GB2632096A - A method and a device for testing a model - Google Patents

Abstract

A method for testing a multivariate model for generating a value of an analyte concentration using spectroscopic techniques, the method comprising: inputting a spectrum of an interferent 8 and processing the spectrum in combination with a spectrum of the analyte 4 to be measured to generate an output value for the analyte to be measured using the model. In addition, a method of producing a multivariate analysis model for a device for non-invasive spectroscopic analyte measure, the method comprising, during testing of the model, spectrally spiking an input spectrum to the model and determining the output from the model in response. In addition, a device for testing a model for non-invasive measurement of analyte concentration using spectroscopic techniques, arranged to receive a spectrum of known concentration of an analyte and a spectrum of an interferent, processing the receives spectra and generating an output for the analyte.

Description

A Method and a Device for Testing a Model The present invention relates to a method and device of testing a multi variate model used for non-invasive measurement of analyte concentration using spectroscopic techniques.

In embodiments, the method can be used to investigate the influence of the presence of certain substances on the model's performance in quantitating an analyte. A multivariate model herein may be defined as a processing tool that establishes a connection between multiple input variables and an output variable.

For example, the model could be arranged to receive as inputs a number of variables that can be measured or determined and then to process these and produce as an output, a value for blood glucose concentration or the glucose concentration of interstitial fluid.

Diabetes mellitus, in its different forms, is affecting an increasing number of -)c) individuals and placing undue strain on national health care budgets. Estimates (from 2015) state that 415 million people worldwide suffer from diabetes whilst this number is predicted to increase to 642 million by 2040.

For control of treatment the self-monitoring of blood glucose is recommended, which is usually performed with an invasive finger-prick method. In Type 1 diabetes patients dosing of insulin is frequently based on 4 to 6 blood glucose determinations per day. For these reasons, it has been a long-term goal to develop truly non-invasive techniques to measure the blood glucose levels of diabetes patients The current clinical trend favours indwelling electrochemical sensors that allow for continuous glucose monitoring in a minimally-invasive way. However, a skin puncture is still required, with associated discomfort for the user and increased risk of infection. These sensors also suffer from a biocompatibility issue that limits their life to a few weeks.

For many decades it has been a goal to develop non-invasive techniques to measure the blood glucose levels of diabetes patients but practical solutions for general use have so far not been developed. The majority of approaches have been based on optical measurement of glucose in tissues such as the skin. Amongst these, spectroscopic techniques, such as fluorescence, absorbance and Raman have attracted considerable attention. Despite the fact that inelastic Raman scattering is a weak process and thus results in a poor signal, a number of factors render it an attractive option as a spectroscopic technique for measurement of glucose, and indeed other lit analyte, concentrations in the skin of a user. They include the high chemical specificity, minimal interference from tissue water content and only a modest fluorescence background. These render the technique one of the most promising candidates for noninvasive glucose monitoring.

Since the first feasibility study of measuring blood glucose with near-infrared Raman spectroscopy in 1997 several groups have substantiated the fundamental effectiveness of the technology by quantitative measurements of glucose levels in vivo. However, these reports may be considered proof-of-concept only in the sense that all measurements were performed in a controlled environment whilst the predictive capabilities of the calibration model were assessed by cross-validation only.

In earlier publications and patent applications, the current applicant has described the design and development of a table-top, confocal near-infrared Raman instrument for intermittent glucose determination. The instrument uses a principle of critical-depth Raman spectroscopy, where measurements are taken from interstitial fluid within a defined region of the skin. It is worth noting that in contrast to previous technology that also utilizes a confocal setup to probe in the living part of the skin, the work of the current applicant is the first of its kind to systematically study the relation between probing depth and prospective performance of the Raman-based glucometer, thus allowing definition of a critical depth from which the Raman signal should be acquired.

