CN-122025167-A - Expert mixed blood glucose prediction method based on hypergraph wavelet convolution

Abstract

The invention belongs to the technical field of time sequence data prediction, and relates to an expert mixed blood sugar prediction method based on hypergraph wavelet convolution. The method comprises the steps of 1) constructing a self-adaptive sparse hypergraph, namely, adaptively constructing a sparse hypergraph of blood glucose sequence data by utilizing the co-occurrence relation of nodes in the hyperedge, dynamically modeling high-order coupling relation between blood glucose values at different time points, 2) constructing a hypergraph wavelet convolution, namely, designing multi-level wavelet transformation of a hypergraph structure, dynamically learning multi-scale characteristic information in a high-order neighborhood of the blood glucose values at each time point, 3) constructing a special gating hybrid prediction, namely, coupling the output of an expert prediction network based on a spline function by utilizing a confidence score, fitting a prediction mapping from historical blood glucose levels to future blood glucose values, and identifying a blood glucose change mode of a patient, and 4) calculating method overall loss, namely, evaluating the quality of the self-adaptive sparse hypergraph and the error of a blood glucose prediction result by utilizing the hypergraph construction loss and the blood glucose prediction loss, minimizing the overall loss optimization parameter, and fitting a blood glucose data prediction function.

Inventors

• LIU YINGSHU
• ZHANG JIANING
• SHI XIAOYAN
• WANG YACHEN
• GAO JING
• LI PENG
• LI XINYU
• LIU XUHAN
• ZHU YAN

Assignees

• 大连理工大学附属中心医院(大连市中心医院)

Dates

Publication Date

20260512

Application Date

20260121

Claims (5)

