WO2017129634A1

WO2017129634A1 - Estimating a substance concentration using multiple non-invasive sensors

Info

Publication number: WO2017129634A1
Application number: PCT/EP2017/051573
Authority: WO
Inventors: Terje VEGAR KARSTANG
Original assignee: Prediktor Medical As
Priority date: 2016-01-25
Filing date: 2017-01-25
Publication date: 2017-08-03
Also published as: EP3408772A1; CN107924715A

Abstract

A method for determining a level of a substance in a material is disclosed. The method comprises: obtaining a first measured spectrum from a first noninvasive sensor, the first measured spectrum relating to a first characteristic of the material; obtaining a second measured spectrum from a second noninvasive sensor, the second measured spectrum relating to a second characteristic of the material; calculating a first intermediate estimate ŷ _x of the concentration of a chemical constituent in the material based on the first measured spectrum; calculating a second intermediate estimate ŷ ₂ of the concentration of a chemical constituent in the material based on the second measured spectrum; determining a first and second residual spectrum r₁, r₂, the residual spectra defined as the deviation between the measured spectrum and an estimated spectrum, the estimated spectrum being the component of the measured spectrum corresponding to the intermediate estimate of the concentration of the chemical constituent; and calculating an improved estimated concentration of the chemical constituent in the material based on the first intermediate estimate ŷ ₁, the second intermediate estimate ŷ ₂ and the residual spectra r₁, r₂.

Description

ESTIMATING A SUBSTANCE CONCENTRATION USING MULTIPLE NONINVASIVE SENSORS

TECHNICAL FIELD

The present invention relates to methods and devices for use in estimating the concentration of a chemical constituent in a material, preferably without the need for invasive or direct measurement.

BACKGROUND

A great deal of control systems operate on the basis of knowing the levels of various substances within a material, often where the material is made of multiple different substances. This is typically measured by taking a sample of the material. This sample is then used in various chemical and physical tests to determine its composition. Such direct measurements may be termed "invasive" methods.

However, in some situations it is undesirable or infeasible to directly measure the level of the substance in the material by taking a sample. This may be because of the cost, difficulty or danger of retrieving the samples.

A further difficulty arises when the level of the substance changes over time, for example due to metabolism or decay. This necessitates continually taking invasive samples in order to have accurate information about the level. However, it may be undesirable to repeatedly take samples of the material for conventional analysis, as taking such samples may damage the material, and the analysis tends to destroy the sample.

In particular, many modern medical techniques rely on knowing the concentration of a chemical constituent in blood. For example, people with diabetes need to regularly monitor their blood glucose levels. Conventionally, this would involve taking hourly blood samples from the person using invasive means. For example, a needle may be inserted into a vein in the arm of the person, and a sample of the blood extracted into a syringe or by using a finger pricking device. This sample can then be analysed directly by a suitable device.

This process has the advantage of providing a high accuracy reading of the blood glucose level. However, in practice there are many disadvantages of using such invasive means. In particular, they are inconvenient and tend to require significant training to safely use. More problematically, even when used properly, taking blood samples may cause tissue damage, and can result in infection.

Because of these disadvantages, there has been research into non-invasive methods for taking samples. For example, the use of optical sensors may provide an estimate of the level of the substance (such as glucose) in the material (such as blood). However, while such non-invasive methods are more convenient, they tend to be of low accuracy. This may be because, for example, a sensor measures not only the desired substance, but also other undesired substances in the material.

There is consequently a need for an improved non-invasive approach to measuring, or at least estimating, the level of a substance, in particular the concentration of a chemical constituent, in a material.

SUMMARY OF INVENTION

In a first aspect, there is provided a method for determining a concentration of a chemical constituent in a material, comprising:

A method for determining a concentration of a chemical constituent in a material, comprising: obtaining a first measured spectrum from a first noninvasive sensor, the first measured spectrum relating to a first characteristic of the material; obtaining a second measured spectrum from a second noninvasive sensor, the second measured spectrum relating to a second characteristic of the material; calculating a first intermediate estimate of the concentration of a chemical constituent in the material based on the first measured spectrum; calculating a second intermediate estimate of the concentration of a chemical constituent in the material based on the second measured spectrum; determining a first and second residual spectrum, the residual spectra defined as the deviation between the measured spectrum and an estimated spectrum, the estimated spectrum being the component of the measured spectrum corresponding to the intermediate estimate of the concentration of the chemical constituent; and calculating an improved estimated concentration of the chemical constituent in the material based on the first intermediate estimate, the second intermediate estimate and the residual spectra r-ι, r₂.

