Targeted interference subtraction applied to near-infrared measurement of analytes

• US 20030023148A1
• Filed: 06/25/2002
• Published: 01/30/2003
• Est. Priority Date: 07/22/1999
• Status: Active Grant

• Current Assignee: GLT Acquisition Corporation
• Original Assignee: Sensys Medical, Inc. (Sensys Medical Limited)
• Inventors: Ruchti, Timothy L., Blank, Thomas B., Lorenz, Alexander D.
• Granted Patent: US 6,697,654 B2
• Time in Patent Office: Days
• US Class Current: 600/300
• CPC Class Codes: A61B 5/1075 for measuring dimensions by non-invasive methods, e.g. for determining thickness of tissue layer A61B8/0858 takes precedence
• CPC Class Codes: A61B 5/14532 for measuring glucose, e.g. by tissue impedance measurement
• CPC Class Codes: A61B 5/1455 using optical sensors, e.g. spectral photometrical oximeters
• CPC Class Codes: A61B 5/1495 Calibrating or testing of in-vivo probes
• CPC Class Codes: A61B 5/7264 Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems neural networks per se G06N3/00; expert systems per se G06N5/00
• CPC Class Codes: G01N 21/4785 Standardising light scatter apparatus; Standards therefor

Abstract

Methods and apparatus for estimating and removing spectral interference improve precision and robustness of non-invasive analyte measurement using Near-Infrared (NIR) spectroscopy. The estimation of spectral interference is accomplished, either, through multivariate modeling or discrete factor analysis, using a calibration set of samples in which the interference is orthogonal to the analyte signal of interest, or where the shape of the interference is known. Each of the methods results in a multivariate model in which the spectral interference is estimated for a new sample and removed by vector subtraction. Independent models based on classes of sample variability are used to collapse spectral interference and determine more accurately which model is best equipped to estimate the signal of interference in the new sample. Principal components analysis and other commonly known analytical techniques can be used to determine class membership.

Claims

1. An apparatus for modeling and removing targeted interfering signals from spectral measurements comprising:

means for measuring a spectrum of a sample; and

at least one interference model, wherein said interference is approximately orthogonal to an analyte signal of interest, said interference model adapted to remove spectral interference from said measured spectrum.

2. The apparatus of claim 1, wherein said spectrum is a noninvasive measurement and wherein said sample comprises a tissue sample.

3. The apparatus of claim 1 wherein said means for measuring a spectrum comprises:

a radiation source;

means for coupling radiation to the sample;

means for collecting radiation that is any of diffusely scattered and transmitted from the sample;

means for dispersing said collected radiation;

means for detecting said dispersed radiation at predetermined bands of wavelengths and converting said detected radiation to a proportional voltage; and

means for digitizing said voltage.

4. The apparatus of claim 3, wherein said light source comprises a broadband light source.

5. The apparatus of claim 3, further comprising a band pass filter, said band pass filter adapted to limit radiation reaching said sample to a targeted wavelength range.

6. The apparatus of claim 3, said apparatus further comprising a blocking element, said blocking element adapted to intermittently block transmission of radiation toward said sample so that any of a baseline and a detector dark current can be detected.

7. The apparatus of claim 3, wherein said means for coupling radiation to said sample comprises at least one fiber optic and a focusing element.

8. The apparatus of claim 3, wherein said means for collecting light comprises at least one fiber optic.

9. The apparatus of claim 3, further comprising a probe, said probe adapted to establish contact of said means for coupling light and said means for collecting light with surface of said tissue sample.

10. The apparatus of claim 3, wherein said means for dispersing collected radiation comprises a grating.

11. The apparatus of claim 3, further comprising a coupling element adapted to couple said means for collecting radiation to an aperture, said aperture adapted to directed radiation toward said means for dispersing radiation.

12. The apparatus of claim 11, wherein said aperture comprises a slit.

13. The apparatus of claim 3, wherein said means for detecting radiation comprises a detector array with associated amplifiers, said array and said amplifiers adapted to convert said detected radiation to a voltage proportionate to a detected signal.

14. The apparatus of claim 3, wherein said means for digitizing said voltage comprises an analog-to-digital converter.

15. The apparatus of claim 1, wherein said interference model is calculated from one or more calibration sets.

16. The apparatus of claim 1, wherein said interference model estimates interference based on one of:

multivariate modeling, wherein The first method models known interference by regressing spectral data against measured property values of the interference; and

discrete factor analysis, wherein signal of interference is estimated by performing an analysis of spectral scores and loadings and determining which factor best represents the interference; and

wherein signal is removed by reconstructing the spectrum using the spectral scores and loadings less the factor representing the interference.

