A post-meal blood glucose prediction system based on physiological information Gaussian process and meta-learning

Abstract

The present invention discloses a postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning, including a Gaussian process blood glucose prediction module with a neural network mean function, an optimization module of physiological information fusion, and a meta-learning module; the neural network mean function in the Gaussian process blood glucose prediction module with a neural network mean function is used to capture nonlinear glucose dynamics; the optimization module of physiological information fusion describes the prior knowledge of glucose dynamics through partial differential inequalities, and embeds it into the model optimization process in the form of a cost function; the meta-learning module improves the rapid learning and generalization capabilities of the model through a meta-learning method. The present invention adopts the above-mentioned postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning, which can effectively integrate the known dynamic changes of blood glucose for data-driven learning, realize individualized postprandial blood glucose prediction under small samples, and quantitatively estimate the uncertainty of the prediction results.

Description

Technical Field

[0001] The present invention relates to the technical field of blood sugar prediction, and in particular to a postprandial blood sugar prediction system based on physiological information Gaussian process and meta-learning. Background Art

[0002] Recent advances in smart measurement and monitoring technologies have enhanced the availability of medical data, which is beneficial for the diagnosis and treatment of diseases and supports personalized medical services. Specifically, diabetes is a chronic metabolic disorder characterized by high blood sugar, which is mostly caused by the inability to synthesize insulin (type 1 diabetes) or reduced insulin secretion and reduced sensitivity to insulin action (type 2 diabetes). Due to long-term and persistent high blood sugar levels, serious complications such as retinopathy and nephropathy often occur.

[0003] Exogenous insulin injection is the final treatment for diabetes and is responsible for blood sugar control. Multiple daily injections (MDI) are the traditional treatment method, which requires patients to manually inject basal insulin doses and mealtime doses through an insulin pen; or inject insulin through a continuous subcutaneous insulin infusion pump (CSII). With the development of blood sugar regulation technology, insulin doses can be determined by artificial pancreas (AP) and decision support systems, both of which are based on continuous glucose monitoring (CGM) measurements. The former can provide micro-injection doses calculated by the control algorithm one by one, while the latter mainly adjusts the insulin dose based on the commonly used standard dose calculator.

[0004] Despite the different structures, both artificial pancreas and decision support systems have a major part in blood glucose prediction, which may be part of the control algorithm or used for monitoring and early warning of high and low blood glucose. In particular, model-based algorithms are the mainstream method for control-related predictions, among which model predictive control is favored in artificial pancreas systems, which predicts the trajectory through a model that describes the dynamic response of blood glucose and determines the control action through iterative optimization within a limited time.

[0005] However, the obvious drawback of model prediction is that it is difficult to accurately model the nonlinear time-varying glucose metabolism process, and personalized parameter tuning is also time-consuming. For the above reasons, many studies have adopted data-driven methods, which are mainly divided into time series methods and machine learning methods. The former is committed to using historical blood glucose data to predict future performance through statistical analysis. However, the limited flexibility of time series methods and strict assumptions (such as stability and linearity) often lead to limited prediction performance. In comparison, machine learning methods have shown their advantages, which mainly include feedforward neural networks, recursive neural networks, support vector machines, and hybrid models.

[0006] It is worth noting that although many studies are committed to using artificial intelligence methods to achieve blood glucose prediction. Most methods only consider short-term prediction results of 30 minutes and 60 minutes, and are mainly used for AP control. However, knowing long-term blood glucose changes is crucial for mealtime insulin dosage decisions, but very few studies focus on long-term postprandial blood glucose prediction to improve decision support systems. Compared with artificial pancreas, MDI therapy has a smaller burden of use, and most diabetic patients treated with insulin prefer MDI therapy. Therefore, providing long-term postprandial blood glucose prediction information is critical for diabetes treatment, because mealtime insulin dosage significantly affects the overall blood glucose level of the day. However, long-term postprandial prediction using data-driven methods is quite challenging because some of them are black box models that lack interpretability and cannot gain doctors' confidence in clinical use. Another noteworthy issue is that the available data for model learning is limited, which may lead to insufficient grasp of underlying physiological processes, resulting in biased blood glucose predictions. Summary of the invention

