A deep neural network blood glucose prediction method based on metabolic causal knowledge

Abstract

The present application belongs to the technical field of blood glucose prediction, and relates to a deep neural network blood glucose prediction method based on metabolic causal knowledge: active samples of insulin data, meal data and historical blood glucose level time series in a training set are screened by using a convolutional recurrent neural network, and an insulin action subnetwork, a meal action subnetwork and an auxiliary action subnetwork including three one-dimensional convolution and maximum pooling layers, LSTM layers and full connection layers are respectively constructed; the above subnetworks are integrated into an integrated network; the integrated network is trained based on the training set, and early stopping and best model saving mechanisms are used in combination with a validation set to obtain a deep neural network model with the best effect, and the deep neural network model is used to predict future blood glucose. The deep neural network model has the beneficial effect that while predicting future blood glucose levels, the explainability of the blood glucose prediction result is enhanced through the structural design of the deep neural network model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blood glucose prediction, and particularly relates to a deep neural network blood glucose prediction method based on metabolic causal knowledge. BACKGROUND

[0002] With the increasing number of global diabetic patients, the use of artificial intelligence technology to realize the prediction and management of blood glucose of diabetic patients has attracted widespread attention from scholars in all walks of life. Since the 20th century, the rapid development of continuous blood glucose monitoring technology and data storage technology, the volume of blood glucose metabolism related data of diabetic patients has been increasing, which makes it possible to use deep learning technology to construct blood glucose metabolism model and predict the future trend of blood glucose of patients.

[0003] In early related research, Pérez-Gandía et al. constructed a blood glucose prediction model based on artificial neural network, and verified the possibility of predicting future blood glucose of diabetic patients by using historical blood glucose data and neural network. Martinsson et al. compared and analyzed the influence of recurrent neural network structure and parameters on the accuracy of blood glucose prediction, and gave the distribution interval of future blood glucose of patients. Idrissi et al. compared and analyzed the accuracy of convolutional neural network and recurrent neural network in blood glucose prediction, and the experimental results showed that by carefully designing and adjusting the structure and parameters of convolutional neural network, a blood glucose prediction model with higher prediction accuracy can be obtained. Although neural network has strong ability to fit nonlinear process, the prediction model constructed by using continuous blood glucose monitoring data only ignores the influence of insulin infusion and carbohydrate intake on the dynamic change of blood glucose, therefore, the prediction accuracy of neural network model based on blood glucose monitoring single variable still needs to be improved.

[0004] With the continuous accumulation of other variable data in the process of blood glucose metabolism, in order to be able to deeply mine the data patterns among the key variables of blood glucose metabolism, the blood glucose prediction model based on deep neural network also becomes more complex.

[0005] Due to the differences in insulin metabolism and action, carbohydrate absorption, physiological state, hormone level and other factors, the blood glucose metabolism process of diabetic patients shows strong individual difference, in order to enable the deep neural network model to more accurately describe the blood glucose metabolism process of a specific patient, transfer learning and meta learning and other technologies are widely used in the individualization of deep neural network model for blood glucose prediction.

[0006] The deep neural network based on a large amount of blood glucose metabolism data and personalized data has shown better performance than mechanism models and linear data-driven models in blood glucose prediction tasks, but since the deep neural network completely ignores metabolic causal knowledge and the causal relationship between blood glucose metabolism-related variables in the design and training process, the blood glucose prediction deep neural network model obtained by data fitting is often difficult to accurately describe the influence of different key factors on future blood glucose changes, especially when there is a strong correlation between large-dose insulin infusion and dietary intake. In the case where the causal relationship between the future blood glucose change and the related key variables cannot be accurately described, the deep neural network model is difficult to be used for early warning of abnormal blood glucose events such as hyperglycemia or hypoglycemia and analysis of possible causes, and is also difficult to be further used for managing blood glucose levels of diabetic patients, which seriously affects the significance and practical application value of the deep neural network in blood glucose management. SUMMARY

[0007] TECHNICAL PROBLEM

[0008] In view of the above-mentioned defects and deficiencies of the prior art, the present application provides a deep neural network blood glucose prediction method based on metabolic causal knowledge, which solves the technical problem that the blood glucose prediction model based on the deep neural network completely ignores metabolic causal knowledge in the construction process, and the blood glucose prediction deep neural network model obtained only by data fitting cannot reflect the causal relationship between blood glucose metabolism-related variables, making it difficult to be used in actual scenarios such as blood glucose control, and limiting the practical application value of the blood glucose prediction deep neural network model in the clinical decision support system.

