1.

Introduction

Monitoring and quantification of glucose level in blood is an important ongoing field of research in clinical analysis.^{1, 2} Specifically, monitoring blood glucose by near infrared spectroscopy (NIRS) has become one of the most active areas in biomedical optics and pharmaceutical research investigated over 20 years by many research groups.^{3, 4, 5, 6} More recently, interest in NIRS is on the rise for strict continuous control of blood glucose levels for patients in an intensive care unit (ICU) and operating room (OR) known as tight glycemic control (TGC).⁷ In TGC, blood glucose is frequently (four to six daily measurements) or continuously monitored in order to provide information needed to administer the proper amount of insulin or glucose in order to avoid hypoglycemia or hyperglycemia while maintaining normoglycemia (80 to 110 mg/dL).⁸ Recently, Tang report on the effect of 30 drugs used in hospitals on glucose measurement with handheld glucose meters and a portable glucose analyzer.⁹ To date, however, very little research is available describing the influence of drugs, found in many patients while treated in ICU, on NIRS measurements. The Food and Drug Administration of the United States also requires understanding regarding the effect of drugs on accuracy and precision for blood glucose assays using NIRS. In general, the measurement of blood glucose by any technique is inherently complex because of the wide range of potentially interfering components in blood and body tissues. Therefore, a study designed to screen the influence of interferent substances at clinically relevant concentrations is necessary.

The clinical and laboratory standards institute (CLSI) refer to the term interference as a cause of clinically significant bias in the measured analyte concentration due to the effect of another component or property of the sample.¹⁰ By clinical significance we mean the importance of an error to its potential to alter a physician's diagnosis, treatment, or management of a patient. One of our goals in this study was to evaluate the effects of test interferent compounds on NIRS glucose prediction errors. The main test interferences for this evaluation are saccharide compounds, suspected as having the greatest potential to interfere with NIR glucose analysis. Saccharide compounds can be divided into three main groups: 1. monosaccharides (e.g., fructose, galactose, mannitol, and xylose), 2. disaccharides (e.g., maltose), and 3. polysaccharides (e.g., dextran and hetastarch). These saccharides are suspect since they have similar spectroscopic fingerprints to glucose and may be present physiologically at large relative concentrations. The selection of specific interferents and test concentration levels for interference testing are gathered from a variety of reference literature sources. Although not comprehensive in these literature, sources for interference information include the following: common handheld meter interferences, selected interferences from the EP7-A2 guidelines (high concentrations), and commonly used drugs in ICUs.

To reduce or eliminate the effect of these saccharide compounds on glucose measurements, a preprocessing method such as orthogonal signal correction (OSC) can be used. OSC aims to remove strong structured variation from the spectral data, X, that is unrelated, or orthogonal, to the concentration Y.¹¹ To achieve such orthogonality, X is decomposed into scores (t), which are required to be orthogonal to Y, and loading (p) vectors. Multivariate calibration by partial least squares (PLS) is accomplished with X calibrated against scores and new scores and loadings are recalculated.¹² Next, new scores and loadings, which contain information not related to the concentration, are subtracted from X in order to correct the spectra. For new spectral data, X _new, a new score vector is calculated and multiplied with the transpose of the loading vector previously found. Finally, the product of the two is subtracted from X _new. This type of mathematical treatment has been applied here with one more element; during calibration and validation modeling the bias (intercept) and slope from the prediction data sets, that are not included in the calibration model, were determined. This process was repeated 100 times (∼10% of the model size across the entire set) in order to compute statistic power on the bias and slope parameter. For each new validation spectra, both the prediction vector, [TeX:] $\hat b$ $\widehat{b}$ , and the bias and slope from the calibration process were automatically applied yielding a standardized prediction. A flow chart describing the difference between OSC and slope and bias OSC (sbOSC) is depicted in Fig. 1.

Fig. 1

Flow chart describing the difference between OSC procedures to sbOSC. (a) Calibration, (b) Validation. sbOSC contribution achieved on the validation side.

