Diabetes insulin pump intensified treatment dose prediction model constructed based on artificial intelligence algorithm and application thereof

Abstract

The invention discloses a diabetes insulin pump fortified treatment dose prediction model constructed based on an artificial intelligence algorithm, and the diabetes insulin pump fortified treatment dose prediction model only needs to obtain the age, waistline, body mass index and glycosylated hemoglobin when a patient is equipped with an insulin pump and the fasting blood-glucose level before the patient is equipped with the insulin pump. The insulin basal rate of the patient on the first day of the insulin pump and the large dose before three meals on the first day can be obtained, and then the basal rate and the large dose before three meals of the patient on each day during the insulin pump can be predicted only by combining the blood glucose value of the patient on eight o'clock every day. The diabetes insulin pump intensified treatment dose prediction model can help clinicians to formulate daily insulin dose advice and formulate individualized insulin dose during treatment with a pump according to the condition of a diabetic patient, so that the blood sugar of the diabetic patient is controlled to reach the standard, and hypoglycemia events are reduced.

Description

Technical Field

[0001] This invention relates to the field of biomedical technology, specifically to a dose prediction model for intensive insulin pump therapy for diabetes based on artificial intelligence algorithms and its application. Background Technology

[0002] Due to factors such as economic development, changes in dietary structure, sedentary lifestyles and reduced physical activity, obesity, aging population, and environmental pollution, type 2 diabetes is becoming increasingly prevalent worldwide, with its rising prevalence particularly evident in my country, a country experiencing rapid economic growth. Diabetes mellitus complicated by macrovascular complications (stroke, myocardial infarction, lower extremity vascular occlusion) and microvascular complications (retinopathy, nephropathy, neuropathy) leads to organ dysfunction, often progressive and irreversible, resulting in high rates of disability and mortality, placing a heavy burden on individuals, families, and society.

[0003] Therefore, to reduce overall healthcare costs, on the one hand, it is necessary to strengthen public education on diabetes prevention and control, especially among high-risk groups (obesity, family history of diabetes, prediabetes, etc.) to prevent the onset of type 2 diabetes and reduce the proportion of newly diagnosed diabetes patients each year. On the other hand, for newly diagnosed type 2 diabetes patients, especially those with high blood glucose at diagnosis (HbA1c ≥ 9%), there is evidence that short-term intensive insulin pump therapy can effectively reverse type 2 diabetes. After treatment, patients only need to maintain long-term drug-free remission through a healthy lifestyle and simple diet and exercise control. For non-newly diagnosed type 2 diabetes patients whose blood glucose is poorly controlled by hypoglycemic agents, short-term intensive insulin pump therapy can also partially restore pancreatic β-cell function and insulin sensitivity. After intensive therapy, the blood glucose control regimen can be simplified, and good blood glucose control can be achieved.

[0004] The essence of short-term intensive insulin pump therapy lies in its short duration (2-3 weeks), intensive approach (normalizing blood glucose), and long-term adherence to a healthy lifestyle. The principle behind newly diagnosed type 2 diabetes patients achieving long-term drug-free remission through short-term intensive insulin pump therapy is twofold. First, sufficient insulin rapidly normalizes the patient's blood glucose (fasting blood glucose <6 mmol / L, 2-hour postprandial blood glucose <8 mmol / L) and maintains normal blood glucose for 2 weeks. During this period, exogenous insulin inhibits endogenous insulin secretion, keeping pancreatic β-cells in a non-secreting state, allowing them sufficient rest, repair, and even transdifferentiation, thus improving their secretory function. Second, blood glucose normalization can improve insulin resistance caused by chronic hyperglycemia and hyperlipidemia, thereby effectively improving insulin sensitivity.

[0005] The effectiveness of short-term intensive insulin pump therapy depends primarily on the physician's understanding and experience with intensive therapy. Extensive evidence currently confirms that normalizing blood glucose levels during intensive therapy is the most crucial condition for improving pancreatic β-cell function and insulin sensitivity. In my country, endocrinology departments in major hospitals typically have between 20 and 50 insulin pumps, but most are used merely as a substitute for insulin injections, failing to achieve truly rapid and sustained normalization of blood glucose. Furthermore, due to limitations in hospital stays and bed turnover, as well as patient-related factors, the duration of pump use is often short, failing to allow pancreatic β-cells sufficient rest. Secondly, hypoglycemic reactions are not uncommon during short-term intensive insulin pump therapy, often mild and easily corrected. This is often a signal of restored insulin sensitivity and a basis for insulin dosage reduction. Many physicians' fear of hypoglycemia and reluctance to lower blood glucose levels is also one reason for treatment failure.

[0006] Furthermore, most current models for predicting insulin dosage require consideration of numerous complex factors, such as whether the patient has hypertension, hyperlipidemia, and cytokine levels (including AST and Cr levels), to predict insulin dosage in diabetic patients. Each time insulin dosage is predicted, the patient's cytokine levels need to be collected again and combined with the patient's clinical information to determine the appropriate insulin dose. In addition to being complex to operate, the model also requires continuous sampling and testing of the patient, resulting in significantly higher costs and time.

[0007] Therefore, there is an urgent need for a predictive model for insulin pump intensive therapy dosage to achieve more accurate and effective insulin pump intensive therapy, which can promote blood glucose normalization while avoiding hypoglycemic events during treatment. Summary of the Invention

[0008] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and to provide a dose prediction model for intensive insulin pump therapy for diabetes based on artificial intelligence algorithms and its application.

