A Bayesian framework for personalized model identification and prediction of future blood glucose levels in type 1 diabetes using readily available patient data.

By using an individualized nonlinear glucose-insulin kinetic physiological model and Bayesian estimation techniques, combined with particle filtering for real-time prediction, the problem of accurate prediction of future blood glucose concentration in patients with type 1 diabetes has been solved, achieving more efficient blood glucose management.

CN116635094BActive Publication Date: 2026-11-17DEXCOM INC
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202180064934.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-01-25
Filing Date
2021-11-12
Publication Date
2026-11-17
Estimated Expiration
2041-11-12

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively predict future blood glucose levels in patients with type 1 diabetes, especially when faced with patient-to-patient variability and confounding factors. Traditional models fail to provide accurate blood glucose predictions, leading to poor treatment outcomes.

Method used

We employ an individualized nonlinear glucose-insulin kinetic physiological model, combined with Markov chain Monte Carlo Bayes estimation and particle filtering techniques, to update blood glucose concentration predictions in real time through personalized model parameter estimation and residual model enhancement.

Benefits of technology

It improves the accuracy and timeliness of future blood glucose concentration prediction, significantly outperforming traditional black-box models, especially in terms of time gain.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116635094B_ABST
    Figure CN116635094B_ABST
Patent Text Reader

Abstract

A method for predicting future blood glucose concentration in an individual patient, the method comprising: selecting an individualized nonlinear glucose-insulin kinetic physiological model, the selected model having multiple model parameters whose values ​​are to be determined; estimating the value of each of the multiple model parameters by applying parameter estimation techniques to prior information and data of the individual patient to obtain posterior information, a first subset of the model parameters having values ​​estimated according to prior population data, and a second subset of the model parameters having values ​​personalized for the individual patient; and applying a nonlinear prediction technique to the selected model using the estimated value of each of the model parameters to obtain a predicted blood glucose concentration for the individual patient at a future time.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Cross-references to related applications

[0002] This application claims the benefit of U.S. Provisional Patent Application No. 63 / 141345, filed January 25, 2021. This application also relates to U.S. Applications Nos. 17 / 112870, 17 / 112828, and 17 / 112814, all filed December 4, 2020. The contents of these applications are incorporated herein by reference in their entirety and are therefore expressly formed part of this specification. Background Technology

[0003] Type 1 diabetes (T1D) is a chronic autoimmune disease caused by the progressive destruction of beta cells in the pancreas, preventing the body from producing endogenous insulin. As a result, blood glucose (BG) levels often exceed the hyperglycemic threshold (BG > 180 mg / dL). If this occurs frequently and persists, it can lead to several serious long-term cardiovascular complications, as well as kidney disease and neuropathy. To lower blood glucose (BG) levels, exogenous insulin must be administered several times daily. Unfortunately, excessive exogenous insulin administration can cause hypoglycemia, i.e., BG < 70 mg / dL. Hypoglycemia can cause fainting, dizziness, coma, and even death, making it dangerous even in the short term.

[0004] Effective treatment for type 1 diabetes (T1D) relies on frequent blood glucose (BG) monitoring via traditional finger-prick devices or more modern, minimally invasive continuous glucose monitoring (CGM) sensors, and this is not always easy to achieve. In fact, from the patient's perspective, T1D management is a lifelong learning process, learning how several daily factors (e.g., illness, diet, and physical activity) affect BG levels and how to use interventions (e.g., reducing carbohydrate intake and, of course, insulin administration) to keep BG within a safe range.

[0005] In this regard, the research community has made many efforts to provide new tools that can help patients with T1D. Among them, CGM-based algorithms that can predict future BG concentrations in real time by making active treatment decisions based on expected future glucose levels rather than current glucose levels have the potential to significantly improve the treatment outcomes of T1D. Summary of the Invention

[0006] In a first aspect, a method for predicting future blood glucose concentrations in an individual patient is provided, the method comprising: selecting an individualized nonlinear glucose-insulin kinetic physiological model, the selected model having multiple model parameters whose values ​​are to be determined; estimating the value of each of the multiple model parameters by applying parameter estimation techniques to prior information and data of the individual patient to obtain posterior information, a first subset of the model parameters having values ​​estimated according to prior population data, and a second subset of the model parameters having values ​​personalized for the individual patient; and applying a nonlinear prediction technique to the selected model using the estimated value of each of the model parameters to obtain a predicted blood glucose concentration for the individual patient at a future time.

[0007] In the implementation of the first aspect, the parameter estimation technique is Markov Chain Monte Carlo (MCMC) Bayesian estimator.

[0008] In the implementation of the first aspect, the parameter estimation technique is selected from the group consisting of the maximum a posteriori method, the maximum likelihood method, and the prediction error minimization method.

[0009] In the implementation of the first aspect, the nonlinear prediction technique is particle filtering.

[0010] In the implementation of the first aspect, the nonlinear prediction technique is selected from the group consisting of extended Kalman filtering and unscented Kalman filtering.

[0011] In the implementation of the first aspect, the selected individualized nonlinear physiological models include the subcutaneous insulin absorber model, the oral glucose absorber model, and the glucose-insulin kinetic model.

[0012] In the implementation of the first aspect, the selected individualized nonlinear physiological model uses the individual patient's past blood glucose concentration levels, carbohydrate intake information, and exogenous insulin data as inputs.

[0013] In an implementation of the first aspect, the subcutaneous insulin absorption submodel includes a first model parameter and a second model parameter, the first model parameter specifying a value for insulin distribution, and the second model parameter specifying a delay in the appearance of insulin in the first compartment, the first and second model parameters being located in a first subset of model parameters having values ​​estimated according to prior population data.

[0014] In an implementation of the first aspect, the subcutaneous insulin absorption submodel includes a third model parameter and a fourth model parameter, the third model parameter specifying a diffusion rate constant from the first compartment to the second compartment, and the fourth model parameter specifying a subcutaneous insulin absorption rate constant from the second compartment to the plasma, the third and fourth model parameters being located in a second subset of model parameters having values ​​personalized for individual patients.

