Diabetes data supervision method and system based on real-time acquisition

Abstract

The invention provides a diabetes data supervision method and system based on real-time acquisition, and relates to the technical field of diabetes data analysis, and the method comprises the steps: collecting static characteristic parameters of a patient, the static characteristic parameters comprising personal information and treatment information of the patient; periodically collecting dynamic characteristic parameters of the patient, wherein the dynamic characteristic parameters comprise blood glucose data of the patient; feature vectors are constructed based on the static feature parameters and the dynamic feature parameters, the feature vectors comprise prediction correction feature vectors and sign feature vectors, and the larger the values in the prediction correction feature vectors and the sign feature vectors are, the larger the negative effects on diabetes are; and performing nephropathy lesion risk assessment through a nephropathy prediction module according to the feature vectors. The method provided by the invention has the advantage of providing accurate nephropathy lesion risk assessment under the influence of diabetes mellitus, which is more adaptive to individualized difference.

Description

Technical Field

[0001] The present invention relates to the technical field of diabetes data analysis, and in particular to a diabetes data supervision method and system based on real-time collection. Background Art

[0002] Diabetes is a metabolic disease characterized by chronic hyperglycemia. Long-term hyperglycemia can cause damage to multiple organs throughout the body. Among them, diabetic nephropathy is one of the most serious complications in diabetic patients.

[0003] Existing kidney disease risk assessment methods usually perform static analysis based on a single physiological indicator or data at a specific time point, ignoring individual differences in patients and the long-term impact of dynamic blood sugar changes in diabetic patients, while blood sugar fluctuations are a key factor in the occurrence of diabetic complications. At the same time, in existing blood sugar risk prediction models, the extracted patient data and characteristics are often not fully expressed and utilized. If kidney disease is predicted based on existing blood sugar risk prediction models, individual differences are easily ignored, and kidney disease risk assessment is likely to be inaccurate.

[0004] Therefore, it is necessary to continue to optimize the existing diabetes data supervision program to achieve more accurate risk assessment of kidney disease under the influence of diabetes that is more adaptable to individual differences. Summary of the invention

[0005] The object of the present invention is to provide a diabetes data monitoring method and system based on real-time collection, which can provide an accurate risk assessment of kidney disease under the influence of diabetes that is more adaptable to individual differences.

[0006] The present invention is achieved through the following technical solutions: A diabetes data supervision method based on real-time collection includes the following steps: Collect the patient's static characteristic parameters, which include the patient's personal information and treatment information; Periodically collecting dynamic characteristic parameters of the patient, the dynamic characteristic parameters including the patient's blood sugar data; Constructing a feature vector based on static feature parameters and dynamic feature parameters, the feature vector includes a prediction correction feature vector and a physical sign feature vector, the larger the value in the prediction correction feature vector and the physical sign feature vector, the greater the negative impact on diabetes; The risk of renal lesions is assessed through the renal disease prediction module based on the feature vector.

[0007] Preferably, the personal information includes age, height, weight and habitual exercise time; and the treatment information includes duration of illness.

[0008] Preferably, the prediction correction feature vector is constructed based on the static feature parameters, and the method includes: Obtaining personal correction parameters based on the personal information : ; Among them, age represents age, w represents weight in kilograms, h represents height in meters, and t represents the average daily exercise time in minutes. and are adjustment coefficients, b and q are adjustment biases, and the adjustment coefficients and the adjustment biases are empirical parameters; Acquire treatment correction parameters according to the treatment information : ; in, represents the duration of illness and is expressed in years; Construct the predicted correction feature vector : .

[0009] Preferably, the method for obtaining the adjustment coefficient and the adjustment bias is: Based on big data, the age, height, weight and habitual exercise time of multiple diabetic patients were obtained; Get the maximum age and minimum value , the maximum average daily exercise duration and minimum value , the maximum value of body mass index and minimum value ; The equation is established and solved to obtain the adjustment coefficient and the adjustment bias: .

