and to check out the outliers data.In Table 1, a total of 200 glucose  Comparing with the differences between 20 groups of glucose NIR absorption spectra, their differences achieve significant level (T-test, p < 0.05).

Process of Human noninvasive measurement experiment:
In order to get training data covering wide variety scope of human blood glucose concentration, the blood glucose tests were made on volunteers at certain time before or after their meal with the interval time of half an hour.After measurement of blood glucose, multi-sensors were attached to measurement sites to get responding spectra information [10].A total of 142 samples for each type of measurements were got in this experiment and the blood glucose concentration ranges from 3.2 mmol/L to 10.8 mmol/L.

Data preparation for BP-ANN training and testing:
Similarly to the data preparation of calibration model, experimental data in noninvasive blood glucose measurement were given in Table 2. T-test is applied to test the spectra difference of different groups of blood glucose concentration, p of T-test is smaller than 0.05, which demonstrates that their differences reach significant levels.

Building Calibration Model
In both the glucose samples experiment and noninvasive blood glucose measurement experiment, 7 kinds of calibration model were built to realize the function of continuous monitoring.In A + B + C method of glucose experiment, 20 spectra data samples were used to build the calibration model.Figure 7 shows the relation between glucose concentration and glucose absorption spectra, from the figure we can see this validation owns a fine prediction ability for unknown received spectra information.In Table 3, the performances of different measure-ment methods were compared, in the A + B + C method, the RMSEP is 4.412 mg/dL and CC is 0.90, which shows a great improvement in prediction performance to other single or two wavelength methods.Compare with statistic values of three single wavelength and combination of two wavelength methods, single wavelength A shows best sensitivity

Results of Human Blood Glucose
Noninvasive Measurement

BP-ANN Measurement Model
In the human noninvasive blood glucose measurement experiment, a total of 142 samples were obtained, 122 couples of blood glucose absorption spectra data and corresponding blood glucose reference values were used for the BP network training.As shown in Figure 6, the network can reach goal after 10 training, RMSEP of this model is 0.088 mmol/L and CC is 0.9315 in the method of A + B + C. Comparing with the other six methods, the BP-ANN greatly decreased the prediction error, which can be seen in Table 4.

Testing of Measurement Model
The left 20 samples were used as validation data sets to test performance of the measurement model.20 spectra information data were inputted to the BP-ANN multisensors information fusion model, 20 output values of predicted blood glucose concentration were received at the output layer, the result was given in Table 5.The RMSEP between NIR method and standard method is 0.3143 mg/dL.T-test is applied and p is 0.1728, bigger than 0.05, which demonstrate that the relation between those two values is well related.

Conclusion
In this noninvasive blood glucose monitoring system, two new measurement ideas were employed, which are continuous monitoring and distributed multi-wavelength measurements.The experimental result has proved that this system has an improved advantage in blood glucose prediction, and the multi-wavelength information fusion model has also contributed greatly to the monitoring     property.In the later work, we hope that new multi-wavelength strategy can be used to further improve the prediction accuracy.

Figure 1 .
Figure 1.Framework of the multiwavelength arrays monitoring system.

Figure 4 .
Figure 4. System interface of calculated blood glucose concentration.

Figure 6 .
Figure 6.Training curve in human blood glucose concentration prediction.

Figure 8 .
Figure 8. Relation of blood glucose concentration between measurement method and NIR arrays information fusion method.