17. The apparatus of claim 1, further comprising a processing element, wherein said processing element is programmed to execute a sample variability classification procedure, wherein most appropriate interference model is selected for estimating interference to be removed from said measured spectrum.

18. A method for estimating and removing spectral interference from a measured spectrum comprising the steps of:

providing one or more calibration sets of exemplary spectral measurements and corresponding measured property values;

providing one or more spectral measurements, each of said measurements comprising a sample;

calculating one or more interference models from said one or more calibration sets, wherein said interference is approximately orthogonal to an analyte signal of interest;

estimating said interference; and

removing said interference from said sample.

19. The method of claim 18, wherein said spectrum is denoted by a vector m∈

^N of absorbance values pertaining to a set of N wavelengths λ

∈

^N.

20. The method of claim 18, wherein said measurements are in the wavelength region 700-2500 nm, said wavelength region corresponding to the near infrared (NIR).

21. The method of claim 18, further comprising the step of:

detecting outliers, wherein said outliers are invalid measurements caused by spectral variations due to any of instrument malfunction, poor sampling of a subject and subjects outside of the calibration set.

22. The method of claim 21, wherein said outlier detection step employs principal components analysis and analysis of resulting residuals to detect spectral outliers.

23. The method of claim 22, wherein said outlier detection step comprises the steps of:

projecting a spectrum m onto five eigenvectors contained in a matrix o, said matrix o being previously developed through a principal components analysis of said calibration set, where

$\mathrm{xp}<br><br>\text{\hspace{1em}<br><br>}<br><br>c_o=\underset{k=1}{\overset{5}{\sum <br><br>}}<br><br>m<br><br>\text{\hspace{1em}<br><br>}<br><br>o_k,$

xpc_o being a 1 by 5 vector of scores and where o_k is the k^th column of the matrix o;

determining the residual q, according to

q=m−

xpc_oo¹;

comparing said residual q to three times the standard deviation of the residual of said calibration set; and

reporting said sample as an outlier if q is greater.

24. The method of claim 18, further comprising the step of:

optionally, preprocessing, wherein said preprocessing step includes one or more transformations that attenuate noise and instrumental variation without affecting a signal of interest;

25. The method of claim 24, wherein a preprocessed measurement, X∈

^N, is determined according to;

x=h(ë

,m)

where h;

<sup>N×

2</sup>→

^N is a preprocessing function.

26. The method of claim 18, wherein parameters of said interference models are calculated using multivariate regression and wherein a resulting model comprises a multivariate regression vector.

27. The method of claim 26, wherein said multivariate regression vector is computed using one of:

Partial Least Squares (PLSR);

Principal Component Regression (PCR);

Locally Weighted Regression (LWR);

Multiple Linear Regression (MLR); and

Classical Least Squared (CLS).

28. The method of claim 27, wherein said regression vector is computed using Principal Component Regression.

29. The method of claim 28, wherein a Principal Components analysis is performed on one or more spectral measurements, wherein the collected spectra, x contain n number of samples by p number of wavelengths.

30. The method of claim 29, wherein a n by 1 vector of eigenvalues and a p by n eigenvector matrix is computed using Singular Value Decomposition.

31. The method of claim 30, wherein said regression vector, W, is related to said spectral measurements, x, and interference property values, y, by:

U=V^Tx and W=yU^Tinv(UU^T)V^T,

where V is a matrix containing a selected number of eigenvectors.

32. The method of claim 31, wherein predicted property values of the interference are given by:

33. The method of claim 32, wherein the number of factors, said factors comprising eigenvectors, representing said interference is determined through cross validation on one of the calibration set and an independent set of test samples by iteratively increasing the number of factors used to develop the model and minimizing the standard error of prediction (SEP).

34. The method of claim 33, wherein the SEP is given by:

35. The method of claim 34, wherein acceptability of the model is based on an F-test, wherein said F-test is the ratio of the squared SEP to the variance in the measured property values.

36. The method of claim 30, wherein said eigenvector matrix V and said regression vector W are saved and used to estimate interference of new sample spectra.

37. The method of claim 26, wherein said regression vector, W, is represented by:

W=g(x,y)

where g;

<sup>N×

2</sup>→

^N is the multivariate regression function, x is the preprocessed spectra of the calibration set and y is one of the corresponding measured property values of the interference and a target score vector determined through principal components analysis.

