MDI insulin dose decision-making system with expert preference learning function

Abstract

The invention discloses an MDI insulin dose decision-making system with expert preference learning, which belongs to the technical field of insulin injection therapy and comprises a model learning module fusing physiological information, a closed-loop feedback type optimization decision-making module, a safety constraint module and a human-in-loop decision-making self-adaption module. According to the MDI insulin dosage decision-making system with expert preference learning, expert decision-making experience can be effectively applied to insulin dosage optimization, individualized self-adaption of bottom layer insulin dosage optimization decision-making and top layer decision-making is achieved, and the MDI treatment effect is improved.

Description

Technical Field

[0001] The invention relates to the technical field of insulin injection therapy, and in particular to an MDI insulin dosage decision system with expert preference learning. Background Art

[0002] Diabetes mellitus is a persistent metabolic disorder caused by complete lack of insulin (type 1) or relative insufficiency of insulin secretion or impaired insulin sensitivity (type 2), which can lead to acute symptoms such as hypoglycemia, as well as serious long-term complications such as cardiovascular disease and chronic kidney disease.

[0003] At present, patients with type 1 and type 2 diabetes will use intensive insulin therapy to improve blood sugar management during the appropriate course of the disease, among which multiple daily injections (MDI) are a common method. MDI treatment uses separate injections of basal insulin and mealtime insulin, of which basal insulin is used to maintain fasting and nighttime blood sugar stability, while mealtime insulin is used to counteract the rise in postprandial blood sugar caused by meals. The basal dose is mainly determined by the patient or doctor based on clinical experience and blood sugar monitoring results; for mealtime doses, it is usually necessary to estimate the carbohydrate content of intake, and combine the pre-meal blood sugar level, and calculate the dose by the standard formula based on the insulin to carbohydrate ratio (Carbohydrate Ratio, CR) and the correction factor (Correction Factor, CF). This method has problems such as difficulty in estimating carbohydrate content, inaccurate personalized parameter debugging, and over-reliance on the clinical treatment level of physicians and patient experience. In recent years, with the development of continuous glucose monitoring and intelligent control technology, intelligent decision-making technology for insulin dosage based on data-driven control and learning methods has received enthusiastic attention. The development of intelligent decision-making technology has reduced the dependence on decision-making experience and individualized parameters of doctors and patients to a certain extent, and improved blood sugar management performance. However, in actual interactions with decision-making algorithms, humans often intervene in the performance of decision-making algorithms based on their own preferences or domain expertise (e.g., adjusting recommended doses, giving priorities for control targets) to quickly achieve personalized treatment results. How to effectively learn expert decision preferences during the interaction process and enhance the rationality and accuracy of the doses recommended by decision-making algorithms is an important issue that needs to be addressed to improve decision compliance and achieve individualized blood sugar management.

[0004] According to investigation and understanding, the currently disclosed prior arts are as follows:

[0005] The invention patent with patent application publication number CN201980042748.9 discloses a system for providing dosage recommendations for basal insulin titration. The method uses individualized control targets, historical blood glucose monitoring and insulin injection information to complete basal dose adjustment; the invention patent with patent application publication number CN201910310314.4 discloses a method and system for selecting insulin dosage forms and adjusting dosages, and provides an insulin dosage adjustment model based on the medical records of patients with type 2 diabetes combined with the Xgboost machine learning model; the invention patent with patent application publication number CN109564775A discloses a system and method for optimizing mealtime insulin dosage, and provides a system and method for adjusting the short-acting dosage of the subject's expected meal using the insulin infusion data recorded in the insulin pen.

[0006] The invention patent with patent application publication number CN202010930702.5 discloses a Gaussian process-based individualized decision-making system for pre-meal insulin dosage, which uses historical blood sugar management information to establish a post-meal blood sugar prediction model, designs a risk-sensitive control framework, and completes pre-meal insulin dosage decisions; the invention patent with patent application publication number CN202110117772.3 discloses an expert experience-assisted pre-meal insulin dosage learning optimization decision-making system based on the above patents, which can evaluate the credibility of Gaussian process regression prediction of post-meal blood sugar under a small number of samples and decide whether to assist with expert decision-making. However, none of the above patents consider how to effectively learn expert decision-making preferences, and then guide the decision-making method / system to improve the acceptance rate of actual use, while quickly achieving the desired blood sugar management. Summary of the invention

[0007] The purpose of the present invention is to provide an MDI insulin dosage decision system with expert preference learning, which can effectively apply expert decision-making experience to insulin dosage optimization, realize bottom-level insulin dosage optimization decision and top-level decision individualization adaptation, and improve the therapeutic effect of MDI.

