James Lambert, Michael Storrie-Lombardi,
and Mark Borchert* Jet Propulsion Laboratory
California Institute of Technology
Pasadena, CA 91109
*Department of Ophthalmology
USC School of Medicine
Los Angeles, CA 90027
We have elicited a reliable glucose signature in mammalian physiological ranges using near infrared Raman laser excitation at 785 nm and multivariate analysis. In a recent series of experiments we measured glucose levels in an artificial aqueous humor in the range from 0.5 to 13X normal values. Data were obtained in 100 mL samples to mimic the volume constraints imposed by the human and rabbit anterior chamber of the eye. Feature extraction and data analysis were accomplished using linear multivariate analysis techniques (partial least squares fit). The spectra of the artificial aqueous humor closely approximate spectra acquired from rabbit aqueous humor.
Keywords: Raman, spectroscopy, glucose, multivariate analysis, diabetes mellitus, rabbit, aqueous humor, eye.
Non-invasive measurement of blood glucose by any method including optical spectroscopy techniques has remained an elusive target for at least two decades. Blood, tissue, and most excreted fluids contain numerous substances which confound glucose spectral signatures. On the other hand, Aqueous humor (AH) filling the anterior chamber of the eye (between the lens and cornea) contains relatively few molecules capable of interfering with the spectroscopic detection of glucose. These are primarily lactate, ascorbate, and urea [1]. This fact, and its optically accessible location behind the cornea make AH an obvious choice as a site on which to attempt non-invasive analysis of glucose.
A further advantage of aqueous humor is that its glucose concentration appears linearly related to plasma glucose concentration in animal studies. Estimates of the time constant for the equilibration of blood and AH glucose derived from these studies range from 20 to 60 minutes [2, 3]. Furthermore, these time constants are not affected by diabetes [3]. Lactate and urea levels in AH are also felt to vary with blood levels, while ascorbate is concentrated in the AH by active transport mechanisms.
The potential for non-invasive measurement of blood glucose using Raman spectroscopy on AH has been suggested before [4]. Previous work has demonstrated that the principle AH metabolites, including glucose, can be distinguished in water solutions containing mixed metabolites [5,6]. In addition, techniques have been described which could increase laser Raman sensitivity so that these metabolites could be measured at laser intensities which can be used safely in the eye [7,8]. Reliable measurement, however, of these metabolites at physiologic levels with Raman spectroscopy has yet to be described.
The purpose of this study is to demonstrate that the Raman spectra of a solution of mixed metabolites approximates that of aqueous humor, and that physiologic concentrations of glucose in such a solution of mixed metabolites can be measured with Raman spectroscopy.
Raman spectroscopy offers the possibility of remotely obtaining a measurement of glucose in vivo because, in contrast to infrared spectroscopy, its spectral signature is not obscured by water. In addition, Raman spectral bands are considerably narrower than those produced in classical infrared spectral experiments and Raman excitation in the near infrared region (700-1300 nm) encounters minimal fluorescence in aqueous media.
The majority of the photons incident on a target molecule scatter with unchanged frequency. A small proportion of light scatters with a shift in photon energy. This Raman shift occurs when photon energy transfers to (or from) the molecule during an inelastic collision. The vibrational spectra produced as a result of Raman scattering reveals the state of the atomic nuclei and chemical bonding within a molecule, as well as the interactions between the molecule and its local chemical environment.
Attempts to employ Raman techniques to directly measure glucose concentration in serum, plasma, and whole blood have met with encouraging success in vitro [9,10]. However, efforts to utilize these (and other) techniques in vivo for transcutaneous measurement of whole blood glucose levels have met with considerable difficulty. This is partly because whole blood and most tissue are highly absorptive, containing many fluorescent and Raman-active confounders.
