Noninvasive Blood Glucose Measurements
with Near Infrared Spectroscopy
Burmeister and Mark A. Arnold
Department of Chemistry and Optical Science and Technology Center Advanced Technology Laboratories
University of Iowa, Iowa City, Iowa 52242
Gary W. Small
Center for Intelligent Chemical Instrumentation Department of Chemistry
Athens, Ohio 45701
One approach for measuring blood glucose in a painless and noninvasive manner involves passing a selected beam of harmless near infrared light though a vascular-equilibrated region of the human body.1,2 In concept, the in vivo blood glucose concentration can be obtained by analyzing the spectral information contained within the light exiting the body. For this approach to work, the spectral signature of glucose must be unique relative to those of all other chemical components within the human body and this glucose specific information must be obtained with sufficiently high signal-to-noise to permit reliable differentiation between glucose dependent signals and signals generated by other matrix components.
Although data processing is capable of enhancing the signal-to-noise ratio (SNR) of near infrared spectroscopic data and sophisticated multivariate data processing algorithms (i.e., partial least squares (PLS) regression and/or artificial neural networks (ANN)) are necessary to selectively extract the glucose-dependent spectral information, the quality of the raw spectral information drives the ultimate analytical performance and the successful implementation of this noninvasive approach.
The purpose of this article is to highlight the optical and spectroscopic considerations relevant to collecting high quality noninvasive near infrared spectra from human subjects. Issues covered in this article include: viable spectral regions within the near infrared for noninvasive clinical measurements, the principal matrix components that attenuate light as it passes through the body, creation of an in vitro model to help characterize noninvasive spectra from human subjects, and the experimental parameters that are the most critical for successful noninvasive clinical measurements. Issues of data processing and multivariate calibration methods are discussed in the accompanying companion article entitled Data Handling Issues for Near-Infrared Glucose Measurements.
Spectral Regions. The entire near infrared region of the electromagnetic spectrum encompasses light with wavelengths ranging from 0.7 to 2.5 microns (14,286 - 4000 cm-1 wavenumbers). For the most part, near infrared spectroscopic information corresponds to harmonics of overtones and combinations of fundamental vibrational transitions more frequently associated with mid-infrared spectroscopy. Overtone and combination absorptions are principally seen for CH, OH, and NH molecular groups. The energetics associated with these transitions result in absorption bands that are broad and featureless with low absorptivities. The ambiguity of CH, OH, and NH groups within biological systems and the physical nature of these transitions result in complex, overlapping spectra. The chemical environment surrounding these CH, OH, and NH groups controls the exact position and shape of these near infrared absorption features, thereby resulting in unique spectral signatures for the molecular species of interest. Selective analytical measurements rely on the uniqueness of these spectra.
Water is a critical matrix component for near infrared spectra of aqueous based clinical samples, such as the human body. The high concentration of water in clinical samples coupled with the relatively strong absorptivity of OH groups result in large water absorbance bands. The strong absorbance of water dictates using the regions between these water bands where sufficient amounts of light are transmitted. The following three regions are generally accessible: 1) the combination region: 2.0 - 2.5 microns (5000 - 4000 cm-1); 2) the first overtone region: 1.54 - 1.82 microns (6500 - 5500 cm-1); and 3) the short-wavelength near infrared (sw-NIR) region: 0.7 - 1.33 microns (14,286 - 7500 cm-1). Glucose has three absorption bands in both the combination region (centered at 2.10, 2.27, and 2.32 microns) and the first overtone region (centered at 1.73, 1.69, and 1.61 microns). Although glucose absorption bands are difficult to measure in the sw-NIR owing to their extremely low absorptivities, bands centered at 0.76, 0.92, and 1.00 microns are reported.3
Noninvasive Human Spectra. The intensity of near infrared light passing through the human body decreases considerably as a function of distance.
Scattering and absorption are responsible for this attenuation. Photons are scattered by the dense, heterogeneous tissue and this scattering phenomenon is the basis for noninvasive technology to image objects located within soft tissues.4,5 Photons are also absorbed by chemical species within the optical path. Water and body fatty are the principal near infrared absorbers as near infrared light pass through the human body.6 Figure 1 shows single beam spectra collected by passing light in the combination (Figure 1A) and overtone (Figure 1B) regions through human tissue. In these experiments, the near infrared light was transmitted through a section of the webbing tissue between the thumb and forefinger of the volunteer. In both cases, light is heavily absorbed at the frequency extremes and a band of light with significant radiant power is observed in the center portion of these spectra. A detailed analysis of these spectra reveals that, to a first approximation, the relative percentages of water and fat within the tissue dictate the shape and quality of noninvasive spectra.6 Noninvasive human spectra are sensitive to the thickness and composition of the tissue through which the incident light is transmitted.