[0009] TECHNICAL SCHEME

[0010] In order to achieve the above-mentioned purpose, the main technical scheme adopted by the present application comprises:

[0011] In a first aspect, the present application provides a deep neural network blood glucose prediction method based on metabolic causal knowledge, comprising:

[0012] Obtaining pre-processed insulin data, meal data and historical blood glucose level time series and dividing them into a training set, a validation set and a test set;

[0013] Using a convolutional recurrent neural network Screening active samples of insulin data, meal data and historical blood glucose level time series in the training set, and constructing insulin action subnetwork, meal action subnetwork and auxiliary action subnetwork including three one-dimensional convolution and max pooling layers, LSTM layers and fully connected layers, respectively; wherein, the historical blood glucose monitoring data is extracted from the active samples of the insulin data and insulin infusion information two channels as input data of the insulin action sub-network, and the model output of the insulin action sub-network is: wherein, is an adjustment coefficient; the recorded meal data is extracted from the active samples of the meal data two channels as input data of the insulin action sub-network, and the model output of the insulin action sub-network is: wherein, is an adjustment coefficient; the historical blood glucose monitoring data is extracted from the active samples of the historical blood glucose level time series as input data of the auxiliary action sub-network, and the model output of the auxiliary action sub-network is: ;

[0014] The integrated network is obtained by integrating the insulin action sub-network, the meal action sub-network and the auxiliary action sub-network.

[0015] The integrated network is trained based on the training set, and the early stopping and best model saving mechanism are used in combination with the validation set to obtain the deep neural network model with the best effect.

[0016] The test set is predicted based on the deep neural network model to obtain the blood glucose prediction result and the evaluation index, and the future blood glucose is predicted using the deep neural network model.

[0017] Optionally, a convolutional recurrent neural network The active samples of the insulin data in the training set are screened, and an insulin action sub-network including three one-dimensional convolution and max pooling layers, LSTM layers and fully connected layers is constructed, which further includes:

[0018] After removing the input data of the last time step, the convolution and max pooling layer performs step-by-step spatial dimension compression and channel dimension expansion on the input data, and outputs multi-scale features;

[0019] The time sequence characteristics of the multi-scale features after dimension rearrangement are learned using a long short-term memory network.

[0020] The multi-scale features are mapped to blood glucose prediction values by a fully connected network in a step-by-step dimension reduction manner and are outputted.

[0021] The Sigmoid function is used as an output gating function to constrain the sub-network output, and the response of blood glucose to insulin infusion is mapped to the negative half-axis interval.

[0022] Optionally, a convolutional recurrent neural network The active samples of the meal data in the training set are screened, and a meal action sub-network including three one-dimensional convolution and max pooling layers, LSTM layers and fully connected layers is constructed, which further includes:

[0023] After removing the input data of the last time step, the convolution and max-pooling layer performs step-by-step spatial dimension compression and channel dimension expansion on the input data, and outputs multi-scale features;

[0024] The long short-term memory network is used to learn the time sequence characteristics of the multi-scale features after dimension rearrangement;

[0025] The multi-scale features are mapped to blood glucose prediction values by step-by-step dimension reduction through a fully connected network and output.

[0026] The Sigmoid function is used as an output gate function to constrain the output of the subnetwork, and the response of the meal to the insulin infusion is mapped to the non-negative interval.

[0027] Optionally, a convolutional recurrent neural network is used Active samples of the historical blood glucose level time series in the training set are screened, and an auxiliary action subnetwork including three one-dimensional convolution and max-pooling layers, LSTM layers, and fully connected layers is constructed, which also includes:

[0028] After removing the data of the last time step, the convolution and max-pooling layer performs step-by-step spatial dimension compression and channel dimension expansion on the input data, and outputs multi-scale features;

[0029] The long short-term memory network is used to learn the time sequence characteristics of the multi-scale features after dimension rearrangement;

[0030] The multi-scale features are mapped to blood glucose prediction values by step-by-step dimension reduction through a fully connected network and output.

[0031] Optionally, the insulin action subnetwork, the meal action subnetwork, and the auxiliary action subnetwork are integrated to obtain an integrated network, which includes:

[0032] The insulin action subnetwork, the meal action subnetwork, and the auxiliary action subnetwork are integrated to obtain an integrated network according to the following formula:

[0033] , wherein, and represent the prediction value of the future blood glucose concentration and the measured value of the current time blood glucose concentration, respectively.

[0034] Optionally, the method further includes using a stepwise learning rate decreasing mechanism to train the deep learning network.

[0035] Optionally, the method further includes setting the patience parameter of the deep learning network to 40 rounds, automatically terminating the training when the performance of the deep neural network does not improve and exceeds the patience parameter, and loading the best model weight.

[0036] Optionally, the batch size of the deep learning network is set to 512, the maximum training round is set to 400 rounds, the Adam optimizer is selected for adaptive learning rate adjustment, and the mean square error is selected as the loss function.

[0037] In a second aspect, the present application provides a computer readable storage medium, which stores a computer program, and the program is executed to realize the blood glucose prediction method based on metabolic causal knowledge of any one of the first aspect.

[0038] In a third aspect, the present application provides a storage device, which comprises a storage medium and a processor, and the storage medium stores a computer program, and the program is executed by the processor to realize the blood glucose prediction method based on metabolic causal knowledge of any one of the first aspect.