[0010] Preferably, the method for obtaining the vital sign feature vector through the dynamic feature parameter includes: Collect patients in time period N blood sugar values ​​within and form a blood sugar time series; Setting the Hyperglycemia Threshold and hypoglycemia threshold ; Obtain the value of the blood sugar time series not less than The number of parameters , and record the corresponding blood sugar acquisition time and form a first blood sugar recording time series in chronological order; Obtain the value in the blood sugar time series not higher than The number of parameters , and record the corresponding blood sugar acquisition time and form a second blood sugar recording time series in chronological order; Obtaining the coefficient of variation CV of the blood glucose time series, where the coefficient of variation is the ratio of the standard deviation to the mean; Obtain the difference between every two adjacent values ​​in the first blood glucose record time series, and take the minimum value ; Obtain the difference between every two adjacent values ​​in the second blood glucose record time series, and take the minimum value ; Get the physical sign feature vector : .

[0011] Preferably, the kidney disease prediction module comprises: An input layer, used for inputting the feature vector; A feature fusion layer, used for performing multi-modal fusion on the feature vectors to obtain a fused feature matrix; The output layer is used to output the prediction result of whether kidney disease occurs based on the fused feature matrix.

[0012] Preferably, the method for performing multimodal fusion on the feature vector is: Correct the feature vector for prediction and the sign feature vector Perform a linear transformation: ; ; , ; , ; ; in, and After linear transformation, and , and is the linearly changing weight, and is a linearly varying bias, is the linearized target dimension, for The dimension of for The dimension of , max is the maximum value function, and Respectively represent the dimensions and A real matrix of represents a one-dimensional vector of dimension d; Get the concatenated feature vector S: ; Get attention weights : ; in, represents the softmax function, is the attention weight matrix, For attention bias; Get the fusion feature matrix : ; in, Is the matrix dot product operator.

[0013] Preferably, the method of outputting the prediction result of whether renal lesions occur according to the fused feature vector is: The fusion feature matrix Perform linear transformation to obtain intermediate parameters : ; in, is the transformation weight and has dimension , is the transformation bias and is a constant; Obtain the probability of kidney disease through sigmoid : ; Among them, e is a natural constant.

[0014] Preferably, when the kidney disease prediction module makes a prediction, the label used is determined by the glomerular filtration rate. When the glomerular filtration rate is lower than a preset threshold, the label is that kidney disease lesions have occurred, otherwise the label is that kidney disease lesions have not occurred.

[0015] A diabetes data supervision system based on real-time collection is applied to the above-mentioned diabetes data supervision method based on real-time collection, comprising: Static collection module, used to collect static characteristic parameters of patients, including personal information and treatment information of patients; A dynamic acquisition module, used for periodically acquiring dynamic characteristic parameters of a patient, wherein the dynamic characteristic parameters include blood sugar data of the patient; A feature construction module, used to construct a feature vector based on static feature parameters and dynamic feature parameters, wherein the feature vector includes a prediction correction feature vector and a physical sign feature vector, wherein a larger value in the prediction correction feature vector and the physical sign feature vector represents a greater negative impact on diabetes; The evaluation module is used to evaluate the risk of renal lesions through a renal disease prediction module according to the feature vector.

[0016] The technical solution of the present invention has at least the following advantages and beneficial effects: The present invention collects static characteristic parameters and dynamic characteristic parameters of diabetic patients and constructs a comprehensive characteristic vector based on various types of data, which helps to more comprehensively reflect the individual health status and disease risk factors; The data selected and features constructed by the present invention include not only blood sugar fluctuations, peaks and troughs of blood sugar peaks, but also the impact of individual exercise and the impact of obesity and age, which improves the comprehensiveness, individual targeting and accuracy of feature expression; The present invention constructs static features and dynamic features into prediction correction feature vectors and physical sign feature vectors respectively, so that the patient's own blood sugar data characteristics and the degree of influence of individual differences on risk can be quantified, further improving the reliability of subsequent application to the model for prediction; The present invention analyzes the fused feature vectors through the kidney disease prediction module, which can comprehensively consider the long-term blood sugar control situation and the differences in individual physique and living habits, and more accurately output the predicted kidney disease risk situation, assisting clinicians in early screening, early diagnosis and personalized intervention; The invention is reasonably designed, suitable for different types of diabetic patients, can overcome individual differences to perform targeted data processing, and has wider applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic diagram of a flow chart of a diabetes data monitoring method based on real-time collection provided in Example 1 of the present invention; Figure 2 A schematic diagram of the principle of a diabetes data monitoring system based on real-time data collection provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0019] Example 1 This embodiment provides a diabetes data monitoring method based on real-time collection, see Figure 1 , including the following steps: Collect the patient's static characteristic parameters, which include the patient's personal information and treatment information; Periodically collecting dynamic characteristic parameters of the patient, the dynamic characteristic parameters including the patient's blood sugar data; Constructing a feature vector based on static feature parameters and dynamic feature parameters, the feature vector includes a prediction correction feature vector and a physical sign feature vector, the larger the value in the prediction correction feature vector and the physical sign feature vector, the greater the negative impact on diabetes; The risk of renal lesions is assessed based on the feature vector using the renal disease prediction module.