[0009] The primary objective of this invention is to provide a dose prediction model for intensive insulin pump therapy for diabetes based on artificial intelligence algorithms.

[0010] A second objective of this invention is to provide the application of any of the above-described insulin pump intensive therapy dose prediction models in guiding insulin pump intensive therapy.

[0011] A third objective of this invention is to provide a predictive system for guiding intensive insulin pump therapy for diabetes.

[0012] A fourth objective of this invention is to provide a computer-readable storage medium.

[0013] To achieve the above objectives, the present invention is implemented through the following solution:

[0014] In order to achieve more accurate and effective insulin pump intensive therapy, promote blood glucose normalization while minimizing the occurrence of hypoglycemic events, this invention requests protection for a diabetes insulin pump intensive therapy dose prediction model based on artificial intelligence algorithm, which includes prediction module 1, prediction module 2, prediction module 3 and result output module.

[0015] The prediction module 1 obtains the patient's age, waist circumference, body mass index, glycated hemoglobin and fasting blood glucose level before insulin pump initiation as input to the trained machine learning model 1, and obtains the patient's basal insulin rate and pre-meal bolus dose on the first day of insulin pump initiation.

[0016] The three pre-meal high doses include the high dose before breakfast, the high dose before lunch, and the high dose before dinner.

[0017] The prediction module 2 obtains the patient's 8 blood glucose values ​​on the first day of insulin pump use, and combines the patient's age, waist circumference, body mass index, glycated hemoglobin, fasting blood glucose level before insulin pump use, basal insulin rate on the first day of insulin pump use, and pre-meal bolus doses on the first day as input to the trained machine learning model 2 to obtain the patient's basal insulin rate and pre-meal bolus doses on the second day of insulin pump use.

[0018] The prediction module 3 obtains the patient's 8-point blood glucose values ​​on day n-2 of insulin pump use, and combines these with the patient's age, waist circumference, body mass index, glycated hemoglobin, fasting blood glucose level before insulin pump use, basal insulin rate on day n-2 of insulin pump use, pre-meal bolus dose on day n-2, 8-point blood glucose levels on day n-1 of insulin pump use, basal insulin rate on day n-1, and pre-meal bolus dose on day n-1, as input to the trained machine learning model 3, to obtain the patient's basal insulin rate and pre-meal bolus dose on day n of insulin pump use; where n is an integer ≥ 3.

[0019] The result output module is used to output the results obtained by prediction module 1, prediction module 2 and prediction module 3.

[0020] Among the combined indicators, the fasting blood glucose level before insulin pump administration is the fasting blood glucose level measured at the last time before insulin pump administration.

[0021] Preferably, the patient is a type 2 diabetic patient.

[0022] Preferably, the 8 blood glucose values ​​include blood glucose values ​​before and 2 hours after each of the three meals, before bedtime, and at 3 a.m.

[0023] Preferably, the training method for the trained machine learning model 1 is as follows:

[0024] Patients who underwent intensive insulin pump therapy were used as the training set samples. The basal insulin rate and the pre-meal bolus dose on the first day of insulin pump use in the training set samples were used as the prediction targets of the machine learning model. The machine learning model was trained using the training set samples to obtain the trained machine learning model 1.

[0025] The training method for the trained machine learning model 2 is as follows: the basal insulin rate and the pre-meal bolus dose on the second day of insulin pump use in the training set samples are used as the prediction targets of the machine learning model. The machine learning model is trained using the training set samples to obtain the trained machine learning model 2.

[0026] The training method for the trained machine learning model 3 is as follows: the basal insulin rate on day n and the pre-meal bolus dose on day n in the training set samples are used as the prediction targets of the machine learning model. The machine learning model is trained using the training set samples to obtain the trained machine learning model 3; where n is an integer ≥ 3.

[0027] More preferably, the machine learning model 1 is an LSTM model, the machine learning model 2 is an LSTM model, and the machine learning model 3 is an LSTM model.

[0028] More preferably, the training is K-fold cross-validation.

[0029] More preferably, the K-fold cross-validation is 5-fold cross-validation.

[0030] The diabetes insulin pump intensive therapy dose prediction model only needs to obtain the patient's age, waist circumference, body mass index, glycated hemoglobin, and fasting blood glucose level before insulin pump initiation to obtain the basal insulin rate and pre-meal maximal dose on the first day of insulin pump initiation. Then, by combining the patient's 8 blood glucose values ​​monitored during insulin pump initiation (which can be derived from continuous glucose monitoring or fingertip capillary blood glucose measurement) and the insulin dose of the previous 1 or 2 days (including basal rate and pre-meal maximal dose), accurate prediction of the basal insulin rate and pre-meal maximal dose for each day during insulin pump initiation can be achieved. When using the diabetes insulin pump intensive therapy dose prediction model, only one set of the patient's physical indicators and the patient's 8 blood glucose values ​​per day are needed to accurately predict the daily basal insulin rate and pre-meal maximal dose during insulin pump initiation. This reduces the number of indicators combined (reducing the amount of data) while achieving more accurate and effective insulin pump intensive therapy.

[0031] This invention also claims protection for the application of any of the above-described insulin pump intensive therapy dose prediction models in guiding insulin pump intensive therapy.

[0032] Preferably, the guided insulin pump intensive therapy refers to the basal insulin rate and pre-meal bolus dose during the guided insulin pump intensive therapy process.

