Keywords

1 Introduction

Currently, the glucose is measured by pricking the fingertip and extracting the blood sample using a tiny disposable lancet, the extracted blood needs to be placed on one test sensor strip and inserted into the glucose meter that will show the numerical glucose value within 2 s. Although this method is the most painless way to measure glucose level, the patients are recommended to do the self-test at least 5–7 times per day depending on the diabetic type. So the main issue of the invasive finger pricking for glucose test are inconvenient, a little painful and high cost. Additionally, the lancet prick has a high risk of infection.

During the last decade, there have been many non-invasive methods to improve the glucose level measurement and the optical method is regarded as the main approach in most researches. The target of this article is to explore the relationship between glucose concentration and ultrasonic and near infrared light, try to predict the glucose level within acceptable range by doing the calibration scheme and validation.

2 Theory

Light attenuates due to the absorption and scattering of human tissue. The attenuation of light can be expressed as:

$$ I = I_{0} {\text{e}}^{{ - \mu_{eff} L}} $$
(1)

which is the light transport theory. This equation describes light propagation in human tissue through a set of spectroscopic properties. And in this equation I is the reflected light intensity, I0 is incident light intensity.

$$ \mu_{eff} = \sqrt {3\mu_{a} \left( {\mu_{a} + \mu_{s}^{\prime } } \right)} $$
(2)

\( \mu_{eff} \) is shown as Eq. (2) interacts with the absorption coefficient \( \mu_{a} \) and reduced scattering coefficient \( \mu_{a}^{{\prime }} \), where \( \mu_{a}^{{\prime }} = \mu_{s} [1 - g] \). And the absorption coefficient can be described as the absorbance per unit path length equation that contains the molar absorption coefficient ∈ and the molar concentration \( C:2.303 \in C\;{\text{cm}}^{ - 1} \) [1]. The changes in glucose concentration can affect the light scattered intensity from human tissue. The higher glucose concentration, the more glucose molecule in the blood vessel. So it leads less scattering, less optical path and less absorption. It means the absorbance decreases with the increase of glucose concentration [2].

3 Experimental Method

3.1 Measurement Method

In this paper, both Ultrasonic transducer and Near infrared spectrometer are applied to blood glucose measurement.

3.1.1 Ultrasonic Transducer

Currently, the commonly used frequency range for ultrasonic diagnostic apparatus is 2–10 MHz, where 3–5 MHz has been widely used. In this experiment, a pair of 5 MHz Ultrasonic transducers with transmitter and receiver are selected to connect with the signal generator and signal analyzer.

Figure 1 shows the block diagram of the ultrasonic transducer experiment. There are two volunteers in this experiment, one male subject and one female subject. Those two ultrasonic transducers apply to four measurement targets: fingertip, palm, wrist and arm. Those targets are measured six times a day, 90 min before and after breakfast, lunch and dinner. Meanwhile the real glucose level is measured by traditional glucose meter as reference value. In this experiment, the room temperature is also considered to explore the influence for the relationship between glucose and ultrasonic.

Fig. 1
figure 1

The block diagram of the ultrasonic transducer experiment

Full size image

3.1.2 Near Infrared Spectrometer

In this part, Near Infrared Spectrometer (AOFT: Acousto Optic Tunable Filter method) is applied to middle fingertip of left hand by reflection method. The device and experimental schematic are shown in Fig. 2. Wavelength range is 1200–2500 nm. This device includes the spectrometer and laptop where corresponding software is installed. The resolution is 1 nm, scan number is 4 times and integration number is 5 times.

Fig. 2
figure 2

The spectrometer device and experimental schematic representation (left) and the structure of AOFT unit in the spectrometer device (right)

Full size image

There are two data pre-process methods, one is SNV (standard normal variate transformation) [3] processing that can eliminate solid particle size, surface scattering and the influence of NIR diffuse reflection spectrum from optical path difference. And another one is differential [4] calculation that can erase baseline and other background interference, recognize overlapping peak, improve resolution and sensitivity effectively. And We also apply squalane oil to fingertip to erase some noise since the fingertip can absorb infrared easily.

In this experiment oral glucose tolerance test is used to help achieve the wide blood glucose level. Oral Glucose Tolerance Test: the subjects are fasting for more than 10 h and no breakfast before the experiment day. (1) Prepare the sugar water that 250 ml solution with 75 g of sugar. (2) Measure the baseline glucose level first and then smear Squalane oil on test fingertip. (3) After 5 min drink the 250 ml solution with 75 g of sugar. (4) And then start the near infrared experiment, save the absorbance spectral data and measure glucose level by glucose meter every 10 min for two hours, totally 16 times. And each time we collect 30 times spectral data to do the average calculation in order to reduce noise and improve signal-to-noise ratio.