[0007] The purpose of the present invention is to provide a postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning. Under the framework of meta-learning, the prior blood glucose change law is integrated into the optimization learning process of the model, which can not only take into account the consistent group characteristics of diabetic patients, but also learn the specificity of blood glucose changes among different individuals, thereby realizing individualized long-term postprandial blood glucose prediction under small samples.

[0008] To achieve the above-mentioned object, the present invention provides a postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning, comprising a Gaussian process blood glucose prediction module with a neural network mean function, an optimization module for physiological information fusion, and a meta-learning module;

[0009] The neural network mean function in the Gaussian process blood glucose prediction module with a neural network mean function is used to capture nonlinear glucose dynamics, which quantifies the uncertainty information of blood glucose prediction under the Bayesian framework and further evaluates the potential risks of hyperglycemia and hypoglycemia;

[0010] The optimization module of physiological information fusion describes the prior knowledge of glucose dynamics through partial differential inequalities and embeds it into the model optimization process in the form of a cost function, thereby guiding the model training to the expected direction;

[0011] The meta-learning module improves the rapid learning and generalization capabilities of the model through the meta-learning method MAML. In the meta-training stage, through iterative learning of inner and outer loops, an initial model is first obtained, and then, multiple individualized blood glucose prediction models are obtained in the meta-testing stage.

[0012] Preferably, the Gaussian process GP blood sugar prediction module with a neural network mean function converts G p =[G pre ,G post ] is defined as the sequence of glucose measurements for a meal event p;

[0013] Among them G pre =[g 1 ,…,g m ] consists of m pre-meal glucose data, G post =[g m+1 ,…,g l ] contains the corresponding postprandial glucose data, input sample Where N k =(lm)n p is the sample size of patient k, u p represents the insulin infusion dose for meal event p, d p represents the corresponding carbohydrate intake, Δt: = t m+j -t m is the time difference between glucose and meal to be predicted, and the sequence G p Divided into multiple samples, the corresponding output labels are

[0014] Will and Omit the superscript and write it directly as x j and j , the joint distribution of each finite subset of random variables is described by a multivariate Gaussian distribution The random function Mean function Covariance function

[0015] Consider the measured output y j =h(x j )+∈ contains noise interference ∈, the multivariate Gaussian distribution is rewritten as Among them, ∈ follows the independent Distribution, δ(x j ,x' j ) is the Kronecker function, let For the training data set, by maximizing the marginal log-likelihood MLL function:

[0016]

[0017] To determine the hyperparameters φ and θ of the mean function and covariance function, K(X,X) is the covariance matrix of X, with element Kjk = k(x j ,x k ), the squared exponential SE covariance kernel function in the GP model is

[0018] Hyperparameters of SE kernel function Contains signal variance Noise variance and length scale That is, the diagonal matrix Σ -1 The elements of are normalized before training. After the model training is completed, the new input x is obtained by posterior inference based on the joint prior Gaussian distribution. * The predicted output

[0019]

[0020] The final posterior distribution of is:

[0021]

[0022]

[0023]

[0024] Where K(X,x * )=[k(x * ,x 1 ),…,k(x * ,x N )] T , obtain the predicted mean and estimated uncertainty, and then determine the long-term postprandial glucose sequence by continuously feeding the glucose mean predicted in the previous step into the input of the current prediction

[0025] Preferably, a neural network mean function is introduced into GP, and a long short-term memory LSTM neural network is used to describe the average level of glucose. A unit, a forget gate, an input gate and an output gate are introduced on the basis of a standard recurrent neural network:

[0026] Forget gate f = β(w f ·[s t-1 ,x j ]+b f ) is used to determine the hidden state s t-1 What information in needs to be forgotten after the sigmoid function β(·) is mapped?