In the current applicant's International application number W02011/83111 (granted in many jurisdictions) there is described a method and apparatus for non-invasive in vivo measurement by Raman spectroscopy of glucose present in interstitial fluid in skin. Amongst other aspects there is described apparatus for non-invasive in vivo measurement by Raman spectroscopy of glucose present in interstitial fluid in the skin of a subject, comprising a light source, optical components defining a light path from said light source to a measurement location, a light detection unit, optical components defining a return path for Raman scattered light from said measurement location to said light detection unit, and a skin engaging member having a distal surface for defining the position of said optical components defining the return path with respect to a surface of said skin in use, and wherein said optical components defining a return path for Raman scattered light selectively transmit to said light detection unit light scattered from near said measurement location such that at least 50% of Raman scattered light received at the light detection unit originates at depths from 60 to 400 pm beyond said distal surface of the skin engaging member.

In the current applicant's co-pending and granted patent portfolio a number of apparatuses are described for use in determining a blood glucose concentration using Raman spectroscopy.

Several parameters are preferably needed to meet user expectations in order to reach practical utility for a non-invasive glucose monitor (NIGM). These include for example, accuracy, cost, size, ease of use, calibration requirement and calibration stability.

Calibration stability has long been a unique selling point for the most successful continuous glucose monitors while the more conventional finger prick devices have sought to lower the cost per measurement. While it is expected that eventually NIGM technologies will be able to outperform all other technologies on cost per measurement since there is no need for consumables, this can only be practically realised if there are low requirements for calibration or none at all.

There is a desire to ensure that use of a multivariate model is stable and not subject to unacceptable variation of output in dependence on interferents.

According to a first aspect of the present invention, there is provided a method of testing a multivariate model for generating a value of an analyte concentration using spectroscopic techniques, the method comprising inputting a spectrum of an interferent and processing the spectrum in combination with a spectrum of the analyte to be measured to generate an output value for the analyte to be measured using the model.

The method preferably comprises: (a) providing a multivariate model for generating an output value of an analyte concentration based on a spectroscopic measurement; (b) inputting to the model a spectrum of known concentration of the to analyte to be measured; (c) inputting a spectrum of an interferent; and (d) processing the received spectra of the analyte to be measured and the interferent and generating an output value for the analyte to be measured using the model.

A method is provided that enables a multivariate model to be tested, which further enables determination of whether or not an interferent in question might be contraindicated for use with the model. A technical benefit is provided in that a simple and robust way of testing multivariate models is achieved, which enables the identification of contraindicating interferents.

In an example, the method comprises repeating the steps (a) to (d) to generate an indication of the effect of the spectrum of the interferent on the determined value of the analyte to be measured.

In an example, the analyte to be measured is glucose.

In an example, the interferent is a topical substance on the outside of the skin of a user.

In an example, the interferent is an interstitial substance found in the interstitial fluid of a user.

In an example, the interferent is an interstitial substance derived from outside the body of a user.

In an example, the method comprises, in dependence on the generated output value for the analyte to be measured updating the model.

According to a second aspect of the present invention, there is provided a method of producing a multivariate analysis model for a device for non-invasive measurement of analyte using spectroscopic techniques, the method comprising, during testing of the model, spectrally spiking an input spectrum to the model; and determining the output from the model in response to the spectrally spiked input.

According to a third aspect of the present invention, there is provided a device for non-invasive measurement of analyte concentration using spectroscopic techniques, the device having: an optical source; a spectrometer for receiving a generated spectrum for measurement of analyte concentration; and the device receiving the spectrum and inputting the spectrum to a measurement model, wherein the model has been tested using the method of the first aspect of the present invention.

According to a fourth aspect of the present invention, there is provided a device for testing a model for non-invasive measurement of analyte concentration using spectroscopic techniques, the device being arranged to: (a) receive a spectrum of known concentration of an analyte to be measured and a spectrum of an interferent; and (b) process the received spectra of the analyte to be measured and the interferent and generating an output value for the analyte to be measured using the model.

Thus, a device is provided to execute the method of the first aspect of the present invention. Typically, the device could be implemented using a computer processor arranged to receive the input spectra of the analyte to be measured and the interferent or the single superposed spectrum and to process it as described herein.