Aqueous humor (AH), on the other hand, is relatively non-absorptive, and contains few Raman-active molecules. The four dominant, Raman-active molecules in AH are (concentrations are for rabbit AH) glucose (97 mg/dl), lactate (84 mg/dl), urea (36 mg/dl), and ascorbate (16 mg/dl) [1]. There is also a small amount of protein (26 mg/dl) that may produce fluorescence activity in sufficient strength to adversely affect Raman signal to noise ratios. Raman spectra from aqueous humor specimens of rabbits and humans, as well as spectra obtained through fresh excised rabbit corneas, have demonstrated detectable peaks of activity attributed to glucose, lactate, urea, amino acids, and proteins [10].
For our work, we have chosen a Raman excitation wavelength in the near infrared region to diminish extraneous biological fluorescence and minimize tissue damage. The price for these advantages is that Raman scattering efficiency decreases inversely with wavelength to the fourth power. We used an external cavity stabilized laser diode emitting at 785 nm and a Kaiser Optical Systems f/1.8 fluorescent imaging spectrograph with holographic filter and HoloPlex transmission grating. The holographic probe head was mounted on an Olympus BX60 microscope with 10X objective. Data were collected using a Princeton Instruments camera with an EEV back illuminated, NIR optimized 1024x256 CCD array operated at -80°C.
An artificial aqueous humor was designed to provide random fluctuations in concentration for the four major AH metabolites across a range of concentrations from 0.5X to 13X normal values for rabbit (Table 1). Metabolite levels in this range can be seen in hypoglycemia and diabetes (glucose), renal failure (urea), and myocardial infraction (lactate). The analytes were dissolved in pH buffered physiological (0.9%) saline. Variation in the other three analytes can dramatically alter glucose estimation. To develop a tool for estimating AH glucose levels we obtained spectra from the 20 randomly generated mixtures depicted in Table 1. Concentrations of the four principal constituents of the aqueous humor (AH) were randomly mixed in physiological buffered saline. Individual component levels range from 0.5X to 13X levels expected in rabbit AH. Mixtures are listed in order of increasing glucose concentration. Correlation coefficients across the 6 possible combinations ranged from -0.37 (ascorbate | urea) to 0.44 (lactate | glucose). Two aliquots were taken from each mixture and analyzed separately to produce a total of 40 samples and to permit estimation of test/re-test accuracy.
Table I: Metabolite Concentration in Artificial Aqueous Test Mixtures (mg/dl) |
|||||||||
Mixture |
Glucose |
Lactate |
Ascorbate |
Urea |
Mixture |
Glucose |
Lactate |
Ascorbate |
Urea |
1 |
50 |
80 |
1000 |
50 |
11 |
400 |
1100 |
1300 |
100 |
2 |
60 |
700 |
600 |
120 |
12 |
500 |
1300 |
500 |
170 |
3 |
80 |
60 |
1200 |
400 |
13 |
600 |
900 |
300 |
1200 |
4 |
100 |
100 |
400 |
250 |
14 |
700 |
50 |
140 |
1000 |
5 |
120 |
120 |
50 |
140 |
15 |
800 |
1200 |
100 |
800 |
6 |
140 |
250 |
800 |
500 |
16 |
900 |
300 |
60 |
200 |
7 |
170 |
140 |
80 |
60 |
17 |
1000 |
400 |
250 |
700 |
8 |
200 |
170 |
200 |
900 |
18 |
1100 |
1000 |
1100 |
80 |
9 |
250 |
800 |
170 |
600 |
19 |
1200 |
1000 |
900 |
80 |
10 |
300 |
500 |
120 |
1100 |
20 |
1300 |
600 |
700 |
1300 |
Samples were placed in quartz cuvettes designed to limit sample volume to 100 mL and to permit direct access to the test solution without traversing quartz walls or coverslips. Data acquisition and multivariate analysis were accomplished using Holograms and Grams, commercial software packages provided by Kaiser Optical Systems, Inc. and Galactic Industries Corporation, respectively. The integration time for each spectrum was 100 seconds with an average power delivered to sample of 50 mW. Since we are interested in minimal exposure times for future in vivo measurements, only a single spectrum was collected for each sample. However, each mixture of concentrations was duplicated and two independent measurements performed on each aliquot.
