1. Introduction

The blood glucose level of diabetic patients should be closely monitored to provide guidance values for medication or insulin injection. The traditional finger-prick method not only results in physical and psychological pain but also the danger of infection due to frequent pricking, which cannot be used to monitor blood glucose continuously. Therefore, developing a method to continuously monitor glucose levels, which is crucial to diabetes prevention and treatment, is particularly important.

The noninvasive quantification of blood glucose in vivo is ideal and has thus been studied for many years. Previous studies have focused on spectroscopic methods, especially the near-infrared (NIR) spectroscopy [1]. However, the weak glucose absorption coefficient in this overtone band makes it very difficult to extract a net signal due to significant disturbances of interfering components, such as hemoglobin [2]. Mid-IR spectroscopy is known as a highly sensitive and selective method to determine the concentration of an identify target substances, and the characteristic absorption band in the mid-IR region enables the distinct recognition of a target substance, such as glucose, in the presence of various chemical species. Because glucose has a strong and sharp fundamental frequency vibration at 950~1200 cm⁻¹ (8.333~10.526$\mu m$), the mid-IR spectrum analysis of this band has been widely investigated to measure glucose in vitro [3–5]. However, the strong absorption of water in the band has been the biggest challenge for noninvasive glucose detection in vivo using mid-IR light. Due to the strong water absorption, traditional mid-IR light coupled with an ATR crystal only penetrates several micrometers into human skin [6]. A glucose signal was not observed at this depth. Developed mid-IR laser light sources, such as the quantum cascade laser (QCL), seem to provide a possibility to obtain a stronger glucose signal in the deeper skin regions in combination with scattering spectroscopy [7]. However, the backscattering signal of QCL from the dermis layer of human skin was attenuated on the order of 10⁴, which indicates that the incident light should at least have a power of hundreds of milliwatt (mW) [7]. The maximum permissible exposure (MPE) for human skin is 100 mW/cm² [8], and incidence light at this power would lead to tissue heating and/or damage during in vivo measurements. Furthermore, a glucose signal was not obvious when examining the scattering light of human skin in vivo [7]. In addition, the C-C bending vibrations at 1035 cm⁻¹ and 1077 cm⁻¹ (and/or symmetric stretch of PO₂^- at 1080 cm⁻¹, C-OP stretch at 1047 cm⁻¹) in the spectrum of human skin do not only relate to glucose, as the glucose C-C bands are a minor component compared with those of the other chemical components of tissue, such as proteins and fats [9, 10]. These interferences of absorption peaks due to vibrations and/or stretches from other skin components also pose a significant challenge to the noninvasive measurement of glucose in vivo using scattering light in the mid-IR band. Studies have reported the use of photoacoustic spectroscopy, but this method also suffers from a large incidence power density, weak signal detected and low detection limit [11, 12]. Although transmission cells with long optical lengths have been used to successfully determine the glucose concentration in vitro [13–16], the biggest challenge of this method is the precise quantification of an effective sensing optical length to noninvasively measure glucose in vivo. To date, reports that detail the measurement of glucose in vivo using this method have not been published.

This paper introduces a minimally invasive method to determine the glucose level in vivo using mid-IR spectroscopy in combination with a tunable carbon dioxide (CO₂) laser and small fiber-based ATR sensor. The potential of this method was evaluated in vitro. The newly fabricated small fiber-optic ATR sensor is intended to be subcutaneously implanted in tissue to continuously monitor the body’s glucose levels within the interstitial fluid (ISF), where the glucose concentration is considered to be strongly correlated with the concentration of glucose in blood [17]. Several strong, prominent and isolated glucose absorption peaks in the “finger print” band allow it to be distinguished from other interfering species in human blood in vitro [3, 5, 18–20]. Compared with other components in the blood, a relatively lower percentage of protein was observed in the ISF due to the filtration through capillary walls. Moreover, a biocompatible semipermeable membrane with a selectable molecular weight cut off that is designed to be the protective cover to separate the implanted sensor from the tissue can also be used to further filter large biological molecules within the ISF [21]. Thus, glucose can be more easily measured in the ISF than in the blood using mid-IR spectroscopy. Lactate (absorbance peak at 1040 cm⁻¹), creatinine (absorbance peak at 1045 cm⁻¹ and 1099 cm⁻¹), bicarbonate (absorbance peak at 1011 cm⁻¹) and phosphate (absorbance peak at 990 cm⁻¹ and 1080 cm⁻¹) are identified as the main interferences [4, 13, 15, 22–25] in the determination of glucose in the ISF using mid-IR spectroscopy in the band of 950~1200 cm⁻¹. The results obtained by M. Pleitez [24] et al. indicated that albumin (absorbance peak at 1052 cm⁻¹ and 1084 cm⁻¹ in the band of 950~1100 cm⁻¹) is the predominant protein component of ISF, and the other protein components are present at much smaller concentrations. Albumin and lactate can be neglected for the selected wavelengths if glucose is to be determined from the ISF by infrared spectroscopy due to their spectral properties between 950 cm⁻¹ and 1200 cm⁻¹ [24, 25]. Creatinine and bicarbonate also cannot be considered due to the low absorption coefficient of creatinine at 1045 cm⁻¹ and selected measurement wavelength far away from 1011 cm⁻¹. The significant absorbance of phosphate at 1080 cm⁻¹ is the predominant interference in the determination of glucose from ISF by mid-IR spectral analysis.