[0039] Advantages

[0040] The present application has the following advantages: the blood glucose prediction method based on metabolic causal knowledge of the present application proposes a deep neural network architecture based on separation of key processes of blood glucose metabolism, which enables each sub neural network model to learn the influence of different influencing factors on future blood glucose changes of diabetic patients; by designing a sub network model based on a gating unit, combining physiological knowledge of blood glucose metabolism, and using the gating unit to enable each sub network model to accurately describe the response of blood glucose to key factors such as large-dose insulin infusion and dietary intake; the effectiveness of the proposed model is verified by simulation data, and the experimental results show that the proposed model can accurately reflect the causal relationship between key variables of blood glucose metabolism. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The deep neural network architecture based on metabolic causal knowledge provided for the embodiments of the present application is shown in the figure;

[0042] Figure 2 The flowchart of the blood glucose prediction method based on metabolic causal knowledge of the deep neural network provided for the embodiments of the present application is shown in the figure;

[0043] Figure 3 The CRNN structure diagram provided for the embodiments of the present application is shown in the figure;

[0044] Figure 4 The CRNN data flow diagram provided for the embodiments of the present application is shown in the figure;

[0045] Figure 5 The insulin action sub network model diagram provided for the embodiments of the present application is shown in the figure;

[0046] Figure 6 The insulin action sub network data flow diagram provided for the embodiments of the present application is shown in the figure;

[0047] Figure 7 A meal effect sub-network model diagram provided for an embodiment of the present application;

[0048] Figure 8 A meal effect sub-network data flow diagram provided for an embodiment of the present application;

[0049] Figure 9 An auxiliary effect sub-network model diagram provided for an embodiment of the present application; Figure 10 An auxiliary effect sub-network data flow diagram provided for an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to better explain the present application, so as to be understood, the present application is described in detail below by specific embodiments, combined with the accompanying drawings.

[0051] In the blood glucose metabolism process of diabetic patients, the glucose concentration in the blood is affected by many factors such as diet, insulin, exercise, stress, etc. In order to describe the blood glucose metabolism process in more detail, the blood glucose prediction model based on deep neural network often takes the historical data of blood glucose metabolism related variables in a period of time as the input information of the model, and the model output is the future blood glucose level. The relationship between the future blood glucose level and the historical observation data of the blood glucose metabolism variables is established by minimizing the modeling error and the back propagation of the gradient. When the model is complex enough, this black box model can well fit the blood glucose metabolism data, and then provide relatively accurate prediction. However, since the deep neural network model completely ignores the relatively clear mechanism between some variables in the blood glucose metabolism process in the modeling process, the blood glucose prediction model established by minimizing the prediction error often cannot accurately describe the causal relationship between the blood glucose metabolism related variables, especially in the case where there is a strong correlation between the input variables of the model. Therefore, the present application proposes a deep neural network architecture based on the mechanism of blood glucose metabolism, the structure of which is shown in Figure 1 As shown in the figure, it mainly includes four links: (1) Key variable separation: According to the knowledge of blood glucose metabolism mechanism, separate the key variables of blood glucose metabolism with clear mechanism; (2) Construction of convolutional recurrent neural network: According to the characteristics of the input data, construct a convolutional recurrent neural network for prediction; (3) Construction of gated sub-network model: For variables with clear mechanism, respectively establish network model for describing the influence of key variables on blood glucose metabolism; (4) Sub-model integration: Integrate the influence of sub-models with clear and fuzzy mechanism on blood glucose change, realize the prediction of future blood glucose level. Through the above links, the black box model based on deep neural network can effectively mine the causal relationship between the key variables of blood glucose metabolism, and provide blood glucose level prediction results with certain interpretability, which has important significance for the application of deep neural network in blood glucose management.

[0052] While the exemplary embodiments of the present application are shown in the drawings, it is understood that the present application can be carried out in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0053] In a first aspect, with reference to Figure 2 The embodiment provides a deep neural network blood glucose prediction method based on metabolic causal knowledge, including:

[0054] S1, obtaining pre-processed insulin data, meal data and historical blood glucose level time series and dividing into a training set, a validation set and a test set.

[0055] The insulin data and the meal data are variables separated according to metabolic causal knowledge, and the mechanism of action on blood glucose is clear.

[0056] The data set division adopts a time sequence reservation strategy, divides the complete data into a training set (the first 65%), a validation set (the middle 15%) and a test set (the last 20%) in chronological order, maintains the continuity of the time sequence, and conforms to the actual clinical application scene. The data processing link adopts a sliding window to generate sequence samples, and is classified and processed according to the existence of blood glucose, insulin and food and other effective input channels.

[0057] The regulation of blood glucose level in the human body is affected by multiple hormones, tissues and organs, at present, only part of the regulation mechanism has been revealed, the present application will make full use of this part of the blood glucose metabolism physiological knowledge to analyze the key variables of blood glucose metabolism and their effects, so as to obtain a more clear deep neural network architecture and improve the explainability of blood glucose prediction.