[0020] As a preferred solution of this embodiment, the personal information includes age, height, weight and habitual exercise time; the treatment information includes the duration of illness.

[0021] This embodiment collects static feature parameters and dynamic feature parameters of diabetic patients, and constructs a comprehensive feature vector with more individual representativeness and more comprehensive feature expression by integrating multi-dimensional and multi-type individual health data, so that the kidney disease prediction module in the subsequent steps can not only focus on the abnormal situation of a single indicator when assessing the risk of kidney disease, but also comprehensively analyze the patient's long-term physiological state and disease development trend. In the specific design, the static feature parameters mainly cover the patient's personal information including age, height, weight and habitual exercise time, and the treatment information mainly covers the duration of illness. The above parameters are usually relatively stable in the short term. The dynamic feature parameters are the blood sugar data continuously recorded by the patient within a certain period, and the blood sugar data usually has a certain short-term volatility. Through the combination of several data of this embodiment, more representative features can be extracted more accurately for patients in each different situation.

[0022] That is to say, based on the scheme of this embodiment, the static characteristic parameters can express the impact of the personal condition of the diabetic patient on the kidney disease lesions, and the static characteristic parameters detect blood sugar data in a traditional way, and the kidney disease lesions can usually be predicted by the characteristics of the blood sugar data. This embodiment can further correct and optimize the prediction results that only rely on blood sugar data by additionally obtaining dynamic characteristic parameters.

[0023] In summary, the joint feature establishment and prediction based on the multi-source data set in this embodiment can help to comprehensively reflect the overall health status and potential kidney disease risk factors of diabetic patients in different conditions, and provide a solid data foundation for the accurate prediction and management of the disease in a more targeted and reliable manner.

[0024] In this embodiment, the prediction correction feature vector is constructed based on the static feature parameters, and the method includes: Obtaining personal correction parameters based on the personal information : ; Among them, age represents age, w represents weight in kilograms, h represents height in meters, and t represents the average daily exercise time in minutes. and are adjustment coefficients, b and q are adjustment biases, and the adjustment coefficient and the adjustment bias are empirical parameters; Acquire treatment correction parameters according to the treatment information : ; in, represents the duration of illness and is expressed in years; Construct the predicted correction feature vector : .

[0025] Furthermore, the method for obtaining the adjustment coefficient and the adjustment bias is: Based on big data, the age, height, weight and habitual exercise time of multiple diabetic patients were obtained; Get the maximum age and minimum value , the maximum average daily exercise duration and minimum value , the maximum value of body mass index and minimum value ; The equation is established and solved to obtain the adjustment coefficient and the adjustment bias: .

