Postprandial long window blood glucose profile prediction method based on unknown exogenous input estimation

Abstract

The application discloses a postprandial long-time window blood glucose state trend prediction method based on unknown exogenous input estimation, and belongs to the technical field of blood glucose prediction, and comprises the following steps: S1, a blood glucose metabolism dynamics model is established for the glucose-insulin metabolism physiological process of a diabetes blood glucose patient; S2, a fixed-order simultaneous input and state set value observer is built according to the blood glucose metabolism dynamics model; S3, the observer is used to observe the unknown exogenous input of the patient, the life rule of the patient is summarized, the unknown exogenous input of the current meal is predicted, and accurate blood glucose prediction is carried out in combination with known treatment information. The postprandial long-time window blood glucose state trend prediction method based on unknown exogenous input estimation can consider the influence of factors, such as the life rule and body rhythm of the patient, which are difficult to be quantified or measured, on the blood glucose of the patient in the blood glucose prediction process, effectively captures the personalized blood glucose dynamic rule of the patient, and realizes the postprandial long-time window blood glucose prediction in a small sample scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood glucose prediction, in particular to a postprandial long-time window blood glucose trend prediction method based on unknown exogenous input estimation. BACKGROUND

[0002] With the development of modern diabetes technology, intelligent blood glucose management methods have gradually been popularized in clinical practice. In order to achieve effective blood glucose control for diabetic patients, the blood glucose of the patients should be continuously monitored, and abnormal phenomena of the blood glucose of the patients should be discovered and predicted in time and corresponding intervention or prevention measures should be taken.

[0003] Accurate blood glucose prediction is the basis for realizing effective blood glucose closed-loop control. Based on known blood glucose monitoring information, the blood glucose level and blood glucose abnormal events that may occur in the future of the patient are predicted, and according to the prediction results, corresponding blood glucose management measures can be implemented, such as insulin injection dose or meal intake adjustment to control the blood glucose of the patient within a safe range and maintain the health of the patient.

[0004] Current researches on blood glucose prediction for diabetes mainly have the following three directions: model prediction, data-driven model prediction and hybrid model prediction. The model prediction method establishes a dynamic model according to the physiological law of human blood glucose-insulin metabolism, and the commonly used models include Bergman minimum model, Hovorka model, etc. These models are composed of differential equations with different degrees of complexity. The prediction method based on the model can ensure that the prediction result is consistent with the physiological law, but the flexibility is weak in the use process. In order to realize high-precision prediction, a large-scale model is usually required, and there are many physiological parameters to be determined in the model, which are difficult to identify, calculate and use in real scenarios.

[0005] The data-driven model is based on the blood glucose monitoring data, insulin treatment data, meal information of the patient and other available patient information of the patient to build a model, which is also commonly known as a black box method. Such a model usually needs the support of machine learning technology, and common methods include time series models (ARX, ARMAX, ARIMAX, etc.), fuzzy logic models, Gaussian mixture models, regular learning, reinforcement learning models, Kalman filters, artificial neural network models, etc. The hybrid model combines two or more different methods in the data preprocessing, feature extraction or learning stage, and many hybrid models combine physiological models with machine learning techniques to improve performance.

[0006] Although existing methods have explored blood glucose prediction from different aspects, the postprandial long-time prediction problem under small sample conditions has not been fully solved. In particular, how to comprehensively consider the combined effects of various unknown disturbances that are difficult to measure directly on blood glucose trend is still a major challenge.

[0007] In one aspect, in actual clinical medical scenarios, blood glucose prediction often needs to be performed with only a small amount of individualized blood glucose and other data of the patient. In current research, blood glucose model training often requires a large amount of training data, but in actual clinical situations, the amount of blood glucose data of each patient is limited and difficult to meet the model training needs. Due to insufficient individual data, some studies choose to use a population data set for model training, but in actual situations, there are large individual differences in blood glucose trends among patients, and the model trained by the population data is difficult to achieve high flexibility of individualized blood glucose prediction.