[0015] In one implementation of the first aspect, the selected individualized nonlinear physiological model is a maximal physiological model, in which the parameters for selecting model parameters are ignored.

[0016] In the implementation of the first aspect, the predicted blood glucose concentration is updated in real time using data from a continuous glucose monitor (CGM), carbohydrate intake information, and exogenous insulin data.

[0017] In an implementation of the first aspect, the method further includes enhancing the selected individualized nonlinear physiological model with a residual model to describe the residual modeling error.

[0018] In the implementation of the first aspect, the residual modeling error is modeled as an autoregressive composite moving average (ARIMA) of order (5,5,1).

[0019] In an implementation of the first aspect, the residual model has model parameters estimated using the prediction error method (PEM).

[0020] In the implementation of the first aspect, the predicted blood glucose concentration is predicted at multiple different times. Attached Figure Description

[0021] Figure 1 This is a high-level flowchart illustrating the steps involved in the individualized model-based prediction scheme described in this paper.

[0022] Figure 2 An example of a subcutaneous insulin absorption subsystem protocol is shown.

[0023] Figure 3 An example of an oral glucose absorption subsystem protocol is shown.

[0024] Figure 4 An example of a glucose-insulin kinetics subsystem scheme is shown.

[0025] Figure 5 An example of a residual model scheme is shown.

[0026] Figure 6A The coefficients of determination (COD) obtained using ARX, ARIMAX, and PF at different pH values ​​from 15 min to 120 min are shown; and Figure 6BThe root mean square error (RMSE) obtained using ARX, ARIMAX, and PF is shown for different pH values ​​from 15 min to 120 min. Detailed Implementation

[0027] I. Introduction

[0028] Despite the several different methodologies proposed for real-time blood glucose (BG) prediction, the problem remains unresolved due to several unknown confounding factors that alter BG levels and the high patient-to-patient / intra-patient variability in glucose physiology. Existing prediction techniques rely on a mathematical model describing the relationship between a set of input features and BG. Choosing a suitable model for glucose-insulin regulation is a crucial step, and the options range from fully data-driven "black box" models to models based on mechanistic / semi-mechanistic nonlinear descriptions of metabolic physiology. Compared to black box models, physiological models have the clear advantage of accurately capturing the underlying glucose-insulin metabolic mechanisms. Among the nonlinear physiological models available in the T1D literature, the so-called minimal models have proven to be too rigid and simplistic to provide effective glucose predictions. On the other hand, large models have several equations and numerous parameters, making them difficult to identify and more suitable for computer simulations than model-based predictions.

[0029] Figure 1 This is a flowchart illustrating the steps involved in the personalized model-based prediction scheme described herein. As shown, the first step 10 in the scheme is to select a suitable model structure from the various available options. The selected model is a personalized nonlinear physiological model of glucose-insulin regulation. As described below, the selected model has model parameters that can be personalized based on readily available patient records such as carbohydrate intake and insulin infusion data to simulate the specific physiological functions of an individual.

[0030] exist Figure 1 In the second step 20 of the illustrated prediction scheme, parameters selected from the model parameters to be personalized for individual patients are estimated using parameter estimation techniques. In one implementation, the parameter estimation technique used employs a Bayesian method implemented by Markov Chain Monte Carlo (MCMC) to estimate a large number of parameters in the presence of complex nonlinear dynamics. Next, in the third step 30, the obtained personalized model is used in a nonlinear prediction scheme based on, for example, particle filtering methodology.

[0031] As will be described in further detail below, a computer-simulated clinical trial was conducted to test the predictive protocol described herein, using data from 100 virtual T1D patients generated by simulation models described in the following literature: R. Visentin, E. Campos- M. Schiavon, D. Lv, M. Vetteretti, M. Breton, B. B. Kovatchev, C. Dalla Man and C. Cobelli, “The uva / padova type 1 diabetes simulator goes from single meal to single day,” J. Diabetes Sci. Technol., Vol. 12, No. 2, pp. 273–281, 2018, the full text of which is incorporated herein by reference. For comparison, two baseline black-box models were considered. The metrics chosen for evaluation were root mean square error, coefficient of determination, and time gain. The results show that the proposed method outperforms both baseline methodologies in predicting future glucose concentrations, particularly in terms of time gain.

[0032] II. Physiological Model of Glucose-Insulin Regulation

[0033] This section describes an example of a nonlinear glucose-insulin dynamic physiological model for personalized glucose prediction. The example begins with a maximum physiological model implemented in a recent version of the UVa / Padova T1D simulator (T1DS), as described in the references cited above by Visentin et al., and simplifies this maximum physiological model as much as possible to reduce the number of parameters to be estimated for individualization while maintaining its ability to achieve good glucose predictions.

[0034] The proposed model has two inputs: insulin infusion I(t) and carbohydrate intake CHO(t), and one output: interstitial glucose concentration IG(t). The model consists of three subsystems describing subcutaneous insulin absorption, oral glucose absorption, and glucose-insulin kinetics.

[0035] A. Subcutaneous insulin absorption subsystem

[0036] The subcutaneous insulin absorption model is a slightly simplified version of the absorption model used in the T1DS model described in the following literature: M. Schiavon, C. Dalla Man and C. Cobelli, “Modeling subcutaneous absorption of fast-acting insulin in type 1 diabetes,” IEEE Trans. Biomed. Eng., Vol. 65, No. 9, pp. 2079–2086, 2018. Figure 2As shown in the diagram. The model consists of three compartments. Exogenous insulin I is infused into the first compartment, where it appears after a delayed β-wave. In the first compartment, where insulin is in a non-monomeric state, it is converted to a monomeric state and then diffuses into the plasma. The model equation is:

[0037]

[0038] Among them, I sc1 (mU / kg) and I sc2 (mU / kg) represent insulin in non-monomer and monomeric states, respectively; I p (mU / l) is the plasma insulin concentration; k d (min -1 ) is the diffusion rate constant from the first chamber to the second chamber; k a2 (min -1 ) is the rate constant of subcutaneous insulin absorption from the second compartment to the plasma; k e (min -1 ) is the partial clearance rate; V I (l / kg) is the insulin distribution volume; β(min) is the delay in insulin appearance in the first compartment. This simplification, compared to the model used in T1DS, ignores the fact that a portion of non-monomer insulin can directly reach the plasma. This absorption pathway... Figure 1 It is shown in gray and is related to the rate constant k. a1 The correlation is due to the small number of insulin scores and patients experiencing this condition. Prior information regarding the model parameters reported in Table 1 has been obtained from the literature. Specifically, V... I β and k were set as population values, namely 0.126 l / kg and 8 min, respectively. Furthermore, k... d Restricted to k d ≥k a2 This is because these two combinations are interchangeable. The unknown model parameter is θ. ins =[k a2 k d ].

