RETROSPECTIVE RETROFITTING METHOD TO GENERATE A CONTINUOUS GLUCOSE CONCENTRATION PROFILE BY EXPLOITING CONTINUOUS GLUCOSE MONITORING SENSOR DATA AND BLOOD GLUCOSE MEASUREMENTS

Abstract

Continuous Glucose Monitoring (CGM) devices provide glucose concentration measurements in the subcutaneous tissue with limited accuracy and precision. Therefore, CGM readings cannot be incorporated in a straightforward manner in outcome metrics of clinical trials e.g. aimed to assess new glycaemic-regulation therapies. To define those outcome metrics, frequent Blood Glucose (BG) reference measurements are still needed, with consequent relevant difficulties in outpatient settings. Here we propose a “retrofitting” algorithm that produces a quasi continuous time BG profile by simultaneously exploiting the high accuracy of available BG references (possibly very sparsely collected) and the high temporal resolution of CGM data (usually noisy and affected by significant bias). The inputs of the algorithm are: a CGM time series; some reference BG measurements; a model of blood to interstitial glucose kinetics; and a model of the deterioration in time of sensor accuracy, together with (if available) a priori information (e.g. probabilistic distribution) on the parameters of the model. The algorithm first checks for the presence of possible artifacts or outliers on both CGM datastream and BG references, and then rescales the CGM time series by exploiting a retrospective calibration approach based on a regularized deconvolution method subject to the constraint of returning a profile laying within the confidence interval of the reference BG measurements. As output, the retrofitting algorithm produces an improved “retrofitted” quasi-continuous glucose concentration signal that is better (in terms of both accuracy and precision) than the CGM trace originally measured by the sensor. In clinical trials, the so-obtained retrofitted traces can be used to calculate solid outcome measures, avoiding the need of increasing the data collection burden at the patient level.

Claims

1. A glucose monitoring system for monitoring a glucose level in a user, comprising: (a) a continuous glucose monitoring (CGM) device configured to monitor glucose levels in the user’s interstitial fluids and to generate a CGM time series representative thereof; (b) a blood glucose (BG) data source configured to generate BG references representative of the user’s blood glucose levels at discrete time intervals; (c) a data preprocessing module, responsive to the CGM time series and to the blood glucose data, configured to detect outliers and artifacts in both the CGM time series and the BG references and configured to generate a preprocessed CGM signal corresponding to the CGM time series from which any outliers and artifacts are discarded and to generate a preprocessed BG signal corresponding to the BG references from which any outliers and artifacts are discarded; (d) a retrospective CGM calibration module, responsive to the preprocessed CGM signal and the preprocessed BG signal, configured to perform a retrospective calibration of the preprocessed CGM signal to compensate for systematic underestimation and overestimation of CGM time series with respect to reference BG values due to: blood-to- interstitial glucose kinetics, sensor drift, errors in CGM sensor calibration, and changes in sensor sensitivity, the retrospective CGM calibration module also configured to generate a retrospectively calibrated CGM signal representative thereof by rescaling the calibrated CGM signal so as to stay within a confidence interval of the BG values; and (e) a constrained inverse problem solver module, responsive to the retrospectively calibrated CGM signal, configured deconvolute the retrospectively calibrated CGM signal based on a model of blood-to-interstitial glucose kinetics, and configured to generate a retrofitted glucose concentration profile with a predetermined confidence interval.

2. The glucose monitoring system of claim 1, wherein the data preprocessing module is also responsive to other selected relevant inputs.

3. The glucose monitoring system of claim 2, wherein the other selected relevant inputs are selected from a group consisting of: data regarding meals consumed, data regarding treatment taken for hypoglycemia, data regarding drugs taken, data regarding physical activity, and data regarding stress.

4. The glucose monitoring system of claim 1, wherein the retrospective CGM calibration module is also responsive to other selected relevant inputs.

