Mid-infrared photoacoustic spectroscopy using a quantum cascade laser for non-invasive blood component analysis

We developed a photoacoustic spectroscopic method using mid-IR light for non-invasive analysis of blood components in living bodies. The ultra-low-volume photoacoustic cell enabled highly sensitive measurement, and, using a glucose-containing gel, the photoacoustic spectrum showed an almost linear relationship with the absorption spectrum. The optimum modulation frequency was determined both theoretically and experimentally using the photoacoustic spectra of glucose gels obtained at different modulation frequencies. The photoacoustic spectrum of the human wrist was measured at the same time as blood glucose levels were measured by blood sampling. Discriminant analysis of whether the blood glucose level was higher or lower than 130 mg dl−1 was relatively accurate (70.8%). The wavelengths used for discrimination were those absorbed by insulin and lipids, the levels of which change according to the blood glucose levels, and that absorbed by glucose.


Introduction
Analysis of blood components including proteins, lipids, and carbohydrates, and metabolites thereof, is important for maintenance of health. Such analyzes are usually based on blood tests; it would be extremely useful if they were both immediate and noninvasive. Although methods based on optical spectroscopy including Raman 1,2) and fluorescent spectroscopy 3) have been used for non-invasive blood component analysis, absorption spectroscopy is superior in terms of simplicity, sensitivity, and adaptability to various materials.
Various non-invasive absorption spectroscopic methods using near-and mid-IR light have been developed; mid-IR spectroscopy is more sensitive because it detects absorption corresponding to the fundamental vibrations of proteins, carbohydrates, lipids, amino acids, and water. 4) Mid-IR spectroscopy has been used for in vitro monitoring of blood glucose levels in diabetics. 5) Several studies performed in vivo analysis of blood components using mid-IR attenuated total reflection (ATR) spectroscopy, which yielded human blood glucose levels. [6][7][8][9][10] However, the penetration depth of the ATR evanescent field for living tissue is limited to a few micrometers because of the extremely high absorption coefficient of water in the mid-IR region. Thus, measurements on biological tissues are limited to non-corneal areas such as the lip mucosa.
Recent mid-IR photothermal spectroscopic methods measured the heat generated by light absorption. The penetration depth of mid-IR light is about 25 μm, and these methods detect components in interstitial fluid under a stratum corneum 15-20 μm in thickness. One such method is photothermal deflection spectroscopy (PTD), 11,12) in which a laser beam is directed onto the sample. The heat generated by absorption of mid-IR light changes the refractive index of a sensing prism in contact with the sample, in turn deflecting a second probing laser beam that is detected by a position sensing detector. Photoacoustic spectroscopy (PAS) detects acoustic waves produced by heat generated on absorption of light. [13][14][15] Non-invasive blood component analytical systems based on PAS are simpler, less expensive, and smaller than PTD systems. However, PAS sensitivity is slightly inferior to that of PTD.
Here, we investigated the practical feasibility of the PAS system for blood glucose measurement. To obtain sufficient sensitivity for blood glucose measurement, we improved PAS sensitivity using a very small PA cell. Then we identified the measurement conditions affording maximum sensitivity and stability, and evaluated performance using a biological phantom and human subjects.