[0039] B. Oral glucose absorption subsystem

[0040] The oral glucose absorption subsystem model from the following literature is a simplified version of the model used in T1DS: C. Dalla Man, M. Camilleri and C. Cobelli, “A system model of oral glucose absorption: Validation on gold standard data,” IEEE Trans. Biomed. Eng., Vol. 53, No. 12, pp. 2472–2478, 2006, and this model is used in… Figure 3 As shown in the diagram. This model describes the gastrointestinal tract as a three-compartment system: the first two compartments are responsible for food (solid and ground) in the stomach, while the third compartment models the upper part of the small intestine that has absorbed CHO. The model equation is:

[0041]

[0042] Among them, Q sto1 (mg / kg) and Q sto2 (mg / kg) represent the amounts of solid and liquid glucose in the stomach, respectively; Q gut (mg / kg) is the glucose concentration in the intestine; k gri (min- 1 ) is the grinding rate constant; k empt (min- 1 ) is the gastric emptying rate constant; k abs (min- 1 ) is the intestinal absorption rate constant; CHO (mg / kg / min) is the carbohydrate intake rate. The rate of glucose appearance in plasma, Ra (mg / kg / min), is estimated based on model (2) as follows:

[0043] Ra(t) = fk abs ·Q gut (t) (3)

[0044] Where f (dimensionless) is the fraction of intestinal contents absorbed into the blood plasma.

[0045] Compared to the model used in T1DS, the simplification lies in assuming a constant gastric emptying rate, thus ignoring its dependence on gastric contents (in Figure 3 (shown in gray).

[0046] Prior information for model (2) was obtained from the literature and is detailed in Table 1. Specifically, f is set to 0.9 and k is constrained. gri =k empt Furthermore, k abs Limited to k abs ≤k emptThis is because these two combinations are interchangeable. Therefore, the unknown model parameters are θo. ral =[k abs ,k empt ].

[0047] C. Glucose-insulin kinetic subsystem

[0048] Regarding the model used in the T1DS simulator, we use a more concise description: the glucose-insulin kinetics subsystem. This subsystem is based on a known two-compartment model that describes the effect of plasma insulin action and glucose occurrence rate on plasma glucose concentration, as described in: R.N. Bergman, C.R. Bowden, and C. Cobelli, “Quantitative estimation of insulin sensitivity,” Am. J. Physiol., Vol. 236, No. 6, pp. e667–e677, 1979. The model is further equipped with a third compartment to describe glucose transport from plasma to the interstitium, where glucose is measured by sensors. This model in… Figure 4 As shown in the figure. The model equation is:

[0049]

[0050] Wherein, G (mg / dl) is the plasma glucose concentration, and X (min) is the plasma glucose concentration. -1 ) refers to the role of insulin in glucose processing and production; SG (min -1 Glucose efficacy describes the ability of glucose itself to promote glucose disposal and inhibit glucose production; G b (mg / dl) is the basal glucose concentration in plasma; V G (dl / kg) is the glucose distribution volume; p2(min) -1 ) is the rate constant of insulin action kinetics; SI (ml / μU·min) is insulin sensitivity; I pb (mU / l) is the basal insulin concentration in plasma; IG (mg / dl) is the interstitial glucose concentration; α (min) is the time delay between the plasma and interstitial glucose concentration compartments.

[0051] Given that it has a constant unit ρ, The aforementioned model struggles to capture hypoglycemia, likely due to an inadequate description of insulin action, which has been shown to increase when glucose levels drop below a certain threshold. Therefore, the term ρ(G) is introduced into the above model according to the principles proposed in the following literature: C. Dalla Man, F. Micheletto, D. Lv, M. Breton, B. Kovatchev and C. Cobelli, “The uva / padova type 1 diabetes simulator: New features,” J. Diabetes Sci. Technol., Vol. 8, No. 1, pp. 26–34, 2017.

[0052]

[0053] Among them, G th The hypoglycemic threshold (set to 60 mg / dl) is used, and r1 (dimensionless) and r2 (dimensionless) are two model parameters that do not have a direct physiological interpretation.

[0054] To explain patient-specific intraday variability in insulin sensitivity and to model the so-called dawn phenomenon, parameters SI and G... b Considered to change over time within a day:

[0055]

[0056]

[0057] The prior information on the parameter distributions reported in detail in Table 1 was obtained from the literature. Specifically, r1, r2, and V... G The values ​​have been fixed as population values, namely 1.44 dl / kg, 0.81 dl / kg, and 1.45 dl / kg. The unknown model parameters for the glucose-insulin subsystem are...

[0058] D. Holistic Physiological Model

[0059] In summary, the overall physiological model is obtained by combining the three sub-models introduced so far.

[0060]

[0061] Where, x phy (t) is the state vector, defined as follows:

[0062] x phy (t): = [x ins ,|x oral ,|x glu] T

[0063] =[I sc1 I sc2 I p ,|Q sto1 Q sto2 Q gut ,|G,X,IG] T ;

[0064] u phy (t) := [I(t), CHO(t)] is the input vector;

[0065] f phy (·) is the state update function obtained by combining (1)-(5).

[0066] f phy (·) depends on a set of unknown parameters

[0067] θ phy :=[θ ins θ oral θ glu ]

[0068] The estimation of this set of parameters will be discussed in the next section.

[0069] Table 1: Prior information on model parameters

[0070]

[0071] N(μ,σ) represents a normal distribution of mean μ and standard deviation σ. LOGN(μ,σ) represents a log-normal distribution of mean μ and standard deviation σ. GAMMA(a,b) represents a gamma distribution with shape parameter a and scale parameter b.

