Automatic process for sample selection during multivariate calibration

• US 6,876,931 B2
• Filed: 08/02/2002
• Issued: 04/05/2005
• Est. Priority Date: 08/03/2001
• Status: Expired due to Fees

• Current Assignee: GLT Acquisition Corporation
• Original Assignee: Sensys Medical, Inc. (Sensys Medical Limited)
• Inventors: Ruchti, Timothy L., Blank, Thomas B., Lorenz, Alexander D.
• Primary Examiner(s): Raymond, Edward
• Application Number:
  US10/211,142
• Publication Number:
  US 20030109998A1
• Time in Patent Office: 977 Days
• Field of Search: 702/22, 702/23, 702/75, 702/76, 702/106, 600/316, 600/310, 600/322, 128/633
• US Class Current: 702/22
• 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: G01N 21/274 Calibration, base line adjustment, drift correction
• CPC Class Codes: G01N 21/359 using near infrared light

Abstract

A process for enhancing a multivariate calibration through optimization of a calibration data set operates on a large calibration set of samples that includes measurements and associated reference values to automatically select an optimal sub-set of samples that enables calculation of an optimized calibration model. The process is automatic and bases sample selection on two basic criteria: enhancement of correlation between a partner variable extracted from the independent variable and the dependent variable and reduction of correlation between the dependent variable and interference. The method includes two fundamental steps: evaluation, assigning a measurement of calibration suitability to a subset of data; and optimization, selecting an optimal subset of data as directed by the measurement of suitability. The process is particularly applied in enhancing and automating the calibration process for non-invasive measurement glucose measurement but can be applied in any system involving the calculation of multivariate models from empirical data sets.

Claims

1. A method for optimizing a calibration set comprising the steps of:

providing a calibration set of spectral samples and associated reference values;

assigning a measure of fitness for calibration to individual samples or subsets of samples from the calibration set according to a cost function; and

determining a subset of samples having an optimal measure of fitness by creating new groupings of samples until a subset results that provides a target performance, wherein spurious variations are not correlated to a net analyte signal.

2. A method as in claim 1, wherein said cost function comprises a measure of the quality or magnitude of accessible net analyte signal.

3. A method as in claim 1, wherein said cost function comprises a measure of a lack of correlation between a dependent variable and ancillary or interfeting variables manifested in x, x representing at least one spectral sample, that correlates to y, y representing a reference value.

4. A method as in claim 1 wherein said cost function includes any of:

a standard error of calibration;

a standard error of a monitoring or test set;

range of a dependent variable, wherein the dependent variable comprises one or more of said reference values;

the number of selected samples;

the total variation of the x-block, wherein the x-block includes said spectral samples; and

additional parameters that are indicative of an optimal calibration set.

5. A method as in claim 1, wherein the cost function involves the assignment of a cost or performance measurement for an overall set according to:

$J=\underset{k=1}{\overset{m}{\text{\hspace{1em}<br><br>}}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\mathrm{\alpha <br><br>}}_k<br><br>{\stackrel{.}{J}}_k<br><br>\left(x_s,y_s\right)$

where J Is the cost associated with a particular subset of samples, {dot over (J)}_k(x,y) represents the k^th cost associated with the subset of samples represented by x_s and y_s, and x_k represents the weight or emphasis associated with the k^th figure of merit.

6. A method as in claim 1, wherein an individual cost is computed for each sample that is present within a particular subset according to:

$J_p=\underset{k=1}{\overset{m}{\text{\hspace{1em}<br><br>}}}<br><br>\text{\hspace{1em}<br><br>}<br><br>{\mathrm{\alpha <br><br>}}_k<br><br>{\stackrel{.}{J}}_{k,p}<br><br>\left(x_p,y_p\right)$

where J_p is the cost associated with the p^th sample, α

_k represents the weight associated with the k^th cost function, {dot over (J)}_k,p(x_p, y_p) represents the k^th sub-cost function associated with the sample.

