A developmental noninvasive method of determining the concentration of glucose in blood is based on (1) the acquisition of a near-infrared (NIR) Raman spectrum from the aqueous humor of an eye, (2) analyzing the spectrum by a combination of techniques described below, and (3) recognition that the glucose level in the aqueous humor of the eye is about 80 percent of that in the blood 30 minutes before the spectrum was acquired. More specifically, what the analysis yields is a probabilistic indication that the glucose concentration represented by the Raman spectrum lies in one of three ranges of physiological interest; hypoglycemic (5.8 mM). The method involves less NIR laser power and shorter data-collection times than have been used in previous efforts to use Raman scattering to measure glucose concentrations in blood.

alt
Concentrations of 24 Glucose Solutions as estimated from Raman spectra by the method described in the text agreed fairly well with the actual concentrations. In this case, the concentrations estimated by a neural network correspond to a nonlinear least-squares fit to the actual concentrations, with a regression coefficient of 0.955.

One reason for choosing the aqueous humor of the eye as the target for Raman spectroscopy is, of course, that the interior of the eye is optically accessible. Another reason is that whole blood contains numerous optically active materials, the spectra of which obscure the spectral signature of glucose. In comparison with whole blood, the aqueous humor of the eye contains many fewer Raman-active substances to complicate the overall Raman spectrum and the interpretation thereof.

The Raman-active substances in the aqueous humor of the eye are glucose, ascorbate, lactate, urea, and small amounts of protein. The Raman spectrum (absorbance vs. wavelength) of each of these is affected by interactions with all of the other substances present in the aqueous medium, resulting in both linear and nonlinear variations in Raman spectra with concentrations. Thus, even though only four or five Raman-active constituents are present in significant concentrations, it is necessary to use both linear and nonlinear multifactor analytical techniques to obtain accurate estimates of glucose concentrations from the total Raman spectrum.

In the present method, sets of measured total Raman spectra are first subjected to the classical multivariate-analysis technique known as principal-component analysis. This technique yields a reconstruction of input spectra as linear combinations of feature vectors (eigenvectors of a covariance matrix) that account for the maximum variance in the input spectra. This reconstruction is optimum in the least-squares sense and involves the fewest parameters.

Each feature vector is fed as input to an artificial neural network that is configured to generate either one output or three outputs indicative of the corresponding glucose concentration. Prior to use in this way, the neural network must be trained by use of feature vectors representing known concentrations. Thereafter, if the neural network is configured for one output, then when it is presented with a feature vector from an unknown concentration, the output signal level should represent an estimate of the concentration; alternatively, if the neural network is configured for three outputs, then the signal level at each output terminal should represent a Bayesian a posteriori probability that the glucose concentration lies in one of the three ranges of physiological interest.

In vitro experiments on aqueous solutions of glucose at concentrations from 1.01 to 10.1 mM have demonstrated the feasibility of the method. In these experiments, specimens were illuminated with 100 mW of power at a wavelength of 785 nm from a cavity-stabilized laser diode. The Raman spectrum excited by this illumination was measured by use of a holographic imaging spectrographic probe head containing a liquid-nitrogen-cooled charge-coupled device (CCD). The output of the CCD was digitized and processed as described above, using commercial data-acquisition, multifactor-analysis, and neural-network software. The figure depicts some results obtained with a single-output neural network.

This work was done by Michael Storrie-Lombardi, James Lambert, and Mark Borchert of Caltech for NASA's Jet Propulsion Laboratory. For further information, access the Technical Support Package (TSP) free on-line at www.nasatech.com/tsp under the Bio-Medical category. NPO-20414