By using this approach, it becomes possible to obtain an estimate of the level of a substance in a material. Because multiple non-invasive sensors are used with the readings combined (sensor fusion), the estimate has an improved accuracy. This approach can be done without the need for an invasive sample of the material to be taken, thereby avoiding the difficulties and risks associated with obtaining such a sample. Moreover, because no such sample is needed, it can be reasonably straightforward to regularly monitor the levels. This provides a more convenient method for determining the internal state of a system which relies on the material. Furthermore the use of the residual spectrum - the difference between a the measured spectrum and an estimated spectrum (which is fitted to the measured spectrum via a calibration model) - provides a straightforward means to improve the accuracy of the final estimation.

Calculating the intermediate estimates of the level of the substance in the material can be performed in any suitable manner. However, in a preferred embodiment, calculating a first intermediate estimate y_x of the concentration of a chemical constituent in the material comprises: providing the first measured spectrum u)i as input to a predetermined first calibration model describing the relationship between the corresponding characteristic and concentration of the chemical constituent; and receiving an output from the first calibration model as the first intermediate estimate y_x and wherein calculating a second intermediate estimate y₂ of the concentration of the chemical constituent in the material comprises: providing the second measured spectrum ω₂ as input to a predetermined second calibration model describing the relationship between the corresponding characteristic and concentration of the chemical constituent; and receiving an output from the calibration model as the second intermediate estimate y₂ level of the substance in the material.

In preferred embodiments the calibration model is a latent variable model, such as a partial least squares regression model, having a form ω = tP + r, where an estimated spectrum tP is fitted to the measured spectrum ω, the difference in the fit being the residual spectrum r. The loading matrix P may be estimated from a separate calibration data set using regression. The calibration model may then be utilised, following a sensor reading from a material, to fit the estimated spectrum (corresponding to the estimated concentration of the chemical constituent) to a measured spectrum ω and accordingly output an intermediate estimate of the concentration according to that sensor. In this way, an initial reading from the sensor can be used to provide an intermediate estimate of the concentration using a calibration model specific to that sensor in a straightforward manner. The residual spectrum may then provide a means to further adjust the estimate using a function of the residual as a correction term.

The calibration model may be configured to model the change in the level of the substance in the material over time. A model is easily adaptable and updatable, so that the calculation of the estimated level can be kept highly accurate.

It can be useful for the estimating to take account of possible errors in measurement and of the effect of the other substances in the material on the reading - not calibrated for in the calibration model of each sensor. Thus, the method may further comprise: calculating a first correction term, /(ΙΊ), and a second correction term, /(r₂), such that: y = <xi ( i + /(^ri)) + 9i = ^a2 (y₂ + /<¾)) + g₂ where: y_x is the intermediate estimate from the first sensor; y₂ is the second intermediate estimate from the second sensor; x is a first scaling factor; a₂ is a second scaling factor; gi is a first error term; g₂ is a second error term; and y is an actual measurement of the level of the substance in the material. The first correction term can be formulated as

and the second correction term can be formulated as

/(r₂) = S₂rsd(r₂) where:

S_t is a first coefficient; δ₂ is a second coefficient; and r₁ is a first residual spectrum; r₂ is a second residual spectrum; and rsd is the residual standard deviation of a residual spectrum.

This allows for the estimate for be calculated much more accurately, by using information from the residual spectra, which relates to the response of the measurement due to substances which are not calibrated for in the initial calibration model. In practice it is not possible to develop a calibration model specific to the exact multiple constituent substances of a measured material, therefore use of a function of the residual spectrum may be used to adjust the estimate to account for this.

In some cases, there may be a further inaccuracy between an actual measurement and the corrected estimate. Thus in preferred embodiments, the method further comprises: calculating a bias, β, such that: y = ay + β where: y is the estimated level of the substance in the material according to the model; and a is a third scaling factor.