In Vitro Model for Noninvasive Human Spectra. An in vitro model is available for simulating noninvasive human spectra.6 This model is composed of individual layers of fatty tissue and an aqueous buffer. Fatty tissue from any number of animal sources effectively simulate the features in a noninvasive human spectrum generated from body fat. The aqueous buffer layer simulates the water based spectral features. The power of this in vitro model is the ability to examine in a controlled and systematic manner the critical physical and chemical parameters that affect noninvasive blood glucose measurements.
The relative thickness of the water and fat layers controls how well spectra from the in vitro model match actual in vivo, noninvasive spectra.
A regression method is used to identify the best combination of tissue thickness to accurately simulate human in vivo spectra. In this method, absorbance spectra from pure samples of the aqueous buffer and beef fat are combined according to equation 1:
where AH, Aw, and Af correspond to absorbance spectra for these pure samples of human webbing, water, and beef fat, respectively, and ßi values correspond to the respective regression coefficients. Absorbance values (A) are defined according to equation 2:
where I and Io represent transmitted light intensities with and without the sample of interest, respectively. Absorbances are used owing to their additive nature according to the Beer-Lambert relationship. Application of this method involves starting with an in vivo absorbance spectrum from the subject of interest and then fitting this spectrum by adjusting the relative amounts of the pure water and fat absorbance spectra in such a way to minimize the sum of the square of residuals. Model layer thicknesses are computed as the product of the regression coefficient and the known thickness of the corresponding pure samples. The regression model presented here represents a slightly simplified version where absorbance from muscle protein is ignored. This simplification is well justified for noninvasive spectra over the overtone region. Muscle protein absorbance must be considered, however, in the combination region.6
Analysis of in vivo spectra collected from a small population of human volunteers indicates that typical noninvasive spectra can be accurately modeled with a tissue phantom composed of water layers ranging in thickness from 5.0 to 6.3 mm and fat layers ranging in thickness from 1.4 to 4.2 mm.
Figure 2 shows the results of a typical regression analysis for the two-layer model (water and fat) and for a three layer model (water, fat, and muscle). The in vitro phantom obtained from the two-layer model corresponds to a 6.29 mm thick layer of water combined with a 1.37 mm thick layer of beef fat.
The utility of our proposed in vitro model is exemplified in a simple experiment to assess the effect of optical path length on the ability to measure glucose from near infrared overtone spectra. In this experiment, the thickness of the aqueous layer increases from 5.6 to 6.2 mm while the fat layer thickness is held constant at 1.6 mm. The resulting single beam spectra are presented in Figure 3 along with a glucose absorbance spectrum for comparison. The raw intensities of these single beam spectra have been normalized to their respective maximum value to highlight differences in band shape. The glucose absorption band centered at 5920 cm-1 is available for the quantification of glucose from these spectra.
The effect of optical path length has been determined by collecting two series of overtone spectra with the 5.6 and 6.2 mm aqueous layer thicknesses, respectively. The same 1.6 mm thick fat layer was used in both cases. Both spectral data sets included triplicate spectra collected for 80 unique aqueous glucose standard solutions that spanned a concentration range from 1 to 27 mM (18-486 mg/dL). These spectra were collected with a Nicolet 740 FT-IR spectrometer over the 7000 - 5000 cm-1 spectral range. This spectrometer was equipped with a 400-watt tungsten-halogen lamp source, gold-coated optics, calcium fluoride beam splitter, and a one-millimeter diameter InGaAs detector with a 1.9 µm cutoff. A 1.49 µm cutoff long pass filter and a 1.85 µm short pass filter were used to isolate the spectral range of interest. A one-inch diameter, 25 mm focal-length convex lens was used to focus the incident radiation through the aqueous layer and onto the detector. Space constraints required positioning the fat layer in front of the focussing lens.
Temperature of the aqueous layer was maintained at 37.0 ± 0.1°C throughout.
Results from the partial least squares (PLS) multivariate calibration models for these two data sets are summarized in Figure 4. The presented concentration correlation plots are obtained by plotting the glucose concentration predicted from the PLS calibration model relative to the known glucose concentration in the corresponding standard solution. In these plots, open circles correspond to the spectra used to build the calibration model and solid circles represent results from a separate set of spectra used solely for validation purposes. Both concentration correlation plots are superimposed on the Clarke error grid which provides an indication of clinical usefulness of the analytical measurements.7 The Clarke error grid breaks the correlation space into five regions that assess measurement accuracy on the basis of validity of the corresponding clinical decision. Correlation points falling within the A region correspond to the correct clinical decision being made based on the similarity between the actual and predicted glucose concentrations. The E regions, on the other hand, represent zones within the correlation space where the exact opposite clinical decision is made from the predicted glucose value. The E region in the upper left in Figure 4, for example, corresponds to the situation where the predicted glucose concentration is high, but the actual glucose concentration is low. On the basis of high readings from the measurement, a physician might prescribe insulin to treat hyperglycemia. Such therapy is exactly opposite that needed because the patient is actually hypoglycemic. Severity of the incorrect clinical decision increases as one progresses toward the E regions. Ideally, glucose values predicted from the near infrared spectra will fall along the unity line and remain within the A region of the error grid.
