The actual situation is that statistical analysis is carried out on collected original data of insulin pump treatment for diabetes in the trimethyl hospital, the clinical effect of the insulin pump after initial dose treatment actually formulated by endocrine professional doctors is taken as a research object, and the clinical effect is divided into 'control excellent', 'control in' and 'control poor' according to the deviation of the actual fasting blood glucose of a patient from a blood glucose control target after the patient receives the dose treatment. Wherein, the fasting blood glucose in the blood glucose control target interval is defined as 'control optimal'; a deviation of blood glucose from the control target interval, but less than 2mmol/l, is defined as "in control", and an occurrence of hypoglycemia or a deviation from the upper limit of the control target by 2mmol/l is defined as "poor control". The results show that the blood glucose results in the original data are 21.3% excellent, 32.5% excellent and 46.2% poor. The results show that even the endocrine professional in the third hospital, given the patient's initial insulin dose empirically, only a few patients achieve a second sky fasting glucose optimum. Therefore, the initial dose of the insulin pump actually prescribed by the endocrine professional in the raw data is mostly not the originally optimal dose for that patient. The physical, pathological, diabetic and biochemical information of each patient are different, and theoretically, each patient corresponds to an independent optimal dosage. The optimal dosage is a necessary condition for supervising learning (super learning) in model training and improving the prediction accuracy of the model. The existing basic amount of the original insulin pump in the original data needs to be corrected to be used as a sample for machine learning, the step is the key for improving the prediction accuracy, the evaluation needs to be carried out according to the disease information of the diabetic patient, the blood sugar control target of the patient is determined according to the clinical conditions of different patients, and the basic amount of the original insulin pump in the database is corrected.

The existing theoretical dose formulation process is divided into two specific cases:

1) calculation of insulin doses for patients not treated with insulin are set according to the insulin doses for the different diabetes types:

type 1 diabetes mellitus: an insulin pump theoretical dose (U) is weight (kg) × (0.4-0.5);

type 2 diabetes: the theoretical dosage (U) of insulin pump is weight (kg) × (0.5-1.0).

2) A patient who has received medication is pumped with the theoretical pump insulin dose (U) (total amount of insulin used throughout the day + equivalent dose of non-insulin hypoglycemic agent) multiplied by a corresponding factor as described in the background.

The actual clinical endocrinologist formulates the initial process of insulin pump: the dose is set according to the theory, and the actually formulated insulin dose is obtained by combining the increase and decrease of clinical experience.

Namely: the clinician sets the insulin pump basic amount as the theoretical dose plus the clinical experience increment and decrement

Specifically, the type 2 diabetic can evaluate the disease information of the diabetic according to the Chinese guide for the prevention and treatment of type 2 diabetes, and the blood sugar control target is formulated according to the clinical conditions of different patients, wherein the control target comprises general control, strict control and loose control. Wherein the general control requires that the glycosylated hemoglobin HbA1c is less than 7%, the strict control requires that the glycosylated hemoglobin HbA1c is less than 6.5%, and the loose control requires that the glycosylated hemoglobin HbA1c is less than 8%. Similarly, type 1 diabetic patients and gestational diabetic patients define individualized glycemic control goals according to the corresponding clinical guidelines.

The principle and process of insulin pump basal volume correction is as follows:

the initial insulin dose has the following logical relationship with the next day of fasting glucose: initial insulin dose (basal amount) appropriate → fasting glycemia is excellent the next day; the initial insulin dose (basal amount) is slightly larger or smaller → the second day of fasting plasma glucose, the initial insulin dose (basal amount) is excessive or insufficient → the second day of fasting plasma glucose is poor. Therefore, we need to correct the data of the original insulin pump basic amount in fasting blood sugar and poor patients in the database before training.

