2.1.

Device

We developed a diffuse-reflectance spectrometer to measure the NIR spectra of RBC samples. Figure 1 is a schematic diagram of the instrument. The instrument consisted of a 150-watt halogen lamp light source (Philips, TYPE 6550), an optical fiber bundle that included both detector and illuminator fibers (Fujikura, cladding diameter 200 μm, core diameter 175 μm, $\mathrm{NA}=0.2$), and mini spectrometer containing a grating and a 256-channel linear array photo detector (Hamamatsu Photonics K. K., C9914GB). The optical signal from the spectrometer was transferred to a computer (Panasonic, Let’s Note), where the optical signal was processed to predict the glucose concentrations.

Fig. 1

The schematic diagram of the diffuse-reflectance spectrometer used in the present study.

The detector and illuminator optical fibers were fixed at the measuring tip (see Fig. 1) of the optical fiber bundle. Figure 2 shows the geometry of the optical fibers at the measuring tip. The tip of the optical fiber was designed such that twelve illuminator fibers surrounded one central detector fiber in a circle. The distance between the detector fiber and each of the illuminator fibers was 650 μm. This design maintained a sufficient average NIR radiation path length and allowed the illuminator to emit high-intensity radiation.

Fig. 2

The geometry of detector and illuminator fibers at the measurement tip of the sample.

For spectral measurement, the measuring tip was directly submerged into the RCC. Inside the spectrometer, radiation from the halogen lamp was gathered through a concave mirror, transmitted through the illuminator fibers, and irradiated into the samples. Then, in the RCC the diffuse-reflected radiation was scattered many times in the turbid media, and only the radiation that reached the detector fiber was transmitted back to the spectrometer. Finally, the transmitted signal was sent to the computer, where the absorbance spectra were calculated. The reference spectra needed to form the absorbance spectra were taken by irradiating a standard 10% reflectance target (Labsphere) at the beginning of the experiment. Although this spectrometer is able to record measurements in wavelength regions between 1100 and 2200 nm, only spectral signals taken in the 1430 to 1850 nm range were used for glucose prediction. The accumulation time of each linear array photoreceptor was 50 ms, and 200 spectral data were averaged per measurement.

The quality of the spectra was also assessed by taking a spectral measurement every 5 min for 120 min and then conducting a root-mean-square (rms) noise analysis of 100% lines on the collected data. Results showed that the average rms of 100% line of the wavelength region between 1430 and 1850 nm was 87.4 μAU, which we predicted to be sufficient for the purpose of the study.

2.2.

Red Blood Cell Samples

The RBC samples used for this experiment were prepared by the Japanese Red Cross Society (JRCS) using two packages of Red Cells Concentrates-Leukocytes Reduced “Nisseki” (RCC-LR) that did not qualify for transfusion after a blood test. The blood from the two packages was not mixed together in this experiment.

To make the RBC sample, the disqualified RCC-LR was centrifuged to separate the RBC portion from the rest of the blood. A typical RCC-LR contains glucose concentrations up to $600\mathrm{mg}/\mathrm{dL}$ and Hct values around 50 to 60%. The collected RBC portion was first adjusted into samples with Hct levels of 50 or 60% by diluting it with physiologic saline solution. Then, within each Hct group, the samples were further adjusted into samples with glucose concentrations of 0, 100, 200, 300, 400, 500, or $600\mathrm{mg}/\mathrm{dL}$. Because the Hct and glucose levels were hand-adjusted, it was likely for the actual Hct and glucose values to deviate from the ideal values. Therefore, the actual Hct and glucose levels were measured using an automated hematology analyzer (Sysmex, K-4500) and Glucose CII-test Wako (Wako Pure Chemicals, Osaka), respectively, and used in the data analysis. Since the measurements were taken from the supernatant of the samples, the effect of haemolysis on glucose measurement was considered negligible.

After the RBC samples were separated into their respective Hct and glucose levels, 25 mL of each sample were transferred to a glass container, inside which the measurement took place. To create different temperature conditions, these containers were first placed in a 25 °C water bath and taken out at the time of measurement. Then, following the first round of measurements, they were transferred to a 27 °C water bath for the second round of measurements. All experimental steps, from preparation of the RBC samples to measurement of the spectra, were completed within one hour; hence, the properties of each sample were assumed to be constant throughout the experiment. Furthermore, the Hct levels and glucose concentrations were not correlated to each other ($r=-0.08$).