5. The glucose monitoring system of claim 4, wherein the other selected relevant inputs are selected from a group consisting of: data regarding meals consumed, data regarding treatment taken for hypoglycemia, data regarding drugs taken, data regarding physical activity, and data regarding stress.

6. The glucose monitoring system of claim 1, wherein the retrospective CGM calibration module is also responsive to: a CGM drift/degradation model and a priori information on model parameters.

7. The glucose monitoring system of claim 1, wherein the retrospective CGM calibration module is also responsive to a blood-to-interstitial glucose kinetics model.

8. The glucose monitoring system of claim 1, wherein the constrained inverse problem solver module is also responsive to other selected relevant inputs.

9. The glucose monitoring system of claim 8, wherein the other selected relevant inputs are selected from a group consisting of: data regarding meals consumed, data regarding treatment taken for hypoglycemia, data regarding drugs taken, data regarding physical activity, and data regarding stress.

10. The glucose monitoring system of claim 1, wherein the constrained inverse problem solver module is also responsive to a blood-to-interstitial glucose kinetics model.

11. A method for monitoring a glucose level in a user, comprising the steps of: (a) continuously monitoring glucose levels in the user’s interstitial fluids and to generating a continuous glucose monitoring (CGM) time series representative thereof; (b) generating blood glucose (BG) references representative of the user’s blood glucose levels at discrete time intervals; (c) detecting outliers and artifacts in both the CGM time series and the BG references and generating a preprocessed CGM signal corresponding to the CGM time series from which any outliers and artifacts are discarded and generating a preprocessed BG signal corresponding to the BG references from which any outliers and artifacts are discarded; (d) performing a retrospective calibration of the preprocessed CGM signal, employing the preprocessed BG signal, thereby compensating for systematic underestimation and overestimation of CGM time series with respect to reference BG values due to: blood-to- interstitial glucose kinetics, sensor drift, errors in CGM sensor calibration, and changes in sensor sensitivity, the retrospective CGM calibration module and generating a retrospectively calibrated CGM signal representative thereof by rescaling the calibrated CGM signal so as to stay within a confidence interval of the BG values; and (e) deconvoluting the retrospectively calibrated CGM signal based on a model of blood-to- interstitial glucose kinetics, and thereby generating a retrofitted glucose concentration profile with a predetermined confidence interval.

12. The method for monitoring a glucose level in a user of claim 11, further comprising the step of employing other selected relevant inputs in the step of detecting outliers and artifacts.

13. The method for monitoring a glucose level in a user of claim 12, wherein the other selected relevant inputs are selected from a group consisting of: data regarding meals consumed, data regarding treatment taken for hypoglycemia, data regarding drugs taken, data regarding physical activity, and data regarding stress.

14. The method for monitoring a glucose level in a user of claim 11, further comprising the step of employing other selected relevant inputs in the step of performing a retrospective calibration.

15. The method for monitoring a glucose level in a user of claim 14, wherein the other selected relevant inputs are selected from a group consisting of: data regarding meals consumed, data regarding treatment taken for hypoglycemia, data regarding drugs taken, data regarding physical activity, and data regarding stress.

16. The method for monitoring a glucose level in a user of claim 11, further comprising the step of employing a CGM drift/degradation model and a priori information on model parameters in performing the step of performing a retrospective calibration.

17. The method for monitoring a glucose level in a user of claim 11, further comprising the step of employing a blood-to-interstitial glucose kinetics model in performing the step of performing a retrospective calibration.

18. The method for monitoring a glucose level in a user of claim 11, further comprising the step of employing other selected relevant inputs in the step of deconvoluting the retrospectively calibrated CGM signal.

19. The method for monitoring a glucose level in a user of claim 18, wherein the other selected relevant inputs are selected from a group consisting of: data regarding meals consumed, data regarding treatment taken for hypoglycemia, data regarding drugs taken, data regarding physical activity, and data regarding stress.

20. The method for monitoring a glucose level in a user of claim 11, further comprising the step of employing a blood-to-interstitial glucose kinetics model in the performing the step of deconvoluting the retrospectively calibrated CGM signal.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0018] FIG. 1 is a graph showing SMBG measurements versus continuous-time glycemia.