In this work, the ability to determine glucose concentration in the presence of saccharide interferences in NIRS measurements by using Fourier transform near-infrared spectroscopy (FT-NIRS) with a modification to OSC, for the first time to our knowledge, was investigated. Detailed results demonstrate the use of a single calibration for quantitative determination of glucose levels in whole blood samples from 14 healthy anonymous patients. These patient samples contained maximum physiological levels of six different saccharides. The measurement error effect of the individual interferents was determined based on the maximum interference effect (MIE) of each substance. Furthermore, the detailed use of FT-NIRS in transmission mode coupled with sbOSC and a single PLS calibration model for determination of glucose in whole blood will be described.

2.

Material and Methods

2.4.

Calibration Model

The goal during NIR calibration is to increase analytical specificity or selectivity for glucose and decrease the effects of interferents. For that, a PLS1 model was used to develop a multivariate calibration model for predicting glucose concentration from the NIR spectra.¹⁴ Spectra here refer to the absorbance data given by the modified Beer–Lambert law¹⁵

Eq. 1

[TeX:] \documentclass[12pt]{minimal}\begin{document}\begin{equation} \mu _a = \log _{10} \left({\frac{{I_{bck} }}{{I_{bld} }}} \right), \end{equation}\end{document} $${\mu }_a={\log }_{10}\left(\frac{I_{bck}}{I_{bld}}\right),$$

where μ_a is the absorption coefficient, I _bck is the background intensity, and I _bld is the light intensity from the blood.

Calibrations for glucose must be designed such that they are sensitive to changes in glucose concentration but are insensitive to interferent chemicals. A method for making calibrations more selective to the analyte of interest and less selective toward an interferent includes the use of “interferent compensation.” This is accomplished by including low (nonzero), medium, and high concentration samples into the initial calibration and testing the sensitivity of the calibration to glucose and to the interferent following the incorporation of such samples into the initial calibration. A standard factorial design for including such samples is depicted in Fig. 2. This is implemented by taking the same basic blood sample with three different Hct levels (24, 36, and 48 vol.%) and spiking it to three levels of glucose (80, 140, and 200 mg/dL), and three levels of interferent (low, mid, and high expected physiological concentration). Table 1 summarizes these ranges of expected concentration for each inteferent. These intervals of values are near those commonly used in ICU and in OR. In addition, a combination of spectra from sample types with “real-world” effects (instrument type, sample temperature, flow rate, and flowcell pathlength) was added to the calibration model. Hence, our calibration data is formulated from a mixture of two groups of samples. The first group (N = 1007) is the real-world samples while the second (N = 361) contains the saccharide samples at constant: flow rate (0.5 ml/min), temperature (25ºC) and flowcell pathlength (1 mm) state.

Fig. 2

Simple factorial design samples file to accommodate for single inteferent for each instrument used in this study. L- low, M- mid, H-high.

Fig. 6

Representative cross-validation performance (a) and predicted glucose concentration for a validation set (b) versus actual values observed during modeling. In (c), the slope and bias correction were applied to the validation set, noting the prediction performance is on samples that did not participate in the calibration modeling process.

Table 1

Three level of interferent concentration used in this work. To convert mg glucose/dL to mmol/L, divide by 18.015

Interferent Type	Low (mg/dL)	Mid (mg/dL)	High (mg/dL)
Fructose	6	12	18
Galactose	20	40	60
Mannitol	200	400	600
Xylose	10	15	20
Dextran	1000	2000	3000
Maltose	60	120	180

Outlier detection based on both Mahalanobis distance (MD) and spectral F-ratio (SFR) metrics were applied to ensure the quality of the data used to construct the calibration model. A criterion of μ ± 3σ was used as an outlier screening for each metric. μ and σ represents the mean and the standard deviation receptively of the MD (or SFR). Samples that exhibit MD and SFR above this criterion were considered an outlier and removed from the calibration model. OSC was applied for preprocessing of the NIR spectra before being used in the PLS model. As stated in Sec. 1, the main goal of OSC is to determine and remove from the spectral data, X, the part of information that is not correlated with the concentration, Y, within a limited number of orthogonal scores (t) and loadings (p). Filtered data are obtained after iterative removal of the orthogonal components as follows:

