Classifications

• • G—PHYSICS
  • G01—MEASURING; TESTING
  • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
  • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
  • G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
  • G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
  • G01N21/27—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands using photo-electric detection ; circuits for computing concentration
  • G01N21/274—Calibration, base line adjustment, drift correction
• • G—PHYSICS
  • G01—MEASURING; TESTING
  • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
  • G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
  • G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
  • G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
  • G01N21/31—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
  • G01N21/35—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
  • G01N21/359—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light using near infrared light
• • G—PHYSICS
  • G01—MEASURING; TESTING
  • G01J—MEASUREMENT OF INTENSITY, VELOCITY, SPECTRAL CONTENT, POLARISATION, PHASE OR PULSE CHARACTERISTICS OF INFRARED, VISIBLE OR ULTRAVIOLET LIGHT; COLORIMETRY; RADIATION PYROMETRY
  • G01J3/00—Spectrometry; Spectrophotometry; Monochromators; Measuring colours
  • G01J3/28—Investigating the spectrum
  • G01J2003/2866—Markers; Calibrating of scan

Definitions

• Spectral analysis is widely used in identifying and quantitating analytes in a sample of a material.
• One form of spectral analysis measures the amount of electromagnetic radiation which is absorbed by a sample.
• an infrared spectrophotometer directs a beam of infrared radiation towards a sample, and then measures the amount of radiation absorbed by the sample over a range of infrared wavelengths.
• An absorbance spectrum may then be plotted which depicts sample absorbance as a function of wavelength.
• the shape of the absorbance spectrum including relative magnitudes and wavelengths of peak absorbances, serves as a characteristic 'fingerprint' of particular analytes in the sample.
• the absorbance spectrum may furnish information useful in identifying analytes present in a sample.
• the absorbance spectrum can also be of use for quantitative analysis of the concentration of individual analytes in the sample.
• the absorbance of an analyte in a sample is approximately proportional to the concentration of the analyte in the sample.
• the concentration of the analyte may be determined by comparing the absorbance of the sample to the absorbance of a reference sample at the same wavelengths, where the reference sample contains a known concentration of the analyte.
• One fundamental goal of a near-infrared spectroscopic method for biological fluid analyte concentration measurements such as blood glucose levels is to collect high quality data.
• outliers result from statistical errors or systematic errors
• outlier detection identifies samples containing such errors with sufficient confidence that such samples can be considered unique with respect to the sampled population. Inclusion of a small number of outliers within a set of measurements can degrade or destroy a calibration model that would otherwise be obtained by the measurements.
• a first source of error is related to sample preparation. Blood serum samples require a great deal of preparation before chemometric analysis. During this preparation, a number of factors can affect the sample. For example, the amount of time that blood samples are allowed to clot may affect sample continuity in terms of fibrinogen content. The level of clotting also impacts the quality of centrifugation and ultimately the decanting of serum from cells. Samples prepared for clinical assays determine the quality of the data used for reference and calibration, so that great care must be exercised with the samples since this data will ultimately define the limit of prediction abilities.
• a second source of error may result from the spectral measurement process.
• the use of a flowcell for sample containment during data acquisition is susceptible to problems such as bubbles in the optical path as well as dilution effects from reference saline solution carryover. These dilution effects are usually negligible, but bubbles in the optical path are not infrequent and have a severe impact on data quality.
• errors produced by mechanical or electronic problems occurring within the analysis instrumentation can have important effects on data quality.
• a third source of error is also related to the reference tests. Errors due to out-of-specification instrumental controls and low sample volume during clinical assays have similar effects to errors related to sample preparation, described above.
• a fourth source of error and probably the most difficult to identify and control, relates to sources of the samples, that is, to the individuals providing the biological fluids.
• a sample taken from an individual may at first seem to be quite unique with respect to a previously sampled population, but may in fact be an ordinary sample when a larger sample population is considered, that is, a putative unique sample may be only an artifact of undersampling.
• outliers from a data set can be accomplished in a qualitative and subjective sense by graphical inspection of plotted data in those cases when the dimensionality is low, that is, where the number of data points associated with each measurement is small. In those instances where the number of data points associated with each measurement is large, however, outlier detection may be more quickly and efficiently be accomplished by a number of automatable procedures such as residual analysis. However, such procedures are often subject to a number of errors, or at least subject to errors in interpretation, especially in the relatively high dimensional spaces that are typically associated with multifactorial chemometric analyses.
• chemometric applications for biological fluid analyte measurement require multiple measurements taken from a number of individual test subjects over a period of time.
• natural variations in samples and unintended errors can diminish the accuracy of results. Further, these errors are magnified by the relatively small number of biological fluid samples that can economically be drawn and tested.