[0027] Input gate i = β(w i ·[st-1 ,x j ]+b i ) and tanh layer g = tanh(w g ·[s t-1 ,x j ]+b g ) together determine the current unit state C t =f⊙C t-1 What information is stored in +i⊙g? The operator ⊙ represents the Hadamard product.

[0028] The current hidden state H t It is obtained based on the output gate o, and the calculation formula is o = β (w o ·[s t-1 ,x j ]+b o ), H t =o⊙tanh(C t ), w f ,w i ,w g ,w o is the corresponding weight matrix, b f ,b i ,b g ,b o is the corresponding bias vector; let is the mapping of LSTM, w=[w f ,w i ,w g ,w o ],b=[b f ,b i ,b g ,b o ], then the GP model with neural network is expressed as:

[0029] Where φ = [w, b], in the Bayesian learning framework, the parameters of the LSTM average function and kernel function are obtained by Obtained, postprandial blood glucose will be predicted based on the posterior distribution.

[0030] Preferably, the optimization module for physiological information fusion adopts the partial differential inequality Describe the effects of insulin and carbohydrates on blood sugar changes and design a cost function Embed a priori partial differential inequalities during training, where

[0031]

[0032]

[0033]

[0034] λ 1 and λ 2 is the weight coefficient, and the cost function is optimized using the gradient descent method. The gradient corresponding to the parameter is After multiple iterations, the model is obtained.

[0035] Preferably, the meta-learning module uses the MAML method to quickly learn new tasks without being restricted by model type, and defines the meta-training and meta-testing datasets as and

[0036] in Divided into T sup and T que Set, a total of K 1 Tasks, V sup , V que All by K 2 tasks. In the meta-training phase, the tasks The inner loop update is based on Inner loop Where Φ = [φ,θ], and the outer loop is updated as in α 1 and α 2 They are the inner loop learning rate and the outer loop learning rate respectively;

[0037] In the meta-testing phase, the gradient descent algorithm is used according to the obtained basic model. Quickly learn a task-specific model for patient k, where α 3 is the corresponding learning rate, prediction validation is based on the dataset V que Based on.

[0038] Therefore, the present invention adopts the above-mentioned postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning, which can realize individualized postprandial blood glucose prediction, and its beneficial effects are as follows:

[0039] (1) Considering the complexity of the glucose metabolism process and the coupling effect of other external interferences, the present invention proposes a Gaussian process model for long-term postprandial blood glucose prediction. The neural network mean function designed in the model enhances the flexibility of the model to better capture the dynamic changes of glucose; in addition, under the Bayesian framework, the uncertainty information of blood glucose prediction is quantified, which can be used to further evaluate the potential risks of hyperglycemia and hypoglycemia.

[0040] (2) In order to ensure that the content of model learning conforms to the basic laws of blood glucose changes under the influence of meal size and insulin dosage, the present invention develops a physiological information optimization step for model learning to guide model training in the expected direction; in addition, the meta-learning method can use relatively less data to learn a personalized blood glucose prediction model to overcome the problem of individual differences and achieve rapid adaptation.

[0041] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a schematic diagram of the overall structure of an embodiment of a postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning of the present invention;

[0043] Figure 2 It is a three-meal blood glucose prediction curve of a virtual patient 8 in an embodiment of a postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning of the present invention;

[0044] Figure 3 These are the P-EGA results of different methods of an embodiment of a postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning of the present invention. DETAILED DESCRIPTION

[0045] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0046] Unless otherwise defined, technical or scientific terms used in the present invention shall have the common meanings understood by one having ordinary skills in the field to which the present invention belongs.