[0058] Insulin is one of the main hormones in the human body for glucose metabolism, which can promote the utilization of glucose in peripheral tissues and inhibit the production of glucose in the liver, and the two effects work together to reduce the glucose level in the blood. Specifically, for diabetic patients, exogenous insulin is generally absorbed into the blood circulation system through the subcutaneous tissue, and its target cells are muscle cells, liver cells and adipocytes. Insulin binds to insulin receptors on the cell membrane to activate intracellular PI3K / Akt signaling pathways, and promotes the translocation of glucose transporter 4 (GLUT4 transporter) to the cell membrane. Studies have shown that the surface GLUT4 transporter can increase 10-40 times under the action of insulin, thereby accelerating the consumption and utilization of glucose in the blood by muscle and liver cells, and reducing the plasma glucose concentration. After decades of research, the mechanism of insulin's hypoglycemic effect has become relatively clear. GLUT4 transporter mainly transports glucose by facilitated diffusion, which is affected not only by insulin metabolism but also by glucose concentration. Therefore, in order to accurately describe the effect of insulin on blood glucose changes, the infusion of insulin and the blood glucose concentration information need to be considered in the process of separating the effect of insulin and establishing the model.

[0059] In addition to insulin injection, dietary carbohydrate intake is also one of the key factors affecting future blood glucose changes in diabetic patients. Monosaccharides after food digestion and decomposition are absorbed by small intestinal mucosal cells, among which glucose and galactose are mainly transported into intestinal epithelial cells by sodium-dependent active transport, and fructose is absorbed by facilitated diffusion. The absorbed monosaccharides are released from the intestinal epithelial cells into the capillaries and then transported to the liver through the portal vein system. Among the monosaccharides reaching the liver, fructose and galactose are converted to glucose by the metabolic pathway of liver cells. The mechanism of the effect of carbohydrates on future blood glucose changes in type 1 diabetic patients is relatively clear, so the present application quantitatively analyzes the role of carbohydrates in blood glucose metabolism and establishes a sub-model to describe the response of future blood glucose to dietary intake.

[0060] Using the above blood glucose metabolism mechanism, the effects of key factors affecting future blood glucose levels (large-dose insulin infusion and carbohydrate intake) can be separated from the many factors affecting blood glucose changes, and the response of future blood glucose changes to these key variables can be accurately described through the establishment of a sub-model, so that the blood glucose prediction model based on deep neural network has certain interpretability.

[0061] The present application designs a CRNN network applied to a sub-network model, and the structure thereof is as shown in Figure 3The data input into the CRNN network is single-channel or multi-channel time series data. In order to extract the complex features of the input data, the CRNN network extracts the spatial features of the input data through a three-layer 1D convolutional neural network, and uses a 1D max-pooling operation to compress the obtained features. Generally, when extracting deep features of the input data, the convolutional neural network and the pooling are used multiple times. Then, the generated features are rearranged in dimensions, and the long short-term memory (LSTM) network is used to learn the time series dependence in the input data. Finally, the features are mapped to the predicted value through the fully connected network. The whole process combines the local feature extraction capability of the convolutional neural network and the time series modeling capability of the recurrent neural network, and captures the complex nonlinear relationship between blood glucose and key factors affecting blood glucose.

[0062] The parameter configuration of the CRNN model is shown in Table 1, which fully considers the time series characteristics and local correlation of blood glucose metabolism data. Through the convolutional layer to extract multi-scale features, the LSTM layer to model long-range dependence, and the fully connected layer to realize feature mapping, the final output is the predicted value of the future blood glucose change of the patient. The CRNN parameter configuration is shown in Table 1.

[0063] Table 1 Parameter configuration table of CRNN model

[0064]

[0065] The data flow process of the CRNN model is shown in Figure 4 The process can be divided into multi-stage feature transformation and dimension rearrangement operation. The input time series data is compressed in space dimension and expanded in channel dimension through convolution-pooling layer, and then the dimension is rearranged to adapt to the time series modeling requirement of the LSTM layer, and finally the scalar prediction value is mapped through the fully connected layer. This process realizes the deep fusion of local feature extraction and time series dynamic modeling through dynamic adjustment of feature dimension, and directly reflects the ability of the model to capture multi-level and multi-scale characteristics of blood glucose metabolism data.

[0066] Specifically, the present embodiment constructs a subnetwork model and integrates the subnetwork model into a deep neural network model, and superimposes the prediction results of insulin, diet and other factors. The model can comprehensively consider the interaction of multiple factors. This integration strategy solves the problem that a single model is not sufficient for multi-variable processing, and makes the prediction result more close to the real metabolism scene.

[0067] S2, using a convolutional recurrent neural network Active samples of insulin data, meal data and historical blood glucose level time series in the training set are screened, and an insulin action subnetwork, a meal action subnetwork and an auxiliary action subnetwork including three one-dimensional convolution and max-pooling layers, LSTM layers and fully connected layers are constructed.