7. A method as in claim 6, wherein m represents the total number of sub-cost functions.

8. A method as in claim 6, wherein the overall cost is the summation of the individual figures of merit associated with the samples of a particular subset.

9. A method as in claim 1, wherein determining a subset of samples having an optimal measure of fitness comprises determining a subset that either minimizes cost or maximized a figure of merit.

10. A method as in claim 1, wherein determining a subset of samples having an optimal measure of fitness is performed in an iterative manner.

11. A method as in claim 10, wherein determining a subset of samples having an optimal measure of fitness comprises using cost of each sample, or the cost of all samples is used at each iteration to remove at least one sample until said target level of performance is achieved.

12. A method as in claim 11, wherein hard inclusion and exclusion limits are set to the final cost of each sample so that they are either included in or excluded from the calibration set.

13. A method as in claim 11, wherein determining a subset of samples having an optimal measure of fitness comprises establishing a set of decision rules in a hierarchical framework in which each individual cost or figure of merit is applied separately and in a particular order.

14. A method as in claim 11, wherein determining a subset of samples having an optimal measure of fitness comprises:

beginning with a subset that includes at least one sample; and

iteratively adding one or more samples into the calibration set until a desired performance level is achieved.

15. A method as in claim 11, wherein determining a subset of samples having an optimal measure of fitness comprises selecting a subset of optimal or near optimal samples according to any of the techniques of:

dynamic programming;

gradient searching;

random searching;

genetic algorithms; and

evolutionary programming.

16. A method as in claim 15, wherein determining a subset of samples having an optimal measure of fitness according to a genetic algorithm comprises performing the following steps iteratively until a target level of performance is achieved or until at least one population of solutions converges:

creating one or more populations of possible solutions, wherein each solution is a “

chromosome”

or a vector in which each element represents a sample, so that each solution defines a set of calibration samples;

evaluating each solution and assigning a measure of fitness according to the cost function;

selecting solutions in a random manner according to a probability related to the level of fitness;

randomly combining various solutions to form a new population of possible solutions;

randomly removing samples from, or introducing new samples into various solutions; and

optionally, when multiple populations are employed, exchanging various solutions from each population to enhance the diversity of each population.

17. A method as in claim 1, wherein providing a calibration set of spectral samples and associated reference values comprises:

collecting data.

18. A method as in claim 17, wherein determining a subset of samples having an optimal measure of fitness comprises:

A. calculating at least one partner variable;

B. determining interference;

C. determining an inaccessible portion, c_noise, of an analyte signal;

D. determining an accessible portion, c_signal, of the analyte signal;

E. estimating a signal-to-noise-ratio (SNR) according to;

$\mathrm{SNR}=\frac{{\mathrm{nas}}_{\mathrm{signal}}}{{\mathrm{nas}}_{\mathrm{noise}}};<br><br>$

F. estimating a correlation coefficient between said partner variable and the accessible portion of the analyte signal;

G. iteratively performing steps A-F, leaving at least one sample out at each iteration;

H. selecting a sample to remove; and

I. continuing steps A-H until said target performance is achieved.

19. A method as in claim 18, wherein collecting data comprises collecting paired data points each consisting of a measurement of an independent variable(s), x, and a target or dependent variable, y, from a reference device.

20. A method as in claim 19, wherein x comprises at least one tissue absorbance spectrum collected on a near-infrared spectrometer and y comprises at least one corresponding glucose measurement taken on a glucose analyzer.

21. A method as in claim 20, wherein said calibration set comprises a N-by-M matrix, {overscore (x)} and an N-by-1 matrix, {overscore (y)};

wherein said calibration set is used in multivariate calibration to determine a model for calculating future analyte values, {overscore (y)}, given future analyzer measurements.

22. A method as in claim 21, wherein step A comprises calculating an N-by-P partner variable, {overscore (z)}, from the independent variable, {overscore (x)}.