[0019] FIG. 2 is a graph showing a comparison of CGM time series (blue dots) vs SMBG.

[0020] FIG. 3 is a block scheme representing how the retrofitting algorithm works.

[0021] FIG. 4 includes two graphs showing examples of BG and CGM outliers or unreliable values.

[0022] FIG. 5 is a block diagram depicting the retrospective CGM calibration procedure.

[0023] FIG. 6 is a graph showing an example of retrospectively calibrated CGM versus originally measured CGM time series.

[0024] FIG. 7 is a graph showing retrofitted glucose concentration time series versus originally measured CGM time series.

[0025] FIG. 8 includes two graphs showing results for a first patient.

[0026] FIG. 9 includes two graphs showing results for a second patient.

[0027] To facilitate understanding, identical reference numerals have been used, where possible, to designate identical elements that are common to the figures. It is contemplated that elements and features of one embodiment may be beneficially incorporated in other embodiments without further recitation.

DETAILED DESCRIPTION

[0028] A preferred embodiment of the invention is now described in detail. Referring to the drawings, like numbers indicate like parts throughout the views. Unless otherwise specifically indicated in the disclosure that follows, the drawings are not necessarily drawn to scale. As used in the description herein and throughout the claims, the following terms take the meanings explicitly associated herein, unless the context clearly dictates otherwise: the meaning of “a,” “an,” and “the” includes plural reference, the meaning of “in” includes “in” and “on.”

[0029] FIG. 3 is a block scheme representing how the retrofitting algorithm works. On the right side, signals and models giving an input to the retrofitting algorithm as shown. In the middle section the core of the retrofitting procedure comprising three main sub-blocks is depicted: A data preprocessing and outlier detection, B. retrospective CGM calibration, and C. constrained inverse problem solver. On the left side these signals returned in output all shown. Starting from the schematization in sub-blocks of the retrofitting procedure depicted in FIG. 3, we present a possible and preferred embodiment.

Sub-Block A Data Preprocessing and Outlier Detection

[0030] The aim of this sub-block (sub-block A in FIG. 3) is to detect unreliable data and outliers. The block receives in input CGM and BG time series that will be used to create the retrofitting profile. The block contains an algorithm that is employed to detect outliers and/or unreliable values on both BG and CGM time series. An example of outlier is a single sample that is not consistent with the trends of the time series. An example of unreliable values is two, or more, consecutive values whose distance in time is not able to explain, from a physiological or technological point of view, their distance in glucose space. Outliers and anomalous data are identified and excluded from the successive steps of analysis. The outputs are: discarded BG, discarded CGM, preprocessed BG, and preprocessed CGM data.

[0031] A possible algorithm for this sub-block should calculate the first-order time derivative of the time series, using finite differences or any statistically-based estimation procedure (e.g. a Bayesian smoothing procedure). The time series of the first-order differences is inspected for the presence of outliers and unreliable values following a given criterion. For example, a possible criterion for the detection of outliers checks any couple of two consecutive derivatives. If both amplitudes are greater than a given threshold X and have opposite sign, then the value in the middle is labeled as outlier. As far as the detection of unreliable values is concerned, a possible criterion checks every couple of two consecutive values. Every couple of values that are taken Y minutes apart each other and whose distance in the glucose space is greater than Z mg/dl, where Z=g(Y), are labelled as unreliable. FIG. 4 shows examples of BG (triangles) and CGM (dots) outliers or unreliable values that are eliminated in this step. In the left diagram two BG unreliable values are circled. In the right diagram two CGM outliers are circled.

[0032] The detection of outliers or unreliable data can be improved also exploiting other relevant inputs that could help in describing fluctuations in glucose dynamics (e.g. meals quantity and scheduling, hypo treatments, drugs, physical activity and stress information, etc.) opportunely modeled by other relevant models (e.g. models for meal absorption, insulin action, etc.), when available.