[0072] III. Offline Bayesian Estimation of Personalized Model Parameters Using Markov Chain Monte Carlo Theory

[0073] By using training data Y:={CGM(t) k ),t k =k·T s ,k=1,…,D} and U:={u phy (t k ),t k =k·T s Identify the unknown model parameters θ for each patient in the group k = 1, ..., D. phy This allows for model personalization, where D is the number of available data points.

[0074] Identification is performed using a Bayesian method, and specifically, in this work, θ phyIt is estimated by its posterior mean, which is defined as follows:

[0075]

[0076] Given that the posterior mean is θ phy The minimum unbiased estimator of variance.

[0077] According to Bayes' theorem, the posterior density function p is obtained. θ|Y,U (θ|Y,U), which is:

[0078]

[0079] Where, p Y|θ,I (Y|θ,U) is the likelihood function, which is the probability of observing a certain Y given the parameter vector θ and the input U.

[0080] Since the integral in (9) is difficult to analyze, it must be approximated using MCMC. Specifically, this is achieved by creating a Markov chain based on the posterior distribution p. θ|Y,U (θ|Y,U) Generate N samples θ i For each i = 1, ..., N, the static distribution of the Markov chain is exactly the posterior distribution (target distribution). Then, these samples θ i Used to perform Monte Carlo integration to obtain θ phy Point estimates:

[0081]

[0082] To construct such a chain, the single-component Metropolis-Hastington (SCMH) algorithm was used.

[0083] A. Implementation Details

[0084] θ phy Divided into five groups θ phy :=[θ1,θ2,θ3,θ4,θ5], that is θ3:=[SI D ],θ4:=[p2,k a2 ,k d ],θ5:=[k empt ,k abs ].

[0085] This partitioning scheme was chosen because it improves the MC mixing and allows breaking the correlation between SI and p2, which is important according to the literature.

[0086] The algorithm's iteration i consists of five steps p = 1, ..., 5, and each step is determined by approving / rejecting the proposal density function q.p Sample φ extracted from (·|·) p To update θ phy The p-th partition θ p Specifically, as specified in the SHCM procedure, the probability of approval occurring is α.

[0087]

[0088] Given other components θ -p Value θ -p =θ i,-p ,π(θ p |θ i,-p ) and θ p It is proportional to the posterior:

[0089] π(θ p |θ i,-p ) = p Y|θ,U (Y|θ p ,θ i,-p ,U)p θ (θ p |θ i,-p ,U)

[0090] θ i,-p Including θ p θ other than phy All other components. More precisely, θ i,-p Includes the latest version of each component available at the current stage of the algorithm: θ i,-p =[θ i,1 ,…,θ i,p-1 ,θ i-1,p+1 ,θ i-1,5 When processing the p-th component in iteration i, the components up to p-1 have been updated, while the other components from p+1 to 5 have not been updated, so their values ​​calculated in the previous iteration i-1 are used.

[0091] For the proposed distribution involved, a Gaussian distribution is used, which is derived from θ in the previous chain iteration. p The assumed value is central.

[0092]

[0093] Where, Σ p This is the adjustment parameter for the acceptance rate of the regulating chain. Σ p Set as a diagonal matrix, its components are partitions p. The estimated conditional standard deviation of each element is multiplied by a scaling factor as proposed by H. Haario, E. Saksman, and J. Tamminen, “Adaptive proposal distribution for random walk metropolisal algorithm,” Computational Statistics, Vol. 14, pp. 375-395, 1999. Adaptive SHCM is achieved by running two exploratory MCMCs (nIter = 600 iterations) to compute the estimate and updating it every 1500 iterations.

[0094] Finally, the convergence of MCMC is verified using the known Raftery-Lewis criterion, providing the number of iterations necessary to ensure that the Markov chain represents the target posterior distribution.

[0095] The adaptive single-component Metropolis-Hastington algorithm is summarized in algorithm (1).

[0096]

[0097] IV. Residual Model

[0098] As a final step, by using Figure 5 The residual model scheme uses random error e(t) to disrupt the output of the physiological model IG(t) to obtain the measured output CGM(t).

[0099] CGM(t) = iG(t) + e(t). (12)

[0100] e(t)(mg / dl) describes the residual modeling error and the measurement error introduced by the CGM sensor. The autoregressive integrated moving average (ARIMA) modeled as an order (5,5,1) is a colored stochastic process.

[0101]

[0102] Among them, {a i} i=1,…,6 and {b i} i=1,…,5 These are the coefficients of the model, and ∈(t)(mg / dl) is a model with zero mean and constant standard deviation SD. ∈ The white noise. Due to the consideration of ARIMA, one of the zeros in the denominator is in 1.

[0103] The model order was chosen to keep the model as simple as possible while still being able to describe the residuals e(t) well. The model parameters were estimated using the standard prediction error method (PEM).

[0104] A. Enhanced Model

[0105] The state-space representation of the residual model is enhanced (8), and then the final model is obtained.

[0106]

[0107] Wherein, by using t k =k·T s Backward Euler method for f phy (x phy ,u phy ,t,θ phy Discretize to obtain f d Furthermore, it is assumed that this is corrupted by the model error v_phy, which is a model with covariance. The iid Gaussian process. The model is enhanced with the state vector associated with the state space representation of the ARIMA process describing the residual (13), where matrices A, B and C are as defined in GEPBox, GMJenkins, GCReinsel and GMLjung, “Timeseries analysis: Forecasting and control”, 5th edition, Hoboken, NJ: Wiley and Sons, 2015.

[0108]

[0109] C = [1 0 0 0 0]

[0110] Finally, by defining the enhanced state x(t) k ):=[x phy (t k )|e(t k )] T and enhance error v(t) k ):=[v phy (t k )|B·∈(t k )] T Combine equations (14a) and (14b).