A cross-correlational analysis of the spectra identified 38 wavenumbers or spectral bins correlating significantly with glucose concentration. (Table 2) This made it possible to implement a partial least squares algorithm for data reduction and calibration using fewer test measurements than data samples. This is particularly important when attempting to build a robust spectral prediction algorithm capable of identifying the effect of multi-metabolite concentration variable on spectral signatures [11].
Table
II: Wavenumbers Selected for |
|||
213 |
846 |
1539 |
2718 |
426 |
906 |
1850 |
2751 |
442 |
1010 |
1900 |
2778 |
513 |
1053 |
1950 |
2827 |
535 |
1059 |
2000 |
2871 |
628 |
1119 |
2050 |
2904 |
644 |
1255 |
2100 |
2980 |
671 |
1332 |
2150 |
3362 |
693 |
1359 |
2200 |
|
742 |
1457 |
2250 |
|
For our multivariate analysis algorithm, we have chosen the partial least squares (PLS) technique, a multivariate analysis and spectral decomposition algorithm that, unlike principal component analysis, uses concentration information to calculate the eigenvectors. In our design, test and training samples were put into a ‘’round-robin" or autocorrelation training mode to iteratively employ all but one of the sample set in the minimization and eigenvector extraction process. Hence, the system trains on all but one of the samples, estimates the glucose level in that sample, then rotates the test sample back into the general pool and repeats the cycle until all samples have served as an unknown test subject. In this manner, the system is masked for the concentration of analytes in each unknown sample.
Rabbit AH was obtained from three animals within five minutes of sacrifice by other investigators. These animals had experienced experimental myocardial infarction 48 hours prior to euthanasia. They were sacrificed with a rapid exsanguination technique. Rabbit AH samples were kept on ice until glucose levels could be measured and Raman spectroscopy performed. Glucose concentration in rabbit AH samples was measured with a commercial glucometer (Glucometer Elite, Bayer) and confirmed against concentration standards.
The individual Raman signatures for the four principal metabolites found
in AH (glucose, lactate, ascorbate, and urea) are quite distinct (first four spectra of
Figure 1). When the four analytes are mixed at concentrations approximating the
normal levels, the composite signature shows marked similarity to that of the rabbit AH
(last two spectra of Figure 1). The rabbit AH spectra also contains 
evidence of broadband fluorescence and elevated lactate activity
(secondary to the myocardial infarction). Nevertheless, we can still detect common peaks
attributable to lactate, urea, and glucose. The similarity between spectra of artificial
AH and rabbit AH is quite striking.
In our experiments with artificial AH, the unknown glucose concentration was extracted with no preprocessing. The glucose concentration was estimated from the artificial AH mixtures using the raw spectra while the concentrations of the various metabolites was randomly varied over a broad range of clinical conditions.
Following our determination of the spectral
regions producing significant correlations to glucose concentration, we ran a series of
experiments employing both the full 3200 wavenumbers acquired and the restricted minimal
set. There was no diminution in algorithm accuracy using the 38 selected bins instead of
the entire spectrum. The partial least squares algorithm proved capable of predicting
glucose concentrations of unknown samples across a wide range of clinically significant
metabolic states. 
Figure 2 demonstrates the excellent predictive ability of the current system, with the PLS algorithm producing correlation coefficients of 0.99 for both predicted vs actual glucose concentrations and for test/re-test accuracy.
The rabbit AH signature (last spectrum of Figure 1) is complicated by both the expected elevation of lactate secondary to severe myocardial infarction, and the drugs introduced as part of the experimental procedure. Potential Raman-active molecules include aspirin, ketamine, xylazine, pentobarbital, and heparin. We are currently investigating the Raman response of these and several other potentially confounding substances. Nevertheless, expected Raman activity appears at a variety of sites including 1004, 1126, and 2896 cm-1.
The actual rabbit AH glucose levels were measured to be higher than normal (332 ± 19 mg/dl in the left eyes; 328 ± 37 mg/dl in the right eyes). The broadband activity apparent between 800 and 1,500 cm-1 and between 2,300 and 2800 cm-1 prohibits from artificial AH to estimate rabbit AH glucose. In vivo samples are now being acquired to generate in vivo calibration standards.