Table 1 shows a summary of the experimental conditions and the measured Hct, temperature, and glucose values for each sample. RBC from Donor A was separated into four measurement conditions: 1. Hct 50% at 25 °C; 2. Hct 50% at 27 °C; 3. Hct 60% at 25 °C; and 4. Hct 60% at 27 °C, as shown in Table 1. RBC from Donor B was adjusted to two measurement conditions: 1. Hct 50% at 25 °C and 2. Hct 50% at 27 °C. Each measurement condition had RBC samples with seven glucose levels ranging from 0 to $600\mathrm{mg}/\mathrm{dL}$, making 42 independent blood samples in total. Three spectral measurements were taken for each sample, producing a total of 126 spectral readings available for analysis.

Table 1

Experimental conditions.

2.3.

Prediction Method

Partial least squares (PLS) regression analysis was used to form calibration models for glucose quantification based on the measured data.²² Because the number of data points for each condition was limited, the cross validation method was utilized.²³ Furthermore, the effect of multiplicative scatter correction (MSC) preprocessing method, which is known to reduce the scattering effect for quantitative prediction, was observed using the spectral data from each condition.²⁴

3.1.

Glucose Sensing in Samples from the Same Condition

The glucose concentration in each sample was estimated from all the spectral data taken from its respective condition. Table 2 summarizes the cross validation and calibration regression results of samples from conditions 1 to 6. Fairly robust correlations were obtained between predicted and measured glucose levels for all groups. The average correlation coefficient of cross validation (rCVal) was 0.994, and the standard error of cross validation (SECV) was $17.1\mathrm{mg}/\mathrm{dL}$. Moreover, the average correlation coefficient of calibration (rCal) was 0.998, and the standard error of calibration (SEC) was $9.9\mathrm{mg}/\mathrm{dL}$. The numbers of PLS factors used for all conditions, excluding condition 1, were four or five.

Table 2

Prediction of glucose concentrations in spectral region between 1430 and 1850 nm for RBC samples from conditions 1 to 6.

Our results showed improvement in prediction performances compared to previous studies on optical glucose sensing in blood. For example, Amerov et al. obtained rCVal of 0.980 and SECV of $21.6\mathrm{mg}/\mathrm{dL}$ in the wavelength region between 1550 and 1754 nm in their research utilizing diffuse-transmittance spectroscopy.¹⁹ Compared to their study, the SECV was improved by around $4.5\mathrm{mg}/\mathrm{dL}$ in the present study. Although different types of measurement systems were utilized, their study could serve as a basis of comparison for results from our study because similar experimental setups and a similar wavelength region were used.

A closer look at the results from a single condition also shows that correlations obtained are fairly high. For example, Fig. 4(a) is a regression graph that illustrates the results obtained from samples in condition 4. The correlation was fairly high, with rCVal of 0.997 and SECV of $13.3\mathrm{mg}/\mathrm{dL}$. Furthermore, Fig. 4(b) is a regression vector obtained from the same set of samples. The vector had a positive peak around 1580 nm, which is near the peak absorbance spectra of glucose. In addition, Kim and Yoon also observed two positive peaks around 1580 nm and 1750 nm in their regression vector,²⁰ which are similar to those seen in Fig. 4(b). These evidences support the validity of this calibration model.

Fig. 4

Analysis for diffuse-reflectance measurement between 1430 and 1850 nm for samples from condition 4: (a) cross-validation prediction of glucose concentrations, and (b) regression vector of cross validation model (PLS 4) and absorbance of aqueous glucose ($10\mathrm{g}/\mathrm{dL}$).

3.2.

Glucose Sensing in Samples from Different Conditions

Several datasets were constructed by pooling together spectral data from various conditions, as shown in Table 3. These calibration sets were then used to predict glucose levels to test whether a good calibration model could be constructed from data taken from samples with various conditions. The effects of donor, Hct, and temperature differences on glucose prediction were observed.

Table 3

Calibration models for glucose prediction in wavelength regions between 1430 and 1850 nm using datasets that consist of spectral data from different conditions.