[0008] On the other hand, postprandial blood glucose is influenced by many factors for a long period of time, including many factors that are difficult to quantify and measure, such as the patient's exercise activity, additional meals, body rhythm, stress level, etc. These factors have a difficult time directly reflecting the patient's blood glucose in physiological models or data-driven models, which may in turn lead to poor blood glucose prediction results. The above factors can be collectively referred to as unknown exogenous inputs that affect blood glucose, which are closely related to the patient's daily life habits. People's daily activities and body rhythms in life often follow relatively stable habits, and diabetes patients, due to the existence of routine blood glucose management behaviors, their life habits may be more regular. Patients often have to perform regular blood glucose management behaviors such as meals, insulin injection, blood glucose check, etc. at specific time intervals, and their eating activities and daily exercise behaviors outside the fixed meal times are also limited by blood glucose control goals. The regularity of the above life habits allows us to accurately summarize the patient's life regularity in a small sample and predict the unknown exogenous inputs in the future based on the above regularity.

[0009] Based on the above, a postprandial long-time window blood glucose trend prediction method based on small sample is proposed, which integrates various blood glucose influencing factors that are difficult to quantify and measure into unknown exogenous inputs in prediction. SUMMARY

[0010] The purpose of the present application is to provide a postprandial long-time window blood glucose trend prediction method based on unknown exogenous input estimation to solve the problems in the background art.

[0011] To achieve the above-mentioned purpose, the present application provides a postprandial long-time window blood glucose trend prediction method based on unknown exogenous input estimation, comprising the following steps:

[0012] S1, a blood glucose metabolism dynamics model is established for the glucose-insulin metabolism physiological process of a diabetes blood glucose patient to simulate the physiological response of the patient's blood glucose to carbohydrate intake and insulin treatment strategies;

[0013] S2, according to the structure of the blood glucose metabolism kinetics model in S1, a fixed-order simultaneous input and state set observer is built, which can observe the unknown exogenous input affecting blood glucose in the meal time window according to the known patient blood glucose data, so as to reflect the patient's life habits and body rhythm, and update the unknown exogenous input affecting blood glucose in the meal time window;

[0014] S3, the observer is used to observe the unknown exogenous input of the patient's past data, the observation results are combined to summarize the patient's life habits, the weighted moving average method is used to predict the unknown exogenous input of the current meal, and the known treatment information is combined to accurately predict the blood glucose.

[0015] Preferably, the blood glucose metabolism kinetics model is composed of a core sub-model, a meal absorption sub-model and an insulin transport sub-model, wherein the core sub-model reflects the interaction of glucose and insulin, and the kinetic equation is:

[0016]

[0017] In the formula, G(t) is the plasma glucose concentration, G b is the basal glucose concentration related to the basal rate of insulin delivery of the patient, I p (t) is the plasma insulin mass, I b is the basal plasma insulin concentration, R a (t) is the plasma glucose appearance rate, X(t) is the insulin action, S g represents the glucose effectiveness, S I is the insulin sensitivity, which represents the ability of insulin to control glucose production and utilization, p2 is the delay between insulin concentration and insulin action, V g and V I are the glucose and insulin distribution volumes respectively, and BW is the body weight of the subject.

[0018] The meal absorption sub-model reflects the patient's meal intake behavior, and the kinetic equation is:

[0019]

[0020] R a (t)=k q1 ·f·Q1(t)+K q2 ·f·Q2(t);

[0021] In the formula, Q1(t) and Q2(t) are the meal amounts in the first and second meal absorption compartments, k q1 , k q12 and k q2 are rate constants, f is the meal bioavailability, u m(t) is the meal input;

[0022] The kinetic equations of the insulin transport model are:

[0023]

[0024] where I sc1 (t) and I sc2 (t) represent the mass of insulin in the two subcutaneous transport compartments, I p (t) is the mass of plasma insulin, k c1 , k c2 , and k c12 are rate constants representing the subcutaneous insulin transport, k cl is the insulin clearance rate, u i (t) is the subcutaneous insulin input.