This allows for the estimate to be made more accurate by adjusting for the bias over time. Here y may be the estimate from a physical model, such as a metabolism model, which may be combined with the first and second intermediate estimates to further improve the accuracy of the final estimate.

This may occur by updating the model based on one or more of the first correction term, the second correction term and the bias, thereby improving the accuracy of the model.

Although actual (non-estimated) measurements are not essential, they can be useful to ensure that the estimated measurements are accurate. Thus one or more of the first scaling factor, the second scaling factor, the third scaling factor and the actual measurement of the level of the substance in the material are calculated by means of measurement of the level of the substance in the material. In contrast, one or more of the first correction term, the second correction term and the bias term are calculated by means of a recursive least squares regression or a Kalman filter.

The approach can be used for any substance and any material. However, it has been found particularly useful where the substance is glucose and the material is the blood of a user. This avoids the need for conventional invasive glucose measurement. In such cases, the first non-invasive sensor comprises an optical sensor configured to measure absorbance of light and the second sensor noninvasive comprises a bio-impedance sensor configured to measure the impedance of skin of the user. Such an arrangement provides measurements which lead to an accurate estimate.

Where glucose is measured, it is possible to predict how other factors (such as eating or exercise) will influence glucose. Thus in some embodiments, the method further comprises: receiving a third reading representing a quantity of food ingested by the user; and receiving a fourth reading representing a quantity of exercise performed by the user; wherein calculating an estimated level of the substance in the material further comprises: inputting the third reading and the fourth reading into the model; wherein the output of the model is further based on the third reading and the fourth reading. The model can also be adapted to model the metabolism of the user. This allows the estimate to have an improved predictive power.

In a second aspect, there is provided a device configured to perform the method of the first aspect.

BRIEF DESCRIPTION OF FIGURES

Examples of the present invention will now be described with reference to the accompanying drawings, in which:

Figure 1 shows a method for non-invasively obtaining an estimated concentration of a chemical constituent in a material;

Figure 2 shows a method for making use of a sensor fusion model in obtaining the estimated level of the substance in the material; Figure 3 shows the measured, estimated and ideal residual spectra corresponding to one sensor;

Figure 4 shows a method for calibrating the sensor fusion model using actual measurements; and

Figure 5 shows a modified method for calibrating the sensor fusion model which based on a physical model.

DETAILED DESCRIPTION

Overview

An overview of a method for non-invasively obtaining an estimated concentration of a chemical constituent in a material is shown in Figure 1. The method is described in terms of two sensors (and consequently two readings). However, any number of sensors may be present. As a rule, a greater number of sensors will tend to provide a more accurate estimate. Each of the sensors should be based on a different measurement principle or measuring position, and preferably measure different characteristics.

At step 110, a first reading is obtained from a first sensor. The first reading relates to a first characteristic of the material.

At step 120, a second reading is obtained from a second sensor. The second reading relates to a second characteristic of the material. The second characteristic is different from the first characteristic.

Preferably the reading takes the form of a spectrum comprising a series of intensities across a plurality of values of the sensor parameter. For example one such sensor may be an optical sensor, where the characteristic being read is the absorbance of near infra-red light, the material and the input parameter is the wavelength of light output by the sensor. Another example is a bioimpedance sensor where the sensor parameter is the frequency used in the impedance measurement and the sensed characteristic is the impedance of the material. In this case the reading is a spectrum comprising the bioimpedance intensities across a range of measured frequencies.

In this embodiment the measured spectrum may consist of 10 intensity readings at 10 distinct values of the sensor parameter. The form of the measured spectrum ω received by the sensor is dependent on the concentration of the chemical constituents within the measured material. Some part of this measured spectrum is due to the particular chemical constituent of interest. The remaining part of the measured spectrum may be due to the response of other constituents of the measured material.

Preferably the sensors are chosen such that the desired substance in the material is particularly responsive to the measured characteristic. That is the sensors are chosen such that the measured spectrum changes significantly with the concentration of the substance of interest in the material and a significant portion of the measured spectrum is due to the chemical constituent to be measured. The sensor may be configured to maximise any such effect. For example, if the desired substance is glucose, it is known that the absorbance of light in blood differs based on the level of glucose in the blood. Because of this, an optical sensor may be selected to measure the characteristic of absorbance of light in the blood. Alternatively, it is known that the impedance measured at a user's skin differs based on the glucose level of their blood. Because of this, a bio-impedance sensor may be selected to measure the characteristic of impedance.