[0068] wherein the historical blood glucose monitoring data is extracted from active samples of insulin data and insulin infusion information two channels as input data of the insulin action subnetwork, and the model output of the insulin action subnetwork is: wherein, represents the influence of insulin on future blood glucose changes, and respectively represent the historical blood glucose monitoring data and the insulin infusion information, denotes the CRNN network, is a regulation coefficient used to adjust the response strength of blood glucose to insulin, which can be set according to parameters such as insulin sensitivity of the patient.

[0069] recorded meal data is extracted from active samples of meal data as input data of the meal action subnetwork, and the model output of the meal action subnetwork is: wherein, represents the influence of meal intake on future blood glucose changes, is the recorded meal data, is a regulation coefficient used to adjust the response strength of blood glucose to diet, which can be set according to the glycemic index of food to constrain the amplitude of future blood glucose changes.

[0070] The historical blood glucose monitoring data is extracted from active samples of the historical blood glucose level time series as input data of the auxiliary action subnetwork, and the model output of the auxiliary action subnetwork is: wherein, represents the influence of sleep, exercise and stress on future blood glucose changes, represents the historical blood glucose monitoring data, and the effects of exercise, stress, sleep and other factors on blood glucose metabolism are mined.

[0071] The insulin action subnetwork, the meal action subnetwork and the auxiliary action subnetwork are integrated to obtain an integrated network;

[0072] The integrated network is trained based on a training set, and an early stop and best model saving mechanism are used in combination with a validation set to obtain a deep neural network model with the best effect;

[0073] Based on the deep neural network model, the test set is predicted to obtain blood glucose prediction results and evaluation indexes, and the deep neural network model is used to predict future blood glucose.

[0074] Optionally, a convolutional recurrent neural network The active samples of insulin data in the training set are screened, and an insulin action subnetwork including three one-dimensional convolution and max-pooling layers, LSTM layers, and fully connected layers is constructed, and the insulin action subnetwork further includes:

[0075] After removing the input data of the last time step, the convolution and max-pooling layers perform step-by-step spatial dimension compression and channel dimension expansion on the input data, and output multi-scale features;

[0076] The long short-term memory network is used to learn the time sequence characteristics of the multi-scale features after dimension rearrangement;

[0077] The multi-scale features are mapped to blood glucose prediction values through step-by-step dimension reduction of the fully connected network and output;

[0078] The Sigmoid function is used as an output gating function to constrain the subnetwork output, and the response of blood glucose to insulin infusion is mapped to the negative half-axis interval.

[0079] Optionally, a convolutional recurrent neural network The active samples of meal data in the training set are screened, and a meal action subnetwork including three one-dimensional convolution and max-pooling layers, LSTM layers, and fully connected layers is constructed, and the meal action subnetwork further includes:

[0080] After removing the input data of the last time step, the convolution and max-pooling layers perform step-by-step spatial dimension compression and channel dimension expansion on the input data, and output multi-scale features;

[0081] The long short-term memory network is used to learn the time sequence characteristics of the multi-scale features after dimension rearrangement;

[0082] The multi-scale features are mapped to blood glucose prediction values through step-by-step dimension reduction of the fully connected network and output;

[0083] The Sigmoid function is used as an output gating function to constrain the subnetwork output, and the response of blood glucose to insulin infusion is mapped to the negative half-axis interval.

[0084] Optionally, a convolutional recurrent neural network The active samples of meal data in the training set are screened, and a meal action subnetwork including three one-dimensional convolution and max-pooling layers, LSTM layers, and fully connected layers is constructed, and the meal action subnetwork further includes:

[0085] After removing the input data of the last time step, the convolution and max-pooling layers perform step-by-step spatial dimension compression and channel dimension expansion on the input data, and output multi-scale features;

[0086] The long short-term memory network is used to learn the time sequence characteristics of the multi-scale features after dimension rearrangement;

[0087] The multi-scale features are mapped to blood glucose prediction values by a fully connected network and output.

[0088] In the daily blood glucose management of type 1 diabetes patients, the transport, utilization and storage of glucose in the blood by liver, muscle and fat cells are promoted by the injection of insulin. In order to enable the deep neural network to capture the hypoglycemic effect of insulin, the present application designs an insulin action sub-network model based on the built CRNN network, and the structure is as shown in Figure 5 First, the samples with active insulin variables are selected from the data; then the blood glucose level data and insulin infusion data of these samples are extracted as the input of the insulin action sub-network model; and then the data is predicted through the CRNN network to predict the influence of insulin on future blood glucose changes.

[0089] As the only hormone in the human body that can reduce blood glucose concentration, the glucose concentration in the blood will gradually decrease with the infusion of insulin. In order to depict the action of insulin, the output of the insulin action sub-network needs to be limited to non-positive numbers. Further, with the increase of insulin concentration in the body, the glucose transport protein on the cell surface tends to be saturated, so that the intake and consumption of glucose in the blood by liver, muscle, fat and other cells tends to be saturated, that is, there is a physiological upper limit to the influence of insulin infusion on future blood glucose concentration of the patient, which needs to be explicitly represented in the neural network. Considering the above physiological mechanisms, the present application uses a sigmoid function as the output gating function to map the response of blood glucose to insulin infusion to the negative half-axis interval, and at the same time, uses the coefficient to adjust the amplitude of the response.