[0025] Preferably, the specific steps of establishing a blood glucose metabolism kinetic model in S1 are:

[0026] 1) First, linearize the core sub-model, the meal absorption sub-model, and the insulin transport sub-model, meanwhile, there are many factors affecting blood glucose levels in real-world scenarios, in addition to insulin input and meal input, various blood glucose influencing factors that are difficult to quantify and measure, such as patient daily activities, body hormone fluctuations, etc., the influence of these factors on blood glucose does not follow any specific deterministic physiological law;

[0027] The above factors are integrated into the blood glucose model as unknown exogenous input d(t) to consider their influence on blood glucose, without constraining the unknown exogenous input to be any type of signal, without assuming that it follows any model, the continuous linear time-invariant blood glucose state space model after adding the unknown exogenous input is:

[0028]

[0029] y(t) = C c x(t) ;

[0030] where d(t) is the unknown exogenous input signal to be estimated, u i (t) and u m (t) are known insulin input and meal carbohydrate intake, respectively, y(t) is the model output, i.e., blood glucose value, x(t) is the state vector defined as:

[0031] x(t) = [G(t) X(t) I sc1 (t) I sc2 (t) I p (t) Q1(t) Q2(t)] T ;

[0032] Input matrix B ci , B cm , B cd , output matrix C c and state transition matrix A c have the following specific forms respectively:

[0033] B ci = [0 0 1 0 0 0 0] T ;

[0034] B cm = [0 0 0 0 0 1 0] T ;

[0035] B cd = [1 0 0 0 0 0 0] T ;

[0036] C c = [1 0 0 0 0 0 0];

[0037]

[0038] 2) Then the continuous linear time-invariant blood glucose state space model after adding unknown exogenous input is discretized, and the discrete step h = 5 min is preferably set to efficiently capture and monitor the patient's blood glucose behavior, and the discrete blood glucose model is obtained as:

[0039] x k+1 = Ax k + Bu k + Gd k + Wω k ;

[0040] y k = Cx k + Du k + Hd k + v k ;

[0041] In the formula, k is the discrete time step, k = 1, 2, 3,..., T, x k , d k and y k are the sampling sequences of x(t), d(t) and y(t), respectively, x k+1 is the state vector at the k+1 sampling time; u k = [u ik ; u mk ], is the known input vector, and the specific input time and value of the patient's insulin input and carbohydrate input are obtained by reading the patient's case, u ik and u mk are u i(t), u m (t) is the sampling sequence; ω k v is the process noise, v k is the measurement noise, which simulates the bounded noise existing in the actual blood glucose detection device, and is set as ‖ω k ‖≤η ω = 10 -4 , ‖v k ‖≤η v = 10 -2 ;

[0042] In the discrete blood glucose model, the matrices A, B, G, W, C, D, and H are respectively:

[0043] A = h·A c + I7; B i = h·B ci ; B m = h·B cm ; B = [B i B m ]; G = h·B cd ; W = I7; C = C c ; D = 0 1×2 ; H = 0;

[0044] In the formula, I7 represents a 7×7 unit matrix.

[0045] Preferably, in S2, the specific steps of constructing a fixed-order simultaneous input and state set value observer are as follows:

[0046] 1) Singular value decomposition is performed on the matrix H in S1:

[0047]

[0048] Definition is the two unitary matrices obtained after singular value decomposition of H, ∑ is the diagonal matrix obtained after singular value decomposition, and the rank of the H matrix is p H , U1 and V1 are matrices composed of the first p H columns of the U matrix and the V matrix, respectively, and U2 and V2 are matrices composed of the p H +1 to the last column of the U matrix and the V matrix, respectively.

[0049] 2) Orthogonal decomposition is performed on the unknown exogenous input by using the matrix obtained by singular value decomposition, and the exogenous input is decomposed into two components d 1,k and d 2,k , and the expression is as follows:

[0050]

[0051] 3) The two components after decomposition are brought into the discrete blood glucose model, and the matrix The discrete blood glucose model output y is transformed using non-singular transformation k Decoupling, after decoupling, two output components z are obtained 1,k and z 2,k , the decoupled blood glucose model system is as follows:

[0052]

[0053] Where,

[0054] The decoupled system structure separates the blood glucose model parts that are affected by unknown exogenous inputs and those that are not affected by unknown exogenous inputs. The decoupled matrices will be used for subsequent observer calculations.