Based on the actual insulin dosage of the first day of insulin pump treatment of the diabetic patient in the blood sugar control target and the second sky abdominal blood sugar value after insulin pump treatment in the database, correcting the original insulin pump basic dosage of the diabetic patient in the database, wherein:

in addition, when the patient is gestational diabetes, the correction method is as follows:

it should be noted that, by correcting, the original insulin initial dose in the database is corrected up and down according to the actual blood sugar deviation from the control target. The scheme provides a logic scheme for large-data batch processing, but the blood sugar control target cannot be generalized and is individualized due to the diversity and complexity of diabetes diseases. In order to further improve the accuracy, the diabetes specialist can perform manual review and re-correction, the senior professional endocrine specialist can look over the original insulin pump basic amount, the insulin amount during meal, the blood sugar before meal and at night, the treated fasting blood sugar and the specific medical history retrospectively, and the clinical experience is combined to perform manual adjustment and education, so that the insulin pump basic amount in the database is just like the guidance of the specialist, and the optimal dosage of the patient is approached as much as possible after optimization.

EXAMPLE five

In this embodiment, in the training of the BP neural network, a training sample of a training set is used as a training object of the BP neural network, a test sample of a test set is used as a verification object of a model, and supervised machine learning training is performed, as shown in fig. 3, the method specifically includes the following steps:

setting the number N of layers and the number M of nodes of hidden neurons in the multilayer perceptron, wherein: n is more than or equal to 2 and less than or equal to 3, M is more than or equal to 5 and less than or equal to 35, a training stopping rule of the multilayer perceptron is set to be that the error cannot be further reduced, a training sample of a training set is used as a training object of the multilayer perceptron, a test sample of a test set is used as a verification object of the multilayer perceptron, an artificial neural network is used for identifying and training and judging whether the accuracy rate meets the requirement, when the accuracy rate at least meets 90%, a prediction model is built based on an identification training result meeting the accuracy rate requirement, and when the accuracy rate does not meet 90%, the parameters of the multilayer perceptron are adjusted again and retrained.

It should be noted that the parameter adjustment method of the multilayer perceptron includes the following steps:

3) when the accuracy meets the requirement, stopping parameter adjustment;

Specifically, the parameter adjustment method can traverse all parameters within the range until the accuracy rate meets 90% or more.

In this embodiment, the number N of layers of hidden neurons in the multilayer perceptron is set to 2, the number M of nodes is set to 30, and the training stopping rule of the multilayer perceptron is set to be such that the error cannot be further reduced, and the accuracy of the training stopping rule can reach 93.5%.

EXAMPLE six

The construction of the prediction model comprises the steps of adjusting the fitting degree of the model and constructing the prediction model based on the recognition training result meeting the accuracy requirement.

In this embodiment, the prediction model includes an untreated drug model and a treated drug model, as shown in fig. 4, where:

the input variables of the non-drug treatment model comprise body information, disease information, diabetes treatment information and biochemical inspection information, wherein the diabetes treatment information is a default value, and the output of the non-drug treatment model is an initial recommended dose of the insulin pump;

the input variables of the used drug therapy model comprise physical information, disease information, diabetes therapy information and biochemical test information, wherein the diabetes therapy information comprises the use condition of the previous insulin and the use condition of other non-insulin hypoglycemic drugs, and the output of the used drug therapy model is the initial recommended dose of the insulin pump.

It should be noted that the required data can be obtained through on-site inquiry, physical examination and laboratory blood drawing examination, and the initial recommended dosage of the insulin pump can be obtained through model calculation, as shown in fig. 5.

EXAMPLE seven

The insulin pump initial recommended dose includes a basal amount of insulin and a prandial amount of insulin, the general ratio of the basal amount of insulin to the prandial amount of insulin being 1:1, and both employing a mean subcutaneous infusion rate, wherein: the insulin basic amount adopts a multi-section structure, and the multi-section structure comprises one of a three-section structure, a four-section structure or a multi-section structure. Therefore, after acquiring the insulin basal amount, the insulin meal time amount can be matched quickly, and the proportional relation between the insulin basal amount and the insulin meal time amount is shown in fig. 6.

The invention selects the insulin pump basic quantity as the machine learning object based on the following theoretical facts:

1. the suitability of the insulin pump basal amount and the control of the fasting blood sugar have certain logical relations. The control range of fasting blood glucose in the blood glucose control target in the 'guide' has strict limitation (4.4-7.0 mmol/l), and the requirement on postprandial blood glucose (< 10mmol/l) is relatively wide.