Comparison of the scatter of the points in Figures 4 A and B clearly show the superior analytical performance with longer optical path lengths.
Quantitatively, we can compute the standard error of prediction (SEP) for the validation data in both cases. The SEP drops from 1.60 mM to 0.49 mM when increasing the path length from 5.6 to 6.2 mm. It must be noted that the spectral noise is essentially identical for these two data sets (23.4 and 23.5 micro-absorbance units for the 5.6 and 6.2 mm path length data, respectively). Improvements are expected from the Beer-Lambert relationship which states that longer path lengths will provide larger signals, thereby enhancing the signal-to-noise ratio of the measurement and improving analytical performance.
Experimental Parameters. It is important to recognize that noninvasive blood glucose measurements are simply absorbance measurements in a complex matrix. As such, one must be able to differentiate the amount of light absorbed by glucose from spectral noise. In this regard sample thickness is a critical experimental parameter because tissue thickness affects both glucose sensitivity and spectral noise. As the sample thickness increases more light is absorbed for a given concentration of glucose, thereby enhancing sensitivity and lowering the detection limit. On the other hand, fewer photons successfully traverse a thicker layer of tissue, thereby reducing the measured light intensity and increasing spectral noise. A compromise is required to maximize sensitivity to glucose while minimizing spectral noise.
Successful noninvasive clinical measurements require the ability to collect reproducible noninvasive spectra from human subjects. Between run variations must be avoided in the thickness, composition, and temperature of the sampling site. This demand for spectral reproducibility makes the human-to-spectrometer interface critical. The temperature of the interface must be controlled to minimize thermal induced spectral shifts. The compressibility of human tissue further complicates the interface which must fix the amount and thickness of the tissue being sampled while avoiding excess pressure which can degrade tissue integrity.
Finally, near infrared spectroscopy is capable of measuring clinically relevant levels of glucose in complex biological matrices.8-14 The challenge is to make this measurement from spectra collected noninvasively from human subjects. Success demands low spectral noise, long optical path lengths and a reproducible human-to-spectrometer interface.
1. Arnold, M. A. in Handbook of Clinical Laboratory Automation, Robotics, and Optimization, Kost, G. J., Ed, John Wiley & Sons; 1996, Chapter 26, pp. 631-647.
2. Heise, H. M.; Marbach, R.; Koschinsky, Th.; Gries; F. A. Artif. Organs 1994, 18, 439-447.
3. Kohl, M.; Essenpreis, M.; Cope, M. Phys. Med. Biol. 1995, 40, 1267-1287.
4. Yodh, A.; Chance, B. Physics Today 1995, 48, 34-40.
5. Sevich-Muraca, E. M. J. Biomed. Opt. 1996, 3, 342-355
6. Burmeister, J. J.; Chung, H.; Arnold, M. A. Photochemistry and Photobiology. 1998, 67, 50-55.
7. Clarke, W. L.; Cox, D.; Gonder-Frederick, L. A.; Carter, W.; Pohl, S. L.
Diabetes Care 1987 10, 622-627.
8. Hall, J.; Pollard, A. Clin. Chem., 38 (1992) 1623.
9. Hazen, K. H.; Arnold, M. A.; Small, G. W.; (1998) submitted for publication.
10. Lu, G.; Zhou, X.; Arnold, M. A.; Small, G. W. Applied Spectroscopy 1997 51, 1330-1339.
11. Riley, M. R.; Rhiel, M.; Zhou, X.; Arnold, M. A.; Murhammer, D. W.
Biotechnology and Bioengineering 1997 55, 11-15.
12. Shaffer, R.E.; Small, G.W.; Arnold, M.A. Analytical Chemistry 1996 68, 2663-2675.
13. Chung, H.; Arnold, M.A.; Rhiel, M.; Murhammer, D. W. Applied Spectroscopy 1996 50, 270-276.
14. Pan, S.; Chung, H.; Arnold, M. A.; Small, G. W. Analytical Chemistry 1996 68, 1124-1135.
Back to In This Issue . . .