[0111] Enhanced model update equation readings

[0112] x(t k+1 )=f(x(t k+1 ),u(t k+1 ),t,θ)+v(t k (15)

[0113] V. Real-time glucose prediction using particle filtering

[0114] Therefore, we focus on estimating the parameters of a physiological model of T1D subjects to capture patient-specific dynamics. This process can be performed “offline” using available information obtained from retrospective patient data.

[0115] This section will demonstrate how to use this personalized model to predict future glucose concentrations in real time. This task must be performed "online," therefore a sequential algorithm is employed, where each time step t... k When the new measured value y(t) k ) = CGM(t k When ) becomes available, update the model state x(t) k The current estimate of glucose concentration is used to infer future glucose concentrations. Specifically, particle filtering (PF), a sequential Bayesian prediction technique capable of handling nonlinear structures in the model, is employed.

[0116] PF is based on the posterior probability function p(x(t) k )|y(t 1:k ),u(t 1:k The recursive update of y(t) is performed, where y(t) 1:k ) are variables y(t1),...,y(t) k The abbreviation of ) and u(t) 1:k ) means u(t1),…,u(t k ).

[0117] The p(x(t) method is executed through two basic steps: single-step advance prediction and measurement update. k-1 )|y(t 1:k-1 ),u(t 1:k-1 Recursive updates of ))

[0118] The single-step advance prediction step assumes that at time t k-1 The posterior probability p(x(t) can be obtained. k-1 )|y(t 1:k-1 ),u(t 1:k-1 And use this posterior probability to infer

[0119] p(x(t k )|y(t 1:k-1 ),u(t 1:k ))

[0120] Then, when the new measured value y(t) k At time t k When it becomes available, use this measurement to calculate the posterior probability in the measurement update step.

[0121] p(x(t k)|y(t 1:k ),u(t 1:k )).

[0122] Then, for each available measurement in the dataset, these two steps are repeated. PF performs these steps using a sampled approximation of the working probability function:

[0123]

[0124] in, It is a set of P points, called "particles", with p(x(t) as the unit of measurement. k-1 )|y(t 1:k-1 ),u(t 1:k-1 This is supported by weights. Each particle is associated with a weight. ∑ p w p (t k-1 ) = 1 is associated, and

[0125]

[0126] in, It represents p(x(t) k )|y(t 1:k-1 ),u(t 1:k P particles, each P particle with weight ∑ p w *p (t k ) = 1 is associated.

[0127] PF’s single-step lead prediction step is particularly convenient when dealing with state-space models such as (15).

[0128] It can be seen

[0129] p(x(t k )|y(t 1:k-1 ),u(t 1:k ))=∫p(x(t k )|x(t k-1 ))p(x(t k-1 )|y(t 1:k-1 ),u(t 1:k-1 ))dx(t k-1 )

[0130] Among them, the state update equation (15) fully describes p(x(t) k )|x(t k-1 )).

[0131] Regarding the measurement update step, it can be proven that it holds true for p(x(t)). k)|y(t 1:k ),u(t 1:k ))∝p(y(t k )|x(t k ),u(t 1:k ))p(x(t k )|y(t 1:k-1 ),u(t 1:k ))

[0132] Among them, p(y(t) k )|x(t k ),u(t 1:k )) is the likelihood function fully specified by (14c).

[0133] For detailed calculation methods of these quantities during implementation, please refer to the "Implementation Details" section.

[0134] The additional result is the posterior probability p(x(t) k )|y(t 1:k ),u(t 1:k The probability is further used by PF to calculate the posterior probability.

[0135]

[0136] It describes the state distribution from i-step lead to PH-step lead in time prediction (where PH is the prediction time domain).

[0137] In particular, for a state-space model such as (15), it can be shown that...

[0138]

[0139] It is fully specified by equation (14c).

[0140] Finally, the point estimate of the future CGM value at times t_(k+i), i=1,…,PH can be derived using the posterior expected value:

[0141]

[0142] A. Implementation Details

[0143] The numerical scheme implemented by PF is presented below to perform single-step lead forecasting, measurement updates, and multi-step lead forecasting.

[0144] Single-step advance prediction step. To recap, at time t... k-1 ,p(x(t) k-1 )|y(t 1:k-1 ),u(t 1:k-1 )) can be with relevant weights ∑p w(t k-1 A group of P particles with p=1 The defined sampling form is obtained, making

[0145]

[0146] PF through p(x(t) k )|x(t k-1 Extract a new set of particles To perform a single-step advance prediction step:

[0147] x p (t k )~p(x p (t k )|x p (t k-1 (16)

[0148] The probability is specified by (15):

[0149] p(x p (t k )|x p (t k-1 ))~N(f(x p (t k-1 ),u,t k-1 ,θ),Σ v ).

[0150] Therefore, in order to extract a new set of particles, we only need to let each particle x p (t k-1 It evolves according to model (15) and destroys it with the realization of noise v.

[0151] Measurement update steps. The PF algorithm updates the x-th particle value for each p-th particle. p (t k The weight w *p (t k ) set to x p (t k )

[0152] w *p (t k )=p(y(t k )|x p (t k ),u(t 1:k The likelihood function estimated above. (17)

[0153] Specifically, based on equation (14c) and the statistics of the random modeling error e, p(y(t) k )|x p(t k ),u(t 1:k Defined as:

[0154] p(y(t k )|x(t k ),u(t 1:k ))=N(y(t k )-y p (t k ),SD ∈ (18)

[0155] y is obtained using (14c). p (t k ).

[0156] Then the weights are normalized so that ∑ p w *p (t k ) = 1.

[0157] This provides a sampled representation of the posterior density.

[0158]

[0159] Resampling step. To improve the accuracy of PF, the measurement update step is performed by updating a new set of particles. Specifically, a new set of P particles is used, i.e., p(x(t) = ... k )|y(t 1:k-1 ),u(t 1:k The sampling representation of the generated replace Make Pr(x) *p (t k )=x p (t k ))=w *p (t k This step is performed using the recognized resampling algorithm described in the following literature: G. Kitagawa, “Monte carlo filter and smoother for non-gaussian non-linear state space models,” J. Comput. Graph. Statist., Vol. 5, No. 1, pp. 1–25, 1996.