Embodiments of the present invention will now be described in detail with reference to the accompanying drawings, in which: Figure 1 is a schematic view of an optical configuration for taking non-invasive measurements of glucose concentration; Figure 2 is a schematic flow diagram showing the steps in a process of spectral spiking for device and/or model testing; Figures 3a and 3b show examples of measurement biases induced by spectral spiking with urea using a glucose measurement model. The test concentration of urea is in this example 6.5mM; Figure 4 shows the quantitative Raman spectra of glucose and urea at a concentration of 5mM; Figure 5 shows an example of a thenar spectrum and the same spectrum spiked with glucose and urea, respectively, at a concentration of 5mM; Figure 6 shows a representation of the average change in glucose measurements when thenar spectra are spiked with a spectrum of glucose (i e the analyte of interest) and urea, respectively, in varying concentrations; Figures 7 and 8 (A to C) show examples corresponding to the views of Figures 4, 5 and 6, with different interferent spectra used as the spiking spectra (alanine in the case of Figure 7 and ibuprofen in the case of Figure 8); Figure 1 is a schematic view of a non-limiting example of an optical configuration for taking non-invasive measurements of glucose concentration using Raman spectroscopy.

The basis for a spectroscopic setup is a light source, e.g., a laser, which is used for illuminating a sample. The light from the light source (the incoming light) will interact with the sample, and often result in an alteration of the light which is transmitted through, emitted by, reflected by and/or scattered by the sample. By collecting the altered light and analyzing its spectral distribution, information about the interaction between the incoming light and the molecular sample can be obtained; hence information about the molecular components can be obtained.

The spectral distribution is typically measured by using a spectrometer. A spectrometer is an optical apparatus that works by separating the light beam directed into the optical apparatus into different frequency components and subsequently measuring the intensity of these components by using e.g., a CCD detector, a CCD array, photodiode or such.

Figure 1 shows a first embodiment of an optical configuration that might be included within an optical probe 201. The probe comprises a first optical fibre 203 for guiding light into the optical probe 201. The light source is normally a laser, and the optical configuration is shown merely as an example of a suitable optical configuration for use with the calibration method described herein.

Upon exiting the first fibre 203, the incoming light 205 is collimated using a first lens 207 and optically filtered by passing through a first filter 209 blocking any percentage between 0 and 100 of frequencies/wavelengths outside the laser frequency/wavelength. Blocking of frequencies outside the laser frequency ensures that e.g. fluorescence generated inside the first fibre 203 is removed from the incoming light 205. The first filter 209 may also block any percentage between 0 and 100 of the laser frequency. This is an advantage if the intensity of the incoming light 205 is too high for the requirements of the sample. The first filter 209 is preferably a band-pass filter, a notch filter, an edge filter or such.

The optical probe 201 further comprises a dichroic mirror 211 that either reflects or transmits any percentage between 0 and 100 of the light, where the percentage of reflected and transmitted light is dependent on the coating on the dichroic mirror 211, the angle at which the light hits the dichroic mirror 211, and the frequency of the light. The dichroic mirror 211 can e.g. be coated such that it reflects the highest percent of the incoming light 205 when the dichroic mirror 211 is positioned at a given angle in relation to the direction of the incoming light 205. Changing the angle between the dichroic mirror 211 and the incoming light 205 will therefore reduce the percent of incoming light 205 reflected by the dichroic mirror 211.

In this example, most of the incoming light 205 is reflected by the dichroic mirror 211 and focused inside the skin 213 of a subject by a second lens 215. The focus point 217 of the incoming light 205 is defined by the focal length 218 of the second lens 215 and the distance distal of the lens of a window 219 and in particular its distal surface which engages the skin in use. The second lens 215 is preferably convex but could also be aspheric or planar.

The present method and device relate to a system in which the effect of interfering substances on the results provided by a non-invasive glucose monitoring system can be understood and allowed for in calibration models.

Interferences or interfering substances are defined as physical conditions or chemical substances that can affect performance such as the safety and/or effectiveness, of a measuring device.

Interference generally falls into the categories of interfering conditions and interfering substances. Interfering conditions include both device dependent conditions that affect the device which may be introduced by the device itself or by the surrounding environment, e.g., temperature, humidity, focal depth etc. User dependent interfering conditions relate to a physical condition that is introduced by the user such as disease or skin phototype.