[0047] like Figure 1 As shown, a postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning includes a Gaussian process blood glucose prediction module with a neural network mean function, an optimization module for physiological information fusion, and a meta-learning module;

[0048] The neural network mean function in the Gaussian process blood glucose prediction module increases the flexibility of the model and can be better used to capture nonlinear glucose dynamics. In addition, under the Bayesian framework, it can quantify the uncertainty information of blood glucose prediction and further evaluate the potential risks of hyperglycemia and hypoglycemia.

[0049] Furthermore, the Gaussian process GP blood glucose prediction module with a neural network mean function converts G p =[G pre ,G post ] is defined as the sequence of glucose measurements for a meal event p;

[0050] Among them Gpre =[g 1 ,…,g m ] consists of m pre-meal glucose data, G post =[g m+1 ,…,g l ] contains the corresponding postprandial glucose data, input sample Where N k =(lm)n p is the sample size of patient k, u p represents the insulin infusion dose for meal event p, d p represents the corresponding carbohydrate intake, Δt: = t m+j -t m is the time difference between glucose and meal to be predicted, and the sequence G p Divided into multiple samples, the corresponding output labels are

[0051] For ease of understanding, and Omit the superscript and write it directly as x j and j , the joint distribution of each finite subset of random variables is described by a multivariate Gaussian distribution The random function Mean function Covariance function

[0052] Considering the measured output y j =h(x j )+∈ contains noise interference ∈, the multivariate Gaussian distribution is rewritten as Among them, ∈ follows the independent Distribution, δ(x j ,x' j ) is the Kronecker function, let For the training data set, by maximizing the marginal log-likelihood MLL function:

[0053]

[0054] To determine the hyperparameters φ and θ of the mean function and covariance function, K(X,X) is the covariance matrix of X, with element K jk = k(x j ,x k ), the squared exponential SE covariance kernel function in the GP model is

[0055] Hyperparameters of SE kernel function Contains signal variance Noise variance and length scale That is, the diagonal matrix Σ -1 The elements of are normalized before training. After the model training is completed, the new input x can be obtained through posterior inference based on the joint prior Gaussian distribution. * The predicted output

[0056]

[0057] The final posterior distribution of is:

[0058]

[0059]

[0060]

[0061] Where K(X,x * )=[k(x * ,x 1 ),…,k(x * ,x N )] T , obtain the predicted mean and estimated uncertainty, and then determine the long-term postprandial glucose sequence by continuously feeding the glucose mean predicted in the previous step into the input of the current prediction

[0062] Furthermore, a neural network mean function is introduced into GP, and a long short-term memory LSTM neural network is used to describe the average level of glucose. On the basis of the standard recurrent neural network, a unit, a forget gate, an input gate and an output gate are introduced:

[0063] Forget gate f = β(w f ·[s t-1 ,x j [+b f ) is used to determine the hidden state s t-1 What information in needs to be forgotten after the sigmoid function β(·) is mapped?

[0064] Input gate i = β(w i ·[s t-1 ,x j ]+b i ) and tanh layer g = tanh(w g ·[s t-1 ,xj ]+b g ) together determine the current unit state C t =f⊙C t-1 What information is stored in +i⊙g? The operator ⊙ represents the Hadamard product.

[0065] The current hidden state H t It is obtained based on the output gate o, and the calculation formula is o = β (w o ·[s t-1 ,x j ]+b o ), H t =o⊙tanh(C t ), w f ,w i ,w g ,w o is the corresponding weight matrix, b f ,b i ,b g ,b o is the corresponding bias vector; let is the mapping of LSTM, w=[w f ,w i ,w g ,w o ],b=[b f ,b i ,b g ,b o ], then the GP model with neural network is expressed as:

[0066] Where φ = [w, b], in the Bayesian learning framework, the parameters of the LSTM average function and kernel function are obtained by Obtained, postprandial blood glucose will be predicted based on the posterior distribution.