[0033] The present invention also claims protection for a prediction system for guiding intensive insulin pump therapy for diabetes, comprising an acquisition unit, a storage unit and a processing unit, wherein the acquisition unit is used to acquire the patient's age, waist circumference, body mass index, glycated hemoglobin and fasting blood glucose level before insulin pump administration.

[0034] The storage unit stores program instructions that can be executed by the processing unit;

[0035] The processing unit includes any of the above-described diabetes insulin pump intensive therapy dose prediction models.

[0036] When the program instruction is executed by the processing unit, the patient's age, waist circumference, body mass index, glycated hemoglobin, and fasting blood glucose level before insulin pump administration, obtained by the acquisition unit, are input into any of the above-mentioned diabetes insulin pump intensive treatment dose prediction models to obtain the results output by prediction module 1, prediction module 2, and prediction module 3 in the diabetes insulin pump intensive treatment dose prediction model.

[0037] The present invention also claims protection for a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described insulin pump intensive therapy dose prediction models for diabetes and / or the above-described prediction systems.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] This invention provides a dose prediction model for intensive insulin pump therapy in diabetes based on an artificial intelligence algorithm. This model only requires the patient's age, waist circumference, body mass index, glycated hemoglobin, and fasting blood glucose level before starting insulin pump therapy to obtain the basal insulin rate and pre-meal bolus doses on the first day of insulin pump therapy. Then, by combining this with the patient's eight blood glucose levels, accurate predictions of the daily basal rate and pre-meal bolus doses during insulin pump therapy can be achieved. The model input is relatively simple and easy to obtain. When using this intensive insulin pump therapy dose prediction model, accurate and effective intensive insulin pump therapy can be achieved, promoting blood glucose normalization while avoiding hypoglycemic events during treatment. Attached Figure Description

[0040] Figure 1 The graphs show the mean squared error and mean absolute error results during the training process of machine learning model 1 in Example 1; A is the mean squared error result graph; B is the mean absolute error result graph.

[0041] Figure 2 The graphs show the mean squared error and mean absolute error results during the training process of machine learning model 2 in Example 1; A is the mean squared error result graph; B is the mean absolute error result graph.

[0042] Figure 3 The graphs show the mean squared error and mean absolute error results during the training process of machine learning model 3 in Example 1; A is the mean squared error result graph; B is the mean absolute error result graph. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Unless otherwise specified, the experimental methods used in the following embodiments are conventional methods.

[0044] The patient information used in the embodiments of the present invention has been disclosed to the patients and their authorization has been obtained before use; all experiments in the embodiments of the present invention have been approved by the ethics committee of the First Affiliated Hospital of Sun Yat-sen University.

[0045] Example 1: Establishment of a dose prediction model for intensive insulin pump therapy in diabetes based on artificial intelligence algorithms

[0046] I. Experimental Methods

[0047] 1. Acquisition and processing of sample data

[0048] Clinical information was obtained from 2,994 patients with type 2 diabetes who received short-term intensive insulin pump therapy at the Department of Endocrinology, The First Affiliated Hospital of Sun Yat-sen University between September 1, 2001 and December 31, 2022. (The screening criteria were: discharge diagnosis of type 2 diabetes, admission and discharge departments were both the Department of Endocrinology, and all patients with type 2 diabetes who received short-term intensive insulin pump therapy during hospitalization).

[0049] Clinical information collection includes:

[0050] (1) Clinical characteristics of patients required to build an insulin dose prediction model: age, waist circumference, body mass index (BMI), glycated hemoglobin (HbA1c), and fasting blood glucose (FBG, i.e., the last fasting blood glucose before insulin pump administration) of patients with type 2 diabetes.

[0051] (2) During the intensive insulin pump therapy in hospital, capillary blood glucose at 8 o'clock every day, including blood glucose before three meals, blood glucose 2 hours after three meals, blood glucose before bedtime and blood glucose at 3 o'clock in the morning; where blood glucose monitoring values ​​were missing due to the patient's out-of-hospital examination, the missing values ​​were supplemented by the nearest mean difference method.

[0052] (3) Daily insulin pump dose orders during insulin pump intensive therapy, including basal insulin rate and pre-meal bolus dose, as well as the start and end times of the orders (i.e., the start and end times of insulin administration).

[0053] Based on the course of disease and pre-admission blood glucose control regimen of patients with type 2 diabetes, patients are classified and labeled. The specific classification rules are as follows:

[0054] a: Newly diagnosed type 2 diabetes: Patients who have never used hypoglycemic treatment after diagnosis, or whose hypoglycemic treatment did not exceed 14 days before admission, and who used insulin pump for short-term intensive treatment after admission (Group A, 2068 cases);

[0055] b: Type 2 diabetes poorly controlled by oral hypoglycemic agents: Non-newly diagnosed type 2 diabetes patients who had been treated with oral hypoglycemic agents for more than 1 month before admission, had poor blood glucose control, and were admitted to the hospital for short-term intensive insulin pump therapy (Group B, 491 cases).

[0056] c: Type 2 diabetes with poor glycemic control using insulin-containing glucose-lowering regimens: Non-newly diagnosed type 2 diabetes patients who had been treated with glucose-lowering regimens including insulin for more than 1 month before admission but whose glycemic control was still poor, and who received short-term intensive insulin pump therapy after admission (C1 group, 435 cases).

