Tae-Woong Koo1, Andrew J. Berger1, Irving Itzkan1, Gary
Horowitz2, and Michael S. Feld1
1George R. Harrison Spectroscopy Laboratory, Massachusetts Institute of Technology (MIT), Cambridge, MA 02139
2Beth Israel Deaconess Medical Center (BIDMC), Boston, MA 02215
Raman spectroscopy, which probes the vibrational modes of molecules, offers a unique tool for noninvasive biological measurements, such as blood analysis and disease diagnosis. In particular, near-infrared Raman combines significant penetration depth (>1mm) with sharply defined lineshapes, making it possible to discriminate between spectral contributions from different analytes even when signals are extremely small.
Due to the low intensity of Raman signals the collection system must be highly efficient. In the last decade, improvements in holographic gratings and filters, and CCD detectors have made compact, high-throughput Raman systems widely available, facilitating many new applications. The schematic diagram of our system is shown in Figure 1.
As with absorption-based glucose detection, we utilize multivariate calibration to extract the Raman signature of glucose. However, we also take advantage of the known spectrum of glucose to make the calibration more robust against noise. A technique we recently developed for this purpose, called hybrid linear analysis (HLA) , has consistently demonstrated prediction ability superior to the well-known method of partial least squares (PLS), which relies solely upon reference concentrations to perform regressions.
We have demonstrated the ability to measure glucose concentrations in blood serum samples from a multipatient data set. Samples from 35 patients were gathered and analyzed with a Hitachi analyzer at the BIDMC. Samples were stored at 4° C between processing at the hospital and analysis on our system. Each spectrum was irradiated for 300 seconds at 300mW with about a 100mm diameter beam. The laser beam was focused to this high intensity to simplify the collection optics. For transcutaneous measurements the beam will be spread over a larger area to reduce the power density below any damage threshold. No data on long-term exposure effects is available. Concentrations were predicted using a leave-one-out HLA cross-validation.
Figure 2 is the resulting prediction plot of glucose concentration in multipatient serum. The x-axis represents the concentration measured by the analyzer; the y-axis represents the concentration predicted by the HLA algorithm. The root mean squared error of prediction (RMSEP) is 18 mg/dL, and the correlation coefficient is 0.92. Other blood analytes such as albumin, total protein, and cholesterol were also measured with clinically acceptable accuracy.
These results imply that Raman spectroscopy can monitor serum glucose even in realistic, highly multivariate data. We are now studying glucose and other analyte measurements in human whole blood samples. Coupling these efforts with studies of light-scattering in the blood-tissue matrix, we will design and test a transcutaneous Raman blood analyzer.
This research was conducted at the Laser Biomedical Research Center at MIT and was supported by Chiron Diagnostics and NIH grant P41-RR02594.
 A. J. Berger, T.-W. Koo, I. Itzkan, and M. S. Feld. An enhanced algorithm for linear multivariate calibration. Analytical Chemistry, 70(3):623-627, Feb. 1998.
Back to In This Issue . . .