Due to the interfering species and broad absorption lines of glucose in the liquid phase, a tunable wavelength laser should be used to measure glucose via mid-IR spectroscopy. The laser source can be employed to improve the measurement resolution because of the high laser peak power and high spectral resolution. Optical parametric oscillator (OPO) laser sources [26, 27], difference-frequency generation (DFG) lasers [28, 29], semiconductor lasers (such as lead salt lasers (LSL) [30, 31] and continuous-wave (CW)-QCL) [13, 32] and tunable carbon dioxide (CO₂) gas lasers [33–35] are most often used in the laboratory and have an emission band containing the region of 8.333~10.526. The OPO and DFG source can be tuned over a wide spectral range, but their system is inherently very complex, and they cannot be easily moved. Furthermore, the emission quality of the DFG laser line in the mid-IR band suffers from discontinuous tunability and a low output power of the pump laser source. Moreover, LSL and CW-QCL exhibit large beam divergences and astigmatism, and they often require cooling to cryogenic temperatures. The low power of LSL (generally 0.1~0.5 mW) especially limits its application in spectrum analysis. A CO₂ laser, which has a relatively compact design, narrow line width, is preferably operated at room temperature and can be tuned over a broad range (9.183-10.860$\mu m$) [35, 36], can be easily and inexpensively operated compared with CW-QCL and has demonstrated a robust ability to determine the glucose level in combination with an ATR sensor [33–37]. However, due to the operation characteristics of CO₂ lasers, simultaneously tuning the spectral line, the stability of the line and power is very difficult. Previous studies have shown room for improvement. In this paper, a method was proposed to control and stabilize the laser emission wavelength and power. A quasi-continuous pulsed CO₂ laser with an emission band of 9.19~9.77 (1024~1088 cm⁻¹) is presented. The two strong absorption peaks of glucose (1080 and 1035 cm⁻¹) were contained in this band.

A long effective sensing optical length was required to improve the measurement sensitivity due to the weak signal of glucose in the body. Furthermore, the sensor also should be sufficiently small to be implanted in tissue. However, striking a balance between small size and a long sensing optical length is impossible for traditional crystal ATR sensors. Nevertheless, fiber-optic ATR sensors can achieve both a long sensing optical length and small size via structure modifications, such as a coiled design. Furthermore, tissue heating can be avoided by the use of fiber-optic ATR sensors to determine glucose in vivo. In this study, several small fiber-optic ATR sensors with long sensing optical lengths were designed, and the effective optical sensing lengths of these sensors were quantified for the first time. An AgCl/Br material with a broad transmission band of 4~18 μm and low loss in mid-infrared region was chosen as the material for fiber-optic ATR sensor because of its excellent characteristics, such as its non-toxic, non-hygroscopic, non-brittle and biocompatible nature [38, 39].

Based on the tunable wavelength laser source and small fiber-optic ATR sensor, a high-precision system to measure was established, and its potential was evaluated in vitro. Clinically acceptable results indicated that the research presented in this paper is encouraging toward the development of a minimally invasive method to monitor glucose continuously in vivo.