In addition, preferably the first and second sensors (and any additional sensors) are selected such that their respective characteristics are not correlated to each other over the same materials and substances. That is, there should be no known simple equation (such as a linear equation) that would map the variation (with concentration of the substance) in the measured spectrum from the first sensor to the variation in the measured spectrum of the second sensor.

While steps 110 and 120 are described sequentially, they may be performed simultaneously.

At step 130, a first intermediate estimate of the level of the substance in the material is calculated. This is based on the first reading obtained at step 110.

At step 140, a second intermediate estimate of the level of the substance in the material is calculated. This is based on the second reading obtained at step 120.

While steps 130 and 140 are described sequentially, they may be performed simultaneously. Steps 130 and 140 may comprise using a formula, algorithm, look-up table, model or the like in order to calculate an intermediate estimate of the concentration based on the reading from the sensor. Alternatively, a machine learning approach (such as a neural network) may be utilised with a suitable training set to determine a relation between the measured spectrum and the estimated concentration of the chemical constituent within the material. The selection of the appropriate approach will be made easily using the common knowledge in the art of the relationships between the measured characteristic, the substance and the material.

In this embodiment steps 130 and 140 each use a model, calibrated to the spectral response of the corresponding sensor, which takes as an input the measured spectrum and outputs an estimated value of the concentration of the chemical constituent in the material - the "intermediate estimate". This model may be referred to as a calibration model since it is calibrated using a calibration data set to relate the spectral response of the sensor to a concentration of the particular chemical constituent for which it is calibrated.

Preferably, the calibration model is a partial least squares regression model, a principal component regression model or other similar modelling approaches based on latent variables.

Such approaches may proceed by fitting an estimated spectrum to the measured spectrum using one of the approaches listed above. The form of the model may be developed using a separate calibration data set.

The calibration model may be used to obtain an estimate of the chemical constituent in the material (based on the fitted estimated spectrum) and a residual spectrum. The residual spectrum is the difference between the measured and estimated spectrum, which is due to other substances being present in the material which have not been calibrated for - the calibration model being calibrated to model the spectral response of the chemical constituent of interest.

The specific form of the calibration model is discuseed further below.

Finally, at step 150, an estimated level of the substance in the material is calculated. This is based on the first intermediate estimated concentration and the second intermediate estimated concentration. The calculation may preferably also use a residual spectrum where the residual spectrum is the component of the measured spectrum which corresponds to the deviation between the measured spectrum and estimated spectrum.

The final estimated concentration of the chemical constituent within the material is based on a sensor fusion model wherein the model uses the estimates from each sensor and the corresponding residual spectra to output an improved, final estimated concentration value.

Calibration model

A first step in obtaining the estimates of a chemical constituent is to identify a calibration model. The identification of this model is done on a separate calibration data set. As will be explained, this model can be a principal component regression model, a partial least squares regression model or other modeling approaches based on latent variables. The calibration model, is then used for obtaining the estimates of the chemical constituents and the residual spectra.

The general form of such a latent variable model is

W = TP + R where W is the calibration data, T is the scores matrix and P is the loading matrix. The residuals are contained in R. The loading matrix P may be estimated from a the calibration data set using regression. The calibration model may then be utilised, following a sensor reading from a material, to fit an estimated spectrum (corresponding to the estimated concentration of the chemical constituent) to a measured spectrum ω and accordingly output an intermediate estimate of the concentration according to that sensor.

The calibration model is applied to the measured spectrum ω from the material using the form of the matrix P determined from the separate calibration data set. The model is used to fit an estimated spectrum to the measured spectrum and output an intermediate estimate of the concentration associated with the calculated estimated spectrum. As mentioned above, the output of each sensor can be seen to be a measure of the corresponding characteristic of the overall material. However, since the material is likely to be made up of multiple substances (which have not been calibrated for in the model and form of P), the measured spectrum output by the sensor (and consequently the intermediate estimate) is therefore likely to be not immediately useful as a measure of the concentration of the chemical constituent in the material.