38. The method of claim 37, wherein for a new sample x′

, said new sample containing interference, a new spectrum {overscore (x)}, with interference removed is given by;

39. The method of claim 18, wherein parameters of said interference models are calculated using discrete factor analysis.

40. The method of claim 39, wherein a signal of interference is estimated by performing an analysis of spectral scores and loadings and determining which factor best represents the interference.

41. The method of claim 40, wherein said signal of interference is removed from said measured spectrum by reconstructing said spectrum using said spectral loadings and scores minus the factor representing the interference.

42. The method of claim 41, wherein said interference modeling step employs one of principal components analysis and another factor-based analytical method.

43. The method of claim 42, wherein a set of spectra x, having m number of samples and n wavelengths is given by:

x=xvv^T,

where v is a n by m matrix of eigenvectors developed from a principal components analysis on the spectra x.

44. The method of claim 42, wherein spectra {overscore (x)}:

, having said interference removed, are given by;

{overscore (x)}=x(I−

V_kv_k^T),

where the k^th column of the eigenvector matrix v represents the interference to be removed.

45. The method of claim 42, wherein a set of spectra, x, having n number of samples by p number of wavelengths are used in a principal components analysis to generate a n by 1 vector of eigenvalues, and a p by n eigenvector matrix.

46. The method of claim 45, wherein said eigenvectors and eigenvalues are computed using Singular Value Decomposition.

47. The method of claim 45, wherein spectral scores, T, are given by:

48. The method of claim 47, wherein said eigenvectors are analyzed for shape and variation similarities with shape and variation manifested at each wavelength where the interference is known.

49. The method of claim 47, wherein spectral scores are analyzed for similar changes in magnitude where there are corresponding property values for said interference or if the magnitude of the interference is known.

50. The method of claim 47, wherein a corresponding eigenvector is stored wherein a factor corresponding to said interference is known, and wherein said stored eigenvector is used to correct future samples.

51. The method of claim 18, further comprising the step of:

classifying samples, wherein a sample is compared to other samples for which interference has already been modeled so that the most appropriate interference model is selected for estimating interference to be removed.

52. The method of claim 51, wherein said sample classification step comprises a principal components analysis of said sample and an analysis of spectral scores through t-squared and t-square limit computation.

53. The method of claim 52, wherein a spectrum m is projected onto p number of eigenvectors contained in a matrix o previously determined through a principal components analysis on said calibration set, where the calculation is given by:

54. The method of claim 53, wherein eigenvalues corresponding to said eigenvector matrix o are used to normalize new scores to unit variance.

55. The method of claim 54, wherein a t-squared value, t, is determined according to:

t=Σ

{overscore (xp)}{overscore (c)}_o*(d−

1),

where ^xpco represents a normalized vector of scores and d is the number of samples in said calibration set, and wherein said t-squared value is a measure of how consistent variability of said spectrum m is with said calibration set of exemplary spectra.

56. The method of claim 55, wherein a t-squared limit is computed by performing an f-test with a confidence limit on the scores matrix of said calibration set so that a sample having a t-squared value exceeding said t-squared limit is considered to have spectral characteristics dissimilar to those in the calibration set, and wherein said confidence limit is optimized according to the application.

57. The method of claim 56, wherein said confidence limit is ninety-five percent.

58. The method of claim 57, wherein said t-squared value is compared to t-square limits for multiple interference models until at least one passes the constraints.

59. The method of claim 58, wherein the model having the lowest t-squared limit is applied to a sample that fits several models.

60. The method of claim 59, wherein a sample failing to meet the constraints of any stored models is classified as an outlier.

61. The method of claim 18, wherein an appropriate model or regression vector is selected and applied to said sample to remove targeted interference.

62. The method of claim 61, wherein removing said interference comprises subtracting it from said sample.

63. The method of claim 18, wherein said step of estimating said interference comprises:

collecting a plurality of baseline samples

removing interference from said baseline samples;

64. The method of claim 63, further comprising the step of;

subtracting interference calculated from said baseline samples from a sample measurement.

65. The method of claim 64, further comprising the step of measuring an analyte value from said sample measurement.

66. The method of claim 65, wherein measurement of said analyte value is biased from a true blood reference value due to non-linearities and un-modeled synergistic effects in data;

and

wherein said bias is constant for all sample measurements that are spectrally similar to said baseline samples;

further comprising the step of adjusting an interference model by adding mean difference between N number of blood reference values and associated model measurements to intercept of said model.

67. The method of claim 18, wherein said calculating and classifying steps are based on any of:

linear discriminant analysis;

SIMCA;

k nearest neighbor;

fuzzy classification; and

artificial neural networks.