• Automated techniques for outlier detection are necessary to assess the suitability of all acquired samples during both research phase and in final uses. The quality of data during clinical studies will define calibration models and the direction of subsequent research thrusts dependent upon results. In an end use, visual inspection of acquired data may or may not be possible.
• the present invention provides a method and apparatus whereby the concentration of an analyte in a sample of a biological fluid may be investigated by spectral analysis of electromagnetic radiation applied to the sample, including collecting calibration data, analyzing the calibration data to identify and remove outliers using the calibration model, constructing a calibration model, collecting unknown sample data, analyzing the unknown sample data to identify and remove outliers, and predicting analyte concentration of non-outliers in the unknown sample data by using the calibration model.
• the analysis of the calibration data set may include data pretreatment, data decomposition to remove redundant data, and identification and removal of outliers as having a low probability of class membership, using generalized distance methods.
• the construction of a calibration model may utilize principal component regression, partial least squares, multiple linear regression, or artificial neural networks, whereby the calibration data set may be reduced to significant factors using principal component analysis or partial least squares scores, enabling calculation of regression coefficients and artificial neural network weights.
• the unknown sample data may be analyzed using data pretreatment, followed by projection into the space defined by the calibration model, and identification and removal of outliers in the unknown sample data as having a low probability of class membership.
• the prediction of analyte concentration of an unknown sample may include projecting data from the unknown sample into the space defined by the calibration model, thereby enabling determination of the analyte concentration.
• a first embodiment of the apparatus of the present invention includes a pump into which a sample is introduced, the pump acting to circulate the sample through tubing to fill a flowcell, with the pump capable of both stopped flow and continuous flow operation.
• a sample compartment housing containing the flowcell and a detector is temperature controlled by a temperature control unit.
• Light from relatively broad bandwidth near-infrared source is directed through a chopper wheel, and the chopper wheel is synchronized by a chopper synchronization unit with respect to the detector, facilitating the apparatus of the present invention to make both light and dark measurements to substantially eliminate electronic noise. Moduloted light then passes through a monochrometer, allowing variance of the wavelength of radiation continuously over an appropriate range.
• the monochromatic light passes through the flowcell and strikes the detector, whereby the amount of light transmitted through the sample is measured.
• Measurement data is stored in a general purpose programmable computer having a general purpose microprocessor, available for further processing according to the present invention.
• the computer may also control operation of the pump, the temperature control unit, the chopper synchronization unit, the chopper wheel, and the monochrometer.
• light from the relatively broad bandwidth light source is directed through the chopper wheel, and thereafter modulated light is passed through a filter wheel, whereby discrete wavelengths of radiation may be selected and transmitted to the flowcell.
• a plurality of narrow bandwidth near-infrared sources such as a plurality of laser diodes, is provided to produce near-infrared radiation at a preselected plurality of wavelengths.
• Light from a selected narrow bandwidth near-infrared source may be pulsed by a driver in synchronization with the detector and directed into the flowcell 106. Synchronization of the selected narrow bandwidth near-infrared source and the detector permits the apparatus to make both light and dark measurements, thereby substantially eliminating significant electronic noise.
• Selection of each of the set of narrow bandwidth near- infrared sources for emission of light to be transmitted into the flowcell may be selected in a convenient order, for instance in order of increasing or decreasing wavelength, by configuring the computer to sequentially pulse each of the set of narrow bandwidth near-infrared sources.
• variations in the intensity of transmitted light as a function of wavelength are converted into digital signals by the detector, with the magnitude of the digital signals determined by the intensity of the transmitted radiation at the wavelength assigned to that particular signal. Thereafter, the digital signals are placed in the memory of the computer for processing as will be described.
• the steps of the method of the present invention includes as a first step collecting data to be used in constructing a calibration model.
• data pretreatment may be performed in order to remove or compensate for spectral artifacts such as scattering (multiplicative) effects, baseline shifts, and instrumental noise.
• Pretreatment of the calibration data may be selected from the group of techniques including calculating mh order derivatives of spectral data, multiplicative scatter correction, /7-point smoothing, mean centering, variance scaling, and the ratiometric method.
• near-infrared spectral calibration data may be formed into a nxp matrix representing n samples, each measured at p wavelengths.
• the nxp matrix may be decomposed by principal component analysis into a set of n, /7-dimensional score vectors formed into a nxn score matrix, and a set of n, -dimensional loading vectors formed into an nxp loading matrix.
• the score vectors are orthogonal and represent projections of the n spectral samples into the space defined by the loading vectors and the major sources of variation.
• Principal component analysis generates a set of n eigenvectors and a set of n eigenvalues, ⁇ . ⁇ 2 ⁇ ... ⁇ n .
• the eigenvalues represent the variance explained by the associated eigenvectors and can be divided into two sets.
• the first q eigenvalues are primary eigenvalues, ⁇ ⁇ 2 ⁇ ... ⁇ ql and account for the significant sources of variations within the data.
• the remaining n-q secondary (error) eigenvalues l q+ 1 ⁇ y. q+2 ⁇ ... ⁇ /. ⁇ 7 account for residual variance or measurement noise.