Interfering substances can either be topical substances which are applied on the outside of the skin, some of which may migrate into the skin, or interstitial substances that are found in the interstitial fluid (ISF) including both those naturally found in the ISF and compounds derived from outside the body.

The present method and device relate primarily to potential interfering substances which includes the presence, or possibly absence, of chemical substances on or in the body of the subject, that may cause a device used to non-invasively measure the glucose concentration, to measure a false high or false low reading.

In systems that rely upon prevailing electrochemical-based glucose measurement technology, univariate regression is used to translate electrical current to a glucose concentration. The univariate approach is very sensitive to spurious chemical activity. This necessitates the issue of glucose specificity and robustness being solved on a hardware level to prevent the generation of interfering signals. For example, the issue of interfering substances may be mitigated by selecting enzymes that are particularly sensitive towards glucose and by coating the electrode with a partially selective membrane which restricts agents that are able to transfer through it.

Following the ISO 15197 standard for self-monitoring blood glucose systems, it is stated that a substance is considered as an interferent if it satisfies either of the following criteria: 1. For glucose concentrations less than 5.56mmo1/1, the average difference between the test sample and the control sample exceeds 0.55mmo1/1; and 2. For glucose concentrations greater than 5.56mmo1/1, the average difference between the test sample and the control sample exceeds 10%.

It is recognized that criteria for categorizing a substance as an interferent may vary between regulatory authorities. For example, the U.S. Food and Drug Administration expects interference evaluation according to the guidance document "Self-Monitoring Blood Glucose Test Systems for Over-the-Counter Use".

In the current system using non-invasive spectroscopic technology for the measurement of an analyte concentration, such as ISF glucose, multivariate analysis is used. Multivariate analysis techniques differ fundamentally from univariate techniques. Multivariate analysis techniques utilise information from multiple variables simultaneously. This allows for more sophisticated calibration and increased robustness by enabling a sensor to analyse and consider multiple sources of variation. For example, by utilising multiple variables simultaneously, multivariate sensing principles can provide increased robustness in the presence of environmental disturbances and other sources of variability, leading to more reliable results.

Multivariate sensing principles can provide more robust and accurate results by analysing multiple variables simultaneously, especially with large data sets that contain many naturally occurring changes in the environment and sample. This allows for a model to better learn and understand how other substances affect the many variables and provide more reliable results. Thus, multivariate sensing principles can account for many sources that can affect the signal and distinguish between changes related to the analyte in question and those related to other molecules in the sample or the environment.

In the case of non-invasive glucose measurement, the hardware is typically a confocal Raman spectrometer that is designed to acquire Raman spectra of the thenar skin. The thenar spectra contain general information relating to the skin and ISF constituents including glucose. However, it is only a small part of the signal that originates from the interaction of the incident laser light with the glucose molecules in the ISF. The multivariate nature of Raman spectra allows for separating and quantifying the glucose signal by multivariate regression techniques.

As is known, multivariate regression models are general mathematical tools that establish a connection between multiple input variables and an output variable. The models are generally achieved and designed by training on paired spectra and reference glucose concentrations that may typically be collected over many days/months and for many subjects and devices. An example of such an arrangement is described in our co-pending application GB2116869.5 and PCT/EP2022/078431.

The use of such multivariate regression models means that it is possible to distinguish spectral variations stemming from biological, environmental and device variations from glucose-induced variations. In other words, the model is tailored to measure glucose through skin measurements on humans.

Referring to the Figure 2, a method for spectral spiking will now be described.

Initially, a glucose model 2 is provided.

A thenar spectrum 4 is provided as an input to the glucose model 2. An initial glucose measurement (G) 6 is derived from the glucose model in dependence on the input thenar spectrum 4.

Next, the thenar spectrum 4 is spiked, i.e., perturbed, by superposition of the Raman spectrum 8 of a potential interferent. A spiked spectrum 10 is thus produced.