The relationship between the measured spectrum, ω, the calculated/estimated spectrum, tP, and the residual spectrum r, may be expressed as ω = tP + r

Here, the estimated spectrum may be expressed as tP, where P is the loading matrix which has orthogonal rows and is estimated from the calibration data set using partial least squares regression or other methods based on latent variables. The scores, t, is estimated as raw responses weighted by the factor loadings P and is defined as the projection coordinates in the multidimensional space.

This relationship between the measured spectrum ω, the estimated spectrum tP (calculated using the calibration model) and the residual spectrum r is best shown with reference to the graph of Figure 3. The measured spectrum ω is shown for a sensor configured according to a parameter. For example this may be the measured absorbance intensities across each of 10 near infra-red wavelengths output by an optical sensor or the measured bioimpedance intensities across a plurality of frequencies. Ideally, all the variation in the spectra should be described by the calibration model. However, because the material itself may respond to the sensor and the material may contain uncalibrated substances (as shown by residual spectrum portion r), the overall measured spectrum ω, combined with the calibration models may not give an accurate estimate of the chemical constituent. In this case, the residual portion of the spectrum is that which corresponds to the unmodelled constituents in the material and the material itself.

The intermediate estimates from each sensor are therefore provided to a sensor fusion model which allows for a more accurate final estimate of the concentration to be found based on the combined output of the sensors.

Sensor Fusion Model

As described above, the individual intermediate estimates from each sensor may not be particularly accurate due to the other components present in the material. Taken alone, it is therefore difficult to estimate the concentration of the chemical constituent with sufficiency accuracy. However, it is possible to incorporate the two measured spectra, ω₁ and ω₂, and an actual measurement of the level of the substance in the material, y, in a sensor fusion model to provide a more accurate estimate, as shown in step 150. The principle here is that since each measured charateritic from the plurality of sensors has a differing relationship with the components present in the material, readings from multiple sensors may be used to more accurately identify the response due to the chemical constituent of interest.

Firstly, as described above with reference to method steps 120 and 130, a calibration model for the relationship between the spectral response and concentration for each sensor is used to calculate a value of the estimated concentration for each sensor. Therefore an estimated value of the concentration y₁ is calculated from the measured spectrum ωι from the first sensor and an estimated value of the concentration y₂ is calculated from the measured spectrum ω₂ from the second sensor. A sensor fusion model may then be used which relates an actual measured value of the chemical concentration y to the estimated values from each sensor.

In preferred embodiments, step 150 comprises making use of one or more sensor fusion models, each of which takes one or more of the intermediate estimates as inputs. These models may be in series, so that the output of a first model can be used as an input to a second model and so on, such that the final model provides the estimated level of the substance in the material as an output. The models may be in parallel, so that the outputs of two or more models are combined, for example by taking an average (mean) value. In some cases, a combination of models in series and parallel may be used.

In addition, the residual spectra r - the difference between the measured and estimated spectra for each sensor - may also be input into the model to provide the final estimate of the chemical concentration in the material. In particular, a correction term which is a function of the residual specttrum, ί( ), may be introduced into the calculation which corresponds to the portion of the intermediate estimate that is caused by other substances in the materials, as described below.

A preferred implementation of step 150 using a model is shown in Figure 2.

At step 210, one or more of the intermediate estimates and the residual spectra are input into the sensor fusion model. At step 220, the model processes the inputs. These may be combined with an internal state of the model. The internal state may be a pre-defined starting state for the model, if the model has not previously received an input. As a result of the processing, the model produces an output which is based on the intermediate estimates.

In a preferred embodiment, step 220 may comprise using a function which models the expected change of the level of the substance in the material over time. This is particularly applicable where the level of the substance changes at a reasonably consistent rate (such as linearly or logarithmically with respect to time). For example, a half-life of a radioactive substance in a substrate or the metabolism of a biological substance in a living being would benefit from such a model.

At step 230, the output from the model is received. This output can be used as the estimated level of the substance in the material.

The model may be a general model which is configured for a wide range of situations and uses. However, in preferred embodiments, the model may be configured to be specific for the particular material or application, in order to give a more accurate result. For example, a general model may map the metabolism of a drug in a nominal average person, whereas a specific model may map the metabolism of a drug in the particular user.