[0055] Preferably, the step of the observer performing observation update in S2 is:

[0056] S21, unknown exogenous input estimation, using the current actual blood glucose monitoring value and the estimated state set to estimate the unknown exogenous input set:

[0057]

[0058] Set during initialization x0 is the initial state of the system at time 0, where is the estimated value of a component of the unknown exogenous input at step k, The initial value of the algorithm is given by Calculated, M1 and M2 are the gain matrices of the observer, is the estimated value of the final updated state of step k based on the information of step k, and are the estimated values ​​of the two components of the unknown exogenous input at step k-1, is the estimated value of the final updated state of step k based on the information of step k-1, is the estimated value of the unknown exogenous input at step k-1;

[0059] S22, time update, transfer and update the state estimation process quantity according to the system dynamics:

[0060]

[0061] Where, is the estimated value of the final updated state of step k-1 based on the information of step k-1, u k-1 is the system input value at time k-1, is the estimated value of the transmission state at step k based on the information at step k;

[0062] S23, measurement update, estimate the final state set according to the current blood glucose measurement value:

[0063]

[0064] In the formula, L and is the gain matrix of the observer, and the minimum H ∞ norm sense optimal observer gain is calculated according to the blood glucose model structure;

[0065] S24, loop iteration S21, S22 and S23, continuously observe unknown exogenous input.

[0066] Preferably, in S3, the calculation formula for predicting the unknown exogenous input of the current meal to be predicted is:

[0067]

[0068] In the formula, T is the time period, d i and d k-T+i are the unknown exogenous input sequences of the i-th meal and the k-T+i-th meal observed by the observer in the current meal type past data, d k+1 is the unknown exogenous input sequence of the k+1-th meal predicted by the weighted moving average method, and the weighted moving average method assigns higher weights to newer data, which can better reflect the latest change rule of the unknown exogenous input affecting the patient's blood glucose.

[0069] Preferably, in S3, the known treatment information includes: the patient's pre-meal blood glucose data, meal data and medication data;

[0070] The known treatment information and the unknown exogenous input result obtained by the prediction calculation formula are jointly substituted into the blood glucose system for multi-step recursive operation to obtain an accurate blood glucose prediction result.

[0071] Therefore, the postprandial long-time window blood glucose trend prediction method based on unknown exogenous input estimation has the following beneficial effects:

[0072] (1) The prediction method proposed by the present application can consider the influence of factors such as patient's life habits and body rhythms which are difficult to quantify or measure on blood glucose in the blood glucose prediction process, effectively capture the patient's blood glucose dynamic rule, and significantly improve the accuracy of postprandial long-time window blood glucose prediction.

[0073] (2) By summarizing the patient's life habits from the unknown exogenous input data and considering this factor in the prediction process, the individualized blood glucose dynamic rule of the patient can be better captured, and more accurate blood glucose prediction can be achieved.

[0074] (3) The blood glucose prediction can be performed only by obtaining a small amount of individualized blood glucose and other data of a patient, and the limitation of requiring a large amount of training data in the traditional method is overcome, and the method is more suitable for a clinical treatment scene.

[0075] The technical solutions of the present application are further described below by means of the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 The flowchart of the embodiment of the present application is shown in the figure.

[0077] Figure 2 The figure shows the observation results of a patient's breakfast meal type for consecutive days without unknown exogenous input in the embodiment of the present application.

[0078] Figure 3 The figure shows the blood glucose prediction results of a patient's breakfast meal for consecutive days in the embodiment of the present application, wherein (a)-(f) respectively represent the blood glucose prediction results of the patient's breakfast meal for the fifth day to the tenth day. DETAILED DESCRIPTION

[0079] The technical solutions of the present application are further described below by means of the accompanying drawings and examples.

[0080] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below by combining the figures in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, but not all the embodiments.

[0081] EMBODIMENT

[0082] The following uses the meal-after long-time window blood glucose trend prediction method based on unknown exogenous input prediction proposed in the present application to test the accuracy of blood glucose prediction.