The spiked spectrum 10 is then similarly input to the glucose model 2 which results in a spiked glucose measurement (G*) 12.

The influence of the potential interferent is assessed by calculating the induced bias in the measurement. The influence can be determined simply based on the difference between G* and G, i.e., A = IG*-GI.

This spectral spiking is performed on a plurality of thenar spectra from different subjects and different reference glucose concentrations which thus results in a distribution of A values for a given potential interferent. The worst-case A values can be compared with clinical acceptability which might typically be included in a known standard, such as ISO 15197.

Although an individual spiking and reading can be used to determine the effect on an interferent spectrum on the desired output of a particular model, in practice, a large number of spectra are spiked and data is gathered.

In data gathered by the applicant, thenar spectra from 11,046 measurement sessions from 139 subjects were gathered and used in the spiking procedure. For an example of the distribution of As, see Figure 3.

Figures 3a and 3b show calculations for spiking with urea, with a glucose model, where the As are presented in box plots with the upper whiskers defining the worst-case biases, which are compared to the ISO 15197 thresholds, shown as the dashed horizontal lines.

The upper whiskers are in this example defined as: [(23 + 1.5(Q3 -(20e3m 114 > 0 Here, Q3 is the third quartile, (Q3-Q1) is the interquartile range and M is the medcouple parameter that robustly quantifies skewness in the distribution. It is worth noting that worst-case biases can be defined in multiple ways.

Q3 + 1.5 (Q3 -120e4M, < The concept of spectral spiking relies on spontaneous Raman scattering being a linear process which means that the strength of the Raman signal from a sample of multiple constituents reads

N

I'Raman ix 'laser a ici* 1=1 In the above expression, /laser represents the intensity of the irradiation laser, a, is the Raman scattering cross-section, i.e., scattering strength, and c is the concentration of the ith constituent.

Thus, the above expression emphasises the superposition of the contributions from the individual constituents.

Accordingly, if a potential interferent is introduced to the measurement sample, the resulting Raman signal will superimpose on the original signal. This is the basic principle behind spectral spiking, where a spectrum of a potential interferent is superimposed on the thenar spectra as if it was present in the skin or interstitial fluid.

For proper spiking of the thenar spectra, it is necessary to know the Raman spectrum, scattering strength, and physiological concentration of a potential interferent.

The physiological concentration can generally be found in scientific literature or databases, while the quantitative Raman spectrum can be measured in a laboratory by dissolving the potential interferent appropriately.

The superposition principle and glucose specificity of the apparatus used in the current example are illustrated by spiking thenar spectra with glucose and, by example, urea, of varying concentrations.

As shown in Figure 4, glucose and urea have quite similar Raman scattering strengths. When they are superimposed on a thenar spectrum in physiological concentrations they represent a perturbation of the original spectrum. See, for example, Figure 5. The variations in the spectra can be seen at the enlarged points for glucose and urea.

The situation is however very different when the spiked spectra are input to a glucose model such as that used in the process of Figure 2, because, as the name suggests, the model responds (as expected) strongly to the glucose spiking, but almost without change to the urea spiking. The sensitivity towards glucose (and insensitivity towards urea) is illustrated in Figure 6, which displays the effect on the average bias in glucose measurements when a glucose or urea spectrum is added to the thenar spectra to represent an interferent at varying concentrations. The spiking varies in concentration from -5 to 5 mM and the results shown in Figure 6 are based on 11,046 measurement sessions from 139 subjects. The error bars represent the standard deviation.

As can be seen, the spiking of the input spectra lead to an almost linear change in glucose measurements when the spiking is itself glucose. The output slope for glucose has a gradient of very nearly 1 (0.96). However, the gradient for the urea spectral spiking is -0.04, i.e., effectively negligible. The slope represents a measure of sensitivity and the fact that the slope of the glucose in that is on the order of 1 supports the utility of spectral spiking for identification of potential interference. In other words, urea does not cause any significant variation in the expected outcome, which demonstrates the robustness of the model for use in measuring glucose levels in patient. -)0

Thus, the present applicants have recognised that introducing spectral spiking works well as a method of determining the robustness of a model for non-invasive measurement of glucose concentrations using multivariate analysis.