2. For patients with poor glycemic control, i.e., patients requiring short or long term treatment with an insulin pump, fasting glucose "contributes" more to the elevation of glycated hemoglobin than postprandial glucose.

3. Based on the clinical experience of high water rising and low water falling, if the fasting blood sugar can be reduced to a normal level, the postprandial blood sugar is easy to control, if the fasting blood sugar is very high, the postprandial blood sugar can reach the standard difficultly, only if the fasting blood sugar and the postprandial blood sugar are well controlled, the glycosylated hemoglobin can reach the standard, and the control of the fasting blood sugar is the primary consideration of blood sugar reduction treatment.

4. Insulin pump commonly used insulin aspart has half-life period of about 90 minutes, and after the insulin dose is acted during meal, the effect disappears quickly, which is not enough to influence the next day of fasting blood glucose. While the insulin pump basal volume is a continuous micro pump, which can reach a pharmacological steady state and affect the fasting glucose the next day for the patient.

Therefore, the invention adopts the strategy of firstly determining the basic amount by artificial intelligence and then matching the meal time according to the formula.

It should be noted that the usage of the insulin pump is relatively flexible, the basic amount means that the insulin pump will pump a small amount of insulin into the body every hour when the patient does not eat food in daily life to simulate the normal insulin secretion mode, each time period can have different dosage, and the maximum 24 hours can be different every hour, namely divided into 24 segments. For adjustment of treatment, 3 or 4 segments are used in clinical work, and the principle is as follows: generally, 8am to 22 pm are the major active hours of a person's day, with the most insulin required; the time from 22 to 3 a.m. is in the sleep period, the insulin requirement is minimum, the time from 3 a.m. to 8 a.night is dawn, the physiological function is gradually recovered, and the insulin requirement is gradually increased. Similarly, some hospitals can also adopt 4 sections and multiple sections in one day, and the principle is that the hourly dosage of the basic amount of each section cannot be greatly different, and the total amount of the basic amount of insulin and the meal time of insulin are close to each other.

Taking a three-stage example, all insulin basal amounts are fixed three-stage time-stage: "8 am-10 pm", "10 pm-3 am", "3 am-8 am". For example, if the base amount of an insulin pump for one patient is 10.5U, the three-stage distribution is "8 am-10pm 0.5U/h", "10 pm-3am 0.3U/h", and "3 am-8am 0.4U/h", as shown in the following table.

It should be noted that the table is only a common example, and the basic amount of each adjustment is increased by 0.1U every hour all day and 24 hours a day, so that the difference between two adjacent basic amounts is 2.4U, and similarly, the basic amount can also be increased by 0.2U every hour, so that 4.8U is increased in 24 hours, and each hospital can have its own experience summary at present, but the general principle is not changed.

Finally, it should be noted that: the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting an initial dose of an insulin pump is characterized by comprising the following steps:

5) constructing a prediction model: adjusting the fitting degree of the model, and constructing a prediction model based on the training result meeting the accuracy requirement;

2. The method of claim 1, wherein said method comprises:

the body information includes: one or more of gender, age, height, weight;

the condition information includes: one or more of diabetes typing, course of diabetes, complications of diabetes;

the diabetes treatment information includes: one or more of previous insulin use cases and previous non-insulin hypoglycemic drug use cases;

the biochemical test information includes: one or more of random blood glucose before insulin pump treatment, fasting blood glucose the next day after insulin pump treatment, glycated hemoglobin, fasting and postprandial C-peptide levels.

3. The method of claim 2, wherein said method comprises:

the conditions of the previous insulin hypoglycemic drugs comprise: the brand of a factory using the insulin in the past, the classification, the usage and the dosage according to the action time of the medicament, the type of the insulin used in the past comprises one or more of quick-acting, short-acting, intermediate-acting, long-acting, super-long-acting and premixing, the usage comprises the daily usage times, and the dosage comprises the international unit value of the insulin used each time;

the previous non-insulin hypoglycemic drug conditions include: previous biguanide drug use and dosage, previous alpha glucosidase inhibitor use and dosage, previous sulfonylurea drug use and dosage, previous sodium-glucose cotransporter 2 inhibitor use and dosage, previous dipeptidyl peptidase IV inhibitor use and dosage, previous glinide drug use and dosage, previous thiazolidinedione drug use and dosage, and previous glucagon-like peptide-1 receptor agonist use and dosage.