Significantly, the elevated glucose level in rabbit AH was expected in response to xylazine. Xylazine is commonly utilized in conjunction with Ketamine as an anesthetic in veterinary surgical procedures. It appears to interfere with the release of insulin by the pancreas. This results in elevated blood glucose levels for two to six hours [13].
We find these initial results most encouraging and propose that Raman spectroscopy of aqueous humor in the near infrared combined with multifactor analysis techniques constitutes a new technology capable of estimating levels of blood glucose and other metabolites non-invasively.
We are currently proceeding to investigate the next three important factors in the development of this technology:
1.) determining the minimum laser power and data acquisition time required for in vivo application;
2.) elucidating the temporal correlattion between glucose levels in AH and steady state blood concentrations;
3.) deconvolving and identifying the broadband fluorescence apparent in rabbit AH.
The authors would like to thank R.J. Bing, A. Kimura, and J. Roseto of the Huntington Memorial Research Institute for supplying the rabbit aqueous humor used in this research.
The research described in this publication was carried out by the Jet Propulsion Laboratory, California Institute of Technology and Childrens Hospital Los Angeles, under a contract with the National Aeronautics and Space Administration.
1. Sears ML. Dynamics of ocular fluids and control of intraocular pressure: formation of aqueous humor. in Principles and Practice of Ophthalmology. Albert DM and Jakobiec FA (eds.) WB Saunders, 1994.
2. Kinsey VE and Reddy DVN. Transport of glucose across blood-aqueous barriers as affected by insulin. J. Physiol. 156:8-16, 1961.
3. DiMattio J. Decreased ascorbic acid entry into cornea of streptozotocin- diabetic rats and guinea-pigs. Exp. Eye Res. 55:337-344, 1992.
4. Tarr RV, and Steffes PG. Non-invasive blood glucose measurement system and method using stimulated Raman spectroscopy. U.S. Patent #5243983, Sept. 14, 1993.
5. Wang SY, Hasty CE, Watson PA, et al. Analysis of metabolites in aqueous solutions by using laser Raman spectroscopy. Applied Optics 32:925-929, 1993.
6. Goetz MJ, Coté GL, Ercken R., et al. Application of a multivariate technique to Raman spectra for quantification of body chemicals. IEEE Trans. Biomed. Eng. 42(7):728-731, 1995.
7. Wicksted JP, Erckens RJ, Motamedi M, and March WF. Raman spectroscopy studies of metabolic concentrations in aqueous solutions and aqueous humor specimens. Applied Spectroscopy 49:987-993, 1995.
8. Ozaki Y, Iriyama K, and Hamaguchi H-O. Multichannel Raman spectroscopy of an intact lens: Raman measurement with laser irradiation below the threshold for retinal damage. Applied Spectroscopy. 41:1245-1247, 1987.
9. Dou, X.M., Yamaguchi, Y., Yamamoto, H., Uenoyama, H. and Ozaki, Y. Biological applications of anti-Stokes Raman spectroscopy-quantitative analysis of glucose in plasma and serum by a highly sensitive multichannel raman spectrometer App. Spect. 50(10):1301-6, 1996.
10. Berger, AJ, Itzkan, I, Feld, MS. Feasibility of measuring blood glucose concentration by near-infrared raman-spectroscopy, Spectrochimica Acta Part A Molecular and Biomolecular Spectroscopy 53(2):287-292, 1997.
11. Erckens, R.J., Motamedi, M., March, W.F., Wicksted, J. Raman spectroscopy for noninvasive characterization of ocular tissue - potential for detection of biological molecules J Raman Spect 28(5):293-9,1997.
12. Storrie-Lombardi, MC. Raman databases, Bayesian pattern matching, and nonlinear artificial neural networks, JPL/Caltech In Situ Workshop on Raman Spectroscopy, June 6, Pasadena, CA, 1997.
13. Arnbjerg, J and Eriksen, T. increased glucose content in the aqueous humour caused by the use of xylazine. Opthal Res 22:265-8,1990.