K. Danzer1, Ch. Fischbacher1, K.U. Jagemann1, K.J.
1Institut für Anorganische und Analytische Chemie, Friedrich-Schiller-Universität, D-07743 Jena, Germany
2Institute for Physical High-Technology Jena (IPHT e.V.), D-07743 Jena, Germany
Diabetic persons have to keep the level of glucose in the blood within narrow target parameters (4 ...7 mmol/L) in order to reduce short-term and long-term complications. Decisions about insulin injections or carbohydrate intake depend on the current blood glucose (BG) concentration.
Thus self-monitoring of BG is an essential component of diabetes treatment programs. All currently available methods are based on the enzymatic oxidation of glucose in blood samples which are placed on test strips or sensors. The strips or sensors can then be evaluated visually, photometrically or amperometrically.
In recent years several techniques have been proposed for minimal-invasive, e.g. subcutaneous implantable bio-sensors  and reverse iontophoresis [2 ], and non-invasive in vivo monitoring of blood and tissue glucose, e.g. near infrared (NIR) and photo-acoustic spectroscopy . This paper describes our work on the NIR diffuse reflection method. The possibility of using diffuse reflectance near-infrared (NIR) spectroscopy for determining BG concentration non-invasively in diabetic patients has been demonstrated by Haaland et al.  and Marbach and Heise [5, 6] as well as by our group [7, 8] using different approaches. A review on the NIR method has been published  and several companies have announced commercial instruments but problems remain to be solved in order to achieve reliable and precise results.
This method has the advantage that no reagents are required for the measurements and that fiber optical components can be used. As a consequence only insulators are in direct contact with the skin. Furthermore the spectrometer can be constructed without moving parts, leading to a rugged design.
The short wavelength region of the near infrared (l » 800 1300 nm) has some properties that are ideal for non-invasive in vivo diagnostics. In this range there is an optical window between the region in which visible light is absorbed by blood and skin pigments and the longer wavelengths in which absorption of water predominates. The NIR spectrum consists of overtones and combinations of the fundamental vibrations mainly of the bonds of carbon, nitrogen and oxygen with hydrogen. NIR spectra of aqueous systems show weak, broad and overlapping bands with random baselines. The position and intensity of the signals vary according to the chemical vicinity (hydrogen bonding effects). The influence of dissolved salts and temperature on the NIR spectra of aqueous systems is well known. Since the normal proportion of glucose in blood and tissue is only about 0.1 % of the water content the spectral variations due to glucose concentration are extremely small. The evaluation of the recorded spectra is further complicated by several influences: water is a strong absorber and also the main component of living tissue. In addition time dependent biological processes take place, e.g. pulse and respiration. Other sources of variation are erroneous spectral recordings caused for example by irregular pressure of the measuring head on the finger. As a consequence multivariate calibration methods  have to be used for evaluating the spectra.
The equipment for recording the spectra is shown in Fig. 1 and consists of a light source, a fiber optical measuring head and a NIR spectrometer. We use a miniaturized spectrometer module developed at the Institute for Physical High-Technology Jena. This spectrometer uses a polychromator with a holographic imaging diffraction grating and an InGaAs photo diode array detector with 128 pixels mounted on a glass block. A very compact design (70 mm x 50 mm x 40 mm) with excellent optical performance was achieved by using a high-dispersion grating with extremely short focal length. The optical resolution (Rayleigh criterion) is about 12 nm in the wavelength region from 820...1320 nm. Spectra are recorded every 50 ms and are accumulated. The reference spectrum is recorded using a ceramic material. The signal is converted with a 16 bit-ADC. The matrices with the coefficients necessary for calibration can be updated via a serial interface. All computations necessary for predicting the glucose concentration are performed on board.
Fiber optical bundles connect light source (tungsten halogen lamp) and spectrometer with the fiber optical measuring head. The fibers illuminating the skin (finger tip) are concentrically arranged around the central part of the bundle which connects to the spectrometer.