[0160] Therefore, all the new particles With the same weight w *p (t k ) = 1 / P are associated, thus making p(x(t) k )|y(t 1:k-1 ),u(t1:k The approximate simplification of )) is as follows:

[0161]

[0162] Multi-step forward prediction. Multi-step forward prediction values ​​can be obtained as follows: First, obtain the probability of the sampling form. This sampling method is from p(x(t) k )|y(t 1:k ),u(t 1:k It begins, and as in the single-step advance prediction step, the P particle propagates i steps ahead.

[0163] Then, a set of particles is calculated using equation (14c). And use it to obtain the sampling form

[0164] Finally, the point estimate y(t) of glucose i-step advance is obtained. k+i ) as a response to p(y(t) k+i )|y(t 1:k ),u(t 1:k+i The average value calculated using the sampling method is:

[0165]

[0166] The PF implementation is summarized in algorithm (2).

[0167]

[0168] VI. Methodological Evaluation

[0169] A. Simulated dataset

[0170] T1DS is an FDA-approved software tool used in preclinical trials as an alternative to closed-loop insulin delivery strategies and open-loop therapies. It provides a population of 100 virtual adult subjects, each characterized by a distinct vector of physiological parameters, thus capturing the between-subject / intra-subject variability of the T1D population. In the specific example described herein, a latest version of the simulator was used, which better interprets between-patient / intra-patient variability (i.e., changes in patient physiology and behavior over time). Two weeks of synthetic data were generated for 100 virtual adults using T1DS to simulate so-called sensor-enhanced insulin pump therapy (SAP). In this treatment regimen, insulin is administered to the patient via an insulin pump, consisting of two parts: the first part is the basal insulin infusion rate, a fixed daily pattern adjusted by the clinician for each patient; the second part consists of multiple insulin doses administered at each meal to counteract corresponding increases in glucose. The amount of insulin injected is proportional to the estimated carbohydrate intake (CHO) and pre-meal blood glucose measurement.

[0171] In the simulated scenario, each subject ate three meals a day, occurring with uniform probability at time intervals of [06:30, 08:30] (breakfast), [12:00, 14:00] (lunch), and [19:30, 21:30] (dinner), with CHO consumption evenly distributed in the ranges of [20, 60] g, [50, 90] g, and [50, 90] g, respectively. A simple and safe protocol was implemented to treat hypoglycemia, namely, consuming 15 g of rapid-acting CHO every 15 minutes when the patient's measured glucose level was below 70 mg / dL. Finally, the measurement error of the CGM sensor was modeled according to the suggestions in the following literature: A. Facchinetti, S. Del Favero, G. Sparacino and C. Cobelli, “Model of glucose sensor error components: Identification and assessment for new dexcom g4 generation devices,” Med. Biol. Eng. Comput., Vol. 53, No. 12, pp. 1259–1269, 2015. The obtained computer dataset was divided into training and test sets, with data from the first week assigned to the former and data from the second week assigned to the latter.

[0172] B. Prediction methods used for comparison

[0173] Two state-of-the-art black-box models were considered as baselines for prediction accuracy: an autoregressive (ARX) model of order 5 with exogenous inputs and an autoregressive integrated moving average (ARIMAX) model of order (5,5,1) with exogenous inputs, both trained on the same training data. Both have two inputs, insulin infusion I(t) and carbohydrate intake CHO(t), and one output, glucose concentration CGM(t).

[0174] C. Evaluation metrics

[0175] The accuracy of predictions was evaluated on a test set by comparing the obtained glucose predictions with actual CGM values ​​using different pH values. Three metrics were considered in this work: root mean square error (RMSE), coefficient of determination (COD), and time gain (TG). These metrics, frequently used in the literature, are defined as follows:

[0176]

[0177] Here, ||·||2 represents the l2 norm. The larger the RMSE, the worse the prediction accuracy.

[0178]

[0179] in, This is the sample mean of the blood glucose measurement time series. COD depends on the RMSE normalized to the sample variance of the signal. In the case of perfect prediction (RMSE = 0 mg / dl), COD is at most COD = 100%, and decreases as RMSE increases.

[0180]

[0181] Among them, delay (s1(t) k ),s2(t k [min] for two signals s1(t) k ) and s2(t k The delay between s1 and s2 is quantized and used to measure the similarity of the two signals based on cross-correlation (xcorr). Specifically, this delay is defined as a time-domain offset τ that maximizes the cross-correlation (xcorr) between the τ-offset versions of s1 and s2:

[0182] Delay (s1(t) k ),s2(t k ))=argmax τ [xcorr(s1(t k ),s2(t k -τ)]

[0183] The higher the TG, the more timely and useful the prediction, and TG=0 means that the model prediction is equivalent to looking at the current glucose level.

[0184] The performance metrics of PF are compared with ARX and ARIMAX using paired t-tests, unless the assumption of normality is rejected by the Lilliefers test (p > 0.05). In this case, the Wilcoxon signed-rank test is used. Reported p-values ​​are calculated using a two-tailed test and are considered statistically significant when < 0.05.

[0185] Table II. Comparison of performance metrics (intermediate values ​​[iqr]) obtained from computer-simulated populations using ARX, ARIMAX, and PF at PH = 30, 60, 90, and 120 min. The best average results are shown in bold.

[0186]

[0187] D. Result

[0188] Figure 6A and Figure 6B RMSE and COD obtained using ARX, ARIMAX, and PF at different pH values ​​from 15 min to 120 min are shown. The scale lines represent the median values ​​of the population, while the area between the 25th and 75th percentiles is shaded. Table II reports the same information, including p-values, at pH = 30, 60, 90, and 120 min.

[0189] Compared to ARX, PF achieved significantly better results for all considered metrics and pH (p<0.05), with greater improvement at higher pH. Specifically, the mean ΔRMSE = RMSE at pH = 30, 60, 90, and 120 min. PF -RMSE ARX The values ​​were -1.47 mg / dl (-6%), -3.85 mg / dl (-15%), -7.83 mg / dl (-23%), and -11.78 mg / dl (-28%), respectively. The same consideration applies when assessing predictive accuracy in terms of mean ΔCOD; at pH = 30, 60, 90, and 120 min, the mean COD increased by 2.19% (2%), 9.86% (14%), 20.30% (40%), and 32.99% (118%), respectively. For ΔTG, PF significantly improved the results at pH = 30, 60, 90, and 120 min by 15.0 min (100%), 40.0 min (200%), 67.5 min (300%), and 95.0 min (380%), respectively.