2. Material and methods

2.1 Dual path laser measurement system

A dual path was employed to overcome the fluctuation of laser power in this study. To avoid heating and/or damaging the sample and exceeding the detector threshold (maximum incidence power: 2 mW) due to high-power (hundreds of mW) during experiment, the laser output power was attenuated by an infrared attenuator (Model 401, Lasnix, Berg, Germany) with an adjustable transmittance from 0.01% to 100%. A Zinc Selenide (ZnSe) beam splitter was used to divide the laser beam into dual paths, one for sample measurement and the other for reference. A ZnSe lens was used to couple the incidence light into the fiber-optic-based ATR sensor, and emergent light was detected by a pyroelcetric detector (LME-353-63, InfraTec GmbH, Dresden, Germany). The reference path light was directly focused on same model detector by a ZnSe lens, and the two detectors were served by two lock-in amplifier circuits of the same model (SR830, Stanford Research Systems, Inc., California, USA), for which a synchronous reference frequency (750 HZ) was offered by the RF circuit of the Laser system. Monitoring the reference detector output eliminated the fluctuation of laser power. Thus, the absorbance change due to glucose concentration is calculated by the following:

2.2 Tunable pulsed CO₂ laser

A method was proposed to control and stabilize the laser emission wavelength and power, and a wavelength tunable CO₂ laser (Merit-G, Access Laser Co., Washington, USA) with a broad quasi-continuous emission spectrum from 9.19 to 9.77$\mu m$(1024~1088 cm⁻¹) and maximum power of 800 mW is presented. The laser was operated in TEM₀₀ mode with air-cooling. The polarization state of the laser beam was perpendicular to the horizontal plane with a waist diameter of <2 mm and divergence angle of <5 mrad. As shown in Fig. 1, the laser was driven by the RF driver (pulsed at 750 HZ). The wavelength tuning was accomplished by adjusting the linear step motor to rotate a grating, and the stabilizing of emission line frequency and power were achieved by modulating the cavity length. The output light was divided by a beam splitter (yellow), and approximately 5% of light was reflected and detected by the IR detector (green) as the sampling signal, which provided feedback to control the working voltage of piezo actuator to modulate the cavity length, and achieve the stabilizing of tuning line frequency and power. The linear motor and its driving can be considered as a module to select the wavelength from the laser emission lines band. The feedback sampling signal and drived piezo actuator can be considered as a module to stabilize the emission line frequency and power. A step linear motor was first used to coarsely adjust the emission wavelength, and the piezo actuator was then used for the fine adjustment. All operations were controlled by the control board with Labview program.

 

Fig. 1 Control schematic of the CO₂ laser

Download Full Size | PDF

2.3 Structural design of fiber-optic ATR sensor

A silver halide fiber with an outer diameter of 700$\mu m$and a core diameter of 630$\mu m$ (A.R.T. Photonics GmbH, Berlin, Germany) was used in this investigation. This fiber was bent into loop sensors. The refractive index of the core and cladding were$n_{core}=2.15$and$n_{clad}=2.13$, respectively. The effective sensing optical length needed to be increased in all the measurement based on fiber-optic ATR sensor to enhance the sensor sensitivity as much as possible. Generally, the geometrical shapes of the sensor elements were changed to increase both the penetration depth and reflection numbers within the element. As shown in Fig. 2, the evanescent field showed a poor penetration depth when the standard cylindrical unclad fiber (Fig. 2(a)) served as a sensor, and the sensitivity was restricted by the fiber sensing length. Furthermore, the flattened and tapered fiber-optic ATR sensor showed increased sensitivity [40, 41], but they are not suitable to be used in a short and small space. The U-shaped (Fig. 2(e)) sensor could be used to increase the number of reflections and the penetration depth at each reflection [42, 43], which enhanced the sensitivity. The semi-circular sensor (Fig. 2(b)), double-coiled sensor (Fig. 2(c)) and triple-coiled sensor (Fig. 2(d)) also provided a method to increase the sensitivity by a mechanism similar to that of the U-shaped sensor. Thus, the effective sensing optical length was significantly increased in a limited and short cell. The bent radius of these sensors was 2.5 mm in this study, and the fiber-optic sensor was sealed in a Plexiglas chamber to construct a flow through the fiber-optic ATR sensor. The chamber had circular openings to pump the sample solutions to the sensor chamber. The HATR (Horizontal ATR) sensor accessory of the FT-IR spectrometer (spectrum TM GX I system, Perkin Elmer, Massachusetts, USA) was employed as a standard cell for comparison purposes, and the experiment results were compared with that of the fiber-optic ATR sensors.