[0083] 1) Data collection and preprocessing: FDA-recognized UVA / Padova T1DM metabolic simulator is used to generate data, including 30-day blood glucose data of 10 virtual patients. The dietary intake is uniformly distributed within a reasonable range, and meal pills are injected at the same time. In addition, the dietary dose is determined using a standard dose calculation formula.

[0084] Before model training, a meal section is created from the collected data. The blood glucose data measured 2 hours before the meal to 4 hours after the meal is extracted to form a meal section, and the blood glucose data is sampled every 5 minutes. The time stamp closest to the meal time is used as an anchor point to distinguish the blood glucose trajectory before and after the meal.

[0085] 2) Performance evaluation index: the root mean square error (RMSE) is used to verify the effectiveness of the algorithm, and the operation formula is as follows:

[0086]

[0087] where n is the total number of blood glucose prediction points, and y i respectively represent the predicted value and the true value of blood glucose. The prediction root mean square error of 2 hours after meal and 4 hours after meal is calculated respectively to evaluate the prediction effect of the algorithm in the long time window after meal.

[0088] 3) Model configuration: to verify the prediction performance of the algorithm in the small sample scene, each patient starts to predict from the 5th day. According to the algorithm running mechanism of the prediction method, the data of each patient is divided into 3 sets according to breakfast, lunch and dinner, which respectively reflect the blood glucose metabolism of breakfast, lunch and dinner. The blood glucose metabolism dynamics model is established, the fixed order simultaneous input and state set observer is built, and the observation update and blood glucose prediction are carried out.

[0089] The body weight BW, insulin sensitivity S I and the plasma glucose concentration basic state G b of the subject in the model are personalized parameters, and the rest of the parameters are determined parameters, which adopt the population average value obtained by experiment. BW is read according to the patient's physical examination results, G b is calculated according to the patient's HbA1c data:

[0090] G b = HbA1c·28.7-46.7;

[0091] If there are some patients whose glycosylated hemoglobin test data are missing, the patient population HbA1c average value is used for calculation. S I is obtained by least square method according to the known blood glucose data of the patient, which can more accurately reflect the blood glucose change rule of the patient.

[0092] After inspection, the discrete blood glucose model of each patient meets the strong detectability condition, that is, the unknown exogenous input and state of each patient meet the strong detectability condition.

[0093] The observation time window of the patient's life rule in this embodiment is 7 days, and the life rule is summarized in this time window, and the prediction of unknown exogenous input is unknown. The average root mean square error of blood glucose prediction of all patients of all meal types is used as the algorithm effect evaluation. To illustrate the effectiveness of introducing unknown exogenous input in prediction, this embodiment compares the prediction effect of the blood glucose prediction method with the introduction of unknown exogenous input and the blood glucose prediction method using only the identified physiological model, and takes the sample proportion with improved prediction effect and the average root mean square error of the algorithm as the evaluation index.

[0094] 4) Result analysis: the effect of the proposed algorithm and the use of pure model prediction is as follows:

[0095]

[0096] The prediction algorithm verification adopts 780 meal data, as shown in the above table, compared with the pure model prediction method without adding unknown exogenous input sequence, after adopting the proposed prediction method, the prediction effect of 98.85% of sample data after 2 hours of meal is improved, and the prediction effect of 97.95% of sample data after 4 hours of meal is improved, which fully illustrates the scientificity and effectiveness of the proposed method in blood glucose prediction. The average RMSE of the proposed blood glucose prediction method is 12.745mg / dL after 2 hours of meal prediction, and the average RMSE of the proposed blood glucose prediction method is 15.654mg / dL after 4 hours of meal prediction, and the prediction accuracy is high. In the prediction task of different patients and different meal types, the method can realize high-precision prediction, and the algorithm stability and universality are strong.

[0097] Figure 2 It is the observation result of a patient's breakfast meal type for consecutive days. As can be seen from the figure, the unknown exogenous input of the same meal type of a patient shows significant similarity characteristics in extreme value performance and trend, and most of the unknown exogenous input sequences are concentrated in a small data interval, which shows that the dynamic trend of the unknown exogenous input has high stability and concentration. The above similarity reflects the regularity of the patient's life affecting the occurrence of blood glucose events, so adding unknown exogenous input in prediction can better reflect the influence of the patient's individualized life habits on blood glucose.