[0026] In the static characteristic parameters selected in this embodiment: with the increase of age, the basic function of the kidney gradually declines (such as the decrease of glomerular filtration rate GFR, the increase of glomerular sclerosis, etc.), and the ability of the kidneys of diabetic patients to withstand the damage caused by continuous high blood sugar is weakened; and abnormal BMI (such as obesity with too high BMI) is an independent risk factor for diabetic nephropathy. For example, overweight or obesity will lead to worsened insulin resistance, and continuous high blood sugar state will increase glomerular high filtration and pressure burden, and it is also easy to induce kidney disease; finally, regular exercise helps to improve insulin sensitivity, stabilize blood sugar levels, etc., which is a favorable factor for preventing diabetic nephropathy; the longer the illness duration, that is, the longer the course of the disease, the greater the risk of kidney disease. Therefore, when this embodiment constructs a prediction and correction feature vector based on static characteristic parameters, the personal correction parameters obtained according to the personal information are mainly that the value of the personal correction parameters will be larger when the age is older, the MBI is larger, and the exercise is less. The setting of the adjustment coefficient and the adjustment bias is to standardize different parameters within a range so that the parameters obtained based on age are within a range. , based on MBI and based on the movement The value ranges of the three parameters are as close as possible. This design helps to eliminate dimensional differences, avoid feature dominance, and also helps to improve the subsequent feature fusion effect. The treatment correction parameter obtained has a larger value the longer the disease lasts.

[0027] On the other hand, the reason why this embodiment divides the feature extraction of static feature parameters into treatment correction parameters and human correction parameters is that the human correction parameters mainly describe individual static physical signs and behavioral habits, that is, the impact on the kidneys is more about basic risks, long-term metabolic levels, exercise regulation and other factors. The treatment correction parameters directly reflect the degree of chronic damage caused by diabetes. In other words, the essential causes of the impact of these two parameters on kidney lesions are different, and grouping for extraction helps the subsequent model to capture features more accurately.

[0028] Next, the characteristics related to blood sugar fluctuation are captured through blood sugar data, and the physical sign characteristic vector is obtained through the dynamic characteristic parameters. The method includes: Collect patients in time period N blood sugar values ​​within and form a blood sugar time series; Setting the Hyperglycemia Threshold and hypoglycemia threshold ; Obtain the value of the blood sugar time series not less than The number of parameters , and record the corresponding blood sugar acquisition time and form a first blood sugar recording time series in chronological order; Obtain the value in the blood sugar time series not higher than The number of parameters , and record the corresponding blood sugar acquisition time and form a second blood sugar recording time series in chronological order; Obtaining the coefficient of variation CV of the blood glucose time series, where the coefficient of variation is the ratio of the standard deviation to the mean; Obtain the difference between every two adjacent values ​​in the first blood glucose record time series, and take the minimum value ; Obtain the difference between every two adjacent values ​​in the second blood glucose record time series, and take the minimum value ; Get the physical sign feature vector : .

[0029] In this embodiment, when the physical sign feature vector is obtained by the dynamic feature parameter, N blood sugar values ​​of the patient within the time period T are obtained and formed into a blood sugar time series, and then the coefficient of variation of blood sugar, the number of times of hyperglycemia and the number of times of hypoglycemia, and the time difference between two adjacent hyperglycemia and the time difference between two adjacent hypoglycemia during this period are obtained as features. Among them, the coefficient of variation helps the subsequent recognition model to perceive the stability of blood sugar control, unstable blood sugar may aggravate the risk of diabetes and its complications such as nephropathy, and excessive hyperglycemia and hypoglycemia may have a negative impact on the kidney, the time difference between two adjacent hyperglycemia and the time difference between two adjacent hypoglycemia can reflect the persistence of this situation, and the time interval is too close, which also means that the negative impact on nephropathy is increased. Therefore, in the physical sign feature vector of this embodiment, the larger the coefficient of variation of blood sugar, the more times of hyperglycemia, the more times of hypoglycemia, and the smaller the time difference between two adjacent hyperglycemia and the smaller the time difference between two adjacent hypoglycemia will make the characteristic parameter value larger. Based on this setting, the ability of the subsequent nephropathy prediction module to perceive the dynamic changes and risk rhythm of blood sugar is improved.

[0030] Finally, the kidney disease prediction module includes: An input layer, used for inputting the feature vector; A feature fusion layer, used for performing multi-modal fusion on the feature vectors to obtain a fused feature matrix; The output layer is used to output the prediction result of whether kidney disease occurs based on the fused feature matrix.