Sub-Block B Retrospective CGM Calibration

[0033] The sub-block (sub-block B in FIG. 3) is aimed to compensate for systematic under/overestimation of CGM time series with respect to reference BG values due to e.g. drift in time, errors in CGM sensor calibration, and changes in sensor sensitivity. This is done by performing a retrospective calibration of the CGM time series. From the sub-block A previously described, the sub- block B receives in input the preprocessed BG and preprocessed CGM data and returns in output a retrospectively calibrated CGM profile which is more accurate (e.g. closer to the reference BG data) than the measured CGM one, since errors due to drifts, suboptimal calibration, etc. have been mitigated.

[0034] A possible algorithm to perform the retrospective calibration divides the CGM time series history into several intervals using the times of calibration events as separators. The history of the available BG references is divided in the same temporal intervals. Under the assumption that calibration parameters remain the same between two consecutive calibrations and that the performance of the sensor degrades in time due to changes in the sensitivity of the sensor, each portion of CGM data is calibrated against all the BG references falling in the same temporal interval. The calibration rule is a regression law f that receives in input the preprocessed CGM time series and returns in output the retrospectively calibrated CGM time series. For example, f = f (a,b,c) where a is a gain parameter, b is an offset, and c is a parameter that takes into account the temporal trend of the data. The parameters of the regression law f are estimated from available data exploiting a model of the blood-to-interstitial fluid glucose transportation dynamics and a model for sensor drift/degradation. The estimation procedure could be able to exploit, if available, a priori information (e.g. probabilistic distribution) on the parameters of the model (e.g. Bayesian estimation). A schematization of the algorithm of sub-block B is reported in FIG. 5. An example of the retrospectively calibrated CGM time series given in output in this step is shown in FIG. 6, in which the retrospectively calibrated CGM is apparently closer (in terms of absolute error) to the reference BG measurements than the originally measured CGM. FIG. 6 shows the same data as in FIG. 1 and FIG. 2, and an example of retrospectively calibrated CGM (dots) vs originally measured CGM (squares) time series is depicted.

[0035] The retrospective CGM calibration can be improved also exploiting other relevant inputs that could help in describing fluctuations in glucose dynamics (e.g. meals quantity and scheduling, hypo treatments, drugs, physical activity and stress information, etc.) opportunely modeled by other relevant models (e.g. models for meal absorption, insulin action, etc.), when available. If available, a priori information (e.g. probabilistic distribution) on the parameters of the model can be provided.

Sub-Block C Constrained Inverse Problem Solver

[0036] The third sub-block of the retrofitting procedure (sub-block C in FIG. 3, middle) performs constrained deconvolution. The inputs of this sub-block are the retrospectively calibrated CGM data given in output by sub-block B, all the BG reference measurements that have not been labeled as outliers or unreliable values together with their confidence interval (e.g., the coefficient of variation or the standard deviation), given in output by sub-block A, and a model of blood-to-interstitium glucose kinetics, e.g. the two-compartment model presented by Rebrin et al. (Am J Physiol 1999). If available, a priori information (e.g. probabilistic distribution) on the parameters of the model can be provided. The block contains an algorithm that performs the constrained deconvolution of the retrospectively calibrated CGM profile, and allows estimating, at (quasi) continuous time the profile of glucose concentration in blood that should have generated that CGM profile. Deconvolution can be based on the regularization approach, see De Nicolao et al. (Automatica 1997). From a mathematical point of view, the deconvolution algorithm finds the solution to the following problem:

[00001]$\begin{array}{l}\widehat{\text{u}}\text{=}\underset{\begin{array}{c}\text{u}\\ \text{BG}-\text{toll<u<BG+toll}\end{array}}{argmin}\left({\left({\text{CGM}}_{\text{retrocalibrated}}-\text{Gu}\right)}^T\right)\\ \left({\displaystyle {.Math.}_{\text{v}}^{-1}\left({\text{CGM}}_{\text{retrocalibrated}}-\text{Gu}\right)+\gamma \cdot {\text{u}}^{\text{T}}\left({\text{F}}^{\text{T}}\text{F}\right)\text{u}}\right)\end{array}$

where CGM.sub.retrocalibrated is a Nxl vector containing samples of the retrospectively calibrated CGM profile produced by sub-block B, G is the NxN matrix obtained by discretizing the blood-to- interstitial fluid glucose kinetics model,