In the first set, data that varied only in the original donor were combined (see Table 3 “donor difference”). The rCVal was 0.971, and the SECV was $36.1\mathrm{mg}/\mathrm{dL}$. For the second dataset, which includes samples with different Hct content (see Table 3 “hematocrit difference”), rCVal was 0.985 and SECV was $29.9\mathrm{mg}/\mathrm{dL}$. For the third dataset, which included spectral data taken at various temperatures (see Table 3 “temperature difference”), rCVal of 0.990 and SECV of $16.5\mathrm{mg}/\mathrm{dL}$ were obtained. The last dataset consists of samples that differed in all parameters, including donor, Hct, and temperature (see Table 3 “donor, hematocrit, temperature difference”). Its rCVal was 0.969, and its SECV was $36.0\mathrm{mg}/\mathrm{dL}$.

The dataset for “temperature difference” exhibited the highest correlation between the reference and predicted glucose concentrations, which is likely due to the fact that samples from this set had the same donor and Hct values and differed only in their measured temperatures. Because the scattering properties of samples with different temperatures were relatively constant, a better calibration model was achieved. In contrast, for datasets that contained samples with varying scattering properties due to Hct or donor differences, the accuracy of glucose prediction decreased.

3.3.

Effects of MSC Preprocessing Method

Spectral data from samples in Tables 2 and 3 were preprocessed by MSC method. Tables 4 and 5 summarize the calibration models created from the MSC preprocessed data. The effect of MSC preprocessing can be seen by comparing Fig. 3(a), which displays all 126 raw spectra, with Fig. 3(b), which illustrates MSC preprocessed spectra from all 126 samples. It is evident that the multiplicative and additive variations in the raw spectra were compensated by the MSC preprocessing method.

Table 4

The effects of MSC preprocessing on the prediction of glucose concentration in a spectral region between 1430 and 1850 nm for RBC samples from conditions 1 to 6.

Table 5

The effects of MSC data preprocessing on glucose prediction in wavelength regions between 1430 and 1830 nm.

Table 4 shows the prediction results using MSC preprocessed data from RBC samples in conditions 1 to 6. After MSC preprocessing, the rCVal and SECV averages for all six conditions were 0.992 and $19.7\mathrm{mg}/\mathrm{dL}$, respectively. Results show that the MSC preprocess did not improve the quantitative determination accuracy of the raw data (average raw $\mathrm{rCVal}=0.994$; $\text{raw}\mathrm{SECV}=17.1\mathrm{mg}/\mathrm{dL}$). These results are most likely due to the small differences in scattering properties between each raw sample, which minimize the effect of the MSC preprocess.

The MSC preprocess was also performed on datasets that combined samples from various conditions (see Table 5). Glucose concentrations were then predicted using the processed data. The rCVal and SECV for the “donor difference” dataset was 0.982 and $28.4\mathrm{mg}/\mathrm{dL}$, respectively, an improvement compared to those of nonpreprocessed dataset, which had an SECV of $36.1\mathrm{mg}/\mathrm{dL}$. This improvement was probably due to the MSC preprocess’s effect, which diminishes the variation in scattering properties of the blood from different donors.

The same effect was seen in the dataset for “hematocrit difference.” The MSC preprocessed rCVal of 0.993 and SECV of $20.9\mathrm{mg}/\mathrm{dL}$ were better than the raw SECV of $29.9\mathrm{mg}/\mathrm{dL}$. In addition, in the dataset for “donor, hematcrit, temp difference,” the MSC preprocessed rCVal of 0.980 and SECV of $29.3\mathrm{mg}/\mathrm{dL}$ were better than the raw rCVal of 0.969 and SECV of $36.0\mathrm{mg}/\mathrm{dL}$. Again, this effect was probably due to the decrease in scattering properties caused by Hct and donor differences following MSC preprocessing.

Interestingly, the dataset for “temperature difference” did not seem to be affected by MSC preprocessing. The rCVal and SECV after MSC preprocessing were 0.989 and $16.6\mathrm{mg}/\mathrm{dL}$, respectively, whereas the rCVal and SECV for the unpreprocessed were 0.990 and $16.5\mathrm{mg}/\mathrm{dL}$, respectively. This is probably because this dataset contained samples that had the same donor and Hct level. The difference in measuring temperature alone did not cause variance in scattering property, which was significant enough for MSC preprocess to pose an effect.

In the present study, the MSC-preprocess method was able to effectively reduce the scattering properties created by different blood properties. The same effect, though, could not be observed in Kim and Yoon’s study.²⁰ One possible explanation for this discrepancy is that different optical measurement systems were used (i.e., diffuse-reflectance as opposed to diffuse-transmission spectroscopy).