1. 1. An expert mixed blood sugar prediction method based on hypergraph wavelet convolution is characterized by comprising the following steps: Step 1, constructing a self-adaptive sparse hypergraph Adopting node embedding, superside sampling, superside updating and supergraph construction for 4 subprocesses, and constructing a self-adaptive sparse supergraph of blood glucose data; step2, constructing wavelet convolution of hypergraph Designing a hypergraph wavelet convolution module, adaptively combining a hypergraph wavelet substrate to represent multi-scale characteristics, modeling a high-order coupling relation in blood glucose sequence data on a plurality of levels, and aggregating multi-scale information of the blood glucose data, wherein the hypergraph wavelet convolution module comprises 3 subprocesses of Laplaciation, substrate construction and aggregation analysis; step 3, constructing special control hybrid prediction The expert gating hybrid module is used for adaptively combining the spline functions and learning the spline function form of the blood glucose prediction mapping, wherein the expert gating hybrid module comprises 2 subprocesses of expert prediction and gating evaluation; Step 4, integral loss The overall loss comprises hypergraph construction loss and blood glucose prediction loss, the quality of the self-adaptive sparse hypergraph and the error of the blood glucose prediction result are respectively evaluated, and the optimization of method parameters is promoted by minimizing the overall loss, so that the fluctuation mode in blood glucose data is fitted.
2. 2. The expert mixed blood glucose prediction method based on hypergraph wavelet convolution according to claim 1, wherein step 1 is specifically as follows: Node embedding, namely firstly modeling the blood sugar value of each moment in a blood sugar sequence of a patient as the node of a hypergraph, then constructing a learnable linear embedding function to map the blood sugar value of each moment to a feature space to obtain the embedded representation of each node in the hypergraph, and giving Blood glucose data set for a patient Wherein A blood glucose test sequence containing T time points for patient d is shown, Indicating the blood sugar level of the d patient at the t time Modeling as nodes of the hypergraph, constructing a learnable linear embedding function as shown in a formula (1), mapping each hypergraph node from a data space to a feature space to obtain node embedding representation in the hypergraph ; (1); Wherein, the Representing a standardized function, and compressing the blood glucose values of all nodes in the hypergraph to a standardized numerical range; And (3) with Respectively mean and standard deviation of blood glucose values; And Representing respectively a learnable linear embedding function Weight matrix and bias vector of (a); The hyperedge sampling comprises defining a sequence composed of nodes which are connected with blood sugar values at multiple moments as hyperedges in a hypergraph, modeling hidden high-order coupling relations in the co-occurrence process among the nodes in the blood sugar sequence, constructing a learnable hyperedge Gaussian distribution by a re-parameterization method, and adaptively fitting hyperedge distribution information, wherein the hyperedge Gaussian distribution in a given feature space is specifically From distribution of Middle sampling Initial representation of individual hyperedges : (2); Wherein, the And Respectively representing mean vector and covariance matrix of Gaussian distribution, then utilizing heavy parameterization method to make the above-mentioned Gaussian distribution Equivalently defined as a learnable gaussian distribution, the calculation process is as follows: (3); Wherein, the And Respectively represent the mean vector and covariance matrix of the standard gaussian distribution, The sampled values representing a standard gaussian distribution, Representing the equivalent relation corresponding to the re-parameterization method; the method comprises modeling membership between the hyperedge and blood sugar node by using attention value between initial representation of hyperedge and node embedded representation, and improving initial representation of hyperedge by means of high-contribution node Embedding representations with nodes Calculating attention value between superedge and node by using learning non-negative embedded function The process is as follows: (4); Wherein, the And (3) with Initial representation of hyperedges respectively Query vector and node embedded representation of (a) Is used to determine the key vector of (1), And The weight matrix and the bias vector of the query vector, And The weight matrix and the bias vector of the key vector respectively, As a function of the non-negative value, T is matrix transposition, attention value Modeling hyperedge initial representation Embedding representations with nodes Membership between the two; the method comprises the steps of obtaining a node and a node, wherein the node is higher in similarity with the node, the contribution to the superside is higher, meanwhile, sparsity constraint is applied to attention values between the superside and the node, noise factors in the membership of the superside and the node are filtered, specifically, the attention values between the superside and the node are ordered, and 3 nodes with the largest contribution are selected to update initial representation of the superside, wherein the process is as follows: (5); Wherein, the Embedding representations for nodes Is used to determine the value vector of (a), And (3) with Respectively a weight matrix and a bias vector; for the correction term initially represented for the hyperedge, For the updated representation of the hyperedge, For the attention value between all supersides and nodes A matrix of components; selecting a function for selecting a superside for ordering The corresponding 3 largest attention values; is a splicing function; Is a multi-layer perceptron; The hypergraph construction comprises the following steps of utilizing the attention value between node embedded representation and updated hyperedge representation to discover the hyperedge to which the node belongs, adaptively finding out the hypergraph structure in the blood glucose sequence of a patient, and realizing the depiction of the high-order coupling relation between blood glucose data, wherein the specific steps are that the node representation is given And updated superside representation Obtaining the attention value between the node and the superside by using the non-negative embedded function The calculation process is as follows: (6); Wherein, the And (3) with Respectively node representation Query vector and hyperedge representation of (a) Is a key vector of (a); To eliminate the influence of random noise, sparsity constraint is applied to the attention value between the node and the superside, namely The attention value ordering between each node and the superside keeps the maximum 3 attention values, and ensures the sparse supergraph Each node of the tree belongs to 3 supersides exactly, redundant supersides which are not connected with any node are removed, and a self-adaptive sparse supergraph is constructed : (7); Wherein, the For the attention value between all nodes and supersides A matrix is formed which is a combination of the two, Reserving a function for ordering for reserving node representations The corresponding 3 largest attention values; is a redundancy elimination function for eliminating redundant supersides in the supergraphs, and finally obtained self-adaptive sparse supergraphs Is a matrix of T rows and E columns, wherein T is the time number of blood sugar data, and since the blood sugar data at each time uniquely corresponds to one node in the hypergraph, T is the number of nodes, E is the number of redundancy removed hyperedges, In each row, 3 elements are non-zero, with the remaining elements being 0.