[0190] Comparing PF with ARIMAX, the former achieved significantly better RMSE and COD at pH > 30 min, with the improvement increasing as pH increased. Specifically, at pH = 60, 90, and 120 min, PF improved ΔRMSE by (p < 0.05) compared to RMSE. PF -RMSE ARIMAX The COD levels were significantly increased by -0.54 mg / dl (-2%), -2.78 mg / dl (-10%), and -5.81 mg / dl (-16%), respectively, and at pH = 60, 90, and 120 min, the ΔCOD was significantly increased by 3.17% (4%), 10.62% (18%), and 20.70% (51%), respectively, by PF (p<0.05). At pH = 30 min, there was no statistically significant difference in performance: p-value (RMSE, pH = 30 min) = 0.22, p-value (COD, pH = 30 min) = 0.28.

[0191] Finally, at pH = 30, 60, 90, and 120 min, PF significantly increased ΔTG (p < 0.05) by 5.0 min (20%), 10.0 min (20%), 10.0 min (13%), and 10.0 min (9%), respectively. Notably, on average, PF allowed TG to equal pH.

[0192] VII. Alternative Site Implementation Plan

[0193] Exemplary examples using specific individualized nonlinear physiological models (such as the uVa / Padova T1D simulator), specific parameter estimation techniques (MCMC), and specific nonlinear prediction schemes (particle filtering methodology) illustrate the combination of Figure 1 The flowchart presented here illustrates a prediction scheme based on a personalized model. More generally, the prediction scheme based on a personalized model can employ a variety of alternative parameter estimation techniques, and various alternative nonlinear prediction schemes can be combined with different personalized nonlinear physiological models. For example, in some implementations, the parameter estimation technique employed may be maximum likelihood estimation or prediction error minimization. Similarly, in some implementations, the nonlinear prediction scheme employed may be extended Kalman filtering or unscented Kalman filtering.

[0194] VIII. Conclusion

[0195] Personalized real-time prediction of glucose concentration in T1D is crucial for effective proactive treatment. A methodology using a personalized physiological model has been proposed to achieve this goal. In some implementations, this gray-box nonlinear model is processed within a Bayesian framework comprising two main phases. In the first phase, we use MCMC to identify patient-specific parameters of the physiological model of T1D subjects, for example. Then, in the second phase, the personalized model is used to infer future glucose data via particle filtering or other nonlinear prediction schemes. The first phase can be performed offline, retrospectively processing historical patient data. The second phase can be run online as new patient data becomes available.

[0196] The various operations of the methods described above can be performed by any suitable means capable of performing these operations, such as various hardware and / or software components, circuits and / or modules. Typically, any operation shown in the figures can be performed by a corresponding functional means capable of performing these operations.

[0197] The various illustrative logic blocks, modules, and circuits described herein can be implemented or executed using a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA) or other programmable logic device (PLD), discrete gate or transistor logic, discrete hardware components, or any combination thereof intended to perform the functions described herein. The general-purpose processor may be a microprocessor, but in alternative embodiments, the processor may be any commercially available processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a combination of multiple microprocessors, a combination of one or more microprocessors combined with a DSP core, or any other such configuration.

[0198] In one or more aspects, the described functions may be implemented using hardware, software, firmware, or any combination thereof. If implemented in software, these functions may be stored or transferred as one or more instructions or code to a non-transitory computer-readable medium. By way of example, and not limitation, such a non-transitory computer-readable medium may include RAM, ROM, EEPROM, CD-ROM or other optical disk storage devices, magnetic disk storage devices, or other magnetic storage devices.

[0199] The methods disclosed herein include one or more steps or actions for implementing the methods. The method steps and / or actions may be interchanged without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specified, the order and / or use of a particular step and / or action may be modified without departing from the scope of the claims.

[0200] Some aspects may include a computer program product for performing the operations presented herein. For example, such a computer program product may include a computer-readable medium having instructions stored thereon (and / or encoded thereon that can be executed by one or more processors to perform the operations described herein. In some aspects, the computer program product may include packaging material.

[0201] Furthermore, it should be understood that modules and / or other suitable means for performing the methods and techniques described herein can be downloaded and / or otherwise obtained by the user terminal and / or base station, if applicable. For example, such a device can be coupled to a server to facilitate the transfer of means for performing the methods described herein. Alternatively, the various methods described herein can be provided via a storage device (e.g., RAM, ROM, physical storage media such as compressed optical discs (CDs) or floppy disks), so that the user terminal and / or base station can obtain the various methods when the storage device is coupled to or provided to the device. Furthermore, any other suitable techniques for providing the methods and techniques described herein to the device can be utilized.

[0202] It should be understood that the claims are not limited to the precise fittings and components described above. Various modifications, alterations, and variations may be made to the arrangement, operation, and details of the methods and apparatus described above without departing from the scope of the claims.

[0203] Unless otherwise defined, all terms (including technical and scientific terms) have their common and conventional meanings to those skilled in the art and are not limited to specific or customary meanings unless expressly defined herein. It should be noted that the use of a particular term in describing certain features or aspects of this disclosure should not be construed as implying that the term is redefined herein to be limited to include any particular characteristic of the feature or aspect of this disclosure associated with that term. Unless expressly stated otherwise, the terms and phrases used in this application, and variations thereof, particularly those used in the appended claims, should be understood as open-ended rather than restrictive. As an example of the foregoing, the term “comprising” should be understood to mean “including, without limitation, and including but not limited to”; the term “comprising” as used herein is synonymous with “including,” “containing,” or “characterized in,” and is inclusive or open-ended, and does not exclude additional unlisted elements or method steps; the term “having” should be interpreted as “having at least”; the term “comprising” should be interpreted as “including but not limited to”; the term “example” is used to provide exemplary instances of the matters discussed, not to exhaustively or restrictively list them; adjectives such as “known,” “normal,” “standard,” and terms with similar meanings should not be construed as limiting the matters described to those available at a particular time or up to a particular time, but should be understood as covering known, normal, or standard techniques known or available at any time now or in the future; and the use of terms such as “preferred,” “ideal,” “desirable,” or “desirable,” and words with similar meanings should not be construed as implying that certain features are critical, necessary, or even important to the structure or function of the invention, but are only intended to emphasize alternative or additional features that may or may not be used in a particular embodiment of the invention. Similarly, unless otherwise explicitly stated, a group of items connected by the conjunction "and" should not be interpreted as requiring each of these items to be present in the group, but rather as "and / or". Likewise, unless otherwise explicitly stated, a group of items connected by the conjunction "or" should not be interpreted as requiring mutual exclusion between the items in the group, but rather as "and / or".