 

Fig. 2 Different shapes of optic-fiber sensors: (a) cylindrical, (b) semi-circular, (c) double-coiled, (d) triple-coiled, and (e) U-shaped.

Download Full Size | PDF

2.4 Noise-equivalent concentration

The NEC can be used quantitatively evaluate short-term noise and long-term system drift. The Allan variance [15] ${\sigma }_{Allan}^2(\lambda (n))$is often used to characterize the noise and system drift as a function of a given time interval$\tau$:

3. Results and discussion

3.1 Laser emission spectrum

The laser emission spectrum was obtained with a Spectrum^TM GX I FT-IR spectrometer (PerkinElmer, Massachusetts, USA). The laser emission line was adjusted first, followed by the incidence through the external window of the FR-IR spectrometer. The DTGS (Deuterated triglycine sulfate) detector operated at room temperature and served as the infrared detector. Each laser emission line was recorded several times with a resolution of 0.1 cm⁻¹ to evaluate the repeatability and stability of emission wavelengths. The laser emission lines in the band of 9.17~9.77$\mu m$contained 36 wavelengths. However, at most five wavelengths can be output at a time, and the five defined wavelengths were individually outputted and then circled (T = 30 S). The duration was 6 s for each emission line, and the time to switch the wavelength was stable (<0.5 s). As shown in Fig. 3(a), the full width at half maximum (FWHM) of the laser emission line at 1081.2 cm⁻¹ was approximately 4 cm⁻¹ (9 nm), and the spectral resolution ($\lambda /\Delta \lambda$) was approximately 2.7 × 10⁴, which is comparable to a quantum cascade laser source [15, 25]. The actual FWHM of CO₂ laser was about 10⁻⁶ cm⁻¹, and the low spectral resolution of the FT-IR spectrometer led to a virtual spectral broadening of the lines. However, the system based on FT-IR spectrometer for laser emission lines measurement can be used to observe the line position. The emission line peaks did not shift during the experiment. The stability of energy at 1081.2 cm⁻¹ was evaluated by the CV (variation coefficient: ratio of standard deviation to mean value) of integrated value from 1076 to 1086 cm⁻¹, and a fluctuation of <0.7% was achieved. The other emission lines (1076.0, 1050.6, 1041.4 and 1037.4 cm⁻¹) were also investigated, and similar characteristics were obtained (data not shown). Figure 3(b) clearly shows that the five laser emission lines, including 1081, 1076, 1051, 1041 and 1037 cm⁻¹ (recorded by Spectrum^TM GX I spectrometer), were all located at the absorption peaks of glucose (1080 and 1035 cm⁻¹) in PBS solution ranging from 950 to 1200 cm⁻¹. The standard glucose absorption spectrum was also recorded by the Spectrum^TM GX I FT-IR spectrometer in combination with the HATR sensor cell at a resolution of 4 cm⁻¹.

 

Fig. 3 (a) Laser emission line at 1081 cm⁻¹. The emission line was recorded several times, and the emission wavelength peak did not shift over time. (b) Normalized CO₂ laser emission lines and glucose absorption spectrum in PBS solution. The laser emission line was performed using spectrum^TM GX I FT-IR spectrometer with resolution of 0.1 cm⁻¹. The glucose absorption spectrum was also recorded by a FT-IR spectrometer in combination with a HATR sensor at resolution of 4 cm⁻¹.