Methods
A schematic of the system is shown in Fig. 1. An external cavity quantum cascade laser (EC-QCL: Hedgehog; Daylight Solutions, USA) for which the wavelength was tunable from 920 to 1200 cm −1 served used as the light source. The QCL emits laser pulses at 500 kHz with a width of 100 ns and average power of about 5 mW. The light is modulated by an optical chopper, expanded by a beam expander, and focused by an off-axis parabolic mirror, and the sample is directly irradiated from below. Acoustic waves emitted from the sample are detected by a condenser microphone (Type4188; Brüel and Kjear, Denmark) with a preamplifier (Type2699; Brüel and Kjear) and the amplified signal is sent to a lock-in amplifier (LI5640; NF Corporation).
For biological samples with high water content, the photoacoustic signal intensity I is: 16) where P is the irradiated laser light power, α is the optical absorption coefficient of the sample, V is the volume of the PA cell, and f is the modulation frequency of the light. Thus, reducing the volume of the PA cell improves sensitivity. We fabricated the PA cell shown in Fig. 2. The brass sensing block has a cavity 2 mm in diameter and 2 mm deep that serves as a PA cell and is connected to the microphone by a hole 1 mm in diameter and 8 mm in length. The volume of this PA cell is much smaller than those of other cells described to date, and the connection was as short and thick as possible (within the structural limitations of the system) to minimize data loss. The bottom of the cell is sealed with a ZnSe window and the sample is pressed tightly against the top opening.
We measured the PA spectra of gelatin with 5% (w/v) aqueous glucose. Glucose-free gelatin served as the reference; the wavenumber step was 2 cm −1 and the acquisition time was 2 min. Figure 3 shows the PAS and ATR spectra of the same sample. The latter were measured using a prism with eight reflective surfaces attached to a Fourier transform (FT) IR spectrometer. The shapes of the two spectra are almost identical, confirming that the PAS spectrum is in an almost linear relationship with the absorption spectrum.
The thermal diffusion length μ s of a temperature wave generated at the surface of a sample after absorption of laser light modulated at frequency f in the direction of the sample depth is: 17,18) where D is the thermal diffusivity of the sample, k is the thermal conductivity, ρ is the specific gravity, and c is the specific heat. Thus, depth profiling is possible by varying the modulation frequency. 19,20) Reference 18 shows that when the thickness of a sample d is sufficiently greater than the light penetration depth μ a , a PA signal proportional to the optical absorption coefficient α is obtained when μ a = 1/α > μ s . From Eq. (1), the lower the modulation frequency f, the higher the signal strength I, however if f is set too low, the condition μ a > μ s is not satisfied as shown in Eq. (2), so μ a ≈ μ s becomes the optimal value. For an epidermal α value ∼400 cm −1 from 1000 to 1100 cm −1 , 13) μ a is 25 μm and the modulation frequency f for μ a = μ s is approximately 60 Hz using the material parameters of Ref. 13.
To verify this, we measured the PAS spectra of gelatin with 5%, 3%, and 1% (w/v) aqueous glucose at different modulation frequencies and calculated the signal-to-noise ratios (SNRs) assuming that variations in the spectra of the glucose-free sample were noise and the absorption peaks in the spectra of the glucose-containing samples were signals. As shown in Fig. 4, the highest SNR was obtained at modulation frequencies of 60-70 Hz, as predicted by the theoretical calculation above.
When evaluating humans, we used the palm side of the wrist where the stratum corneum is relatively thin (10-20 μm). Before measurements, nitrogen gas was blown into the cell for about 5 s to prevent accumulation of water vapor from the skin. When obtaining PA spectra that linearly reflected absorption, a polyurethane gel coated with blackbody spray (TA410KS; Ichinen TASCO, Japan) served as a reference. The measurement conditions were as described above.
An oral glucose tolerance test (OGTT) was performed to explore the correlation between the blood glucose levels measured via blood sampling and photoacoustically. The OGTT was performed as directed by the Japan Diabetes Society; subjects fasted for at least 10 h and then ingested 75 g of glucose solution; blood glucose levels were measured and PA spectra simultaneously obtained every 5 min for 2 h. The one-touch Verioview (Johnson and Johnson, Japan) was used for self-monitoring of blood glucose levels; the subject was a healthy adult male. Figure 5 shows the PA spectra obtained from the wrist measurements and the absorption spectra recorded using a pseudo-black body as the reference. We acquired eight spectra at times when the blood glucose values ranged from 111 to 200 mg dl −1 . Several absorption peaks may be observed, derived from proteins, phospholipids, and cholesterol esters. The possible components contributing to the peaks include PO 2  . However, as a clear glucose spectrum such as that in Fig. 3 was not seen, a partial least squares regression analysis was attempted using the measured blood glucose levels as the objective variables and the PA spectra as the explanatory variables. To this end, we employed commercial software (EXCEL Multivariate Analysis ver. 8.0; Esumi, Japan).