Figures 7A to C and 8A to C show results of the process of multivariate spectral spiking of the glucose prediction model with spectra of alanine and ibuprofen. In both cases the relative shift in gradient compared to that caused by glucose is shown. As can be seen, the different perturbations of thenar spectra have different effects on the glucose measurements. In the examples of Figures 7 and 8, the calibration model is in all cases more sensitive to glucose spiking (as would be expected) than it is to spiking from the chosen tested interferents, alanine and ibuprofen.

Described above is a method of testing a model for non-invasive measurement of analyte concentration using spectroscopic techniques. As explained above, the method comprises providing an initial multivariate model (to be tested) for generating an output value of an analyte concentration based on a spectroscopic measurement. Typically, the model is multivariate regression model for determining the concentration of glucose in blood or interstitial fluid.

The method, in summary comprises inputting to the multivariate model a spectrum of known concentration of the analyte to be measured and a spectrum of an interferent. Typically, a superposed spectrum could be provided as the input made of a superposition of the spectrum of the analyte to be measured and the spectrum of the lit interferent. The received spectra (or superposed spectrum) are processed so as to generate an output value for the analyte to be measured using the multivariate model. The effect of the spectral perturbation or spiking can be seen in the change of the output value and, preferably, by repeating the process for various concertation levels of the analyte to be measured and/or the interferent.

This process works very well in identifying potential interferents and ensuring that the model is sufficiently robust, say, to satisfy particular standards.

Embodiments of the present invention have been described with particular reference to the examples illustrated. However, it will be appreciated that variations and modifications may be made to the examples described within the scope of the present invention.

Claims (12)

1. Claims 1. A method of testing a multivariate model for generating a value of an analyte concentration using spectroscopic techniques, the method comprising: inputting a spectrum of an interferent and processing the spectrum in combination with a spectrum of the analyte to be measured to generate an output value for the analyte to be measured using the model.
2. 2. A method according to claim 1, comprising: a) providing a multivariate model for generating an output value of an analyte concentration based on a spectroscopic measurement; b) inputting to the model a spectrum of known concentration of the analyte to be measured; c) processing the received spectra of the analyte to be measured and the interferent to generate the output value for the analyte to be measured.
3. 3. A method according to claim 2, comprising repeating the steps (a) to (c) to generate an indication of the effect of the spectrum of the interferent on the determined value of the analyte to be measured.
4. 4. A method according to any of claims 1 to 3, in which the analyte to be measured is glucose.
5. 5. A method according to claim 1 to 4, in which the interferent is a topical substance on the outside of the skin of a user.
6. 6. A method according to claim 1 to 5, in which the interferent is an interstitial substances are found in the interstitial fluid of a user.
7. 7. A method according to claim 5, in which the interferent is an interstitial substance is a compound derived from outside the body of a user.
8. 8. A method according to any of claims 1 to 7, comprising, in dependence on the generated output value for the analyte to be measured updating the model.
9. 9. A method of producing a multivariate analysis model for a device for non-invasive measurement of analyte using spectroscopic techniques, the method comprising during testing of the model, spectrally spiking an input spectrum to the model; determining the output from the model in response to the spectrally spiked input.
10. 10. A method according to any of claims 1 to 9, in which the spectrum of known concentration of the analyte to be measured and the spectrum of an interferent are provided as s ingle superposed spectrum to the model to be tested.
11. 11. A device for non-invasive measurement of analyte concentration using spectroscopic techniques, the device having: an optical source a spectrometer for receiving a generated spectrum for measurement of analyte concentration; and the device receiving the spectrum and inputting the spectrum to a measurement model, wherein the model has been tested using the method of any of claims 1 to 10.
12. 12. A device for testing a model for non-invasive measurement of analyte concentration using spectroscopic techniques, the device being arranged to: a) Receive a spectrum of known concentration of an analyte to be measured and a spectrum of an interferent; b) Processing the received spectra of the analyte to be measured and the interferent and generating an output value for the analyte to be measured using the model.