Download Full Size | PDF

3.2 Determination of effective optical sensing length

The effective sensing optical length can be easily calculated for a cylindrical optic-fiber ATR sensor. However, this length is difficult to calculate for bent sensors. The relative effective sensing optical length of these fiber-optic sensors was investigated by employing the HATR accessory as the reference cell. No interferences were introduced during this experiment. The plot of absorbance at 1081 cm⁻¹ versus glucose concentration is shown in Fig. 4. According to the Lambert-Beer law ($A_{{\lambda }_i}={\epsilon }_{{\lambda }_i}cL$), the absolute absorbance of the analyte,$A_{{\lambda }_i}$, correlates with the absorption coefficient,${\epsilon }_{{\lambda }_i}$, at incidence wavelength,${\lambda }_i$, analyte concentration,$c$, and effective optical sensing length,$L$. The effective sensing optical length of the HATR cell was calculated to be 0.0470 mm at 1081 cm⁻¹ for an incidence angle of 45°, length of 48 mm and thickness of 2 mm; the refractive indices of purified water and ZnSe crystal were 1.3333 and 2.4118, respectively. Therefore, the effective sensing optical lengths at 1081 cm⁻¹ were calculated to be 0.0083, 0.0085, 0.0232 and 0.0411 mm for the semi-circular, U-shaped, double-coiled and triple-coiled sensor, respectively.

 

Fig. 4 Plot of absorbance at 1081 cm⁻¹ vs. Glucoseconcentration for fiber-optic sensors and HATRcell combined with tunable CO₂ laser

Download Full Size | PDF

The fiber used in this paper was bent directly without peeling. Therefore, incidence light did not leak from the fiber if the bent radios exceeded 7.5 cm. Thus, the effective sensing optical lengths were nearly identical for the semi-circle and U-shaped sensors. The experimental results were consistent with the geometrical structures, and determination of the optical sensing length assists the fabrication of fiber-optic ATR sensors.

3.3 Determination of NEC

In this study, a PBS solution was used only in the analysis of long-term measurements. The NEC for varying integration times,$\tau$(interval 30 s), is shown in Fig. 5. A minimum NEC should be reached after a certain integration time, but the NEC value slightly increased with the integration time, as shown in Fig. 5. These data suggest that the system was unstable. However, the minimum NEC of the laser experiment system (solid line) was only 3.8 mg/dL for an interval time of 30 S, which is among the most precise glucose measurements using mid-infrared spectroscopy. A NEC value below 10 mg/dL was observed at 1037, 1041, 1051 and 1076 cm⁻¹ when the integration time was below 100 s. The NEC obtained for this experiment setup is comparable that of fiber-optic based laser transmission spectroscopy [25] and better than that of fiber-based lead salt laser spectroscopy [15]. Low NEC values were observed at wavelengths of 1037, 1041, 1051 and 1076 cm⁻¹ but not 1081 cm⁻¹ because the power stability was different at the five laser emission wavelengths, and the power fluctuation was slightly lager at 1081 cm⁻¹.

 

Fig. 5 Plot of NEC vs. sample interval time

Download Full Size | PDF

3.4 Sensitivity of experimental system

The HATR accessory was used as a sensor in this experiment; the slope of the linearly fitted line of the glucose absorbance at 1081 cm⁻¹ versus the glucose concentration is referred to as the sensitivity. Interferences were not introduced during this experiment. The interference of the detector inherent noise could be eliminated when a light source with a higher spectral power was used. Moreover, the absorption law dictates that the monochromaticity of the light positively correlates with the sensitivity of the system. Figure 6 shows that the sensitivity of the laser system was approximately 4 times that of the FT-IR spectrometer due to the higher spectral power density and better emission line resolution of the laser source.

 

Fig. 6 Contrast of the results measured by the laser measurement system and FT-IR spectrometer

Download Full Size | PDF

3.5 Glucose measurement

Fourteen different glucose (Guangfu sci., China) solutions were prepared with a PBS solution to result in concentrations of 5~500 mg/dL, including 24 samples of 5~15 mg/dL $(\Delta c=5mg/dL)$, 20~100 mg/dL $(\Delta c=10mg/dL)$, 125~300 mg/dL $(\Delta c=25mg/dL)$ and 300~500 mg/dL $(\Delta c=50mg/dL)$. The PBS solution was prepared with 136.89 mmol NaCl, 2.68 mmol KCl, 10.14 mmol Na₂HPO4 12H₂O and 1.76 mmol KH₂PO4 (Jiangtian Chemical Technology Co., Ltd., China) dissolved in 1 L deionized water, and the PH value was 7.4 (normal blood pH values varies between 7.35 and 7.45). The impact of the disturbing phosphates, mainly NaH₂PO4 and KH₂PO4, on the glucose determination were investigated by adding 2 mM (Physiological concentration: 0.87~1.45 mM) phosphates to the glucose PBS solution. PBS solutions with different glucose concentrations were analyzed by the CO₂ laser spectrometer in a protocolled but random order. Prior to the measurement of glucose in the solution, the PBS solution absorption spectrum was recorded as the background spectrum. To clearly demonstrate the low concentration of samples, a logarithmic curve of the absorbance at 1037 cm⁻¹ vs. the glucose concentration was introduced and is shown in Fig. 7. These data yielded a good linear correlation coefficient ($R^2=0.9813$for double-coiled sensor and$R^2=0.9914$for triple-coiled sensor, respectively). Slightly larger error bars (standard deviation of different experiments) were observed for low glucose concentration samples because of the weak absorption of trace glucose.