Eq. 2

[TeX:] \documentclass[12pt]{minimal}\begin{document}\begin{equation} X_{OSC} = X - \sum\limits_{i = 1}^n {t_i \cdot p_i ^T }, \end{equation}\end{document} $$X_{OSC}=X-\sum _{i=1}^n{t}_i·p_i^T,$$

where n is the number of times that the treatment is applied. Usually, one factor is commonly used in OSC correction; the second treatment on the corrected X data can remove useful information reducing the predictive analytical ability.¹⁶

After outlier screening and OSC treatment, a PLS calibration matrix is formed. A leave-one-out method of cross-validation was performed on the calibration data set to determine the optimal number of factors (i.e., latent variables). The PLS model size was chosen as the number of factors resulting in the first minima of the standard error of cross-validation (SECV). Common calibration statistics such as SECV, mean relative error, coefficient of determination (R ²), etc. were computed to assess the fitting ability of the model. Once a model has been built, it can be used to predict the concentration of the unknown samples.

3.

Results and Discussion

The aim of the first experiment was to investigate the effect of the saccharide interferences, on the prediction (validation) data sets. A PLS calibration was built with 1007 spectra from the real-world group as described in Sec. 2.4. The number of PLS factors used in the model and the cross-validation glucose prediction plot in the Clarke Error Grid Analysis (CEGA) format of these data sets are presented in Figs. 3a and 3b, respectively. Usually, CEGA is used to quantify clinical accuracy of blood glucose measurement techniques in comparison to a reference value.¹⁸ The grid consists of five zones A–E and provides an assessment of measurement errors and their respective clinical significance. For example, zone A is the region in which the NIR analytical values are within 20% of the reference values and indicates errors without serious clinical implications. As depicted, a linear correlation of R ² = 0.99 between the NIR results and the YSI glucose concentrations, with SECV = 4.1 mg/dL is observed. These results, together with the fact that 100% of the prediction fall within zone A and zero% results on the other zones, confirm both the ability of PLS to calibrate models by using NIR spectra with real-world effects and the reliability of our FT-NIRS modality. Once the calibration is complete, the prediction data sets were build composed by six saccharide interferences: Dextran, Fructose, Galactose, Maltose, Mannitol, and Xylose. For interferent compensation, each interferent was spiked with different concentration levels and different glucose and Hct levels as per a factorial design depicted in Fig. 2; Table 2 shows an example for the Xylose experiment. The NIR predictions are reported under two basic conditions: 1. extrapolation testing without calibration interferent accommodation (compensation) of the test interferent compounds, where the calibration model predicts outside of its calibration space, and 2. interpolation with calibration interferent accommodation (compensation) of the test interferent compounds (i.e., including the extremes within the model), where the model predicts within its calibration space. Figure 4 together with Table 3 presents the results of the prediction in terms of maximum interference effect (MIE). MIE is expressed by

Fig. 3

(a) Plot of the root mean square of standard error of cross-validation (RMSECV) values versus the number of factors used to build the PLS. The arrow indicates the lower RMSECV values results in 27 factors. (b) Cross-validated PLS glucose prediction versus actual glucose concentration with Clark error grid. Cross-validation was performed by leaving one spectrum out of the calibration model at a time.

Fig. 4

Bar graph represents the difference in the MIE values when interferents out (black) or in (white) the calibration model.

Table 2

Example of Xylose sequence measurements. Total sample measurements are equal to 26. Discontinuity in Sample ID results from the measurement of other inteferent according to CLSI Guidelines.

Table 3

Comparison of the prediction results of Maximum Interference Effect (MIE) for six saccharide interferences when exclude and include in the calibration model.