[0090] The data flow process of the insulin action sub-network is as shown in Figure 6 First, the input data is screened, and only the samples with active insulin are retained, and then the blood glucose level and insulin infusion amount channels are extracted from the screened data as the input of the sub-network. The data is adjusted in dimension to meet the needs of the proposed CRNN network, and the last time step of data is removed to avoid interference of future information. Finally, the prediction result is constrained to the negative half-axis interval by the gating function to ensure that the insulin action always shows a decrease in blood glucose concentration. The whole process is through multi-stage feature transformation and physiological knowledge fusion to provide interpretable prediction basis for clinical practice.

[0091] The structure of the meal action sub-network model is as shown in Figure 7 Similar to the insulin action sub-network model, first, the samples with active meal variables are selected from the data; then the meal data of these samples is input into the meal action sub-network model; and then the CRNN network is used to predict the influence of meal action on future blood glucose changes.

[0092] Unlike insulin, meal intake is the primary source of glucose in the body, and the glucose concentration in the blood rises with the intake of carbohydrates. Therefore, the future blood glucose change in response to diet is non-negative. In addition, due to the limitation of the digestive and absorptive capacity of the intestinal tract, the glucose absorption rate tends to saturate as the amount of carbohydrate intake increases. This physiological limit also needs to be explicitly represented in the neural network.

[0093] The data flow process of the meal subnetwork is shown in Figure 8 Similar to the insulin subnetwork, the input data is first screened to retain only samples with meal activity, and then the data of the single channel of meal intake is extracted from the screened data as the input features of the subnetwork. These data are adjusted in dimension to meet the needs of the CRNN network, and the last time step of data is removed to avoid interference from future information. The adjusted data are sequentially subjected to local feature extraction, modeling of time series dependence, and feature mapping by the CRNN network. Finally, the prediction result is constrained to the non-negative interval by the gating function, ensuring that the meal effect always manifests as an increase in blood glucose concentration.

[0094] In addition to insulin infusion and dietary intake, blood glucose is also affected by factors such as exercise, stress, sleep, and physiological state. Generally, for most type 1 diabetes patients, glucose is consumed as fuel for muscle cells during low-to-moderate intensity aerobic exercise, resulting in a rapid decrease in blood glucose concentration. As the intensity of exercise increases, the consumption of glucose by muscle cells from the blood gradually slows down. High-intensity aerobic exercise and anaerobic exercise can cause blood glucose concentration to rise and last for several hours. However, the relationship between blood glucose change and exercise type, duration, and intensity cannot be qualitatively or quantitatively described at present. Therefore, the present application uses the submodel shown in Figure 8 to model the effects of other factors on blood glucose change. The input of the other action submodel is the historical blood glucose level time series of the patient, and the learning of the patient's historical blood glucose monitoring sequence is achieved by the CRNN network.

[0095] The data flow process of the auxiliary subnetwork is shown in Figure 10 Similar to the insulin and meal subnetworks, only the core channel of blood glucose value is retained from the input data, and the known key variables such as insulin and meal are excluded. The screened data are removed from the last time step and input into the CRNN network. Subsequently, local features are extracted and time series dependence is modeled by the CRNN network, and finally the prediction value is output by the fully connected layer.

[0096] The insulin, meal and auxiliary action sub-network models can describe the dynamic changes of blood glucose caused by different factors, wherein, for the action of insulin and meal, the prediction output of the network is associated with the physiological mechanism by using a gating mechanism, so that the prediction result of the key variable sub-network has certain interpretability.

[0097] S3, integrating the insulin action sub-network, the meal action sub-network and the auxiliary action sub-network to obtain an integrated network.

[0098] Optionally, the insulin action sub-network, the meal action sub-network and the auxiliary action sub-network are integrated to obtain an integrated network, comprising:

[0099] The insulin action sub-network, the meal action sub-network and the auxiliary action sub-network are integrated to obtain an integrated network according to the following formula:

[0100] , wherein, and respectively represent the predicted value of the future blood glucose concentration and the measured value of the blood glucose concentration at the current time.

[0101] The dynamic change of blood glucose is the result of the joint action of many factors, in order to predict the future change of blood glucose of the patient, the prediction output of each key variable sub-model is integrated. It can be known from the above formula that the blood glucose prediction model established only needs to learn the change amount of the future blood glucose concentration, which reduces the learning difficulty of the neural network, and at the same time, the prediction output of each sub-network model conforms to the blood glucose metabolism mechanism.

[0102] S4, training the integrated network based on the training set, and combining the validation set to use the early stopping and best model saving mechanism to obtain a deep neural network model with the best effect.