23. A method as in claim 22, wherein said partner variable is either the net analyte signal (NAS) or a feature related to the NAS.

24. A method as in claim 22, wherein the NAS is the part of {overscore (x)} that relates uniquely to {overscore (y)} and is determined through a previously calculated calibration model.

25. A method as in claim 22, wherein step A comprises defining one or more partner variables when a multiplicity of prior calibrations exist or when more than one feature is required to reflect the NAS.

26. A method as in claim 22, wherein {overscore (z)}={overscore (xw)} when a prior calibration model exists in the form of a regression vector, {overscore (w)}.

27. A method as in claim 26, wherein a regression vector from another subject, a variety of subjects or a group of subjects is used to determine P estimates of the net analyte signal present in each row of {overscore (x)}.

28. A method as in claim 22, wherein, if a feature is known that reflects the net analyte signal present in {overscore (x)}, said feature is used by itself or in combination with other features as {overscore (z)} if a calibration does not exist.

29. A method as in claim 23, wherein critical points of first and second derivatives from said near-infrared spectrum are valuable features related to the NAS of glucose.

30. A method as in claim 29, wherein absorbance spectra of any of water, fat and protein are used as features related to the NAS.

31. A method as in claim 30, wherein said critical points are resolved by calculating the first or second derivative of each spectrum and determining the magnitude at the appropriate wavelengths.

32. A method as in claim 31, wherein the appropriate wavelengths include any of 1095±

1-5 nm;

1125±

1-5 nm;

1155±

1-5 nm;

1180±

1-5 nm;

1200±

1-5 nm;

1308±

1-5 nm;

1338±

1-5 nm;

1380±

1-5 nm;

1414±

1-5 nm;

1443±

1-5 nm;

1458±

1-5 nm;

1550±

1-5 nm;

1593±

1-5 nm;

1607±

1-5 nm;

1625±

1-5 nm;

166±

1-5 nm 1687±

1-5 nm;

1706±

1-5 nm;

1727±

1-5 nm;

1747±

1-5 nm;

1783±

1-5 nm;

1874±

1-5 nm;

1784±

1-5 nm; and

1910±

1-5 nm.

33. A method as in claim 22, wherein the interference is any variable other than the dependent variable that is manifested in the independent variable, after processing and feature extraction, that correlates with the target variable over time.

34. A method as in claim 33, wherein interference includes any of:

environmental temperature;

instrument temperature;

sample temperature;

absolute humidity;

instrument drift;

time or sample order; and

sample parameters, wherein sample parameters include any of;

hydration, surface water, skin thickness, blood content in tissues, hematocrit concentration, trigylceride concentration, sodium levels, cholesterol levels, surface hydration, blood pressure, and tissue pH.

35. A method as in claim 33, wherein said variables other than the dependent variable are contained in a vector {overscore (b)} and are identified as the interference.

36. A method as in claim 22, wherein nas_noise is an estimate of the variance {overscore (y)} that is not represented in {overscore (x)}.

37. A method as in claim 36, wherein nas_signal is an estimate of the independence of the analyte values from the interference or the information that can be accessed in the presence of the interference.

38. A method as in claim 37, wherein step F comprises estimating R, the correlation coefficient between nas_signal and {overscore (z)}.

39. A method as in claim 38, wherein, if {overscore (z)} has been more than one column, R is the minimum correlation coefficient between the columns of {overscore (z)} and nas_signal.

40. A method as in claim 38, wherein a result of step G comprises N-dimensional vectors for {overscore (SNR)}, {overscore (nas)}_signal, {overscore (nas)}_noise and {overscore (K)} corresponding to the respective calculation when each particular given sample is removed.

41. A method as in claim 40, wherein step H comprises

selecting the sample to remove that maximizes the N-by-1 measure of fitness, {overscore (M)}, where {overscore (M)} includes any of:

{overscore (SNR)};

{overscore (R)}², squared element-by-element;

{overscore (R)}²{overscore (SNR)}; and

{overscore (nas)}_signal.