[0008] To achieve the above object, the present invention provides an MDI insulin dosage decision system with expert preference learning, including a model learning module integrating physiological information, a closed-loop feedback optimization decision module, a safety constraint module, and a human-in-the-loop decision adaptation module;

[0009] The model learning module that integrates physiological information uses Gaussian process to learn the blood glucose metabolism rules of diabetic patients, and introduces hyperparameter constraints based on blood glucose metabolism rules during the learning process to obtain a prediction model for blood glucose changes after insulin injection that conforms to physiological priors.

[0010] The closed-loop feedback optimization decision module constructs a cost function J(u) based on the blood sugar state prediction value output by the model learning module integrating physiological information and the mealtime insulin or basal insulin dose injected for the same meal type last time, and then solves the cost function by combining the Bayesian optimization algorithm and the Monte Carlo method. Complete this insulin dosage decision:

[0011] The safety constraint module imposes a residual active insulin constraint u on the pre-meal insulin dose recommended by the closed-loop feedback optimization decision module. IOB , determine the final dose u bolus =uu IOB ;

[0012] The human-in-the-loop decision-making adaptation module is triggered by a dose modification event. When the insulin dose recommended by the closed-loop feedback optimization decision module is rejected by an expert or user, and the absolute difference between the modified dose and the recommended dose is greater than the designed threshold Δu, the human-in-the-loop decision-making adaptation module starts to complete the expert decision preference learning.

[0013] Preferably, in the model learning module integrating physiological information, the data set obtained by the k-th decision is recorded as Where X k is an input sample consisting of insulin injection dose, carbohydrate intake information, and blood glucose monitoring data before insulin injection; k is an output sample consisting of blood glucose monitoring data after insulin injection. In order to provide blood glucose prediction information to the decision module, physiological metabolic laws such as insulin injection will reduce blood glucose level and carbon water intake will increase blood glucose level are characterized as hyperparameter constraints in Gaussian process learning. The specific form is as follows:

[0014]

[0015] Among them, θ * To obtain the optimal value; log(p(Y k |X k ,θ)) is the maximum likelihood function used in traditional Gaussian process learning; arg min is log(p(Y k |X k ,θ)) takes the value when it reaches the minimum value; θ is the hyperparameter variable of the Gaussian process; is the imposed insulin glucose-lowering constraint; * is the predicted blood glucose level; u * is the dose of insulin to be injected; is the imposed carbohydrate glycemic constraint; d * is the carbohydrate content of intake;

[0016] The above constraint forms are all related to the mean and variance function forms adopted by the Gaussian process. Finally, the interior point method is used to solve the above constraint optimization problem to complete the model learning.

[0017] Preferably, the model learning module integrating physiological information uses corresponding data sets to respectively learn the changing patterns of blood glucose metabolism before meals and after basal insulin injection.

[0018] Preferably, the closed-loop feedback optimization decision module is specifically:

[0019]

[0020] u low ≤|uu p |≤u up

[0021] 0≤u≤u max

[0022] Among them, γ is the set individual risk sensitivity coefficient; is a vector of blood sugar prediction values ​​output by the blood sugar prediction model, whose elements is a random variable that follows a Gaussian distribution; l is the prediction step length, which is related to the basal or pre-meal insulin dosage decision; G r is a vector of the set postprandial blood sugar or nighttime blood sugar reference values; Q is a penalty matrix that imposes asymmetric penalties on the predicted high and low blood sugar levels; is transposed; express expectations; is Gaussian distribution; m i and are the mean and variance respectively; u up and u low are the upper and lower bounds of the adjustment step size respectively; u max Indicates the maximum allowable dose; u is the recommended dose; u p is the mealtime insulin dose or basal insulin dose of the last same meal type injection; R is the penalty term coefficient set to weigh the degree of trust in the last decision dose; the penalty term R(uu p ) 2 This allows the optimizer to quickly adjust the dose based on feedback from the previous decision-making performance.