[0103] Optionally, the method further comprises: setting the patience parameter of the deep learning network to 40 rounds, automatically terminating the training when the performance of the deep neural network does not improve and exceeds the patience parameter, and loading the model weight with the best performance.

[0104] In view of the common overfitting problem of deep learning, the early stopping and best model saving mechanism is adopted. The patience parameter is set to 40 rounds, and the system automatically terminates the training when the performance of the validation set does not improve and exceeds the patience threshold, and loads the model weight with the best performance of the validation set. This high tolerance is suitable for blood glucose prediction tasks that need to capture complex physiological patterns, which allows the model to explore a wider parameter space and effectively prevents overfitting on the training set.

[0105] S5, predicting the test set based on the deep neural network model to obtain blood glucose prediction results and evaluation indexes, and predicting future blood glucose using the deep neural network model.

[0106] The training process uses mean square error (MSE) as a loss function and early stopping criteria, and the model test stage combines root mean square error (RMSE) and mean absolute error (MAE) for comprehensive evaluation, with the validation set loss as the decisive indicator for model selection.

[0107] The present application improves the practical clinical application value of the blood glucose prediction model through data partitioning strategy, multi-stage training parameter optimization and comprehensive evaluation system.

[0108] Optionally, the method further comprises training the deep learning network using a stepwise learning rate decreasing mechanism.

[0109] In terms of training strategy, a stepwise learning rate decreasing mechanism is adopted to apply three decreasing learning rates to the model for gradient training in sequence. This progressive strategy ensures multi-stage optimization of the model training process: a high learning rate in the initial stage promotes the model to quickly approach the optimal solution region; a moderate learning rate in the middle stage further adjusts the model parameters; and a small learning rate in the final stage performs accurate parameter search near the optimal solution region.

[0110] Optionally, the batch size of the deep learning network is set to 512, the maximum training round is set to 400 rounds, the Adam optimizer is selected for adaptive learning rate adjustment, and the mean square error is used as the loss function.

[0111] The batch training configuration is optimized for time series prediction tasks: the batch size is set to 512 to balance the training speed and memory requirements; the maximum training round is set to 400 to ensure sufficient learning time for the model; the Adam optimizer is selected to provide adaptive learning rate adjustment; and the mean square error (MSE) is used as the loss function suitable for the characteristics of regression prediction tasks.

[0112] The deep neural network blood glucose prediction method based on metabolic causal knowledge provided in this embodiment separates key variables, and the model separates the effects of insulin infusion and carbohydrate intake from complex influencing factors based on the physiological mechanism of blood glucose metabolism, and constructs insulin action subnetwork and meal action subnetwork respectively. This step enables the model to independently learn the dynamic response of insulin lowering blood glucose and diet raising blood glucose, avoiding causal confusion caused by multiple factors.

[0113] In the sub-network design, the output of the model is further constrained through a gating mechanism to ensure that it conforms to known physiological laws. This design effectively solves the problem of "blood glucose rising after insulin infusion" or "blood glucose falling after diet" in traditional models, which is against common sense. It enhances the credibility of the model and makes the prediction results of the key variable sub-networks have certain interpretability. At the same time, other factor sub-networks retain flexible nonlinear modeling capabilities to capture complex influences such as exercise and stress that are difficult to quantify, balancing the advantages of physiological constraints and data-driven.

[0114] Finally, through sub-model integration, the prediction results of insulin, diet and other factors are superimposed, and the model can integrate the interaction of multiple factors. This integration strategy solves the problem of insufficient processing of multiple variables in a single model, making the prediction results closer to the real metabolic scenario.

[0115] In a second aspect, an embodiment of the present application provides a computer readable storage medium having stored thereon a computer program, the program implementing the blood glucose prediction method based on metabolic causal knowledge of any one of the first aspect when executed.

[0116] In a third aspect, an embodiment of the present application provides a storage device comprising a storage medium and a processor, the storage medium storing a computer program, the program implementing the blood glucose prediction method based on metabolic causal knowledge of any one of the first aspect when executed by the processor.

[0117] Those skilled in the art will appreciate that embodiments of the present application can be provided as a method, system or computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) containing computer usable program code.

[0118] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application should also include these modifications and variations.

[0119] Although embodiments of the present application have been shown and described above, it should be understood that the above-described embodiments are exemplary, and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above-described embodiments within the scope of the present application.