A sensor fusion model relates an actual measured value of the chemical concentration y to the estimated values from each sensor. The desired relationship between these values can be formulated as: y = «i(yi + fi(ii)) + g_x = ₂ (y₂ + f₂(r₂)) + g₂ where:

<¾ and a₂ are scaling factors-these will typically be scalar values, but may in some cases be a function respectively; gi and g₂ are error terms which can relate to a measurement inaccuracy; and

/(ii) and /(r₂) are correction terms which may be a function of the residual spectra r. An exemplary calibration method for the sensor fusion model is shown in Figure 4. This calibration step may be performed before or after each estimate calculation (as in step 150), or may only occasionally be performed.

Furthermore, as noted above, this method is described in relation to two sensors. Of course, it will be appreciated that any number of sensors is suitable. The limit to two sensors in the following example is purely for the simplicity of explanation.

At step 310, one or more actual measurements y of the concentration of a chemical constituent in the material are obtained. For example, in the case of blood glucose, a series of invasive measurements of the user's blood glucose level are taken.

Typically, a greater number of actual measurements will lead to more accurate estimations in use. However, a plurality of actual measurements is not essential: a single actual measurement is sufficient.

At step 320, the scaling factors, <¾ and ₂ , are calculated based on the one or more actual measurements. In some cases, this is done by taking sample readings from each of the sensors in combination with a conventional invasive measurement. The scaling factors can be derived using a recursive least squares regression or a Kalman filter.

It should be noted that step 320 need not be performed before each calibration. In some cases, the scaling factors, <¾ and ₂, may be precalculated for a number of subsequent uses. They may alternatively be predetermined, allowing step 320 to be omitted.

At step 330, the correction terms, ^(Λ) and f₂ (r₂), which are a function of the residual spectra, are calculated. T. The correction terms, ^ΟΊ) and f₂ (r₂), can be derived from the residual spectra, using a recursive least squares regression or a Kalman filter as described in the above section.

The correction terms may be used as part of step 150, such that the estimated concentration is calculated based, in part, on the correction terms.

In some embodiments, the correction terms are proportional to the residual standard deviation of the spectra η. These measurements will typically be derived from a sum of each of the components of the material. However, the contribution of each material to the overall value will differ, depending on the type of measurement used.

Thus, if the contributions are uncorrelated with each other, it is possible for the correction term to be formulated as: fiird = SiTsdiri where: δ_ί is a coefficient; and rsd(ri) is the residual standard deviation of the spectrum r^.

The residual standard deviation provides a good approximation of the error due to non-calibration for response of other substances in the material creating the residual spectrum. The prediction error is proportional to the residual standard deviation (rsd) of the fitted spectrum containing an un-modelled chemical component.

Another exemplary model for calibration of the sensor fusion model

In embodiments making use of one or more models, a similar calibration technique may also be used to provide feedback to each model. However, in use, the output of the model may not be precisely the same as an actual measurement. It is possible to calibrate the model to improve its accuracy by feeding back the correction terms to the model.

In this manner, the output, μ, of a model can be formulated as: μ = α_β9_μ + β where α_μ is a scaling factor, which is preferably 1 ; γ_μ is the estimated level of the substance in the material; and β is a bias represents an inaccuracy in the output, μ, of the model in comparison with the actual measurement, y. In this case γ_μ may be an estimated concentration of the chemical constiuent in the material, as derived from a physical model such as a metabolism model

In such a case, the desired relationship between the actual measurement, y, the intermediate estimates, y₁ and y₂ , and the estimate based on the physical model γ_μ can be formulated as: y = «i(yi + f(ii)) = ₂ (y₂ + f(r₂)) = α_μ + β

An exemplary calibration method for use with this form of the sensor fusion model is shown in Figure 5. This is a modified version of the method of Figure 4.

Where the step number is identical to a step in Figure 4, the same description should be used. This is omitted here for brevity.

At step 325, the scaling factor, α_μ, is calculated. This may be calculated in the same manner as the scaling factors α_;.

At step 335, the bias term β is calculated. This may be calculated in the same manner as the other correction terms ^( ).