[0204] All figures used in this specification to represent amounts of ingredients, reaction conditions, etc., should be understood to be modified by the term "about" in all cases. Therefore, unless otherwise stated, the numerical parameters listed herein are approximate values ​​that may vary depending on the desired properties sought. At a minimum, each numerical parameter should be interpreted according to the number of significant figures and ordinary rounding, rather than attempting to limit the application of the principle of equivalence to the scope of any claim in any application claiming priority to this application.

[0205] All references cited herein are incorporated herein by reference in their entirety. Where any publications, patents, or patent applications incorporated herein by reference contradict the disclosure contained herein to some extent, this specification is intended to supersede any such contradictory material and / or claim priority therefrom.

[0206] The subheadings included in this document are for informational purposes only and are intended to help locate sections. These subheadings are not intended to limit the scope of the concepts described therein. Such concepts may apply throughout the specification.

[0207] Furthermore, although the foregoing has been described in considerable detail with illustrations and examples for clarity and understanding, it will be apparent to those skilled in the art that certain changes and modifications can be made. Therefore, this specification and examples should not be construed as limiting the scope of the invention to the specific embodiments and examples described herein, but rather encompassing all modifications and substitutions falling within the true scope and spirit of the invention.

Claims

1. A method for predicting future blood glucose concentrations in an individual patient wearing a continuous glucose monitoring sensor, the method comprising: The processor selects a personalized nonlinear glucose-insulin kinetic physiological model, wherein the selected personalized nonlinear physiological model has multiple model parameters, the values ​​of which are to be determined; The processor estimates population-based values ​​for a first subset of the plurality of model parameters based on prior population data, wherein the estimated population-based values ​​for the first subset are related to at least one of insulin absorption or glucose absorption; The processor estimates personalized values ​​for the second subset of the multiple model parameters by applying parameter estimation techniques to prior information about the probability distribution of each parameter in the second subset of the multiple model parameters and the individual patient data, wherein the individual patient data includes the individual patient's insulin infusion information and the individual patient's carbohydrate intake information; The processor obtains the posterior probability function based on the individual patient's historical glucose measurements from the dynamic blood glucose monitoring sensor, population-based values ​​estimated from a first subset of the plurality of model parameters, and individualized values ​​estimated from a second subset of the plurality of model parameters; The processor receives the individual patient's current glucose measurement from the dynamic glucose monitoring sensor; The processor updates the posterior probability function based on the current glucose measurement. The processor uses the updated posterior probability function to apply particle filtering to the selected individualized nonlinear physiological model to obtain the predicted blood glucose concentration of the individual patient at a future time. The processor generates a notification indicating an impending glucose metabolism abnormality based on the predicted glucose concentration, the notification including an active alert for the impending glucose metabolism abnormality; as well as The processor causes the notification to be output on the device associated with the individual patient.

2. The method according to claim 1, wherein the parameter estimation technique is Markov chain Monte Carlo Bayes estimation.

3. The method according to claim 1, wherein the parameter estimation technique is selected from the group consisting of maximum a posteriori method, maximum likelihood method and prediction error minimization method.

4. The method of claim 1, wherein the notification includes recommended treatment for alleviating the impending glucose metabolism abnormality indicated by the predicted glucose concentration.

5. The method of claim 1, wherein the device associated with the individual patient includes the processor.

6. The method of claim 1, wherein causing includes the processor transmitting the notification to the device associated with the individual patient.

7. The method according to claim 1, wherein the selected individualized nonlinear physiological model includes a subcutaneous insulin absorber model, an oral glucose absorber model, and a glucose-insulin kinetic model.

8. The method of claim 7, wherein the subcutaneous insulin absorption submodel includes a first model parameter and a second model parameter, the first model parameter specifying a value for insulin distribution, the second model parameter specifying a delay in the appearance of insulin in the first compartment, the first and second model parameters being located in a first subset of the plurality of model parameters having the estimated population-based value.

9. The method of claim 8, wherein the subcutaneous insulin absorption submodel comprises a third model parameter and a fourth model parameter, the third model parameter specifying a diffusion rate constant from the first compartment to the second compartment, and the fourth model parameter specifying a subcutaneous insulin absorption rate constant from the second compartment to the plasma, the third and fourth model parameters being located in a second subset of the plurality of model parameters having the estimated personalized values.

10. The method of claim 1, wherein the selected individualized nonlinear physiological model is a maximal physiological model, wherein the parameters for selecting the plurality of model parameters are ignored.

11. The method of claim 1, further comprising enhancing the selected individualized nonlinear physiological model with a residual model to describe residual modeling error.

12. The method of claim 11, wherein the residual modeling error is modeled as an autoregressive composite moving average of order (5,5,1).

13. The method of claim 11, wherein the residual model has model parameters estimated using the prediction error method.

14. The method of claim 1, wherein the predicted blood glucose concentration is predicted at multiple different times.

Citation Information

Patent Citations

  • Glucose prediction using machine learning and time series glucose measurements

    US12205718B2

  • Glucose prediction using machine learning and time series glucose measurements

    US20210375447A1

  • Glucose prediction using machine learning and time series glucose measurements

    US20210375448A1

  • Methods for modeling insulin therapy requirements

    US20110098548A1