3. 3. The expert mixed blood glucose prediction method based on hypergraph wavelet convolution according to claim 1, wherein step 2 is specifically as follows: laplacian, namely calculating a normalized laplacian matrix of the hypergraph by using the degree matrix of the nodes and the degree matrix of the hyperedge, and giving the adaptive sparse hypergraph of the blood sugar data Calculating the degree matrix of the nodes according to the formula (8) Degree matrix with superside : (8); Wherein, the And E is respectively self-adaptive sparse hypergraph The number of intermediate nodes and supersides, ~ Respectively is The element values of row 1 and column j through column T, ~ Respectively is Element values from the t-th row, column 1, to the t-th row, column E, and further, calculating a normalized Laplacian matrix of the hypergraph The following are provided: (9); Wherein, the Is a unit matrix; The base construction comprises the steps of processing a normalized Laplace matrix of the hypergraph by utilizing a spectrum decomposition method, obtaining characteristic values and characteristic vectors of the hypergraph, forming a wavelet base of the hypergraph to model multi-scale characteristic information in the hyperedge, and giving the normalized Laplace matrix The process of spectrum decomposition to obtain its eigenvalues and eigenvectors is as follows: (10); Wherein, the As a function of the spectral decomposition, For the eigenvalues obtained by the spectral decomposition, Compressing the characteristic values of the hypergraph to construct a plurality of groups of functions capable of learning, and generating hypergraph wavelet substrates with various scales: (11); Wherein, the And (3) with Hypergraph wavelet bases of local and global dimensions respectively, And (3) with The scale parameters of the local and global scale, e is the base of the natural logarithm, As a diagonalization function, the vector can be converted into a diagonal matrix; aggregation analysis, namely capturing high-order neighborhood information of each scale in a blood glucose time sequence signal by utilizing a hypergraph wavelet substrate, aggregating and learning multi-scale high-order information of blood glucose data, and giving a hypergraph wavelet substrate group of S scales Normalized blood glucose data on hypergraph nodes The multi-scale high-order neighborhood information aggregation process is as follows: (12); Wherein, the In order to aggregate features at multiple scales, For a scale tradeoff of 1 for the sum, The diagonal matrix is convolved for the hypergraph of the corresponding scale, The weight matrix is convolved for the hypergraph.
4. 4. The expert mixed blood glucose prediction method based on hypergraph wavelet convolution according to claim 1, wherein step 3 is specifically as follows: Expert prediction, namely constructing dynamic spline functions by utilizing Bernoulli distribution, adaptively combining mapping relations between fitting input features and prediction outputs of different basis functions, constructing a robust and interpretable expert prediction network, and standardizing blood sugar data of a given patient Multi-scale aggregation feature output by hypergraph wavelet convolution module Adding the two elements to obtain an input Definition of a learnable spline activation function The following are provided: (13); Wherein, the For the input feature corresponding to time t, siLU is the baseline activation function, And (3) with A weight matrix and a bias vector for the baseline activation function, respectively; Is a spline function; to be distributed from Bernoulli The spline function of the mid-samples combines the weights, Equation (13) dynamically combines spline functions to determine the probability of participating in the operation for the ith spline function Later, a plurality of learnable spline functions are compounded to define an expert prediction network, and mapping between input features and output predictions is learned as shown in a formula (14); (14); Wherein, the For the expert to predict the output of the network, The function composite is represented by a function composite, In order to activate the function, Representing a multi-layer composite of spline functions; Gating evaluation, namely evaluating the confidence level of an expert prediction network by using a gating mixing unit, combining expert prediction results through confidence score to realize fitting of prediction mapping from historical blood glucose level to future blood glucose value, and giving aggregation input Design of gating hybrid cell The following are provided: (15); Wherein, the And (3) with Respectively representing the weight matrix and the bias vector in the gate control mixing unit, For normalizing the exponential function, then assigning a gated blend unit to each expert prediction Evaluating confidence scores of the blood glucose prediction mapping, and linearly combining expert prediction results according to the confidence scores to obtain the blood glucose prediction mapping The following are provided: (16); Wherein, the For the inverse normalization layer, it is responsible for restoring the linear combination of expert predictions to the blood glucose data range.
5. 5. The expert mixed blood glucose prediction method based on hypergraph wavelet convolution according to claim 1, wherein step 4 is specifically as follows: hypergraph construction loss given the final hypergraph node representation obtained by the adaptive sparse hypergraph module Spliced matrix Representation of superside Spliced matrix Adaptive sparse hypergraph The hypergraph construction loss measures hypergraph quality from three aspects of node information quantity, hyperedge diversity and point-edge similarity, and facilitates parameter learning of the self-adaptive sparse hypergraph module, and specifically comprises the following steps: Node information loss Representation from hypergraph nodes Reconstructing blood glucose data Verifying that the information contained in the blood glucose data has been sufficiently embedded into the representation with the mean square error MSE of the reconstructed result and the original blood glucose data: (17); Wherein, the Mapping the embedded representation back to the glycemic data space for the decoding function; Restoring the decoding result to the blood sugar data range as an inverse normalization function; is the vector 2 norm; Loss of superside diversity Constraining different hyperedge representations And (3) with Is minimized: (18); Wherein, the For the inner product between the superside representations, For the vector to be a 2-norm, Is the number of superflimit; point-edge similarity loss The point-edge similarity is adopted to strengthen the consistent corresponding relation between hypergraph node representation and hypergraph representation and self-adaptive sparse hypergraph: (19); wherein Sigmoid is an S-shaped activation function, Is a matrix of all 1's, For the point-edge similarity matrix, Multiplying the matrix element by element; For elements in the self-adaptive sparse hypergraph, the minimum hypergraph construction loss requires that the corresponding point-edge similarity be minimized, namely the feature similarity degree of corresponding nodes and hyperedges be minimized, and conversely, for elements in the self-adaptive sparse hypergraph, which are larger than 0.5, the minimum hypergraph construction loss requires that the corresponding point-edge similarity be maximized, namely the feature similarity degree of corresponding nodes and hyperedges be maximized; Predicted loss of blood glucose given inclusion Real blood glucose value set at each moment Predictive mapping in a hybrid with expert gating module Output predicted blood glucose value set The blood glucose prediction loss adopts an average square error mode to simultaneously calculate the difference between the two in the data and frequency space so as to promote the parameter learning of the hypergraph wavelet convolution module and the expert gating mixing module: (20); Wherein, the And (3) with The real blood glucose value and the predicted blood glucose value at the t time are respectively; And extracting high-frequency and low-frequency parts in the blood glucose label data and the blood glucose prediction data for discrete wavelet transformation.