In order to compute the calibration coefficients the spectra from several diabetes patients with varying BG concentration levels are recorded. Fig. 2 shows spectra from three patients. By visual inspection only baseline variations can be seen. The BG level is determined simultaneously by a conventional glucose analyzer using blood plasma. A typical blood glucose profile is shown in Fig. 3. The matrices of the calibration coefficients are computed by partial least-squares regression (PLS)  and radial-basis neural networks (RBF)  on personal computers using MATLAB  routines.
Since outliers degrade the quality of the calibration set and result in erroneous predictions, outlier detection based on multivariate distances and leverage values is implemented. About 1 % of the spectra of the calibration set are diagnosed as not being representative.
The quality of the predictions was evaluated by computing the leave-one-out cross-validated root mean squared error between the predicted values and the clinical reference method (RMSP). The mean of the RMSP values is 2.2 mmol/L (SD = 0.57 mmol/L) for PLS regression and 2.0 mmol/L (SD = 0.53 mmol/L) for RBF networks (31 glucose profiles). Test data sets lead to satisfactory results for short-time predictions (1 day) but long-term reliability needs improvement. The same effect is observed for investigations of glucose containing Intralipid emulsions which are used as tissue simulating phantom medium. The goal is a non-invasive BG meter which shows reliable long-term results using calibration models for patient groups.
1. G. S. Wilson , Y. Zhang , G. Reach, D. Moatti-Sirat, V. Poitout , D.R. Thevenot, F. Lemonnier, J.C. Klein, Progress toward the development of an implantable sensor for glucose, Clin. Chem. 38, 1613 (1992)
2. G. Rao, P. Glikfeld, R.H. Guy, Reverse Iontophoresis: Development of a Noninvasive Approach for Glucose Monitoring, Pharm. Res. 10, 1751 (1993)
3. G. Spanner , R. Nießner, New concepts for the non-invasive determination of physiological glucose concentrations using modulated laser diodes, Fresenius J Anal Chem. 354, 306 (1996).
4. D.M. Haaland, M.R. Robinson, G.W. Koepp, E.V. Thomas, R.P. Eaton, Reagentless near-infrared determination of glucose in whole blood using multivariate calibration, Appl. Spectroscopy 46, 1575 (1992)
5. R. Marbach, Meßverfahren zur IR-spektroskopischen Blutglucosebestimmung, Fortschr.-Ber. VDI, Vol. 346, Düsseldorf 1993
6. R. Marbach, T. Koschinsky , F.A. Gries, H.M. Heise, Noninvasive blood glucose assay by near-infrared diffuse reflectance spectroscopy of the human inner lip, Appl. Spectroscopy 47, 875 (1993)
7. C. Fischbacher, K.U. Jagemann, K. Danzer, U.A. Müller, L. Papenkordt, J. Schüler, Enhancing calibration models for non-invasive near-infrared spectroscopical blood glucose determination, Fresenius J. Anal. Chem. 359, 78 (1997)
8. U.A. Müller, B. Mertes, K. Danzer, C. Fischbacher, K.U. Jagemann, L. Papenkordt, J. Schüler, Non-invasive blood glucose measurement by means of infrared spectroscopy: calibration models and results, J. Artif. Org. 20, 285 (1997).
9. H.M. Heise, Non-invasive monitoring of metabolites using near infrared spectroscopy: state of the art, Horm. Metab. Res. 28, 527 (1996)
10. H. Martens, T. Naes, Multivariate Calibration, Wiley, Chichester 1989
11. D.M. Haaland, E.V. Thomas, Partial least-squares methods for spectral analyses, 1. Relation to other quantitative calibration methods and the extraction of qualitative information , Anal. Chem. 60, 1193 (1988)
12. S. Chen, C.F.N. Cowan, P.M. Grant, IEEE Transaction Neural Networks , Orthogonal least squares learning for radial basis function networks 2, 302 (1991)
13. MATLAB 4.2b for Windows, The MathWorks, South Natick, MA
Back to In This Issue . . .