[0057] All patients in groups A, B, and C1 (a total of 2994 cases) were compiled as the sample dataset for the diabetes insulin pump intensive therapy dose prediction model, and the feature values ​​of each patient in the sample dataset were obtained. The feature values ​​included: age (at the time of insulin pump initiation), waist circumference (at the time of insulin pump initiation), body mass index (BMI, at the time of insulin pump initiation), HbA1c (before insulin pump initiation), fasting blood glucose (FBG) before insulin pump initiation, 8-point capillary blood glucose (CBG) data during insulin pump initiation (fasting blood glucose, 2 hours after breakfast blood glucose, pre-lunch blood glucose, 2 hours after lunch blood glucose, pre-dinner blood glucose, 2 hours after dinner blood glucose, bedtime blood glucose, and 3 am blood glucose) and insulin dose adjustment order information (basal insulin rate, pre-breakfast maximal dose, pre-lunch maximal dose, and pre-dinner maximal dose, and the start and stop times of the order).

[0058] CBG data refers to the 8 capillary blood glucose values ​​monitored daily during insulin pump use; basal insulin rate includes the basal insulin rate on the day the insulin pump is started and daily during insulin pump use; pre-meal bolus doses include the pre-meal bolus doses (breakfast, lunch, and dinner) on the day the insulin pump is started and daily during insulin pump use.

[0059] For the feature values ​​of each patient in the sample dataset, abnormal data were deleted and samples with too many missing values ​​(such as >50% missing CBG data) were removed. Missing CBG values ​​during intensive insulin pump therapy were filled using the nearest mean imputation method.

[0060] The specific method for filling missing values ​​is as follows: calculate the average value of the nearest valid blood glucose measurement to the missing time point, and fill the missing value with this average value. However, based on the time characteristics of blood glucose changes (such as pre-meal, post-meal, and nighttime patterns), it is necessary to give priority to valid measurements of the same type of time point (such as "before breakfast"), i.e., the same time point on adjacent days: such as the same type of time point on the previous day and the next day.

[0061] 2. Model Establishment

[0062] (1) Establishment of Machine Learning Model 1 (Day 1 Model)

[0063] Machine learning model 1 was used to predict the basal insulin rate and pre-meal bolus doses on the first day of insulin pump use in patients with type 2 diabetes. The model training method is as follows:

[0064] Using the age (at the time of insulin pump initiation), waist circumference (at the time of insulin pump initiation), BMI (at the time of insulin pump initiation), HbA1c (before insulin pump initiation), and fasting blood glucose (FBG) before insulin pump initiation in the sample dataset obtained in step 1 as input to the LSTM neural network model, the basal insulin rate and the pre-meal maximal doses (pre-breakfast, pre-lunch, and pre-dinner) of each type 2 diabetic patient on the first day of insulin pump initiation were used as the prediction targets of the LSTM neural network model. Combining the basal insulin rate and pre-meal maximal doses (actual results, included in the feature values) of each type 2 diabetic patient on the first day of insulin pump initiation, the LSTM neural network model was trained to obtain machine learning model 1 (first day model). Machine learning model 1 was used to predict the basal insulin rate and pre-meal maximal doses of type 2 diabetic patients on the first day of insulin pump initiation.

[0065] The training process employed 5-fold cross-validation, which involved randomly dividing patients in the sample dataset into 5 equal parts. In each training iteration, 4 parts were used as the training set, and the remaining part was used as the validation set to train the model, until all 5 parts were used as the validation set. A total of 200 training epochs were set during the training process, with 200 samples processed in each batch. The Adam optimizer was used during training, with a learning rate set to 0.001.

[0066] Simultaneously record the mean squared error (MSE) and mean absolute error (MAE) during the training process of machine learning model 1.

[0067] (2) Establishment of Machine Learning Model 2 (One-Day Model)

[0068] Machine learning model 2 was used to predict the basal insulin rate and pre-meal bolus doses on day 2 of insulin pump use in patients with type 2 diabetes. The difference between the training methods of machine learning model 2 and machine learning model 1 is as follows:

[0069] The following parameters were used as inputs to the LSTM neural network model for each type 2 diabetes patient in the sample dataset from step 1: age (at the time of insulin pump initiation), waist circumference (at the time of insulin pump initiation), BMI (at the time of insulin pump initiation), HbA1c (before insulin pump initiation), fasting blood glucose (FBG) before insulin pump initiation, blood glucose values ​​at 8 points on day m-1 of insulin pump initiation (fasting blood glucose, 2 hours post-breakfast blood glucose, pre-lunch blood glucose, 2 hours post-lunch blood glucose, pre-dinner blood glucose, 2 hours post-dinner blood glucose, bedtime blood glucose, and 3 AM blood glucose), basal insulin rate on day m-1 of insulin pump initiation, and pre-meal bolus doses on day m-1 of insulin pump initiation. The model was trained using the basal insulin rate on day m of insulin pump initiation and the pre-meal bolus doses (pre-breakfast bolus dose, pre-lunch bolus dose, and pre-dinner bolus dose) of each type 2 diabetes patient. The rest of the training method was exactly the same, resulting in machine learning model 2 (one-day model); where m≥2. Machine learning model 2 was only used to predict the basal insulin rate and pre-meal bolus doses on day 2 of insulin pump initiation for type 2 diabetes patients.

[0070] Simultaneously record the mean squared error (MSE) and mean absolute error (MAE) during the training process of machine learning model 2.