[0023] Preferably, the human-in-the-loop decision-making adaptive module optimizes the key parameters of the closed-loop feedback optimization decision-making module to reduce the difference between the recommended dose u and the actual dose u used by the user. c The difference between the two is as follows:

[0024]

[0025] in, is a known feasible set The key parameter vector of the algorithm in which g l Basal or mealtime insulin decision control target G r The last element of , R is the penalty term coefficient in the closed-loop feedback optimization decision module; Q g and Q R is the penalty term coefficient, which is used to adjust the aggressiveness of parameter optimization; the variable with subscript p represents the parameter value used in this decision before optimization;

[0026] Considering that in clinical treatment, the blood sugar control target is allowed to be within a specific range, which is mainly determined by the patient's age and diabetic complications, g is selected. l As a parameter to be optimized, we can learn the attitude of expert decision-makers towards current blood sugar control; R is selected as another parameter to be optimized to reflect the degree of expert trust in past decision-making experience; by optimizing these two parameters at the same time to minimize the decision-making deviation, we can weigh the degree of expert recognition of past decision-making experience and current control situation in the decision-making process, thereby effectively learning the expert decision-making preference.

[0027] Preferably, the decision-making adaptive module of the person in the loop optimizes the key parameters of the basal and mealtime insulin dosage decisions respectively, and the final control target is the original control target as a whole plus the difference g l -g lp The Bayesian optimization method is used to solve the above optimization problem, and the final parameters obtained will be used in the closed-loop feedback optimization decision module to complete the next insulin dosage decision.

[0028] Therefore, the present invention adopts the above-mentioned MDI insulin dosage decision system with expert preference learning to construct a human-in-the-loop data-driven decision learning framework, actively uses expert / user decision-making experience for MDI insulin dosage optimization, and realizes bottom-level insulin dosage optimization decision and top-level decision individualization adaptation.

[0029] In order to improve the performance of blood glucose management in MDI treatment, the present invention uses patient historical data to design an insulin dosage decision system with expert preference learning. The system characterizes physiological metabolic laws as hyperparameter constraints, establishes a blood glucose metabolism Gaussian process modeling method that integrates physiological priors, and provides reliable blood glucose status prediction information for subsequent decisions; at the same time, based on the above prediction information, the present invention constructs a unified risk-sensitive optimization control framework for basic and mealtime insulin dosage decisions to achieve safe and effective MDI dosage decisions; finally, the present invention optimizes the key parameters of the decision-making algorithm to reduce the difference between the recommended dose and the expert-suggested dose, so as to weigh the expert's trust in past decision-making experience and current control conditions during the decision-making process, thereby effectively learning the expert's decision-making preferences.

[0030] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a method framework diagram of an embodiment of an MDI insulin dosage decision system with expert preference learning of the present invention;

[0032] Figure 2 This is a clinical data verification result of type 1 diabetes patients provided by an embodiment of an MDI insulin dosage decision system with expert preference learning of the present invention. DETAILED DESCRIPTION

[0033] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.

[0034] 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.

[0035] Embodiment 1

[0036] like Figure 1 As shown, an MDI insulin dosage decision system with expert preference learning includes a model learning module integrating physiological information, a closed-loop feedback optimization decision module, a safety constraint module, and a human-in-the-loop decision adaptation module. At the same time, each parameter in the present invention is determined based on clinical data of real diabetic patients using MDI therapy in hospitals.

[0037] The model learning module that integrates physiological information adopts the Gaussian process modeling method, integrates physiological metabolic laws during the training process, effectively learns the blood glucose metabolism laws of diabetic patients, and obtains a blood glucose prediction model after insulin injection.

[0038] Furthermore, the model learning module integrating physiological information includes a data acquisition unit, a data preprocessing unit and a training unit. For the introduction of each unit, please refer to the invention patent No. CN202010930702.5. The model learning steps of basal insulin are similar to those of mealtime insulin, but the collected blood glucose monitoring data information is inconsistent. Since basal insulin has a longer duration of action, in order to provide sufficient blood glucose monitoring information for its decision-making, the sample collects blood glucose monitoring information within 3 hours before basal insulin injection and within 9 hours after injection, and the sampling time interval is 1 hour.