[0098] Figure 3 It is the effect diagram of the proposed method for predicting the blood glucose of a patient after breakfast for consecutive 6 days from the 5th day data, and the prediction result has high matching degree with the actual blood glucose, and can correctly reflect the blood glucose change trend.

[0099] Therefore, the postprandial long-time window blood glucose trend prediction method based on unknown exogenous input estimation can consider the influence of factors such as patient life rules and body rhythm which are difficult to quantify or measure on blood glucose in the blood glucose prediction process, effectively capture the blood glucose dynamic rules of the patient, and significantly improve the accuracy of postprandial long-time window blood glucose prediction. The experimental results show that, compared with the pure physiological model, the prediction effect of 98.85% of sample data after 2 hours of meal is improved, and the prediction effect of 97.95% of sample data after 4 hours of meal is improved, the average RMSE of the prediction after 2 hours of meal is 12.745mg / dL, and the average RMSE of the prediction after 4 hours of meal is 15.654mg / dL, the prediction accuracy is high, which fully illustrates the scientificity and effectiveness of the proposed method in blood glucose prediction.

[0100] By summarizing the patient's life rules through unknown exogenous input data and taking this factor into account in the prediction process, the individualized blood glucose dynamic rules of the patient can be better captured, and more accurate blood glucose prediction can be realized.

[0101] The present application can perform blood glucose prediction with only a small amount of individualized blood glucose and other data of the patient, overcoming the limitation of the conventional method requiring a large amount of training data, and being more suitable for clinical treatment scenarios.

[0102] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not to limit them, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements also cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. A method for predicting blood glucose status in a long postprandial window based on unknown exogenous input estimation, characterized in that: The following steps are involved: S1. A blood glucose metabolism kinetics model is established based on the physiological process of glucose-insulin metabolism in diabetic patients. The blood glucose metabolism kinetics model consists of a core sub-model, a dietary absorption sub-model, and an insulin transport sub-model. The specific steps for establishing the blood glucose metabolism kinetics model are as follows: 1) First, the core sub-model, dietary absorption sub-model, and insulin transport sub-model are linearized and unknown exogenous inputs are added. The continuous linear time-invariant blood glucose state space model after adding unknown exogenous inputs is obtained as follows: y(t)=C c x(t); Where A c is the state transfer matrix, B ci 、B cm and B cd is the input matrix, C c is the output matrix, x(t) is the state vector, u i (t) is subcutaneous insulin input, u m (t) is the dietary input, d(t) is the unknown exogenous input signal to be estimated, and y(t) is the blood glucose value output by the model; 2) Then, the continuous linear time-invariant blood glucose state space model after adding the unknown exogenous input is discretized to obtain the discrete blood glucose model: x k+1 =Ax k +Bu k +Gd k +Wω k ; y k =Cx k +You k +HD k +u k ; Where A, B, G, W, C, D, H are all discretized matrices, k is the discrete time step, k = 1, 2, 3..., T, x k d k and y k are the sampling sequences of x(t), d(t) and y(t), respectively. k+1 is the state vector at the k+1th sampling time, ω k is the process noise, let ||ω k ||≤η ω =10 -4 , v k To measure noise, let ||v k ||≤η u =10 -2 , where u k u is the specific input time and value of the patient's insulin input and carbohydrate input read from the patient's case. k =[u ik ;u mk ],u ik and u mk u i (t) and u m (t) sampling sequence; S2. Based on the structure of the blood glucose metabolism dynamics model in S1, a fixed-order input and state set-valued observer is constructed to observe and update the unknown exogenous inputs that affect blood glucose within the meal time window. The specific steps for constructing the fixed-order input and state set-valued observer are as follows: 1) Perform singular value decomposition on the matrix H in S1: definition are the two unitary matrices obtained after the singular value decomposition of H, ∑ is the diagonal matrix obtained after the singular value decomposition, and the rank of the H matrix is ​​p H , U1 and V1 are the front p of U matrix and V matrix respectively H The matrix consists of columns, U2 and V2 are the pth columns of the U matrix and the V matrix respectively. H +1 to the last column of the matrix; 2) Use the matrix obtained by singular value decomposition to perform orthogonal decomposition on the unknown external input and decompose it into two components d 1,k and d 2,k , the expression is as follows: 3) Bring the two decomposed components into the discrete blood glucose model and define the matrix The discrete blood glucose model output y is transformed using non-singular transformation k Decoupling, after decoupling, two output components z are obtained 1,k and z 2,k , the decoupled blood glucose model system is as follows: x k+1 =Ax k +Bu k +G1d 1,k +G2d 2,k +Wω k 4 y k =Cx k +From k +H1d 1,k +v k ; from 1,k =C1x k +D1u k +∑d 1,k +υ 1,k ; With 2,k =C2x k +D2u k +υ 2,k ; Where, S3. Use an observer to observe the patient's past data for unknown exogenous inputs, combine the observation results to summarize the patient's life patterns, use the weighted moving average method to predict the unknown exogenous inputs of the current meal, and combine known treatment information to make accurate blood sugar predictions.