42. A method as in claim 41, wherein limits are set for the minimum value of {overscore (nas)}_signal and the maximum value for {overscore (nas)}_noise, denoted c_signal and c_noise respectively.

43. A method as in claim 42, further comprising determining said measure of fitness according to:

If nas_signal,k-1 <

c_signal and nas_noise,k-1 >

c_noise then{overscore (M)} = {overscore (R)}²{overscore (SNR)};

else if nas_signal,k-1 >

c_signal and nas_noise,k-1 >

c_noise then{overscore (M)} = {overscore (nas)}_signal;

else{overscore (M)} = {overscore (R)}²end;

where nas<sub>signal,k−

1</sub> and nas<sub>noise,k−

1</sub> refer to the final nas_signal and nas_noise from the prior iteration.

44. A method as in claim 42, further comprising determining said measure of fitness according to:

{overscore (M)}=a₁{overscore (R)}^z+a₂{overscore (SNR)}+a₃{overscore (nas)}_signal;

where a_k are coefficients.

45. A method as in claim 44, wherein the sample selected for removal corresponds to the element of {overscore (M)} having the maximum value.

46. A method as in claim 44, wherein criteria for modifying {overscore (M)} are modifiable according to the information available.

47. A method as in claim 46, wherein a₂=0 when a partner analyte is not available.

48. A method as in claim 46, wherein, a₁=0, a₃=0 and SNR is replaced with nas_noise if measures of interference are not accessible.

49. A method as in claim 41, wherein step I continues until target values for c_signal and c_noise are obtained.

50. A method as in claim 49, further comprising

are preprocessing measurement data;

and

automatically calibrating a calibration using any of the methods of;

linear regression;

multiple linear regression;

principal component regression;

partial least squares regression; and

artificial neural networks.

51. A method as in claim 41, wherein step I continues until a desired number of samples is removed.

52. A method as in claim 41, wherein step I continues until the maximum number of samples is removed.

53. A method as in claim 41, wherein step I continues until removal of further samples would fail to produce a statistically significant increase in {overscore (M)} through an F-test of max({overscore (M)}) versus max({overscore (M)}_k-1).

54. A method as in claim 1, wherein creating new groupings of samples comprises iteratively leaving out groups of samples, said groups are either ordered by sample or randomly selected.

55. A method as in claim 1, wherein creating new groupings of samples comprises iteratively removing samples and iteratively putting previously removed samples back into the calculation.

56. A method as in claim 1, wherein a final calibration set is selected as the one associated having the highest value of said measure of fitness.

57. A method as in claim 1, wherein genetic algorithms or simulated annealing are applied to optimize said measure of fitness in applications with many samples.

58. A method as in claim 1, wherein determining a subset of samples comprises including or excluding samples on the basis of correlation to ancillary parameters.

59. A method as in claim 58, further comprising any of:

calculating said parameters from a spectral measurement, x; and

capturing said parameters from external instrumentation.

60. A method as in claim 59, wherein said external instrumentation comprises a temperature probe.

61. A method as in claim 58, further comprising scaling property values of both a signal being regressed, values of said signal comprising a first data set, and values of a targeted ancillary parameter, values of said parameter comprising a second data set, to mean zero and unit variance according to:

$X=\frac{x-\stackrel{\_}{x}}{{\mathrm{\sigma <br><br>}}_x}$

where X comprises variance scaled data points, x is measured or computed property values, {overscore (x)} is the mean of x, and σ

_x is the standard deviation of x.

62. A method as in claim 61, further comprising computing a difference between the two data sets by subtracting one data set from the other.

63. A method as in claim 62, further comprising removing samples having absolute residuals greater or less than a specified parameter in order to increase or decrease the correlation to the target parameter.