[0039] Furthermore, unlike the method adopted by the training unit of patent number CN202010930702.5, the present invention introduces hyperparameter constraints characterized by the law of blood glucose metabolism in the training process, guiding the Gaussian process method to generate reliable blood glucose prediction information. Specifically, the data set obtained by the data acquisition unit and the data preprocessing unit is denoted as Gaussian process learning is accomplished by maximizing the posterior probability, which is as follows:

[0040]

[0041] In the formula, θ * To obtain the optimal value; log(p(Y k |X k ,θ)) is the maximum likelihood function used in traditional Gaussian process learning; arg min is log(p(Y k |X k ,θ)) takes the value when it reaches the minimum value; θ is the hyperparameter variable of the Gaussian process; Where m(·) is the Gaussian process mean function; K(X k ,X k ) is the covariance matrix, whose element K ij = k(x i ,x j ), where k(·,·) is the Gaussian process covariance function; is the variance of the noise; I is the unit matrix; in this implementation, m(·) and k(·,·) are designed as follows:

[0042]

[0043]

[0044] In the formula is the signal variance, then the Gaussian process hyperparameter After adopting the above-mentioned mean and covariance functions, the constraints of physiological metabolic laws can be derived according to the Gaussian process prediction formula and constant validity conditions: and They are equivalent to hyperparameter constraints of the following form:

[0045]

[0046] Where α i For vector The above constraints are applied to the optimization problem of maximizing the posterior probability, and finally the constrained optimization problem is solved by the interior point method to complete the model learning, and the blood glucose status prediction model is established according to the Gaussian process prediction formula. In addition, the present invention uses a continuous accumulation method to increase the number of samples after each decision is completed to complete the continuous learning and updating of the model.

[0047] The closed-loop feedback optimization decision module constructs the cost function J(u) based on the blood sugar trend prediction value output by the model learning module and the mealtime insulin dose or basal insulin dose injected for the same meal type (breakfast, lunch and dinner) last time, and then combines the Bayesian optimization algorithm and the Monte Carlo method to solve the cost function Complete this insulin dosage decision:

[0048]

[0049] u low ≤|uu p |≤u up ,

[0050] 0≤u≤u max ,

[0051] Among them, γ is the set individual risk sensitivity coefficient, is a vector of blood sugar prediction values ​​output by the blood sugar prediction model, whose elements is a random variable that follows a Gaussian distribution, l is the prediction step length, which is related to the basal or pre-meal insulin dosage decision. In this embodiment, l=9 is taken for the basal insulin dosage decision, and l=8 is taken for the mealtime insulin dosage decision. Q is a penalty matrix, and the penalty value imposed on the predicted high blood sugar level is less than the penalty value imposed on the predicted low blood sugar level; is transposed; express expectations; is Gaussian distribution; m i and are the mean and variance respectively; u up and u low are the upper and lower bounds of the adjustment step size respectively; u max Indicates the maximum allowable dose; u is the recommended dose; u pis the mealtime insulin dose or basal insulin dose of the last same meal type injection, and R is the penalty term coefficient set to weigh the degree of trust in the last decision dose. In this embodiment, R=10. p ) 2 The design allows the optimizer to quickly adjust the dose based on feedback from the previous decision performance.

[0052] G r The vector composed of the set postprandial blood sugar or nighttime blood sugar reference value can be determined according to the patient's age and whether the patient has complications. The complications mainly include coronary atherosclerotic heart disease, retinopathy (proliferative stage), peripheral artery stenosis, renal insufficiency and a history of cerebral infarction within 1 year.

[0053] According to the doctor's adjustment experience, the relationship between the target range of blood sugar control two hours after a meal and before the basic breakfast and the patient's age and the presence or absence of complications is roughly as shown in Table 1, where the blood sugar unit is mg / dL. Based on this experience, this embodiment determines the target blood sugar value after a meal for patients aged ≤40 and without complications to be The target blood sugar level for the basal dose from night to before breakfast is In order to make this the basic control target, patients of other ages and with or without complications are added to this basic control target. The specific method is shown in Table 1. For example, (+10) means adding and All elements plus 10 are determined as the blood sugar control target for the corresponding patient, which is equivalent to the control target for patients aged ≤ 40 and with complications in Table 1.

[0054] Table 1 Blood sugar control targets two hours after meal and before basic breakfast in patients of different ages and with or without complications

[0055]

[0056] The safety constraint module imposes a residual active insulin constraint u on the mealtime insulin dose recommended by the closed-loop feedback optimization decision module. IOB , determine the final dose u bolus =uu IOB. Imposing residual active insulin on Board (IOB) constraints can effectively prevent excessive insulin infusion. This method mainly calculates the changes in residual active insulin in the body after a large dose of insulin before a meal. The calculation of IOB for large doses during meals can refer to the method in the literature (Gondhalekar R, Dassau E, Doyle III FJ Velocity-weighting & velocity-penalty MPC of an artificial pancreas: Improved safety & performance [J]. Automatica, 2018, 91: 105-117).