Results and discussion
We first attempted to draw a correlation plot between blood glucose levels estimated from the spectra and the blood values, but this was not successful. We next used discriminant analysis to check for differences between high and low blood glucose levels. This method considers the variance of each class, and identifies a threshold at which the extent of variance is maximized by dividing the interclass variance by the intraclass variance. The discriminant analysis formula is: where y is the objective variable, x n are the explanatory variables, a n are the discriminant coefficients, and b is a constant. In the present analysis, the objective variables are "high" or "low" blood glucose, and the explanatory variables are the signal intensities at each wavelength. Therefore, by substituting the values into the relational equation, if y is > 0, the blood glucose level is judged to be "high"; if y is < 0, the blood glucose level is judged to be "low." Also, variable selection based on the variable variance method was used to  The Japan Society of Applied Physics by IOP Publishing Ltd extract features that affected high and low blood glucose levels in component analysis. With this method, the most important explanatory variables are selected first, followed by those with the highest coefficients of determination in combination with the explanatory variables, but even those that have been added once can be eliminated if they are deemed unnecessary during the intermediate steps. This method selects a small number of explanatory variables that strongly influence an objective variable. Figure 6 shows the discrimination scores for 144 spectra acquired over 6 d in one healthy adult. A blood glucose level of ⩾130 mg dl −1 was considered "high" and a level <130 mg dl −1 was considered "low"; the accuracy was good at 70.8%. Four wavenumbers were selected by variable selection: 1048, 1032, 1022, and 1072 cm −1 .
Next, discriminant analysis was performed using only 64 of the 144 data points, of which 37 exhibited blood glucose levels of ⩾180 mg dl −1 and 27 had blood glucose levels <100 mg dl −1 . Figure 7 shows the averaged spectra for    The Japan Society of Applied Physics by IOP Publishing Ltd the "high" and "low" blood glucose levels, respectively. The spectral shapes are generally consistent, but there are slight differences, particularly around 1048 cm −1 . Figure 8 shows the discriminant scores after analysis of the 64 spectra. The accuracy was as high as 90.6%. Four wavenumbers, 1052, 1032, 1070, and 1028 cm −1 , were selected; these were almost the same as those chosen when the analysis included all 144 points. Table I lists the components of the biological tissues affected near each wavenumber.
The peaks at 1052 and 1030 cm −1 may be influenced by insulin and lipids, respectively, the levels of which increase as blood glucose levels rise, as well as glucose, the level of which increases as sugar levels rise.

Conclusions
We non-invasively analyzed live biological tissues via PA spectroscopy using light in the mid-IR region. First, an ultralow-volume PA cell was fabricated to ensure high sensitivity; using a glucose-containing gel, we confirmed that the PAS spectrum had an almost linear relationship with the absorption spectrum obtained via FT-IR-based ATR. Then, the optimum modulation frequency was determined both theoretically and experimentally using the PA spectra of glucose gels at different modulation frequencies.
Finally, PA spectrum of the human wrist was measured at the same time as blood glucose levels were measured by blood sampling. No clear correlation between the blood glucose levels and levels estimated via PAS was apparent.
We thus used discriminant analysis to determine whether the blood glucose level was higher or lower than 130 mg dl −1 . Analysis of 144 spectra obtained over a 6 d period achieved a good accuracy of 70.8%. The wavelengths used for discrimination were investigated; absorption by insulin and lipids, the levels of which change according to blood glucose levels, may be involved.
Further human data are required to improve the accuracy of blood glucose estimations. Moreover, the effects of skin moisture and oil, which may compromise measurements, should be further investigated. We will also explore other measurement sites and the effects of ambient and body temperatures. We will initially develop a compact and inexpensive system using multiple single-wavelength QCLs for small numbers of wavelengths that were chosen in this study. Fig. 7. Averaged spectra for the "high" and "low" blood glucose levels.