[0031] As a further preferred solution, the method for performing multimodal fusion on the feature vector is: Correct the feature vector for prediction and the sign feature vector Perform a linear transformation: ; ; , ; , ; ; in, and After linear transformation, and , and is the linearly changing weight, and is a linearly varying bias, is the linearized target dimension, for The dimension of for The dimension of , max is the maximum value function, and Respectively represent the dimensions and A real matrix of represents a one-dimensional vector of dimension d; Get the concatenated feature vector S: ; Get attention weights : ; in, represents the softmax function, is the attention weight matrix, For attention bias; Get the fusion feature matrix : ; in, Is the matrix dot product operator.

[0032] The predicted correction feature vector and the physical sign feature vector obtained in this embodiment have different dimensions and dimensions, and direct splicing or weighting is prone to imbalance. The fusion method adopted in this embodiment enables features of different dimensions and sources to be in a relatively fair expression space when fused. Therefore, the feature fusion of this embodiment can dynamically assign weight values ​​of various features according to the individual differences and feature expressions of different patients, thereby highlighting the key features that have a greater impact on the risk of renal lesions and weakening the impact of interfering features on the prediction results.

[0033] In addition, the method of outputting the prediction result of whether renal disease occurs according to the fused feature vector is: The fusion feature matrix Perform linear transformation to obtain intermediate parameters : ; in, is the transformation weight and has dimension , is the transformation bias and is a constant; Obtain the probability of kidney disease through sigmoid : ; Among them, e is a natural constant.

[0034] This embodiment introduces a sigmoid function into the model to realize the judgment of right and wrong. For example, if the probability of kidney disease is If the output result is greater than 0.5, it is judged that kidney disease will occur, otherwise it is judged that kidney disease will not occur. The kidney disease here can include two situations: kidney disease worsens within a short period of time, such as one month, or a person without kidney disease develops kidney disease.

[0035] It is particularly noted that when the kidney disease prediction module makes predictions, the label used is determined by the glomerular filtration rate. When the glomerular filtration rate is lower than a preset threshold, the label is that kidney disease lesions have occurred, otherwise the label is that kidney disease lesions have not occurred.

[0036] Example 2 A diabetes data monitoring system based on real-time collection is applied to the above-mentioned diabetes data monitoring method based on real-time collection, see Figure 2 ,include: Static collection module, used to collect static characteristic parameters of patients, including personal information and treatment information of patients; A dynamic acquisition module, used for periodically acquiring dynamic characteristic parameters of a patient, wherein the dynamic characteristic parameters include blood sugar data of the patient; A feature construction module, used to construct a feature vector based on static feature parameters and dynamic feature parameters, wherein the feature vector includes a prediction correction feature vector and a physical sign feature vector, wherein a larger value in the prediction correction feature vector and the physical sign feature vector represents a greater negative impact on diabetes; The evaluation module is used to evaluate the risk of renal lesions through a renal disease prediction module according to the feature vector.

[0037] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A diabetes data supervision method based on real-time collection, characterized in that: The following steps are involved: Collect the patient's static characteristic parameters, which include the patient's personal information and treatment information; Periodically collecting dynamic characteristic parameters of the patient, the dynamic characteristic parameters including the patient's blood sugar data; Constructing a feature vector based on static feature parameters and dynamic feature parameters, the feature vector includes a prediction correction feature vector and a physical sign feature vector, the larger the value in the prediction correction feature vector and the physical sign feature vector, the greater the negative impact on diabetes; The risk of renal lesions is assessed through the renal disease prediction module based on the feature vector.

2. A diabetes data monitoring method based on real-time collection according to claim 1, characterized in that: The personal information includes age, height, weight and usual exercise time; the treatment information includes the duration of illness.