Before assessing the feasibility of sbOSC, a new experiment (experiment #2) was conducted in order to validate the efficacy of the standard OSC on interferent containing sample prediction. In this experiment a PLS calibration was constructed with 696 spectra obtained from a combination of a third set of real-world random samples (N = 335) and saccharide interferences (Ns = 361) having different concentrations. The performance of the calibration model before and after the use of OSC is shown in Figs. 5a and 5b, respectively. Inspection of the correlation plots in Fig. 5b in comparison with Fig. 5a reveals that with OSC processing the prediction results are considerably concentrated along the unity line with R ² = 0.99 and a ∼70% decrease in SECV (19.1 mg/dL versus 5.8 mg/dL). In addition, Fig. 4b shows that a high percentage of the NIR predicted data points fall within the zones A and B of the Clarke grid; 99.9% and 0.1% of all NIR predictions fall in the A and B zones, respectively. On the contrary, without using OSC, 84.2%, 15.2%, and 0.6% of the predictions fall in the A, B, and D zones, respectively, resulting in a much higher error expressed as SECV.

Fig. 5

Cross-validated results of glucose prediction versus reference values shown on Clark error grid before (a) and after (b) using OSC.

The next experiment's (experiment #3) purpose was to establish a statistical method for computing the calibration slope (m) and bias (i, intercept) used later during the sbOSC procedure. Eighty percent of the data set of random real-world samples (N = 807) were combined together with saccharide interferences (Ns = 361) in order to build an NIR calibration model. OSC was applied for preprocessing of the NIR spectra of these mixed samples (N _T = 1168) before being used in the calibration modeling step. The validation data sets were composed of the same saccharide type molecules, however, as previously mentioned, with high physiological concentration levels (Table 1, third column). In addition, the spectra of these validation samples were collected on the same two FT-NIRS instruments 35 days after the samples used for calibration. A total of 100 calibration models were developed by selecting different random samples for calibration and validation in order to determine the distribution for the slope and bias. For each set of experiments m and i metrics were computed according to the following expressions:

After identifying the slope and bias factors, a random data set of entire real-world samples (N = 200) were combined together with the saccharide interference samples (N = 361) to construct the PLS model for the calibration. Prediction data sets, not included in the calibration model, were employed as an independent test set. Therefore, our validation process is obtained on “blind” saccharide interferents samples. In addition the prediction equation from Eq. 4, was modified accordingly by

Table 4

Comparison of calibration and validation performance between sbOSC to several background correction methods. Validation results are present as the Maximum Interference Effect (MIE) in mg/dL. The validation data sets are saccharide interferences non-included in the calibration model.

Table 5

Comparison of calibration and validation performance of sbOSC when Dextran excluded during modeling. Validation results are present as the Maximum Interference Effect (MIE) in mg/dL.

4.

Conclusions

There are medicines that can cause hypoglycemia (low blood sugar) or hyperglycemia (high blood sugar) in diabetic and nondiabetic patients. These drugs have a possible effect to interfere with NIR glucose measurements. In this paper, a method for determination of glucose concentrations in blood in presence of saccharide interferences using FT-NIRS and sbOSC as a pretreatment method was demonstrated. sbOSC is based on the traditional OSC platform but in addition uses a universal slope and bias correction on the prediction sets. The slope and bias were computed during calibration using 100 different calibration models and are used later for the validation prediction and routine analysis. A clear advantage of OSC and sbOSC presented in MIE value performance is observed in comparison to other basic preprocessing spectral correction techniques (Table 4). The novelty of this work is our ability to directly quantify glucose value in the presence of saccharide interferences in NIR region using FTS, and the ability of the sbOSC approach to further decrease the MIE values by an additional 10% for each interferent over OSC-PLS. To the best of our knowledge, no previous report of such a study and calibration development has been previously described. The overall results presented in this study clearly show the potential and versatility of our methodology, which could be applied for blood glucose assay in ICU and OR rooms.