[0071] (3) Establishment of Machine Learning Model 3 (Two-Day Model)

[0072] Machine learning model 3 was used to predict the basal insulin rate and pre-meal bolus doses on day n of insulin pump use in patients with type 2 diabetes (n is an integer ≥3). The difference between the training method of machine learning model 3 and that of machine learning model 1 is as follows:

[0073] The following parameters were used as input to the LSTM neural network model for each type 2 diabetes patient in the sample dataset from step 1: age (at the time of insulin pump initiation), waist circumference (at the time of insulin pump initiation), BMI (at the time of insulin pump initiation), HbA1c (before insulin pump initiation), fasting blood glucose (FBG) before insulin pump initiation, 8-point blood glucose value on day n-2 of insulin pump initiation, basal insulin rate on day n-2 of insulin pump initiation, pre-meal bolus doses on day n-2 of insulin pump initiation, 8-point blood glucose value on day n-1 of insulin pump initiation, basal insulin rate on day n-1 of insulin pump initiation, and pre-meal bolus doses on day n-1 of insulin pump initiation. The model was trained using the basal insulin rate and pre-meal bolus doses (pre-breakfast, pre-lunch, and pre-dinner) on day n of insulin pump initiation for each type 2 diabetes patient. The rest of the training method was exactly the same, resulting in machine learning model 3 (two-day model). Machine learning model 3 was used to predict the basal insulin rate and pre-meal bolus doses on day n of insulin pump initiation for type 2 diabetes patients, where n≥3.

[0074] Simultaneously record the mean squared error (MSE) and mean absolute error (MAE) during the training process of machine learning model 3.

[0075] II. Experimental Results

[0076] The mean squared error and mean absolute error results during the training process of machine learning model 1 are shown in the figure below. Figure 1 As shown, Figure 1 In the figure, A represents the mean square error result. Figure 1 B in the figure represents the mean absolute error result.

[0077] The mean squared error and mean absolute error results during the training process of machine learning model 2 are shown in the figure below. Figure 2 As shown, Figure 2 In the figure, A represents the mean square error result. Figure 2 B in the figure represents the mean absolute error result.

[0078] The mean squared error and mean absolute error results during the training process of machine learning model 3 are shown in the figure below. Figure 3 As shown, Figure 3 In the figure, A represents the mean square error result. Figure 3 B in the figure represents the mean absolute error result.

[0079] The results show that during the training of machine learning models 1 to 3, the mean squared error (MSE) and mean absolute error (MAE) of both the training and validation sets decreased with the increase of training epochs. Furthermore, the MSE and MAE were relatively flat in the later stages of model training, indicating that machine learning models 1 to 3 were effectively learning during the training process. In addition, there was no instance of validation loss increasing while training loss decreased, indicating that the models had a good fit.

[0080] Results show that machine learning models 1 through 3 can all accurately predict the target.

[0081] Example 2: A dose prediction model for intensive insulin pump therapy in diabetes based on artificial intelligence algorithms

[0082] An enhanced insulin pump therapy for diabetes based on artificial intelligence algorithms includes prediction module 1, prediction module 2, prediction module 3, and result output module.

[0083] The prediction module 1 is used to predict the basal insulin rate and the pre-meal bolus doses on the first day of insulin pump use in patients with type 2 diabetes. It obtains the patient's age, waist circumference, body mass index (BMI), glycated hemoglobin (HbA1c), and fasting blood glucose level before insulin pump use as inputs to machine learning model 1 in Example 1. Machine learning model 1 is used to obtain the patient's basal insulin rate and pre-meal bolus doses (pre-breakfast bolus, pre-lunch bolus, and pre-dinner bolus) on the first day of insulin pump use.

[0084] Prediction module 2 is used to predict the basal insulin rate and pre-meal bolus dose on the second day after insulin pump administration in patients with type 2 diabetes. It uses the patient's age, waist circumference, body mass index, glycated hemoglobin, fasting blood glucose level before insulin pump administration, basal insulin rate on the first day after insulin pump administration, pre-meal bolus dose on the first day after insulin pump administration, and 8-point blood glucose values ​​on the first day after insulin pump administration as inputs to machine learning model 2 in Example 1. Machine learning model 2 is used to obtain the basal insulin rate and pre-meal bolus dose on the second day after insulin pump administration in patients with type 2 diabetes. The 8-point blood glucose values ​​are obtained through continuous glucose monitoring or fingertip capillary blood glucose measurement.

[0085] Prediction module 3 is used to predict the basal insulin rate and pre-meal bolus dose on day n of insulin pump use in patients with type 2 diabetes (n is an integer ≥3). It takes the patient's age, waist circumference, body mass index, glycated hemoglobin, fasting blood glucose level before insulin pump use, basal insulin rate on day n-2 of insulin pump use, pre-meal bolus dose on day n-2 of insulin pump use, 8-point blood glucose value on day n-2 of insulin pump use, basal insulin rate on day n-1 of insulin pump use, pre-meal bolus dose on day n-1 of insulin pump use, and 8-point blood glucose value on day n-1 of insulin pump use as input to machine learning model 3 in Example 1. Machine learning model 3 is used to obtain the basal insulin rate and pre-meal bolus dose on day n of insulin pump use (n is an integer ≥3). The 8-point blood glucose value is obtained through continuous glucose monitoring or fingertip capillary blood glucose measurement.