• the number of primary eigenvalues q may be determined by an iterative method which compares the q h eigenvalue's variance to the variance of the pooled error eigenvalues via an F-test. Further, reduced eigenvalues may be utilized, which weight the eigenvalues by an amount proportional to the information explained by the associated eigenvectors.
• the q score values for each sample are used to represent the original data during outlier detection, with the original spectra projected into the nxq dimensioned principal component subspace defined by loading the matrix.
• Outliers may be identified using generalized distances, such as
• a generalized distance between a sample and the centroid defined by a set of samples may be determined using the variance-covariance matrix of the set of samples. Where the true variance-covariance matrix and the true centroid of a complete set of samples are unknown, a subset of the complete set may be used to form an approximate variance-covariance matrix and an approximate centroid. Further, by using principal component scores to represent spectral data for each sample, independent variables maximizing the information content may be obtained, insuring an invertible approximate variance-covariance matrix.
• an approximate centroid may be determined as the centroid of a multivariate normal distribution of the set of calibration samples and an approximate variance-covariance matrix of the set of calibration samples, whereby an approximate Mahalanobis distances in units of standard deviations measured between the centroid and each calibration sample may be found.
• Robust distance by utilizing a minimum volume ellipsoid estimator (MVE), robust estimates of an approximate variance-covariance matrix and an approximate centroid may be obtained.
• MVE minimum volume ellipsoid estimator
• a projection algorithm may be used to determine the Robust distance for each calibration sample.
• the probability of class membership may be determined by a number of techniques, including evaluation of a chi-squared distribution function or utilizing Hotelling's T-statistic. Outliers are identified as having relatively large generalized distance which results in a relatively low probability of class membership. Samples whose class membership can be rejected at a confidence level that is greater than approximately 3-5 ⁇ may be considered as outliers. Following identification, outliers in the calibration samples may be removed. The generalized distances of outliers removed from the calibration samples may be examined, to determine whether additional data pretreatment is necessary. In the event that a relatively large number of outliers have very large generalized distances, further pretreatment of the calibration data may be indicated. After such additional pretreatment, the calibration data may again be subjected to analysis. On the other hand, if relatively large numbers of outliers do not have very large generalized distances, then additional data pretreatment may not be appropriate.
• a calibration model may then be constructed utilizing any of a number of techniques, including principal component regression (PCR), partial least squares (PLS), multiple linear regression (MLR), and artificial neural networks (ANN).
• the calibration model will seek to correlate a set of independent variables representing absorbance values of n samples each measured at p wavelengths, with a set of dependent or response variables representing the concentration of an analyte in each of the n samples, by using a p-dimensional regression coefficient vector.
• a calibration model determines regression coefficient vector and is used to predict the concentration of the analyte in other samples, given only the absorbances at the p wavelengths.
• near-infrared spectral data variables are highly correlated and while careful selection of the measurement wavelengths may minimize singularity problems, the spectral regions of interest may suffer from severe overlap and a high number of wavelengths is needed to model a multicomponent system.
• Data compression may be used to address problems with collinearity to determining regression coefficient vector, so that redundant data may be reduced down to significant factors.
• Principal component regression is one technique that incorporates a data compression method.
• the technique of partial least squares may also be used to address the problem of redundant data. With respect to both principal component regression and partial least squares, a determination is made of the appropriate number of score vectors or factors to be included in a calibration model that adequately represents the calibration data.
• the goal of selecting optimal number of factors for regression is to obtain parsimonious models with robust predictive abilities. Including too few factors causes model performance to suffer due to inadequate information during calibration, while including too many factors may also degrade performance.
• Principal components are normally sorted into an order so that the amount of variation explained by each principal component monotonically decreases. Later ordered principal components associated with small eigenvalues may be considered as containing measurement noise. By utilizing only the first q factors and omitting remaining factors, a type of noise rejection may be incorporated within principal component regression.
• the number of principal component analysis or partial least squares scores or factors to use during the regression step may be determined using the standard error of prediction, a measure of the error associated with each set of predictions.
• a piacewise continuous graphical representation may be obtained and utilized to determine the number of factors to retain.
• One criterion for factor selection is to determine the first local minimum.
• Another technique for factor selection uses an F-test to compare standard error of prediction from models using differing numbers of factors.
• data being analyzed may not be amenable to being split into a calibration, training set and a validation, test set.
• the reason may be due to a limited number of available samples or that by splitting data into two sets, one or both of the resulting sets do not adequately represent the sample population.