[0057] The human-in-the-loop decision-making adaptive module is triggered by the dose modification event. When the insulin dose recommended by the closed-loop feedback optimization decision module is rejected by the expert / user, and the absolute difference between the modified dose and the recommended dose is greater than the designed threshold Δu, the human-in-the-loop decision-making adaptive module starts to complete the expert decision preference learning. In this embodiment, Δu=1U is taken. Specifically, the human-in-the-loop decision-making adaptive module optimizes the key parameters of the closed-loop feedback optimization decision module to minimize the difference between the recommended dose u and the actual dose u used by the user. c The difference between the two is as follows:

[0058]

[0059] In the formula is a known feasible set The key parameter vector of the algorithm in which g l Basal or mealtime insulin decision control target G r The last element of , R is the penalty term coefficient in the closed-loop feedback optimization decision module. g and Q R is the penalty term coefficient, which is used to adjust the aggressiveness of parameter optimization. The variable with subscript p represents the parameter value used in this decision before optimization. g =1 / 200, Q R =1 / 200. At the same time, according to actual clinical data testing and doctor experience, the feasible set Θ is set as follows:

[0060]

[0061] In the formula and These are the upper and lower limits of the target range for blood sugar control in Table 1, R min =5, R max=120. The Bayesian optimization method is used to solve the above optimization problem, and the final parameters obtained will be used in the closed-loop feedback optimization decision module to complete the next insulin dosage decision, where the final control target is the original control target plus the difference g l -g lp Furthermore, the human-in-the-loop decision-making adaptive module optimizes the key parameters of mealtime and basal insulin dosage decisions respectively.

[0062] Finally, the present invention uses clinical data from 40 patients with type 1 diabetes who used MDI treatment methods in the hospital to verify the decision-making performance of the system. The patients had regular diets during hospitalization, similar meal times and meal amounts, and basal insulin dosage decisions were scheduled at around 10 o'clock every night. Therefore, we used two days of blood glucose monitoring data and the corresponding doctor-executed dosages to train a postprandial blood glucose prediction model and a basal insulin nighttime blood glucose prediction model that do not require dietary information, and used the clinical data of the remaining hospitalization courses to complete the algorithm's retrospective decision-making, and compared it with the corresponding doctor-executed dosages to analyze the algorithm's decision-making and learning performance.

[0063] It should be noted that after each retrospective decision is made, the sample is expanded into the corresponding data set to continuously update the blood glucose prediction model. In addition, when the absolute value of the difference between the doctor's implementation dose and the algorithm's recommended dose is greater than 1U, the decision-making adaptive module of the human in the loop is activated to complete the doctor's decision preference learning. The overall decision difference change results are shown in Figure 2. Figure 2 As shown in the figure, the 5%, 25%, median, 75% and 95% quantiles are obtained by connecting the discrete decision difference points and smoothing them, and the number of decisions per day is marked above the figure. Figure 2 It can be seen that regardless of whether expert decision preference learning is implemented, the difference between the algorithm recommended dose and the doctor's dose is mostly between -1U and 1U, and after the implementation of expert decision preference learning, the difference between the algorithm recommended dose and the doctor's dose becomes smaller, which is reflected in the smaller and more compact area composed of the 25%-75% and 5%-95% difference quantile lines. The actual clinical data analysis results show that the present invention can decide reasonable basal and mealtime insulin doses in actual scenarios, and effectively learn expert decision preferences in interaction with experts to improve the treatment effect of MDI.

[0064] Therefore, the present invention adopts the above-mentioned MDI insulin dosage decision system with expert preference learning, which can effectively apply expert decision-making experience to insulin dosage optimization, realize bottom-level insulin dosage optimization decision and top-level decision individualization adaptation, and improve the therapeutic effect of MDI.