[0086] The results output module is used to output the results obtained by prediction module 1, prediction module 2 and prediction module 3.

[0087] Prediction modules 1, 2, and 3 can be used individually or in combination in the insulin pump intensive therapy dose prediction model for diabetes. Based on the output of the results module, the basal insulin rate on day 1, the pre-meal bolus dose on day 1, the basal insulin rate on day 2, the pre-meal bolus dose on day 2, the basal insulin rate on day n, and the pre-meal bolus dose on day n (n≥3) can be used to prescribe insulin for patients with type 2 diabetes. The model can be used to intensively treat patients with type 2 diabetes using an insulin pump and measure the patient's blood glucose level at 8 points every day.

[0088] Example 3: Application of a Diabetic Insulin Pump Intensive Therapy Dosage Prediction Model Based on Artificial Intelligence Algorithm

[0089] I. Experimental Methods

[0090] Twenty-four patients with type 2 diabetes who received intensive insulin pump therapy and were hospitalized in the Department of Endocrinology, The First Affiliated Hospital of Sun Yat-sen University from January 1, 2025 to April 30, 2025 were selected as the experimental group (predictive model group) and another 18 patients with type 2 diabetes who were hospitalized during the same period were selected as the control group (experimental group). All samples were approved by the Ethics Committee of The First Affiliated Hospital of Sun Yat-sen University (ethics number

[2024] 836) before the experiment, and all patients signed informed consent forms.

[0091] Clinical characteristics of patients in the experimental and control groups were recorded separately, including age, sex, duration of diabetes, waist circumference, waist-to-hip ratio, body mass index (BMI), systolic blood pressure, diastolic blood pressure, fasting blood glucose before pump administration, glycated hemoglobin, random C-peptide, serum creatinine, estimated glomerular filtration rate, urinary microalbumin / creatinine ratio, alanine aminotransferase, aspartate aminotransferase, lactate dehydrogenase, gamma-glutamyl transferase, plasma albumin, serum uric acid, triglycerides, total cholesterol, low-density cholesterol, high-density cholesterol, white blood cell count, hemoglobin concentration, smoking rate, diabetic retinopathy, diabetic peripheral neuropathy, carotid artery sclerosis, and lower extremity arteriosclerosis. All clinical characteristics are expressed as mean ± standard deviation or median (25th percentile, 75th percentile). Diabetic retinopathy, diabetic peripheral neuropathy, carotid artery sclerosis, and lower extremity arteriosclerosis are expressed as the percentage of patients with these conditions in each group.

[0092] Experimental Group: Age, waist circumference, body mass index (BMI), glycated hemoglobin (HbA1c), and fasting blood glucose levels before insulin pump administration were obtained for each type 2 diabetic patient in the experimental group. The insulin pump intensive treatment dose prediction model based on artificial intelligence algorithms, as shown in Example 2, was used for prediction. Based on the output of the model, the attending physician prescribed medication daily according to the basal rate and pre-meal bolus doses given by the model. Intensive insulin pump therapy was administered to the experimental group, and blood glucose levels were recorded during insulin pump administration. Simultaneously, the average blood glucose, average fasting blood glucose, average 2-hour postprandial blood glucose (the average of blood glucose levels 2 hours after three meals), TIR1 (the proportion of times blood glucose was in the range of 3.9–10.0 mmol / L out of the total monitoring times), and TIR2 (fasting blood glucose <6 mmol / L and <10.0 mmol / L) were calculated during the intensive treatment period. The following data were collected: the proportion of times blood glucose was <8 mmol / L 2 hours after a meal out of the total number of monitoring sessions; TBR1 (the proportion of times blood glucose was <3.9 mmol / L 2 hours after a meal out of the total number of monitoring sessions); TBR2 (the proportion of times blood glucose was <3.0 mmol / L 2 hours after a meal out of the total number of monitoring sessions); mean blood glucose standard deviation (calculated after calculating the blood glucose standard deviation of each sample in the group, and then calculating the mean blood glucose standard deviation of the group); fasting blood glucose standard deviation; 2-hour postprandial blood glucose standard deviation; mean blood glucose variability; fasting blood glucose variability; 2-hour postprandial blood glucose variability; and fasting blood glucose value on the day after removing the insulin pump; insulin dosage during the period when the insulin pump was in use was recorded for the experimental group, including the total insulin dose, basal insulin rate, pre-meal maximal dose, and insulin dosage per kilogram of body weight on the first day of using the insulin pump, the day of maximum insulin dose, and the day of removal of the pump; and the number of days of hospitalization, the number of days with the insulin pump, and weight changes during hospitalization for the experimental group.

[0093] Control group: The control group sample was determined by a clinician with extensive experience in diabetes treatment based on clinical experience and blood glucose control targets. The daily insulin dose for the control group sample was determined, an insulin pump was used, and the daily insulin injection dose was adjusted according to blood glucose levels. Data such as blood glucose levels, insulin dosage, length of hospital stay, number of days with insulin pump, and weight changes during hospitalization were recorded for the control group sample (the same data as those recorded for the experimental group).

[0094] II. Experimental Results

[0095] The clinical characteristics of the experimental group and the control group are shown in Table 1.