• the iterative technique of leave one out cross validation may be used where, during each iteration, a sample is excluded from the calibration set and is used as a test sample. Prediction models using factors determined from calibration samples are then used to make test sample predictions. The test sample is then returned to the calibration set and another sample is excluded. The same process is repeated until all samples have been excluded from the calibration set and predicted by models generated by the calibration samples. All predictions are accumulated to give a standard error of validation.
• the data set for the calibration model may be reduced to significant factors, and regression coefficients for the calibration model may be determined.
• the calibration model may be applied to data collected from samples where concentration of analytes of interest are unknown.
• the unknown sample data may be appropriately pretreated and then projected into the principal component space defined by the calibration model.
• generalized distances for the unknown sample data set may be found, using, for instance, either Mahalanobis or Robust distance as utilized with respect to the calibration data, and the probability of class membership may be estimated using the techniques described above, including evaluation of a chi-squared distribution function or utilizing Hotelling's T-statistic.
• Outliers in the unknown sample data are then identified based upon rejecting class membership at a confidence level that is greater than approximately 3-5 ⁇ .
• the sample is projected into the space defined by the calibration model, and a prediction of the concentration of the analyte made.
• the unknown sample is an outlier, the unknown sample may be rejected and no prediction as to analyte concentration made, although if possible, remeasurement of the unknown sample may be made to verify that the sample is an outlier.
• the steps previously described with respect to the method of the present invention may be configured on the general purpose microprocessor of the computer by employing computer program code segments according to each of such steps.
• FIG. 1 is a schematic block diagram of a first preferred embodiment of the apparatus for biological fluid analyte concentration measurement representing the present invention.
• FIG. 2 is a schematic block diagram of a second preferred embodiment of the apparatus for biological fluid analyte concentration measurement representing the present invention.
• FIG. 3 is a schematic block diagram of a third preferred embodiment of the apparatus for biological fluid analyte concentration measurement representing the present invention.
• FIG. 4 is a flowchart representing initial steps of the method for biological fluid analyte concentration measurement representing the present invention.
• FIG. 5 is a flowchart representing intermediate steps of the method for biological fluid analyte concentration measurement representing the present invention.
• FIG. 6 is a flowchart representing final steps of the method for biological fluid analyte concentration measurement representing the present invention.
• FIG. 7 is a scatter plot of principal component 2 versus principal component 1 of near-infrared spectra from 1 1 1 blood glucose samples in the range of 1580 nm to 1848 nm.
• FIG. 8 is a scatter plot of principal component 2 versus principal component 1 of near-infrared spectra from 1 1 1 blood glucose samples in the range of 2030 nm to 2398 nm.
• FIG. 9 is a scatter plot of principal component 3 versus principal component 2 of near-infrared spectra from 1 1 1 blood glucose samples in the range of 2030 nm to 2398 nm.
• FIG. 10 is a bar graph of calculated Mahalanobis distances for 103 blood glucose samples in the range of 1 100 nm to 2398 nm taken from data depicted in FIGS. 7-9.
• FIG. 1 1 is a scatter plot of predicted blood glucose concentrations from 103 samples using data derived from 2030 nm to 2398 nm, generated from a partial least squares model optimized with twelve factors attaining a standard error of validation of 64.10 mg/dL versus actual blood glucose concentrations.
• FIG. 12 is a scatter plot of predicted blood glucose concentrations from 100 samples using data derived from 2030 nm to 2398 nm, generated from a partial least squares model optimized with eight factors attaining a standard error of validation of 27.43 mg/dL versus actual blood glucose concentrations.
• FIG. 13 is a bar graph of calculated Mahalanobis distances for 100 blood glucose samples in the range of 1580 nm to 1848 nm taken from data depicted in FIGS. 7-9.
• FIG. 14 is a bar graph of calculated Mahalanobis distances for 100 blood glucose samples in the range of 2030 nm to 2398 nm taken from data depicted in FIGS. 7-9.
• FIG. 15 is a scatter plot of predicted blood glucose concentrations from 95 samples using data derived from 2030 nm to 2398 nm, generated from a partial least squares model optimized with eight factors attaining a standard error of validation of 26.97 mg/dL versus actual blood glucose concentrations.
• FIG. 16 is a table representing a summary of outlier detection results for 1 1 1 blood glucose samples over the spectral ranges 1580 nm to 1848 nm and 2030 nm to 2398 nm utilizing the present invention, and indicating possible causes of sample error.
• FIG. 17 is a graph of the standard error of prediction versus the numbers of factors used during regression. DESCRIPTION OF THE PREFERRED EMBODIMENTS
• FIG. 1 depicting a first preferred embodiment of an apparatus for biological fluid analyte concentration measurement.
• a biological fluid sample may be introduced into pump 102 which circulates the sample through tubing 104 to fill flowcell 106.
• Pump 102 may be capable of both stopped flow and continuous flow operation.
• Sample compartment 108 contains flowcell 106 and detector 1 10, and is temperature controlled by temperature control unit 1 12.
• Light from relatively broad bandwidth near-infrared source 1 14 is directed through chopper wheel 1 16.
• Chopper wheel 1 16 is synchronized by chopper synchronization unit 1 18 with respect to detector 1 16, facilitating apparatus 100 to make both light and dark measurements to substantially eliminate electronic noise.
• Modulated light then passes through monochrometer 120, allowing continuous variance of the wavelength of radiation over an appropriate range.
• the monochromatic light passes through flowcell 106 and strikes detector 1 10.