3. A diabetes data monitoring method based on real-time collection according to claim 2, characterized in that: The prediction correction feature vector is constructed based on the static feature parameters, and the method includes: Obtaining personal correction parameters based on the personal information : ； Among them, age represents age, w represents weight in kilograms, h represents height in meters, and t represents the average daily exercise time in minutes. and are adjustment coefficients, b and q are adjustment biases, and the adjustment coefficients and the adjustment biases are empirical parameters; Acquire treatment correction parameters according to the treatment information : ； in, represents the duration of illness and is expressed in years; Construct the predicted corrected feature vector : 。 4. A diabetes data monitoring method based on real-time collection according to claim 3, characterized in that: The method for obtaining the adjustment coefficient and the adjustment bias is: Based on big data, the age, height, weight and habitual exercise time of multiple diabetic patients were obtained; Get the maximum age and minimum value , the maximum average daily exercise duration and minimum value , the maximum value of body mass index and minimum value ; The equation is established and solved to obtain the adjustment coefficient and the adjustment bias: 。 5. A diabetes data monitoring method based on real-time collection according to claim 1, characterized in that: The method for obtaining the physical sign feature vector by using the dynamic feature parameter includes: Collect patients in time period N blood sugar values ​​within and form a blood sugar time series; Setting the Hyperglycemia Threshold and hypoglycemia threshold ; Obtain the value of the blood sugar time series not less than The number of parameters , and record the corresponding blood sugar acquisition time and form a first blood sugar recording time series in chronological order; Obtain the value in the blood sugar time series not higher than The number of parameters , and record the corresponding blood sugar acquisition time and form a second blood sugar recording time series in chronological order; Obtaining the coefficient of variation CV of the blood glucose time series, where the coefficient of variation is the ratio of the standard deviation to the mean; Obtain the difference between every two adjacent values ​​in the first blood glucose record time series, and take the minimum value ; Obtain the difference between every two adjacent values ​​in the second blood glucose record time series, and take the minimum value ; Get the physical sign feature vector : 。 6. A diabetes data monitoring method based on real-time collection according to claim 1, characterized in that: The kidney disease prediction module comprises: An input layer, used for inputting the feature vector; A feature fusion layer, used for performing multi-modal fusion on the feature vectors to obtain a fused feature matrix; The output layer is used to output the prediction result of whether kidney disease occurs based on the fused feature matrix.

7. A diabetes data monitoring method based on real-time collection according to claim 6, characterized in that: The method for performing multimodal fusion on the feature vector is: Correct the feature vector for prediction and the sign feature vector Perform a linear transformation: ； ； ， ； ， ； ； in, and After linear transformation, and , and is the linearly changing weight, and is a linearly varying bias, is the linearized target dimension, for The dimension of for The dimension of , max is the maximum value function, and Representing dimensions and A real matrix of represents a one-dimensional vector of dimension d; Get the concatenated feature vector S: ; Get attention weights : ； in, represents the softmax function, is the attention weight matrix, For attention bias; Get the fusion feature matrix : ； in, Is the matrix dot product operator.

8. A diabetes data monitoring method based on real-time collection according to claim 7, characterized in that: The method of outputting the prediction result of whether renal disease occurs based on the fused feature vector is: The fusion feature matrix Perform linear transformation to obtain intermediate parameters : ； in, is the transformation weight and has dimension , is the transformation bias and is a constant; Obtain the probability of kidney disease through sigmoid : ； Among them, e is a natural constant.

9. A diabetes data monitoring method based on real-time collection according to claim 8, characterized in that: When the kidney disease prediction module makes predictions, the label used is determined by the glomerular filtration rate. When the glomerular filtration rate is lower than a preset threshold, the label is that kidney disease lesions have occurred, otherwise the label is that kidney disease lesions have not occurred.

10. A diabetes data monitoring system based on real-time collection, applied to a diabetes data monitoring method based on real-time collection as claimed in any one of claims 1 to 9, characterized in that: include: Static collection module, used to collect static characteristic parameters of patients, including personal information and treatment information of patients; A dynamic acquisition module, used for periodically acquiring dynamic characteristic parameters of a patient, wherein the dynamic characteristic parameters include blood sugar data of the patient; A feature construction module, used to construct a feature vector based on static feature parameters and dynamic feature parameters, wherein the feature vector includes a prediction correction feature vector and a physical sign feature vector, wherein a larger value in the prediction correction feature vector and the physical sign feature vector represents a greater negative impact on diabetes; The evaluation module is used to evaluate the risk of renal lesions through a renal disease prediction module according to the feature vector.

Images