[0067] The optimization module of physiological information fusion describes the prior knowledge of glucose dynamics through partial differential inequalities and embeds it into the model optimization process in the form of a cost function, thereby guiding the model training in the expected direction;

[0068] Furthermore, the optimization module of physiological information fusion uses partial differential inequality

[0069] Describe the effects of insulin and carbohydrates on blood sugar changes and design a cost function based on the classic MLL function Embed a priori partial differential inequalities during training, where

[0070]

[0071]

[0072]

[0073] λ 1 and λ 2 is the weight coefficient, and the cost function is optimized using the gradient descent method. The gradient corresponding to the parameter is After multiple iterations, the model is obtained.

[0074] The meta-learning module improves the model's rapid learning and generalization capabilities through the meta-learning method MAML method. In the meta-training phase, through iterative learning of the inner and outer loops, the initial model is first obtained, and then, multiple individualized blood glucose prediction models are obtained in the meta-testing phase.

[0075] Furthermore, the meta-learning module is used to overcome the problem of rapid generalization of small sample models, using the MAML method.

[0076] To quickly learn new tasks without being restricted by model type, we define the meta-training and meta-testing datasets as and

[0077] in Divided into T sup and T que Set, a total of K 1 Tasks, V sup , V que All by K 2 tasks. In the meta-training phase, the tasks The inner loop update is based on Inner loop Where Φ=[φ,θ], and the outer loop is updated as in α 1 and α 2 They are the inner loop learning rate and the outer loop learning rate respectively;

[0078] In the meta-testing phase, the gradient descent algorithm is used according to the obtained basic model. Quickly learn a task-specific model for patient k, where α 3 is the corresponding learning rate, prediction validation is based on the dataset V que Based on.

[0079] Embodiment 1

[0080] 1. Data Collection and Preprocessing

[0081] In this example, in order to obtain appropriate parameters and evaluate the performance of the proposed prediction model, the data of 10 virtual patients generated by the UVA / Padova T1DM metabolic simulator were used.

[0082] The plan is: starting from 00:00 on the first day of each patient, it lasts for one month, during which time basal and mealtime insulin are injected. Since patients need to follow doctor's orders while hospitalized, meal times and meal amounts fluctuate within a relatively small range, and the interference of snacks and exercise is ignored. To simulate the actual situation, it is assumed that the occurrence of breakfast, lunch, and dinner conforms to a normal distribution. Here we only focus on glucose pattern recognition during meals, so 90 meal events for each patient are collected and considered to belong to the same task.

[0083] 2. Performance Evaluation Indicators

[0084] For the predictive ability of the method, the commonly used root mean square error (RMSE) and mean absolute error (MAE) indicators were selected. In addition, the mean absolute percentage error (MAPE) was used to consider the different individual scales due to BG variability. For clinical safety considerations, the glucose-specific RMSE (gRMSE) was used, which imposes a high-risk penalty on the prediction results based on Clark error grid analysis (C-EGA). RMSE, MAE, MAPE, and gRMSE are defined as follows:

[0085]

[0086]

[0087]

[0088]

[0089] Where N s is the total number of blood sugar prediction points, P(·,·) is the penalty term, and in the subsequent analysis, and Denotes the average RMSE, MAE, MAPE, and gRMSE of the patients.

[0090] 3. Parkes Error Grid Analysis

[0091] The Parkes Error Grid Analysis (P-EGA) method can visually display the clinical consequences of inaccurate blood glucose predictions, which provides a more nuanced perspective on how measurement errors affect the safety of diabetes management decisions. Figure 3 As shown in the figure, the actual blood glucose value and the predicted blood glucose value are plotted on the error grid, which is divided into 5 areas with different risk levels. The clinical significance of each area is as follows:

[0092] Zone A: No influence on clinical effect.

[0093] Region B: Little or no effect on clinical outcomes.

[0094] Area C: May affect clinical outcomes.

[0095] Zone D: There may be significant medical risks.