• Detector 1 10 measures the amount of light transmitted through the sample. Measurement data is then stored in general purpose programmable computer 124 having a general purpose microprocessor, where the data will be available for further processing as will be described.
• computer 124 may also control operation of pump 102, temperature control unit 1 12, chopper synchronization unit 1 18, chopper wheel 1 16, and monochrometer 120.
• a plurality of narrow bandwidth near-infrared sources 134 such as a plurality of laser diodes, is provided to produce near-infrared radiation at a preselected plurality of wavelengths.
• Light from a selected narrow bandwidth near-infrared source 134 may be pulsed by driver 138 in synchronization with detector 1 10 and directed into flowcell 106.
• Synchronization of the selected narrow bandwidth near-infrared source 134 and detector 1 10 permits apparatus 100 to make both light and dark measurements, thereby substantially eliminating electronic noise.
• Selection of each of the set of narrow bandwidth near- infrared sources 134 for emission of light to be transmitted into flowcell 106 may be selected in a convenient order, for instance in order of increasing or decreasing wavelength, by configuring computer 124 to sequentially pulse each of the set of narrow bandwidth near-infrared sources.
• variations in the intensity of transmitted light as a function of wavelength are converted into digital signals by the detector, with the magnitude of the digital signals determined by the intensity of the transmitted radiation at the wavelength assigned to that particular signal. Thereafter, the digital signals are placed in the memory of computer 124, for processing as will be described.
• step 1 in the method of the present invention refers to collecting data to be used in performing calibration and thereafter constructing a calibration model.
• data pretreatment of step 2 may be performed, as it is often necessary to pretreat raw spectral data prior to data analysis or calibration model building in order to remove or compensate for spectral artifacts such as scattering (multiplicative) effects, baseline shifts, and instrumental noise.
• Pretreatment of the calibration data may be selected from the group of techniques including calculating nth order derivatives of spectral data, multiplicative scatter correction, n-point smoothing, mean centering, variance scaling, and the ratiometric method.
• near-infrared spectral calibration data may be formed into a nxp matrix X representing n samples, each measured at p wavelengths, and may be decomposed by principal component analysis into a set of n, /7-dimensional score vectors formed into a nxn score matrix T, and a set of n, -dimensional loading vectors formed into an nxp loading matrix L, with
• decomposition may be considered decomposing matrix X of rank n into a sum of n rank 1 matrices.
• the score vectors represent projections of the n spectral samples in X into the space defined by the loading vectors.
• the score matrix T represents the major sources of variation found within X, and the column vectors in 7 are orthogonal.
• principal component analysis generates a set of n eigenvectors and a set of n eigenvalues, ⁇ ⁇ 2 ⁇ ... ⁇ n .
• the eigenvalues represent the variance explained by the associated eigenvectors.
• the eigenvalues may be divided into two sets.
• the first q eigenvalues are primary eigenvalues /- ! ⁇ 2 ⁇ ... ⁇ q and account for the significant sources of variations within the data.
• the remaining n-q secondary, or error, eigenvalues ⁇ q+1 ⁇ /. q+2 ⁇ ... ⁇ n account for residual variance or measurement noise.
• the number of primary eigenvalues q may be determined by an iterative method which compares the qr th eigenvalue's variance to the variance of the pooled error eigenvalues via an F-test,
• reduced eigenvalues which weight the eigenvalues by an amount proportional to the information explained by the associated eigenvectors may be utilized with the reduced eigenvalue is defined as
• equation 2 may be expressed as
• the I th sample in the principal component subspace is represented by the q score values of ?,.
• the q score values for each sample are used to represent the original data during outlier detection. In doing so, the original spectra are projected into the nxq dimensioned principal component subspace defined by loading matrix L.
• outliers may be identified using generalized distances, such as Mahalanobis distance or Robust distance.
• a generalized distance between a centroid ⁇ of a set of samples and the I th sample x, may be determined from
• D i [ (x i - ⁇ ) ⁇ - 1 (x i - ⁇ ) t ] 1 2 (5)
• ⁇ is the variance-covariance matrix of the set of samples.
• a subset of the complete set of samples may be used to form an approximate variance-covariance matrix and an approximate centroid.
• independent variables are orthogonal thus maximizing the information content and insuring an invertible approximate variance-covariance matrix.
• Generalized distances may be Mahalanobis distances as described in step 10a of FIG. 4, with an approximate centroid"x determined as the centroid of a multivariate normal distribution of the set of calibration samples and an approximate variance-covariance matrix of the set of calibration samples S.
• An approximate Mahalanobis distances MD in units of standard deviations measured between the centroid and an / th calibration sample x, may thus be determined from
• a projection algorithm may be used to determine the Robust distance RD, for the I th calibration sample from
• X j a p-dimensional vector representing the ? h calibration sample
• v g is a p-dimensional vector representing the g th calibration sample defined by
• M is a p-dimensional vector such that the I th component of M is given by the median of a set formed by the r th component of each of the n vectors x, .
• index / used in equations 10 and 1 1 is determined from
• the probability of class membership may be determined by a number of techniques, including evaluation of a chi-squared distribution function or utilizing Hotelling's T- statistic. As depicted in step 12, outliers are identified as having relatively large generalized distance which results in a relatively low probability of class membership. Generally speaking, samples whose class membership can be rejected at a confidence level in the range of approximately 3-5 ⁇ may be considered as outliers. Following identification, outliers in the calibration samples may be removed as depicted in step 13. Further, as indicated in step 14, the generalized distances of outliers removed from the calibration samples may be examined to determine whether additional data pretreatment is necessary.