Claims

1. A deep neural network blood glucose prediction method based on metabolic causal knowledge, characterized by: include: Obtain pre-processed insulin data, meal data, and historical blood glucose level time series and divide them into training, validation, and test sets; Using convolutional recurrent neural networks Active samples of insulin data, meal data, and historical blood glucose level time series in the training set were screened, and the insulin action sub-network, meal action sub-network, and auxiliary action sub-network were constructed, each consisting of three layers of one-dimensional convolution and maximum pooling layers, LSTM layers, and fully connected layers; historical blood glucose monitoring data was extracted from active samples of insulin data. and insulin infusion information The two channels are used as input data of the insulin action sub-network, and the model output of the insulin action sub-network is for: ,in, is the adjustment coefficient; extract the recorded meal data from the active sample of meal data As the input data of the meal effect sub-network, the model output of the meal effect sub-network for: ,in, is the adjustment coefficient; extract historical blood glucose monitoring data from the active samples of the historical blood glucose level time series as the input data of the auxiliary sub-network, and the model output of the auxiliary sub-network for: ; The insulin action subnetwork, meal action subnetwork and auxiliary action subnetwork are integrated to obtain an integrated network; The integrated network is trained based on the training set, and early stopping and the best model preservation mechanism are used in combination with the validation set to obtain the optimal deep neural network model; The test set is predicted based on the deep neural network model to obtain blood glucose prediction results and evaluation indicators, and the deep neural network model is used to predict future blood glucose.

2. The deep neural network blood glucose prediction method based on metabolic causal knowledge according to claim 1, characterized in that: Using convolutional recurrent neural networks Active samples of insulin data in the training set were screened, and an insulin action subnetwork was constructed, consisting of three one-dimensional convolutional and maximum pooling layers, an LSTM layer, and a fully connected layer. The network also includes: After removing the input data of the last time step, the convolution and maximum pooling layers perform step-by-step spatial dimension compression and channel dimension expansion on the input data, outputting multi-scale features. Utilize long short-term memory networks to learn the temporal characteristics of multi-scale features after dimension rearrangement; Through the fully connected network, the multi-scale features are gradually reduced in dimension and mapped into blood glucose prediction values ​​and output; The sigmoid function is used as the output gating function to constrain the sub-network output and map the blood glucose response to insulin infusion to the negative half-axis interval.

3. The method for predicting blood glucose using a deep neural network based on metabolic causal knowledge according to claim 2, characterized in that: Using convolutional recurrent neural networks Active samples of meal data in the training set are screened, and a meal action subnetwork is constructed, which includes three layers of one-dimensional convolution and maximum pooling layers, an LSTM layer, and a fully connected layer. The network also includes: After removing the input data of the last time step, the convolution and maximum pooling layers perform step-by-step spatial dimension compression and channel dimension expansion on the input data, outputting multi-scale features. Utilize long short-term memory networks to learn the temporal characteristics of multi-scale features after dimension rearrangement; Through the fully connected network, the multi-scale features are gradually reduced in dimension and mapped into blood glucose prediction values ​​and output; The Sigmoid function is used as the output gating function to constrain the sub-network output and map the response of meals to insulin infusion into a non-negative interval.

4. The method for predicting blood glucose using a deep neural network based on metabolic causal knowledge according to claim 3, characterized in that: Using convolutional recurrent neural networks Active samples of the historical blood glucose level time series in the training set were screened, and an auxiliary sub-network consisting of three one-dimensional convolutional and maximum pooling layers, an LSTM layer, and a fully connected layer was constructed. The following also included: After removing the data of the last time step, the convolution and maximum pooling layers perform step-by-step spatial dimension compression and channel dimension expansion on the input data, outputting multi-scale features. Utilize long short-term memory networks to learn the temporal characteristics of multi-scale features after dimension rearrangement; The multi-scale features are mapped to blood glucose prediction values ​​through a fully connected network and then output.

5. The method for predicting blood glucose using a deep neural network based on metabolic causal knowledge according to claim 4, characterized in that: The insulin action subnetwork, meal action subnetwork and auxiliary action subnetwork are integrated to obtain an integrated network, including: The insulin action subnetwork, meal action subnetwork and auxiliary action subnetwork are integrated according to the following formula to obtain the integrated network: ,in, and They represent the predicted value of future blood glucose concentration and the measured value of current blood glucose concentration respectively.

6. The method for predicting blood glucose using a deep neural network based on metabolic causal knowledge according to claim 5, characterized in that: The method also includes: training the deep learning network using a step-by-step learning rate reduction mechanism.

7. The method for predicting blood glucose using a deep neural network based on metabolic causal knowledge according to claim 6, characterized in that: The method also includes: setting a patience parameter of the deep learning network to 40 rounds, automatically terminating training when the performance of the deep neural network no longer improves and exceeds the patience parameter, and reverting to loading the model weights with the best performance.

8. The method for predicting blood glucose using a deep neural network based on metabolic causal knowledge according to claim 7, characterized in that: The batch size of the deep learning network is set to 512, the maximum number of training rounds is set to 400, the Adam optimizer is used for adaptive learning rate adjustment, and the mean square error is used as the loss function.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the deep neural network blood glucose prediction method based on metabolic causal knowledge as described in any one of claims 1 to 7.

10. A storage device comprising a storage medium and a processor, wherein the storage medium stores a computer program, wherein: When the processor executes the computer program, the deep neural network blood glucose prediction method based on metabolic causal knowledge as described in any one of claims 1 to 7 is implemented.

Images