 

Fig. 7 Logarithmic curve of absorbance vs. glucose concentration in PBS solution, which was obtained by the laser spectroscopy in combination with the fiber-optic ATR sensor

Download Full Size | PDF

3.6 Glucose concentration prediction and evaluation

A partial least squares (PLS) regression was used for to quantify the predictability of the different glucose solutions, and a cross validation was performed to optimize the PLS model. The prediction quality of the PLS model was evaluated using the root-mean-square error of prediction (RMSEP). Each sample was measured several times to analyze experimental repeatability. In each experiment, the odd sample original data of a given known concentration were used to establish the calibration sets, and the even sample data were used for prediction. The standard deviation (SD) of prediction was calculated at each individual predicted concentration value. The mean predicted value was compared with the reference concentration (taken for actual value) using Clarke’s error grid [44] to evaluate the prediction accuracy. Five regions were divided in Clarke’s error grid analysis, and the descriptions for respective regions were defined as follows: A zone, clinically accepted accurate reading (less than 20% deviation); B zone, benign or invalid reading; C zone, reading that would result in overcorrecting acceptable blood glucose levels; D zone, reading that would lead to dangerous failure of blood glucose value correction; E zone, absolutely erroneous reading that would result in the opposite correction of blood glucose concentration.

The predicted glucose concentrations versus reference glucose concentrations were plotted on Clarke’s error grid, as shown in Fig. 8. The SD of prediction was labeled as the error bar for each sample and shown in Clarke’s error grid. Figure 8 shows that all predicted values fell in the A zone, which indicates that the established tunable laser spectroscopy based on the fiber-optic ATR sensor can robustly and accurately predict the glucose concentration. The predictions obtained from the semi-circle (top left) and U-shaped (top right) sensors showed a slightly larger than the other two sensors due to the short effective optical length of these two sensors. The short optical length results in a weak signal to be detected, which produces a low signal-to-noise ratio. Therefore, a slightly larger SD of prediction and weak linearity were also found at low glucose concentrations (below 50 mg/dL (hypoglycemic range)), which also should be attributed to the weak signal that resulted from low glucose concentrations.

 

Fig. 8 Plot of prediction glucose concentration versus reference glucose concentration on Clarke’s error grid. Measured using semi-circle (top left), U-shaped (top right), double-coiled (bottom left) and triple-coiled sensor (bottom right), respectively.

Download Full Size | PDF

4. Conclusion and outlook

In conclusion, the application of tunable wavelength CO₂ laser spectroscopy based on fiber-optic ATR sensors to determine the glucose concentration over a physiologically relevant concentration range was reported in this study and demonstrated promising results. The laser measurement was approximately four times as sensitive as FT-IR spectrometry, and a low NEC of 3.8 mg/dL was achieved. Glucose prediction accuracy was clinically acceptable. The experimental results for the degradation of silver halide fiber-optic sensors reported by C. Vrančić [25] et al. indicate that silver halide fiber sensors can be potentially implanted in tissue for 20 days. However, implanted enzyme electrodes are generally useful for three to seven days [45, 46]. In addition, significant drift due to bioelectricity, which is caused by electrolytes in body fluid and electrochemical reactions under hypoxia, reduces the accuracy of glucose determination when using an enzyme electrode sensor. In contrast, a fiber-based ATR sensor is not affected by bioelectricity in the body. In summary, the tunable laser source, high system sensitivity, low NEC and small fiber-optic ATR sensor provide a potential minimally invasive method to monitor glucose in vivo. Further work will focus on improving the biocompatibility and applicability of the small fiber-optic ATR sensor in vivo using biocompatible materials, such as polymer coatings and semi-permeable membranes, to make the sensor suitable for tissue implantation.