• a calibration model may be constructed utilizing any of a number of techniques, including principal component regression (PCR), partial least squares (PLS), multiple linear regression (MLR), and artificial neural networks (ANN).
• PCR principal component regression
• PLS partial least squares
• MLR multiple linear regression
• ANN artificial neural networks
• the calibration model will seek to correlate a set of independent variables representing absorbance values of n samples measured at p wavelengths, symbolically represented by the nxp matrix X, with a set of dependent or response variables representing the concentration of an analyte in each of the n samples, symbolically represented by vector y.
• y is an /7-dimensional vector, or altematively, may be considered to be an nx1 matrix.
• b represents a p-dimensional regression coefficient vector (px1) matrix) and ⁇ is an n-dimensional vector (nxi matrix) representing errors in y.
• the calibration model determines vector b, using
• Knowledge of b is used to predict the concentration of the analyte, y, in unknown samples, given only absorbances at each of the p wavelengths.
• step 16 the determination of (X'X) '1 may be difficult as collinearity is inherent in spectroscopic data.
• near- infrared spectral data variables are highly correlated. While careful selection of the measurement wavelengths may minimize singularity problems, the spectral regions of interest may suffer from severe overlap and a high number of wavelengths is needed to model a multicomponent system.
• Data compression may be used to address problems with collinearity to determining regression coefficient vector b, so that redundant data may be reduced down to significant factors.
• Principal component regression is one technique to determine vector b that incorporates a data compression method.
• the first step in principal component regression is to perform principal component analysis on the calibration data as formed into matrix X.
• the score matrix T represents the major sources of variation found within X, and the column vectors in T are orthogonal.
• T is used in place of X whereby an approximate value of b is found using
• partial least squares (16) as (7T f ) is invertible.
• the techniques of partial least squares may also be used to address the problem of redundant data.
• One difference between partial least squares and principal component regression is the way in which the score matrix T and the loading matrix L are generated.
• NPALS non-linear iterative partial least squares
• loading vectors are extracted one at a time in the order of their contribution to the variance in X. As each loading vector is determined, it is removed from X and the next loading vector is determined. This process is repeated until n loadings have been determined.
• concentration, y block information is used during iterative decomposition of X.
• r values are related to concentration as well as placing useful predictive information into earlier factors as compared to principal component regression.
• determination must be made of the appropriate number of score vectors or factors to be included in a calibration model that adequately represents the calibration data.
• the goal of selecting optimal number of factors for regression is to obtain parsimonious models with robust predictive abilities. Including too few factors causes model performance to suffer due to inadequate information during calibration. Including too many factors may also degrade performance.
• Principal components are normally sorted into an order so that the amount of variation explained by each principal component monotonically decreases. Later ordered principal components associated with small eigenvalues may be considered as containing measurement noise.
• a type of noise rejection may be incorporated within principal component regression.
• the number of principal component analysis or partial least squares scores or factors, q, to use during the regression step may be determined as follows.
• n preliminary calibration models are built. Each preliminary calibration model uses a different number of score vectors selected from the range of 1 through n score vectors. Predictions are then made form the n preliminary calibration models using the standard error of prediction technique
• SEP standard error of prediction
• a piecewise continuous graphical representation such as FIG. 17 may be obtained and utilized to determine the number of factors to retain.
• One criterion for factor selection is to determine the first local minimum. Applying a first local minimum criterion to the data graphed in
• FIG. 17 eight factors would be selected for the calibration model.
• a general interpretation of FIG. 17 is that significant information is being incorporated into the calibration model in factors one through six. As factors seven and eight are included, subtleties in the data are included. For factors nine through fifteen, variations or measurement noise specific to the calibration set are being modeled, so errors increase.
• Another technique for factor selection uses an F-test to compare standard error of prediction from models using differing numbers of factors. An f-test factor optimization would find that the standard error of prediction an eight factor model does not vary significantly from the standard error of prediction of a six factor model, whereby six factors is seen to be optimal.
• data being analyzed may not be amenable to being split into a calibration, training set and a validation, test set.
• the reason may be due to a limited number of available samples or that by splitting data into two sets, one or both of the resulting sets do not adequately represent the sample population.
• the technique of leave one out cross validation may be used in such a situation. Leave one out cross validation is an iterative process, where during each iteration, a sample is excluded from the calibration set and is used as a test sample. Prediction models using 1 through n- 1 factors determined from n- 1 calibration samples are then used to make test sample predictions. The test sample is then returned to the calibration set and another sample is excluded. The same process is repeated until all n samples have been excluded from the calibration set and predicted by models generated by the n- 1 calibration samples. All predictions are accumulated to give the standard error of validation (SEV) given by
