Nowadays, there are about 415 million adults suffering from diabetes^[1]. At present, the treatment for diabetes patients is monitoring their blood glucose level frequently and then taking the appropriate amount of oral hypoglycemic drugs and insulin to control their blood glucose level. Since conventional methods of blood sampling are painful, discontinuous, and have other undesirable features, noninvasive methods are attractive for monitoring blood glucose as well as other analytes^[2].

Various methods, such as Raman spectroscopy^[3–14], near-infrared spectroscopy^[15,16], photo-acoustic spectroscopy^[17,18], and optical coherence tomography^[19,20], have been extensively researched. Raman spectroscopy displays excellent molecular specificity by probing the fundamental vibrational states. Due to the extremely weak Raman scattering signal of water, sharper spectral features, and few spectral overlaps, Raman spectroscopy is particularly suitable for detecting biological analytes.

However, conventional laboratory Raman systems, with their large size, cumbersome nature, and high cost, are limited in clinical applications. Additionally, the stability is not good enough. Conventional Raman detection systems always utilize a convex lens with a certain focal length as the collecting lens, but the focal position will change during the measurement. The problem that Raman signal cannot be stably collected will greatly affect the accuracy and stability of the test.

In this Letter, a thallium-doped grin lens was employed as the collection lens. A specially designed wearable fiber-optic probe was employed to help the stable and convenient collection of the Raman spectrum. A novel method of quantitative analysis of the Raman spectrum that uses the peak area as the main reference factor and the peak intensity as the auxiliary reference factor to calculate the target concentration was proposed. A nonlinearized multivariate dominant factor-based partial least squares (PLS) model was employed for different samples to predict the glucose levels. A glucose solution, 11 rats, and 10 healthy humans were studied as the subjects.

Figure 1 shows the experimental setup. According to the investigation by Enejder^[3,4], a 785 nm diode laser (B&Wtek, BRM-OEM-785-0.3-100-0.22-FC) with a stable power of 300 mW and a linewidth of 0.2 nm was employed as the excitation source and could ensure the harmlessness and high efficiency of the detection system.

A specially designed wearable fiber-optic probe was employed. The probe end packaged the grin lens, while a shading tube was designed to reduce the interference of ambient light and to help fasten the probe with a spire lamella. This probe utilized a thallium-doped grin lens as the collecting lens, which could ensure that the focal position is always on the surface of the measured object while realizing the convergence of wide-band spectra. Only one kind of filter (Semrock, LP02-785RU-25), which could reflect the light below 785 nm while transmitting the light between 805 and 1050 nm, was used in this probe. The function of this filter amounted to the function of a dichroic mirror and a bandpass filter.

Compared with a conventional probe, this probe had the advantages of better testing stability and lower cost and is convenient to wear.

A miniature Raman spectrometer (B&Wtek, BTC675 N-OEM), with dimensions of 124mm×92mm×26mm, an optical resolution of 6cm−1 and a signal-to-noise ratio (SNR) of 4800:1, was employed to detect the Raman spectroscopy. The external optical system was a Czerny–Turner system. A back-illuminated linear-array CCD (Hamamatsu Photonics, S11156-2048-01), with 2048×1pixels, was selected as the detector. A thermoelectrically cooled module, with a cooler temperature of 50°C below the ambient temperature, was employed to reduce the dark noise while increasing the SNR and sensitivity of the detection system.

All of the components were immobilized compactly to achieve the miniaturization of this system.

Glucose solutions which aimed to reflect the blood glucose levels of humans (4.4–11.1 mmol/L) were prepared, with different concentrations of 3.8, 4.1, 4.4, 4.8, 5.1, 5.4, 5.6, 6.7, 7.8, 8.9, 10, 11.1, 12.2, and 13.3 mmol/L. A cuvette containing the glucose solution was put into a dark sample cup to ensure the stabilization of the test. Raman spectra were collected after 20 s. The glucose solution study was employed for verifying the feasibility of in vitro glucose detection using the miniature wearable Raman system.

Eleven Sprague–Dawley (SD) male rats were studied as samples. We here claim that our protocol was approved by XIOPM’s committee on the use of SD rats and human volunteers as experimental subjects. After fasting for 8 h, the rats were anesthetized. Then, the glucose solution was poured into the stomachs of rats. As a result, the blood glucose levels of the rats varied within 1.5 h. Raman spectra were collected transcutaneously every 5 min from the ears of the rats along with glucose reference values provided by a blood glucose meter (Roche, ACCU-CHEK, Active). The laboratory rat study was employed for verifying the feasibility of noninvasive detection of an in vivo blood glucose level using the miniature wearable Raman system.

Ten human volunteers were investigated as samples. Volunteers took a certain dose (2 g/kg) of glucose solution orally. During the next 2 h, Raman spectra were collected transcutaneously every 20 min from the inside of the wrist while glucose reference values were provided by a blood glucose meter. Every volunteer was tested at least 3 times to ensure the repeatability of the experiment. The skin of the human volunteers was not damaged during the test. The human subject study was employed for verifying the feasibility of noninvasive detection of human blood glucose level using the miniature wearable Raman system.