[0096] Table 1 Clinical characteristics

[0097]

[0098] The results showed that there were no significant differences between the experimental group and the control group in terms of body mass index, waist circumference, fasting blood glucose before pump administration, glycated hemoglobin, random C-peptide, blood lipids (triglycerides, total cholesterol, low-density cholesterol and high-density cholesterol), blood uric acid, and diabetic complications (diabetic retinopathy, diabetic peripheral neuropathy, carotid artery sclerosis, lower extremity arteriosclerosis), indicating that the experimental group and the control group were comparable.

[0099] Table 2 shows the number of days of hospitalization, number of days with insulin pump, and weight changes during hospitalization for each sample in the experimental and control groups.

[0100] Table 2. Length of hospital stay, number of days with insulin pump, and weight change during hospitalization for each sample in the experimental and control groups.

[0101] variable Control group (n=18) Experimental group (n=24) p-value Length of stay (days) 8(7，9) 8(7，10) 0.74 Number of days using an insulin pump (days) 6(6，6) 6(6，6) 0.51 Weight change during hospitalization (kg) -0.1(-0.9，0.6) -0.6(-1.5，0.1) 0.23

[0102] The results showed that there were no significant differences in the number of days of hospitalization and the number of days with insulin pumps between the experimental group and the control group.

[0103] Table 3 shows the blood glucose levels during insulin pump administration in both the experimental and control groups.

[0104] Table 3. Blood glucose levels during insulin pump administration in the experimental and control groups.

[0105] Blood glucose levels during pump administration Control group (n=18) Experimental group (n=24) p-value Average blood glucose (mmol / L) 7.41(6.77，8.55) 7.30(6.52，8.57) 0.49 Mean fasting blood glucose (mmol / L) 6.40(5.94，7.00) 5.38(4.94，6.28) 0.01 Average 2-hour postprandial blood glucose (mmol / L) 8.72(7.60，9.42) 8.67(7.49，9.98) 0.85 TIR1 (%) 73.92±5.21 82.21±2.77 0.08 TIR2 (%) 34.4(20.0，47.9) 41.55(29.5，64.1) 0.12 TBR1 (%) 2.92±1.25 1.88±0.49 0.22 TBR2 (%) 0.71±0.35 0.07±0.07 0.04 Mean blood glucose standard deviation (mmol / L) 2.48(2.22，3.34) 2.41(1.96，3.62) 0.81 Fasting blood glucose standard deviation (mmol / L) 0.97(0.61，1.51) 0.82(0.45，1.25) 0.38 Standard deviation of blood glucose 2 hours after a meal (mmol / L) 2.78(2.09，3.83) 2.65(1.92，3.59) 0.62 Mean glycemic variability (%) 0.32(0.27，0.40) 0.37(0.30，0.42) 0.49 Fasting blood glucose variability (%) 0.13(0.10，0.22) 0.13(0.09，0.21) 0.60 Postprandial 2-hour blood glucose variability (%) 0.29(0.25，0.36) 0.30(0.25，0.37) 0.83 Fasting blood glucose (mmol / L) the day after pump removal 7.55±0.62 7.32±0.60 0.44

[0106] Note: TIR1 refers to the proportion of blood glucose monitoring times within the range of 3.9-10.0 mmol / L out of the total number of blood glucose monitoring times; TIR2 refers to the proportion of blood glucose monitoring times within the range of fasting blood glucose <6 mmol / L and 2-hour postprandial blood glucose <8 mmol / L out of the total number of blood glucose monitoring times; TBR1 refers to the proportion of blood glucose monitoring times within the range of 3.9 mmol / L out of the total number of blood glucose monitoring times; TBR2 refers to the proportion of blood glucose monitoring times within the range of 3.0 mmol / L out of the total number of blood glucose monitoring times.

[0107] The results showed that the average fasting blood glucose in the experimental group was significantly lower than that in the control group while using an insulin pump. In terms of trends, the average blood glucose, average 2-hour postprandial blood glucose, and fasting blood glucose on the day after pump removal were all lower in the experimental group than in the control group, while TIR1 and TIR2 were higher in the experimental group than in the control group, although the differences were not statistically significant (due to the small sample size). At the same time, TBR2 (the proportion of moderate hypoglycemic events) in the experimental group was also significantly lower than that in the control group, while TBR1 (the proportion of mild hypoglycemic events) was not significantly different from that in the control group.

[0108] Results show that when using the insulin pump intensive therapy dose prediction model shown in Example 2 to predict the dose of type 2 diabetes patients using insulin pumps, it can not only accurately predict the required insulin dose and achieve better blood glucose control, but also avoid the occurrence of hypoglycemic events during treatment.

[0109] The insulin dosage in the experimental and control groups is shown in Table 4.

[0110] Table 4. Insulin dosage in the experimental and control groups.

[0111]

[0112] The results showed that there was no significant difference in insulin dosage between the experimental group and the control group, and the time to reach the maximum daily dose was also similar in both groups. This indicates that the insulin dose predicted by the insulin pump intensive therapy dose prediction model for diabetes shown in Example 2 is basically consistent with the insulin dose given by the clinical physician's experience in the control group. This shows that the insulin pump intensive therapy dose prediction model for diabetes shown in Example 2 can accurately predict the insulin dose, conforms to clinical standards, and the predicted dose is reasonable.

[0113] Example 4: A predictive system for guiding intensive insulin pump therapy in diabetes.

[0114] A predictive system for guiding intensive insulin pump therapy for diabetes includes an acquisition unit, a storage unit, and a processing unit. The acquisition unit is used to acquire the age, waist circumference, body mass index, glycated hemoglobin, and fasting blood glucose level before insulin pump administration in patients with type 2 diabetes.