2. The method for predicting blood glucose status in a long postprandial window based on unknown exogenous input estimation according to claim 1, characterized in that: In S1, the dynamic equation of the core sub-model is: Where G(t) is the plasma glucose concentration, G b is the basal glucose concentration related to the patient's basal rate of insulin delivery, I p (t) is the mass of plasma insulin, I b is the basal plasma insulin concentration, R a (t) is the rate of plasma glucose appearance, X(t) is the insulin effect, S g Indicates glucose availability, S I is insulin sensitivity, which represents the ability of insulin to control glucose production and use, p2 is the delay between insulin concentration and insulin action, and V g and V I are the glucose and insulin distribution volumes, respectively, and BW is the subject's body weight; The kinetic equation of the dietary absorption submodel is: R a (t)=k q1 ·f·Q1(t)+k q2 ·f·Q2(t); Where Q1(t) and Q2(t) are the amounts of food in the first and second food absorption compartments, and k q1 、k q12 and k q2 is the rate constant, f is the dietary bioavailability, u m (t) is dietary input; The kinetic equation of the insulin transporter model is: Where, I sc1 (t) and I sc2 (t) represents the mass of insulin in the two subcutaneous transport compartments, I p (t) is the mass of plasma insulin, k c1 、k c2 and k c12 is the rate constant for subcutaneous insulin transport, k cl is the insulin clearance rate, u i (t) is subcutaneous insulin infusion.

3. The method for predicting blood glucose status in a long postprandial window based on unknown exogenous input estimation according to claim 1, characterized in that: The steps for the observer in S2 to perform observation update are: S21, unknown exogenous input estimation: using the current actual blood glucose monitoring value and the estimated state set to estimate the unknown exogenous input set; S22, time update: transfer and update the state estimation process quantity according to the system dynamics; S23, measurement update: estimating the final state set based on the current blood glucose measurement value; S24. Iterate S21, S22 and S23 to continuously observe unknown exogenous input.

4. The method for predicting blood glucose status in a long postprandial window based on unknown exogenous input estimation according to claim 1, characterized in that: In S3, the calculation formula for predicting the unknown exogenous input of the current meal using weighted moving average is: Where T is the time period, d i and d k-T+i are the unknown exogenous input sequences of the i-th meal and the k-T+i-th meal observed by the observer in the past data of the current meal type, respectively, and dWMA k+1 is the unknown exogenous input sequence of the k+1th meal predicted using the weighted moving average method.

5. The method for predicting blood glucose status in a long postprandial window based on unknown exogenous input estimation according to claim 4, characterized in that: In said S3, the known treatment information includes: the patient's pre-meal blood sugar data, meal data, and medication data; The known treatment information and the unknown exogenous input results obtained using the prediction calculation formula are substituted into the blood glucose model system for multi-step recursive calculation to obtain accurate blood glucose prediction results.

Images