[0096] Area E may have dangerous consequences.

[0097] 4. Model Configuration

[0098] In previous BG prediction work, a variety of machine learning methods have been considered, among which recurrent neural networks represented by LSTM have performed well in short-term predictions used by AP. In addition, the GP model has also been applied to long-term postprandial blood glucose prediction and has shown good performance in MDI decision-making. Therefore, these two models are included in the comparative experiments. In addition, in order to clearly evaluate the effectiveness of the designed components in the proposed method, the GP based on the neural average function (GPNM for short) and the GPNM model trained based on the MAML method (MetaGPNM for short) are considered. Compared with the proposed method (MetaPIGPNM for short), the utility of the neural average function trained under the physiological information optimization process and personalized meta-learning framework can be evaluated. All models use Python 3.10 and Pytorch 1.12 and run on an AMD Ryzen 76800HS Creator Edition CPU with 32GBRAM.

[0099] V. Results Analysis

[0100] (1) Simulation results

[0101] The BG prediction results based on UVA / Padova simulator data are shown in Table 1:

[0102]

[0103] From the above table, we can see that compared with LSTM and GP, GPNM performs better at the prediction level of 2h and 3h, and only slightly improves the prediction level 4h after meal.

[0104] The prediction accuracy of MetaGPNM is also improved, and the prediction accuracy is further improved at the 2h, 3h and 4h prediction levels, and the average RMSE, MAE, MAPE and gRMSE of all models are the smallest. These results are caused by the physiological-based optimization process designed to introduce prior knowledge of glucose changes, which can constrain the optimization direction of training and ensure that the model learns the correct features consistent with glucose dynamics.

[0105] also, Figure 2 The long-term postprandial blood glucose prediction trajectories of patient 8 at breakfast, lunch, and dinner are shown. Since GPNM and MetaGPNM are variants of MetaPIGPNM and their performance has been evaluated by prediction indicators, we only plot the trajectories of GP, LSTM, and MetaPIGPNM for clarity. Figure 2 It can be seen that although the model error increases with the increase of prediction layer, the prediction curve fits the measured value well, and the actual measured value is also within the given confidence interval, which shows the reliability of the prediction result. In contrast, both GP and LSTM models underestimate or overestimate BG levels, especially when predicting 3h and 4h, which may lead to continuous false alarms of hyperglycemia. For P-EGA, all points are in area A and area B (see Figure 3 ), which indicated clinically acceptable predictions for all models. However, the prediction points of this method were more concentrated in zone A, suggesting that MetaPIGPNM could provide safer results.