[0115] The storage unit stores program instructions that can be executed by the processing unit;

[0116] The processing unit includes the diabetes insulin pump intensive treatment dose prediction model described in Example 2;

[0117] When the program instructions are executed by the processing unit, the patient's age, waist circumference, body mass index, glycated hemoglobin, and fasting blood glucose level before insulin pump administration, obtained by the acquisition unit, are input into the diabetes insulin pump intensive treatment dose prediction model described in Example 2, and the results output by prediction module 1, prediction module 2, and prediction module 3 in the diabetes insulin pump intensive treatment dose prediction model are obtained.

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description and ideas, and it is neither necessary nor possible to exhaustively describe all implementation methods here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A dose prediction model for intensive insulin pump therapy in diabetes based on artificial intelligence algorithms, characterized in that, It includes prediction module 1, prediction module 2, prediction module 3, and result output module; The prediction module 1 obtains the patient's age, waist circumference, body mass index, glycated hemoglobin and fasting blood glucose level before insulin pump initiation as input to the trained machine learning model 1, and obtains the patient's basal insulin rate and pre-meal bolus dose on the first day of insulin pump initiation. The three pre-meal high doses include the high dose before breakfast, the high dose before lunch, and the high dose before dinner. The prediction module 2 obtains the patient's 8 blood glucose values ​​on the first day of insulin pump use, and combines the patient's age, waist circumference, body mass index, glycated hemoglobin, fasting blood glucose level before insulin pump use, basal insulin rate on the first day of insulin pump use, and pre-meal bolus doses on the first day as input to the trained machine learning model 2 to obtain the patient's basal insulin rate and pre-meal bolus doses on the second day of insulin pump use. The prediction module 3 obtains the patient's 8-point blood glucose values ​​on day n-2 of insulin pump use, and combines these with the patient's age, waist circumference, body mass index, glycated hemoglobin, fasting blood glucose level before insulin pump use, basal insulin rate on day n-2 of insulin pump use, pre-meal bolus dose on day n-2, 8-point blood glucose values ​​on day n-1 of insulin pump use, basal insulin rate on day n-1, and pre-meal bolus dose on day n-1, as input to the trained machine learning model 3, to obtain the patient's basal insulin rate and pre-meal bolus dose on day n of insulin pump use; where n is an integer ≥ 3. The result output module is used to output the results obtained by prediction module 1, prediction module 2 and prediction module 3.

2. The insulin pump intensive therapy dose prediction model for diabetes according to claim 1, characterized in that, The patient is a type 2 diabetic.

3. The insulin pump intensive therapy dose prediction model for diabetes according to claim 1, characterized in that, The training method for the trained machine learning model 1 is as follows: Patients who underwent intensive insulin pump therapy were used as the training set samples. The basal insulin rate and the pre-meal bolus dose on the first day of insulin pump use in the training set samples were used as the prediction targets of the machine learning model. The machine learning model was trained using the training set samples to obtain the trained machine learning model 1. The training method for the trained machine learning model 2 is as follows: the basal insulin rate and the pre-meal bolus dose on the second day of insulin pump use in the training set samples are used as the prediction targets of the machine learning model. The machine learning model is trained using the training set samples to obtain the trained machine learning model 2. The training method for the trained machine learning model 3 is as follows: the basal insulin rate on day n and the pre-meal bolus dose on day n in the training set samples are used as the prediction targets of the machine learning model. The machine learning model is trained using the training set samples to obtain the trained machine learning model 3; where n is an integer ≥ 3.

4. The insulin pump intensive therapy dose prediction model for diabetes according to claim 3, characterized in that, The machine learning model 1 is an LSTM model, the machine learning model 2 is an LSTM model, and the machine learning model 3 is an LSTM model.

5. The insulin pump intensive therapy dose prediction model for diabetes according to claim 3, characterized in that, The training method is K-fold cross-validation.

6. The insulin pump intensive therapy dose prediction model for diabetes according to claim 5, characterized in that, The K-fold cross-validation is a 5-fold cross-validation.

7. The application of the insulin pump intensive therapy dose prediction model according to any one of claims 1 to 6 in guiding insulin pump intensive therapy.

8. The application according to claim 7, characterized in that, The guidance for intensive insulin pump therapy refers to the basal insulin rate and pre-meal bolus doses during the intensive insulin pump therapy process.

9. A predictive system for guiding intensive insulin pump therapy in diabetes, comprising an acquisition unit, a storage unit, and a processing unit, characterized in that, The acquisition unit is used to acquire the patient’s age, waist circumference, body mass index, glycated hemoglobin and fasting blood glucose level before insulin pump administration when the patient is using the insulin pump. The storage unit stores program instructions that can be executed by the processing unit; The processing unit includes the diabetes insulin pump intensive therapy dose prediction model as described in any one of claims 1 to 6; When the program instruction is executed by the processing unit, the patient's age, waist circumference, body mass index, glycated hemoglobin, and fasting blood glucose level before insulin pump administration, obtained by the acquisition unit, are input into the diabetes insulin pump intensive treatment dose prediction model according to any one of claims 1 to 6, so as to obtain the results output by prediction module 1, prediction module 2, and prediction module 3 in the diabetes insulin pump intensive treatment dose prediction model.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the diabetes insulin pump intensive therapy dose prediction model according to any one of claims 1 to 6 and / or the prediction system according to claim 9.

Images