[00002]${\text{Σ}}_{\text{v}}^{-1}$

is the inverse of the CGM measurement error covariance NxN matrix, F is a NxN Toeplitz lower triangular matrix that acts as a discrete differentiator, BG is a Nxl vector containing all the BG values received in input, toll is the confidence interval on them, and û is the vector containing the samples of the retrofitted (quasi) continuous glucose concentration time series. In equation (1) it is clear that, for every BG reference value available, the retrofitted glucose concentration profile û should pass close to it, where close is defined by the confidence interval (toll) received in input (constraint). This step allows: eliminating delays/distortions due to glucose transportation from blood to the interstitial fluid by using regularized deconvolution; improving the estimate of the BG signal taking advantage of the BG reference measurements (thanks to the constraints, the retrofitted glucose profile lies within the confidence interval of the available BG references); exploiting a physiological prior on the smoothness of the BG profile to increase precision (i.e. reducing uncertainty due to measurement noise). An example of the output of the constrained deconvolution step, i.e. the retrofitting (quasi) continuous glucose time series, is showed in FIG. 7.

[0037] The constrained deconvolution can be improved also exploiting other relevant inputs that could help in describing fluctuations in glucose dynamics (e.g. meals quantity and scheduling, hypo treatments, drugs, physical activity and stress information, etc.) opportunely modeled by other relevant models (e.g. models for meal absorption, insulin action, etc.), when available. If available, a priori information (e.g. probabilistic distribution) on the parameters of the model can be provided.

Assessment of the Invention

[0038] The retrofitting procedure has been validated on 43 datasets of type 1 diabetic patients. For each patient at least 72 hours of CGM monitoring and frequent BG references were available. BG references have been divided into training-set references, used by the retrofitting algorithm, and test-set references, not used by the retrofitting algorithm and exploited to assess the accuracy of the retrofitted continuous glucose concentration profile.

[0039] FIG. 7 shows the same subject as FIGS. 1, 2, and 6. Retrofitted glucose concentration time series is shown by a line vs originally measured CGM time series shown in dots. FIGS. 8 and 9 (top) show the outcomes of the retrofitting technique in two representative subjects of the dataset (#1 and #3 respectively). CGM trace is depicted in dots, triangles are the BG measurements assigned to the training-set and therefore provided to the retrofitting method. The output of the retrofitting method is the continuous line (the grey area is its confidence interval) and has to be compared with the test BG references (diamonds, depicted with their confidence intervals). By eye inspection, it is evident the improvement in accuracy with respect to CGM or a simple linear interpolation of BG references (dashed line), and also the improvement of smoothness with respect to CGM.

[0040] FIGS. 8 and 9 (bottom) show, for each of the representative patients, a boxplot of the relative errors on test-set. The profile obtained by the retrofitting algorithm is compared with the original CGM values and with linear interpolation of the training-set references. Nominal finger-stick accuracy is also reported for comparison. For both patients, the relative error is significantly reduced with respect to CGM and linear interpolation, resulting in about a 3-fold reduction of the MARD (mean absolute relative deviation), now comparable with that of finger-stick measurements, and a similar reduction of both the 90th and 75th percentiles of error distributions.

[0041] The above described embodiments, while including the preferred embodiment and the best mode of the invention known to the inventor at the time of filing, are given as illustrative examples only. It will be readily appreciated that many deviations may be made from the specific embodiments disclosed in this specification without departing from the spirit and scope of the invention. Accordingly, the scope of the invention is to be determined by the claims below rather than being limited to the specifically described embodiments above.