4. The method of claim 2, wherein said diabetic complication comprises ketoacidosis, lactic acidosis, hyperglycemic hyperosmolar syndrome, diabetes related infections, hypoglycemia, atherosclerotic cardiovascular disease, hypertension, diabetic foot disease, diabetic retinopathy, diabetic peripheral neuropathy, diabetic autonomic neuropathy, and diabetic nephropathy.

5. The method of claim 2, wherein said predictive models comprise an untreated model and a treated model, wherein:

the input variables of the used drug therapy model comprise body information, disease information, diabetes therapy information and biochemical test information, wherein the diabetes therapy information comprises the use condition of the used insulin and the use condition of other non-insulin hypoglycemic drugs, and the output of the used drug therapy model is the initial recommended dose of the insulin pump.

6. The method for predicting initial dose of insulin pump according to claim 1, wherein the neural network training is based on BP neural network algorithm training, comprising the following steps:

setting the number N of layers and the number M of nodes of hidden neurons in the multilayer perceptron, wherein: n is more than or equal to 2 and less than or equal to 3, M is more than or equal to 5 and less than or equal to 35, a training stopping rule of the multilayer perceptron is set to be that the error cannot be further reduced, a training sample of a training set is used as a training object of the multilayer perceptron, a test sample of a test set is used as a verification object of the multilayer perceptron, a BP neural network is used for training and judging whether the accuracy rate meets the requirement or not, when the accuracy rate at least meets 90%, a prediction model is built based on the recognition training result meeting the accuracy rate requirement, and when the accuracy rate does not meet 90%, the parameters of the multilayer perceptron are adjusted again and retrained.

7. The method for predicting initial dose of insulin pump according to claim 6, wherein the method for adjusting parameters of the multi-layer sensor comprises the following steps:

3) when the accuracy meets the requirement, stopping parameter adjustment;

8. The method of claim 3, wherein said method comprises:

the data cleaning in the original data processing process comprises the following steps:

checking and correcting the original data, and discarding incomplete data when missing values exist in the original data; when the original data has an abnormal value, correcting the abnormal value by referring to the original medical record;

the data transformation in the process of processing the original data specifically comprises the following steps:

other non-insulin hypoglycemic agents including different types of oral hypoglycemic agents and glucagon-like peptide-1 receptor agonist injections are normalized and converted into equivalent insulin doses.

9. The method of claim 1, wherein said data correction during raw data processing comprises the steps of:

the general control is to control the fasting blood sugar of the diabetic patient to be 4.4-7.0 mmol/l, and the glycosylated hemoglobin HbA1c is less than 7%;

the strict control is to control the fasting blood sugar of the diabetic patient to be 4.4-6.0 mmol/l, and the glycosylated hemoglobin HbA1c is less than 6.5%;

the loose control is to control the fasting blood sugar of the diabetic patient to be 5.6-8.0 mmol/l, and the glycosylated hemoglobin HbA1c is less than 8%; 2) based on the actual insulin dosage of the first day of insulin pump treatment of the diabetic patient in the blood sugar control target and the second sky abdominal blood sugar value after insulin pump treatment in the database, correcting the original insulin pump basic dosage of the diabetic patient in the database, wherein:

if the second sky abdominal blood glucose value after the insulin pump treatment is more than 9.0, the original insulin pump basic amount of the diabetic patient in the database is increased by 4.8U.

10. The method of claim 9, wherein the insulin pump initial recommended dose comprises a basal insulin amount and a prandial insulin amount, wherein the basal insulin amount and the prandial insulin amount are in a ratio of 1:1 and both adopt a mean subcutaneous infusion rate, and wherein: the insulin basic quantity adopts a multi-section structure, and the multi-section structure comprises one of a three-section structure, a four-section structure or a multi-section structure.