At step 340, one or more of the correction terms ^( ) or the bias β are fed back to the model. In this manner, the internal state of the model may be updated based on the calculated correction terms, in order to improve the accuracy of the model in further use.

Calculation of a final estimate

Once the parameters of the sensor fusion models have been determined using one or more conventional invasive measurements as described above, the sensor fusion model may be used during measurement to output an improved estimate by using the determined parameters to adjust the intermediate estimates, taking into account the correction terms based on the residual spectra. Control systems

In some embodiments, the estimated level being outside of a pre-determined acceptable range is considered an alarm condition. As a consequence of this, an alarm or the like may be raised.

In some embodiments, the estimated level is used for automated control of a system, such that the operation of the system is updated periodically based on the estimated level. For example, users with diabetes may be provided with an artificial pancreas configured to administer insulin to the user. The estimated level may be provided to a control system of the artificial pancreas in order to provide for automated (or at least partly automated) insulin delivery.

Device

A device may be provided which is configured to perform the method described above. The device comprises at least two sensors, a processor and a memory. In use, the sensors will be located so that they can monitor a characteristic of the material. The output of the sensors will be passed to the processor which performs the method noted above. The processor may also implement the model, using the memory to store the internal state of the model.

A Blood Monitor

One particularly beneficial application of the present invention is in the field of blood monitoring. Various health-related characteristics of a person can be determined based on the levels of certain substances in their blood.

For example, blood glucose levels are particularly important to people who have diabetes. If a person with diabetes has overly high blood sugar, they may need to urgently have insulin administered to avoid dangerous health consequences.

Other substances which may be useful for monitoring include alanine, ascorbate, lactate, triacetin and urea. In general, a similar approach can be used for each.

In use, a user is provided with two or more non-invasive sensors, generally on or near the user's skin. Each sensor is configured to measure a different characteristic of the user. The sensors are in communication with a monitoring device.

In preferred embodiments, one sensor is an optical sensor which is configured to measure the absorbance of near-infrared and/or visible light. This can be arranged near a light emitting unit which is arranged to emit near-infrared and/or visible light towards the user. This may be positioned on the user's wrist or finger, or another part on the user's body.

In preferred embodiments, one sensor is a bio-impedance sensor, which is configured to measure the impedance at a portion of the user's skin.

Readings are periodically or continuously obtained from the two or more sensors by the monitoring device. Based on these readings, intermediate estimates of the blood glucose level in the user's blood are calculated in with the method of Figure 2 using well-known mappings between blood glucose levels and sensor readings.

In preferred embodiments, these are then passed to a model. The model can be configured to model the changes in blood glucose over time, for example due to ordinary metabolism in the user.

Periodically, the monitoring device may be re-calibrated based on one or more actual measurements of a user's blood glucose. This allows the calibration method of Figures 4 and 5 to be performed. Preferably, this would be done once a day. A longer period would still allow the approach to function, though there is an increasing risk of inaccurate estimates being calculated.

In some embodiments, the monitoring device may be configured to receive information from one or more other sources. This received information may be provided as further inputs to the model. The model can then be updated to take account of the information.

For example, the monitoring device may be configured to receive exercise information related to exercise performed by the user. This may be received from an accelerometer associated with the monitoring device. The exercise information can then be passed to the model. The model may be configured to reflect that a high level of exercise may, for example, cause the blood glucose level to reduce more quickly with respect to time.

The monitoring device may additionally or alternatively be configured to receive food information related to food consumed by the user. This may involve a lookup service, where a textual description of food is entered by a user and is converted to a number of nutrient parameters such as carbohydrates, protein, fats and the like. The model may be configured to reflect that certain nutrients (such as carbohydrates) tend to cause an increase in blood glucose level shortly after consumption.

In some embodiments, the calculated estimated level may be used to control a medical device, for example an insulin delivery system. Based on the estimated level, the insulin delivery system may calculate and/or deliver an appropriate quantity of insulin to the user.