• step 17 where the subscript (i) represents the I th leave one out iteration which leaves out the I th sample, with the standard error of validation then treated as standard error of prediction.
• data for the calibration model may be reduced to significant factors, and regression coefficients for the calibration model may be determined.
• the calibration model as described above may be applied to data collected from samples where concentration of analytes of interest are unknown, symbolically indicated in FIG. 6 as step 18.
• the unknown sample data may be appropriately pretreated as indicated at step 19, with similar techniques to those described above with respect to pretreatment techniques capable of use with calibration data.
• the sample data may be projected into the principal component space that was previously defined by the calibration model, as indicated in step 20.
• generalized distances for the unknown sample is found using the generalized distance, such as Mahalanobis or Robust distances, that was utilized with respect to the calibration data.
• the probability of class membership may be estimated using the techniques described above, including evaluation of a chi-squared distribution function or utilizing Hotelling's T- statistic.
• unknown sample outliers may then identified based upon rejecting class membership at a confidence level that is in the approximate range of 3-5 ⁇ .
• the unknown sample may be projected into the space defined by the calibration model, and a prediction of the concentration of the analyte may be made.
• the unknown sample should be rejected and no prediction as to analyte concentration is made, although if possible, remeasurement of the unknown sample may be made for reanalysis to verify that the unknown sample is indeed an outlier.
• the method and apparatus of the present invention was applied to blood glucose concentration data obtained from samples from 1 1 1 individuals.
• Six of the samples did not have enough serum to collect a near-infrared spectrum, so that vectors of zeros were used to fill their position within the data matrix in order to maintain succession number integrity during data manipulation.
• the six samples and the associated reference tests were omitted from future analyses. Two other samples were associated with reference test errors and were omitted, leaving 103 samples in the data set.
• FIGS. 7-9 depict separate principal component analyses of two spectral regions performed.
• Vectors of zeros, indicated by reference numeral 200 lie far from the main group of data, as expected.
• samples 28 and 44 are seen to be potential outliers, as are samples 3 and 4 in FIG. 8.
• samples 3, 4, and 44 are potential outliers.
• Mahalanobis distances were calculated for the 103 samples, as shown in FIG. 10, wherein samples 23, 67, and 83 are seen to have Mahalanobis distances much greater than the other samples. Further, in FIGS. 10, 13, and 14, omitted samples are depicted as having zero Mahalanobis distance. A number of additional samples appear in FIG. 10 to be outlier candidates, including samples 3, 4, and 44. The data were subjected to further analyses, as will be described, with samples 23, 67, and 83 omitted, leaving 100 samples in the data set.
• FIGS. 1 1 and 12 depict the detrimental impact of including outlier samples in a data set.
• FIG. 1 1 depicts a scatter plot of predicted blood glucose concentrations from 103 samples using data derived from 2030 nm to 2398 nm generated from a partial least squares model optimized with twelve factors attaining a standard error of validation of 64.10 mg/dL versus actual blood glucose concentrations.
• FIG. 12 depicts a scatter plot of predicted blood glucose concentrations from 100 samples using data derived from
• FIG. 13 and 14 Nine samples were flagged as possible outliers in the 1580 nm to 1848 nm region, and six samples were flagged in the 2030 nm to 2398 nm region as possible outliers. As is apparent from comparison of FIGS. 13 and 14, the flagged samples were different in the two spectral regions. Outliers may be selected to be those flagged samples that are excluded from class membership in either or both spectral ranges, at a confidence level selected to be in the range of 3-5 ⁇ Four of the samples rejected were also identified as possible outliers from the principal component score plots, FIGS. 7-9. Identification of the fifth sample required examination in the higher dimensional space associated with Mahalanobis distances. FIG.
• a set of 89 samples results.
• leave-one-out validation on the set of 89 samples resulted in an SEV of 27.95 mg/dL, a slight increase over the 100 and 95 sample sets. It was separately determined that the six samples omitted in the 89 sample set with respect to the 95 sample set corresponded to samples having a high triglyceride concentration, a high total protein value, or both.
• the presence of the six outliers constituted an artifact of undersampling, that is, if a greater number of representative samples with high triglyceride or total protein concentrations were present in the original set of samples, samples having high triglyceride or total protein concentrations would be less likely to be flagged as outliers. Sensitivity of outlier detection to triglyceride or any other analyte which affects spectral response may be advantageous, however.
• Spectral data may be partitioned such that samples with high triglycerides form a first calibration set while samples with low triglycerides form a second calibration set, so that new samples may be tested with the method and apparatus of the present invention to determine whether the first or second calibration set is representative of the new sample, thus allowing the selection of a prediction model determined from "similar" calibration spectra.
• the method and apparatus of the present invention encompasses altemative biological fluid analyte measurement techniques, including biological fluid analyte concentrations derived using light reflectance, light transmission, and other techniques used in conjunction with invasive, non-invasive, and in-vivo biological fluid analyte measurement techniques.
• measurements of biological fluid analytes may also include triglycerides, cholesterol, and serum proteins, with outlier detection using the method and apparatus of the present invention.