Description

Expert mixed blood glucose prediction method based on hypergraph wavelet convolution Technical Field The invention belongs to the technical field of time sequence data prediction, and relates to an expert mixed blood sugar prediction method based on hypergraph wavelet convolution. Background Diabetes, one of the most common chronic diseases, has become a serious health problem worldwide, and is mainly represented by the long-term hyperglycemic state of the body caused by absolute or relative lack of insulin in patients, and is easy to induce metabolic disorders, tissue damage and other life-threatening short-term or long-term complications, including retinopathy, nervous system damage, kidney dysfunction, cardiovascular complications, macrovascular lesions, cardiomyopathy, ulcers and the like, which provide a great challenge to the global medical care system. The blood sugar prediction method obtains the change mode of blood sugar from the past blood sugar monitoring data of the patient, and accordingly predicts the blood sugar value at the future moment, provides reference for the dietary regulation intake or insulin infusion quantity of the patient, avoids the occurrence of abnormal blood sugar events of the patient, prevents the occurrence and development of complications, improves the quality of life and lightens the medical burden. The existing blood glucose prediction methods are mainly divided into a machine learning method and a deep learning method. The machine learning method utilizes the blood glucose dynamics mechanism in the model learning historical data such as autoregressive integral moving average, K-nearest neighbor, decision tree, random forest and the like, and models the prediction distribution of blood glucose data according to the parameter search space of the induction bias shrinkage model such as differential stable linear combination, similar sample trend uniformity and the like, so as to realize the prediction of the blood glucose level at the future moment. However, the machine learning method generally utilizes a shallow model to fit a linear mode of a human body blood sugar dynamics mechanism, so that nonlinear interaction between an internal blood sugar metabolism mechanism of a human body and an exogenous ingestion factor is difficult to accurately capture, a good prediction effect is achieved only in a blood sugar stationary phase of a patient, and long-time prediction of blood sugar in daily diverse living environments of the patient is difficult to model. The deep learning method utilizes hierarchical structures such as a convolutional neural network, a cyclic neural network and a long-short-term memory network to extract deep features of blood glucose data layer by layer, models a nonlinear mode of a human body blood glucose dynamics mechanism, and realizes long-time blood glucose level prediction. The deep learning method obtains impressive blood sugar prediction results, but only captures the first-order correlation of the blood sugar values pair by pair, ignores the higher-order relation among a plurality of time steps hidden in blood sugar sequence data, and is difficult to accurately model the blood sugar dynamics mechanism hidden in the synergetic/antagonistic effect among a plurality of organs and hormones of a human body. In addition, the deep learning method cannot give mathematical interpretation of blood sugar data deep features, prevents consistency verification between blood sugar prediction results and common sense of physiology, brings inconvenience to personalized adjustment of blood sugar management treatment schemes, is difficult to obtain trust of doctors and patients, and limits wide clinical application. Therefore, a new blood sugar prediction method is urgently needed to be developed, a high-order relation hidden in blood sugar sequence data is modeled, an interpretable change mode of a blood sugar dynamics mechanism is learned, blood sugar of a patient is accurately predicted, and intelligent medical level is improved. Disclosure of Invention In order to solve the problems, the invention provides an expert mixed blood glucose prediction method based on hypergraph wavelet convolution. The invention designs a self-adaptive sparse hypergraph module, builds a sparse hypergraph of blood glucose sequence data to dynamically model a high-order coupling relation between blood glucose values at different moments, designs a hypergraph wavelet convolution module, builds multi-level wavelet transformation of the hypergraph, gathers high-order neighborhood multi-scale characteristic information in the blood glucose sequence data, designs a special gating mixing module, utilizes confidence score coupling to output of an expert prediction network based on spline functions, flexibly fits a prediction mapping from historical blood glucose levels to future blood glucose values, and accurately identifies a blood glucose change mode of a patient. The technical scheme o