Claims

1. A method for determining a concentration of a chemical constituent in a material, comprising:

obtaining a first measured spectrum from a first non-invasive sensor, the first measured spectrum relating to a first characteristic of the material;

obtaining a second measured spectrum from a second non-invasive sensor, the second measured spectrum relating to a second characteristic of the material;

calculating a first intermediate estimate y_x of the concentration of a chemical constituent in the material based on the first measured spectrum;

calculating a second intermediate estimate y₂ of the concentration of a chemical constituent in the material based on the second measured spectrum; determining a first and second residual spectrum r-ι, r₂, the residual spectra defined as the deviation between the measured spectrum and an estimated spectrum, the estimated spectrum being the component of the measured spectrum corresponding to the intermediate estimate of the concentration of the chemical constituent; and

calculating an improved estimated concentration of the chemical constituent in the material based on the first intermediate estimate y₁ the second intermediate estimate y₂ and the residual spectra r-i , r₂.

2. The method of claim 1 , wherein

calculating a first intermediate estimate y_x of the concentration of a chemical constituent in the material comprises:

providing the first measured spectrum ωι as input to a predetermined first calibration model describing the relationship between the corresponding characteristic and concentration of the chemical constituent; and

receiving an output from the first calibration model as the first intermediate estimate y_x and wherein

calculating a second intermediate estimate y₂ of the concentration of the chemical constituent in the material comprises:

providing the second measured spectrum ω₂ as input to a predetermined second calibration model describing the relationship between the corresponding characteristic and concentration of the chemical constituent; and

receiving an output from the calibration model as the second intermediate estimate y₂ level of the substance in the material.

3. The method of claim 2, wherein the calibrationmodel is configured to model the change in the concentration of the chemical constituent in the material over time.

4. The method of any preceding claim further comprising:

calculating a first correction term, /(ΙΊ), and a second correction term, /(r₂), based on the residual spectra r,, using a sensor fusion model: y = «i(yi + + 9i = «₂ (y₂ + /<¾)) + g₂ where:

y_x is the first intermediate estimate from the first sensor;

y₂ is the second intermediate estimate from the second sensor; x is a first scaling factor;

a₂ is a second scaling factor;

gi is a first error term;

g₂ is a second error term; and

y is an actual measurement of the concentration of the chemical constituent in the material.

5. The method of claim 4, wherein:

the first correction term can be formulated as

and the second correction term can be formulated as

/(r₂) = 5₂rsd(r₂)

where:

δ₁ is a first coefficient;

δ₂ is a second coefficient; and

ΐΊ is a first residual spectrum;

r₂ is a second residual spectrum; and

rsd is the residual standard deviation of a residual spectrum.

6. The method of claim 4 or 5, further comprising:

calculating a bias, β, such that:

y = a₃y₃ + β

where:

y₃ is the estimated concentration of the chemical constituent in the material according to a physical model; and

a₃ is a third scaling factor.

7. The method of claim 6, further comprising:

updating the sensor fusion model based on one or more of the first correction term, the second correction term and the bias.

8. The method of claim 6 or 7, wherein one or more of the first scaling factor, the second scaling factor, the third scaling factor and the actual measurement of the concentration of the chemical constituent in the material are calculated by means of measurement of the concentration of the chemical constituent in the material.

9. The method of any of claims 6 to 8, wherein one or more of the first correction term, the second correction term and the bias term are calculated by means of a recursive least squares regression or a Kalman filter.

10. The method of any of the preceding claims, wherein the chemical constituent is glucose and the material is the blood of a user.

11. The method of claim 10, wherein the first non-invasive sensor comprises an optical sensor configured to measure absorbance of light and the second sensor non-invasive comprises a bio-impedance sensor configured to measure the impedance of skin of the user.

12. The method of claim 11 , wherein the absorbance of light is measured across a series of wavelengths in the near infra-red region and the bio- impedance is measured across a series of frequencies.

13. The method of any of claims 10 to 12, further comprising:

receiving a third reading representing a quantity of food ingested by the user; and

receiving a fourth reading representing a quantity of exercise performed by the user;

wherein calculating an estimated concentration of the chemical constituent in the material further comprises:

inputting the third reading and the fourth reading into the model;

wherein the output of the model is further based on the third reading and the fourth reading.

14. The method of any of claims 10 to 13, wherein the model is adapted to model the metabolism of the user.

15. A device for non-invasively measuring the concentration of a chemical constituent in the material configured to perform the method of any preceding claim.