The area of characteristic peaks was used mainly as a calibration reference, while the intensity was used as an auxiliary calibration reference to avoid the interference of baseline fluctuations of Raman spectra. Raman spectra were extracted at the range of 300–3000cm−1 without any pre-processing to avoid errors.

A dominant factor-based multivariate PLS model was employed to predict the glucose concentration^[21–23]. The dominant factor could model the linear and nonlinear relationships according to the physical background, while a statistical-based PLS method was used thereafter to correct the residue errors using the full spectral information. The R2 value and mean absolute error (MAE) are the two indexes used to estimate the fitted result. A higher R2 value and lower MAE mean a better fitted result. The basic principle of the model can be written as Y=f(x)+E1,(1)where Y is the glucose level, f(x) is the dominant factor, and E1 is the residue error. f(x)=k1x1+k2x2+⋯+knxn+kn+1x12+kn+2x22+⋯+k2nxn2+b,(2)where x is the area of the characteristic peaks whose position is similar to the characteristic peaks of the Raman spectrum of glucose. k1,k2,…,k2n could be calculated by the PLS. Y2=g(x)+E2,(3)g(x)=j1x1+j2x2+⋯+jmxm+jm+1x12+jm+2x22+⋯+j2mxm2+c,(4)where x is the intensity of the characteristic peaks whose position is different from the characteristic peaks of the Raman spectrum of glucose. j1,j2,…,j2m could be calculated by the PLS. R2=S1S2=1−S3S2=1−∑i=1n(yi−y^i)2∑i=1n(yi−y¯)2,(5)where S1 is the regression sum of squares, S2 is the total sum of squares, and S3 is the residual sum of squares. MAE=1n∑i=1n|fi−Yi|,(6)where fi is the calculated value, while Yi is the measured value.

Figure 2 exhibits the Raman spectra of the glucose solution. The signal at 1372.7cm−1, which was considered as the CH3− of glucose, is the strongest. According to Fig. 3, the relationship between the intensity of 1372.7cm−1 and the concentration of glucose is uncorrelated. However, the area of 1372.7cm−1 and the concentration of glucose are directly related. Consequently, the areas of peaks were employed as the main reference of calculation. The calibration result shown in Fig. 4 with the MAE is 0.394 mmol/L, and the R2 value is 0.981, which encourages us to do further research.

Figure 5 exhibits the difference between the spectrum of the skin of a rat and the glucose solution. It can be seen that the skin of a rat has similar characteristic peaks with glucose at 406, 502, 586, 670, 767, 1200, 1256, 1447, 1555, and 1656cm−1. Figure 6 shows the predicted blood glucose levels for one of the rat samples using leave-one-out cross validation to confirm the best value of the principal component. The calculated results of the R2 values are 0.954, 0.953, 0.952, and 0.951, showing the accuracy and stability of the calibration model and this miniature wearable Raman system.

Each sample was processed individually in the same method, and the results are shown in Table 1. The mean R2 value for all 11 rats is 0.893, while all of the rat samples have R2 values over 0.85. The results suggest that this miniature wearable Raman system is feasible for in vivo blood glucose detection.

Figure 7 shows the comparison of the Raman spectrum of human skin with those of the glucose and rat. The Raman spectra of human skin were extremely similar to the spectra of rat skin.

Figure 8 shows the calculated results with and without the grin lens for one of the human volunteers. The R2 value is 0.958 with the grin lens, while the R2 value is 0.736 without the grin lens. The result shows that using the grin lens is of great significance to the accuracy and stability of the test.

The results for all of the human volunteers are shown in Table 2. The mean R2 value for all 10 volunteers is 0.844. The results suggest that this miniature wearable Raman system is feasible for human blood glucose detection.

Figures 9(A) and 9(B) exhibit the predicted results for all of the rat samples with an R2 of 0.563, while one of the rat samples has an R2 of 0.960. Figures 9(C) and 9(D) exhibit the predicted results for all of the human volunteers with an R2 of 0.773, while one of the human volunteers has an R2 of 0.935. The result corroborates that modeling for each sample is essential to ensure the purity and accuracy of the calculation.

The reasons for these results may come from two aspects: individual differences and physiological changes. However, we anticipate that the more accurate result can be predicted by the methods of big data statistics and establishing a glucose database.

In conclusion, signal acquisition and data analysis are the 2 major bottlenecks of noninvasive blood glucose detection. In this Letter, a novel method of Raman spectra collection is investigated. A thallium-doped grin lens is employed as the collection lens. A specially designed wearable fiber-optic probe is employed to help the stable and convenient collection of the Raman spectrum. A miniature wearable Raman system is developed to achieve the noninvasive blood glucose detection. The area of characteristic peaks is used as the main calibration reference, while the intensity is used as the auxiliary calibration reference. Additionally, a dominant factor-based multivariate PLS model is constructed to achieve the accurate prediction of blood glucose levels. This serious study investigates in vitro and in vivo glucose detection using a miniature wearable Raman spectroscopy system and demonstrates the feasibility of noninvasive blood glucose detection using this miniature wearable Raman spectroscopy system. The results may be useful for the development of a miniature Raman spectrometer in disease diagnosis and clinical applications of blood glucose as well as other analyte detection.