[0106] Therefore, the present invention adopts the above-mentioned postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning, which integrates the prior knowledge of blood glucose change rules under the influence of meal size and insulin dosage into the optimization learning process, and has good nonlinear modeling ability and can well capture the dynamic changes of glucose; under the framework of meta-learning, relatively small amounts of data can be used to learn personalized blood glucose prediction models to overcome the problem of individual differences and achieve rapid adaptation.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning, characterized by: It includes a Gaussian process blood glucose prediction module with a neural network mean function, an optimization module for physiological information fusion, and a meta-learning module; The neural network mean function in the Gaussian process GP blood glucose prediction module with a neural network mean function is used to capture nonlinear glucose dynamics, which quantifies the uncertainty information of blood glucose prediction under the Bayesian framework and further evaluates the potential risks of hyperglycemia and hypoglycemia; The optimization module of physiological information fusion describes the prior knowledge of glucose dynamics through partial differential inequalities and embeds it into the model optimization process in the form of a cost function, thereby guiding the model training to the expected direction; The meta-learning module improves the rapid learning and generalization capabilities of the model through the meta-learning method MAML. In the meta-training phase, the initial model is first obtained through iterative learning of the inner loop and the outer loop. Then, multiple individualized blood glucose prediction models are obtained in the meta-testing phase. The meta-learning module uses the MAML method to quickly learn new tasks without being restricted by the model type, and defines the meta-training and meta-testing datasets as and ; in Divided into and Set, total tasks, , All by tasks. In the meta-training phase, the tasks The inner loop update is based on Inner loop ,in , denote the hyperparameters of the mean function and kernel function of the Gaussian process, Indicates about , ; Represents the inner loop loss function. Here, we use represents the parameterized Gaussian process; the outer loop is updated as ,in, ; is the meta-loss function, which is in the form of , represents the sum of multiple task losses, which is used to guide the global parameter update. Indicates the parameters updated by the inner layer The model defined, and They are the inner loop learning rate and the outer loop learning rate respectively; In the meta-testing phase, the gradient descent algorithm is used according to the obtained basic model. , quickly learn about patients A task-specific model, where is the corresponding learning rate, prediction validation with data set Based on The neural network mean function is introduced into GP, and the long short-term memory LSTM neural network is used to describe the average level of glucose. On the basis of the standard recurrent neural network, a unit, a forget gate, an input gate and an output gate are introduced; The optimization module of physiological information fusion adopts partial differential inequality , Describes the effects of insulin and carbohydrates on blood sugar changes, including Indicates meal event The insulin infusion dose, Indicates the corresponding carbohydrate intake, For the corresponding j output labels; design cost function , embeds a priori partial differential inequalities during training, where is the log-likelihood function, and They are penalty items designed for insulin and meal intake, and their specific forms are as follows: ， ， ， and is the weight coefficient, and the cost function is optimized using the gradient descent method. The gradient corresponding to the parameter is , ,in, K and They represent the covariance matrix and mean function of the Gaussian process respectively, and the model is obtained after multiple iterations.

2. A postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning according to claim 1, characterized in that: The Gaussian process GP blood sugar prediction module with neural network mean function will Defined as a meal event Glucose measurement sequence; in Depend on Pre-meal glucose data, Contains the corresponding postprandial glucose data, input sample ,in Is a patient The sample size, Indicates meal event The insulin infusion dose, Indicates the corresponding carbohydrate intake, is the time difference between glucose and meal to be predicted, and the sequence Divided into multiple samples, the corresponding output labels are ; Will and Omit the superscript and write it directly as and , the joint distribution of each finite subset of random variables is described by a multivariate Gaussian distribution , where the random function , mean function , covariance function ; Consider measuring output The noise interference contained in , the multivariate Gaussian distribution can be rewritten as ,in, Follow independent distributed, is the Kronecker function, let For the training data set, by maximizing the marginal log-likelihood MLL function: ， To determine the hyperparameters of the mean function and covariance function and ,in , , yes The covariance matrix of , the squared exponential SE covariance kernel function in the GP model is ; Hyperparameters of SE kernel function Contains signal variance , noise variance and length scale , that is, the diagonal matrix The elements of are normalized before training. After the model training is completed, the new input is obtained by posterior inference based on the joint prior Gaussian distribution. The predicted output ; ， The final posterior distribution of is: ； in , obtain the predicted mean and estimated uncertainty, and then determine the long-term postprandial glucose sequence by continuously feeding the glucose mean predicted in the previous step into the input of the current prediction .

3. A postprandial blood glucose prediction system based on physiological information Gaussian process and meta-learning according to claim 2, characterized in that: Forget Gate To determine the hidden state What information in the sigmoid function needs to be After mapping, it is forgotten; Input Gate and layer Determine the current unit status together What information is stored in the operator represents the Hadamard product, Indicates the unit state at the previous moment; Current hidden state According to the output gate The calculation formula is , , is the corresponding weight matrix, is the corresponding bias vector; let is the mapping of LSTM, , , then the GP model with neural network is expressed as: , in , in the Bayesian learning framework, the parameters of the LSTM average function and kernel function are obtained by Obtained, postprandial blood glucose will be predicted based on the posterior distribution.

Images