3.2 Analysis Method

Partial Least Square regression method and back-propagation artificial neural network method are utilized in this article to analysis data. PLSR can find a linear regression model by projecting the predicted variables and the observable variables to a new space [5]. BP-ANN is another calibration model that applies to linear relations as well as nonlinear [6]. It through the output node feedback to hidden layer and input layer is used to adjust the weights value and then repeats that process until get the most ideal result. In this research those two methods can explore the codependent relationship among the volunteers’ glucose from glucose meter and the corresponding experiment data to construct calibration model, and then we use this model to predict other observable variables.

In addition, Clarke’s error grid analysis (EGA) [7] is used for the comparison tool. It can accurately represent the actual situation of the blood glucose measurement by showing results in five regions. So after the prediction, we need to use this method to compare the glucose level from glucose meter and the predicted result from the ultrasonic transducer and near infrared.

4 Result and Discussion

4.1 Ultrasonic Transducer

In ultrasonic transducer experiment, 122 data sets are collected in the ultrasonic experiment, this article takes the first 102 data sets to established prediction calibration model and the left 20 data sets to do validation that in order to test the performance of prediction scheme.

Figure 3 shows EGA result of two subjects in this experiment. The finger shows most of the predicted results fall in the region A, which means the established scheme is capable of making prediction. And the correlation coefficient R of subjects are 0.5419 and 0.4558, the RMSEP (root mean square error of prediction) of the glucose level are 12.1444 and 14.3531 mg/dl in PLSR method. While in BP-ANN method, R of subjects are 0.7332 and 0.6995, RMSEP are 9.3541 and 10.9674 mg/dl. In this situation, the accuracy of ANN method is better than PLSR method.

Fig. 3
figure 3

The EGA result of two subjects from ultrasonic experiment by the PLSR method: circle and ANN diamond: star; female: red, male: black

Full size image
$$ Y_{1} = 2.1958\,x_{1} + 0.2363\,x_{2} - 0.4361\,x_{3} - 0.6979\,x_{4} - 2.9557\,x_{5} + 85.1532 $$
(3)
$$ Y_{2} = 1.7565\,x_{1} - 0.2551\,x_{2} - 0.0910\,x_{3} - 0.0688\,x_{4} + 4.5107\,x_{5} - 9.376 $$
(4)

In the equation, \( Y_{1} \) is the predicted glucose value, \( x_{1} ,x_{2} ,x_{3} ,x_{4} ,x_{5} \) are represent the output value of fingertip, palm, wrist, arm and room temperature respectively. By considering the metabolism rate with the fluctuation of the blood glucose, everyone should have their unique prediction equation to describe the glucose level. In this prediction, fingertip and temperature are more important than other three parameters, while the temperature makes negative influence on subject1 and positive influence on subject2 are opposite relation for two subjects.

4.2 Near Infrared Spectrometer

Figure 4 shows the five subjects’ glucose range between 70 and 210 mg/dl are measured by glucose meter every 10 min from the OGTT and the absorbance spectrum of one subject’s fingertip for 16 times from infrared spectrometer. In order to achieve the prediction of blood glucose concentration by analysis the characteristic absorption spectrum. There are some absorption wavelength areas of glucose, first area is 1450–1490 nm, second area is 1530–1570 nm and third area is 2130–2170 nm. They attribute to combination bands of –CO stretching vibration and –OH stretching and bending vibration of blood glucose.

Fig. 4
figure 4

The five subjects’ glucose range from oral glucose tolerance test and the absorbance spectrum of subject’s fingertip for 16 times

Full size image

This research takes the first eleven times spectral data sets and corresponding glucose concentrations as the predictor to establish the calibration model, takes left five data sets as the tester to validate the model performance in this experiment.

The PLSR and BP-ANN validation results of five subjects are showed in Fig. 5. The points with different colors and different shapes represent the results of five subjects from three wavelength areas in those two figures. All the results fall in the region A, which means they are acceptable scheme for the blood glucose prediction. And Table 1 shows the correlation coefficient R and RMSEP of five subjects in two different methods. By comparing the values, the prediction accuracy of ANN method is better than PLSR method, which is same as the ultrasonic transducer experiment.

Fig. 5
figure 5

The EGA result of five subjects from three characteristic spectrums by PLSR method (left) and ANN method (right), at first area: diamond; second area: circle; third area: star

Full size image
Table 1 The results of calibration curve from the PLS regression and ANN method
Full size table

5 Conclusion

In the paper a non-invasive glucose measurement scheme using ultrasonic transducer and Near IR spectrometer are explored. The predicted results from ultrasonic transducer experiment are less than the international standard ISO 151197 (International standard of SMGB the error is less than or equal to 15 mg/dl to comply with ISO 151197). And in the near infrared experiment, the results show the feasibility of the development of non-invasive glucose measurement scheme based on reflectance through acousto-optic tunable filter. The prediction ability of ANN method is better than PLS regression method, especially in the nonlinear correlation between independent variable and dependent variable.