[0065] 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. An MDI insulin dosage decision system with expert preference learning, characterized by: It includes a model learning module integrating physiological information, a closed-loop feedback optimization decision module, a safety constraint module, and a human-in-the-loop decision-making adaptive module; The model learning module that integrates physiological information uses Gaussian process to learn the blood glucose metabolism rules of diabetic patients, and introduces hyperparameter constraints based on blood glucose metabolism rules during the learning process to obtain a prediction model for blood glucose changes after insulin injection that conforms to physiological priors. The closed-loop feedback optimization decision module constructs a cost function J(u) based on the blood sugar state prediction value output by the model learning module integrating physiological information and the mealtime insulin or basal insulin dose injected for the same meal type last time, and then solves the cost function by combining the Bayesian optimization algorithm and the Monte Carlo method. Complete this insulin dosage decision: The safety constraint module imposes a residual active insulin constraint u on the pre-meal insulin dose recommended by the closed-loop feedback optimization decision module. IOB , determine the final dose u bolus =uu IOB ; The human-in-the-loop decision-making adaptation module is triggered by a dose modification event. When the insulin dose recommended by the closed-loop feedback optimization decision module is rejected by an expert or user, and the absolute difference between the modified dose and the recommended dose is greater than the designed threshold Δu, the human-in-the-loop decision-making adaptation module starts to complete the expert decision preference learning.

2. The MDI insulin dosage decision system with expert preference learning according to claim 1, characterized in that: In the model learning module integrating physiological information, the data set obtained by the k-th decision is recorded as Where X k is an input sample consisting of insulin injection dose, carbohydrate intake information, and blood glucose monitoring data before insulin injection; k is an output sample consisting of blood glucose monitoring data after insulin injection. In order to provide blood glucose prediction information to the decision module, physiological metabolic laws such as insulin injection will reduce blood glucose level and carbon water intake will increase blood glucose level are characterized as hyperparameter constraints in Gaussian process learning. The specific form is as follows: Among them, θ * To obtain the optimal value; log(p(Y k |X k ,θ)) is the maximum likelihood function used in traditional Gaussian process learning; arg min is log(p(Y k |X k ,θ)) takes the value when it reaches the minimum value; θ is the hyperparameter variable of the Gaussian process; is the imposed insulin glucose-lowering constraint; * is the predicted blood glucose level; u * is the dose of insulin to be injected; is the imposed carbohydrate glycemic constraint; d * is the carbohydrate content of intake; The above constraint forms are all related to the mean and variance function forms adopted by the Gaussian process. Finally, the interior point method is used to solve the above constraint optimization problem to complete the model learning.

3. The MDI insulin dosage decision system with expert preference learning according to claim 2, characterized in that: The model learning module integrating physiological information uses corresponding data sets to respectively learn the changing rules of blood glucose metabolism before meals and after basal insulin injection.

4. The MDI insulin dosage decision system with expert preference learning according to claim 3, characterized in that: The closed-loop feedback optimization decision module is specifically: in low ≤|uu p |≤u up 0≤u≤u max Among them, γ is the set individual risk sensitivity coefficient; is a vector of blood sugar prediction values ​​output by the blood sugar prediction model, whose elements is a random variable that follows a Gaussian distribution; l is the prediction step length, which is related to the basal or pre-meal insulin dosage decision; G r is a vector of the set postprandial blood sugar or nighttime blood sugar reference values; Q is a penalty matrix that imposes asymmetric penalties on the predicted high and low blood sugar levels; is transposed; express expectations; is Gaussian distribution; m i and are the mean and variance respectively; u up and u low are the upper and lower bounds of the adjustment step size respectively; u max Indicates the maximum allowable dose; u is the recommended dose; u p is the mealtime insulin dose or basal insulin dose of the last same meal type injection; R is the penalty term coefficient set to weigh the degree of trust in the last decision dose; the penalty term R(uu p ) 2 This allows the optimizer to quickly adjust the dose based on feedback from the previous decision-making performance.

5. The MDI insulin dosage decision system with expert preference learning according to claim 4, characterized in that: The human-in-the-loop decision-making adaptive module optimizes the key parameters of the closed-loop feedback optimization decision-making module to reduce the difference between the recommended dose u and the actual dose u used by the user. c The difference between the two is as follows: in, is a known feasible set The key parameter vector of the algorithm in which g l Basal or mealtime insulin decision control target G r The last element of , R is the penalty term coefficient in the closed-loop feedback optimization decision module; Q g and Q R is the penalty term coefficient, which is used to adjust the aggressiveness of parameter optimization; the variable with subscript p represents the parameter value used in this decision before optimization.

6. The MDI insulin dosage decision system with expert preference learning according to claim 5, characterized in that: The decision-making adaptive module of the human in the loop optimizes the key parameters of the basal and mealtime insulin dosage decisions respectively, and the final control target is the original control target plus the difference g l -g lp The Bayesian optimization method is used to solve the above optimization problem, and the final parameters obtained will be used in the closed-loop feedback optimization decision module to complete the next insulin dosage decision.

Images