Landscapes

• Physics & Mathematics (AREA)
• Spectroscopy & Molecular Physics (AREA)
• Analytical Chemistry (AREA)
• General Physics & Mathematics (AREA)
• Pathology (AREA)
• Health & Medical Sciences (AREA)
• Life Sciences & Earth Sciences (AREA)
• Chemical & Material Sciences (AREA)
• Immunology (AREA)
• Biochemistry (AREA)
• General Health & Medical Sciences (AREA)
• Theoretical Computer Science (AREA)
• Engineering & Computer Science (AREA)
• Mathematical Physics (AREA)
• Investigating Or Analysing Materials By Optical Means (AREA)
• Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
• Investigating Or Analysing Biological Materials (AREA)
• Analysing Materials By The Use Of Radiation (AREA)
• Measuring Pulse, Heart Rate, Blood Pressure Or Blood Flow (AREA)

Applications Claiming Priority (5)

Publications (2)

Family

ID=26669720

Family Applications (1)

Country Status (5)

Cited By (1)

Families Citing this family (22)

Citations (1)

Family Cites Families (2)

• 1996
  • 1996-08-02 CA CA002228844A patent/CA2228844C/en not_active Expired - Fee Related
  • 1996-08-02 WO PCT/US1996/012625 patent/WO1997006418A1/en active Application Filing
  • 1996-08-02 JP JP50854897A patent/JP3323512B2/ja not_active Expired - Fee Related
  • 1996-08-02 EP EP96926226A patent/EP0846253A4/de not_active Withdrawn
  • 1996-08-02 AU AU66448/96A patent/AU711324B2/en not_active Ceased

Patent Citations (1)

Non-Patent Citations (5)

Also Published As

Similar Documents

Legal Events