Medical data feature interpolation method based on graph network

Abstract

The invention relates to the field of medical treatment, in particular to a medical data feature interpolation method based on a graph network, and the method comprises the steps: constructing and training a multi-task graph network interpolation model; constructing to-be-processed data into a medical bipartite graph, and inputting the medical bipartite graph into the trained multi-task graph network interpolation model to obtain interpolated data; the multi-task graph network interpolation model comprises an information extraction module and an MLP, and the information extraction module comprises three cascaded EMRHGRAPHSage modules; according to the method, the defects of a traditional interpolation method during medical data processing are overcome, so that the performance of downstream tasks is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the medical field, and in particular to a medical data feature interpolation method based on a graph network. Background Art

[0002] In reality, medical data is often missing due to various factors in the collection process, such as poor patient compliance, equipment failure or data loss. These missing data not only reduce the accuracy of analysis, but also have a negative impact on clinical decision-making. Data interpolation technology, as an important means to deal with the problem of missing data, has received widespread attention in recent years. However, medical data is highly complex and specific, which makes the application of general data interpolation models in this field unsatisfactory. Medical data usually have the following challenges: First, they are usually high-dimensional and heterogeneous, and the associations between different features (such as physiological indicators, laboratory test results, clinical diagnosis, etc.) are complex and difficult to model; second, the distribution of medical data is often uneven, and some rare clinical conditions or test results are extremely rare in the data, but may play a key role in disease diagnosis; third, the data often contains noise, outliers, and irregular sampling, which further increases the difficulty of interpolation. In this context, traditional interpolation methods such as mean interpolation and K nearest neighbor interpolation are often difficult to effectively deal with the above challenges, resulting in interpolation results that are not sufficiently robust and accurate.

[0003] Although the existing GNN method can capture the complex dependencies between features, it still has shortcomings in making full use of edge information and processing heterogeneous features. Therefore, how to design an interpolation model that has both structured information expression and dynamic dependency modeling capabilities has become a key issue in current research. Summary of the invention

[0004] In order to solve the shortcomings of existing traditional interpolation methods in processing medical data (such as high dimension, heterogeneity and uneven data distribution), the present invention proposes a medical data feature interpolation method based on graph network, including constructing and training a multi-task graph network interpolation model; constructing the data to be processed as a medical bipartite graph and inputting the trained multi-task graph network interpolation model to obtain interpolated data; the multi-task graph network interpolation model includes an information extraction module and MLP, and the information extraction module includes three cascaded EMRHGRAPHSage modules;

[0005] The training process of the multi-task graph network interpolation model includes the following steps:

[0006] S1. Process the acquired medical data into a medical bipartite graph with missing values;

[0007] S2. Input the medical bipartite graph into the information extraction module to obtain an enhanced medical bipartite graph;

[0008] S3. Input the enhanced medical bipartite graph into MLP to obtain the interpolation accuracy and label prediction accuracy;

[0009] S4. Construct a combined loss function based on the edge-level prediction task and the node-level prediction task, calculate the loss through the combined loss function, and back-propagate the training model parameters until the model converges.

[0010] Beneficial effects of the present invention:

[0011] The present invention constructs a data set containing missing data as a bipartite graph with missing information, in which samples and features are respectively regarded as two types of nodes, and the observed feature values ​​are regarded as attribute edges connecting sample and feature nodes. Under this graph structure, the feature interpolation task can be naturally regarded as an edge-level prediction problem, while the label prediction task is regarded as a node-level prediction problem. The present invention solves the problem of feature missing and label prediction in heterogeneous graph structures through modular design. By effectively combining modules such as edge-aware multi-scale attention, heterogeneous edge update and residual network, the model can capture and fuse feature information of nodes and edges from multiple angles. In addition, the present invention adopts the concept of multi-task learning, takes feature interpolation as the main task, and introduces label prediction as an auxiliary task. By designing a combined loss function, the joint optimization of these two tasks is achieved. The design of multi-task learning improves the model's perception of the graph structure, thereby generating higher-quality feature representations and ultimately improving the performance of subsequent tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 is a flow chart of the method of the present invention;

[0013] Figure 2 This is a structural diagram of the model of the present invention;

[0014] Figure 3 Schematic diagram of the edge-aware multi-scale attention (MMA) module flow of the present invention. DETAILED DESCRIPTION

[0015] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0016] The present invention provides a medical data feature interpolation method based on a graph network, comprising constructing and training a multi-task graph network interpolation model; constructing the data to be processed as a medical bipartite graph and inputting the trained multi-task graph network interpolation model to obtain interpolated data; the multi-task graph network interpolation model comprises an information extraction module and an MLP, and the information extraction module comprises three cascaded EMRHGRAPHSage modules.

[0017] like Figure 1 As shown in Figure 1, the training process of the multi-task graph network interpolation model includes the following steps:

[0018] S1. Process the acquired medical data into a medical bipartite graph with missing values.

[0019] Specifically, step S1 processes the acquired medical data into a medical bipartite graph with missing values, including:

[0020] Take each sample in the medical data as a sample node and get the sample node set V A ={u1,u2,…,u m}, m represents the number of samples; each medical feature is taken as a feature node, and the feature node set V is obtained B ={v1,v2,…,v m}, n represents the number of medical characteristic categories; medical characteristics refer to various data and indicators used to describe and analyze individual health conditions in medical research or clinical practice, including demographic information, laboratory test results, lifestyle factors, clinical characteristics, etc.

[0021] Construct a feature matrix A of size m×n. The element A in the i-th row and j-th column of the matrix feature A is ij represents the eigenvalue of the jth feature of the i-th sample; due to the problem of missing eigenvalues ​​caused by missing data, a mask matrix M of size m×n is constructed, and the element M in the i-th row and j-th column of the mask feature M ij is a binary variable, if M ij = 0, it means that the jth feature of the ith sample is missing and A cannot be observed ij If the value of M ij =1, it means that the jth feature of the ith sample is not missing, and A can be observed ij The value of

[0022] According to the feature matrix A and the mask matrix M, a medical bipartite graph G = {V, E}, V = V A ∪V B represents a node set, E={(u i ,v j ,e ij )|u i ∈VA ,v j ∈V B ,M ij =1} represents an edge set, where edges exist only between nodes of different types. ij Represents the sample node u i With feature node v j The edge features between ij =1, then the sample node u i With feature node v j There is an edge between them, and e ij =A ij If A ij If it is a discrete variable, it is converted into a one-hot vector and then assigned to e ij .

[0023] Usually, the dataset comes with labels for downstream tasks. Let the label of the i-th sample be Y i , set a binary variable W for each sample i ∈{0,1}, through the binary variable W i To split the training set / test set, only when W i = 1, Y can be observed in training and testing i In addition, since medical data naturally has missing parts, after generating the mask matrix M, the present invention converts M ij = 1 to construct the training set and test set, and after optimizing and iterating to find the optimal model, M ij = 0 is input into the model to obtain the final interpolation result, and the downstream task (label prediction) is used to verify the accuracy of the interpolation.

[0024] S2. Input the medical bipartite graph into the information extraction module to obtain an enhanced medical bipartite graph.

[0025] Specifically, the EMRHGRAPHSage module includes the EGSAGE module, the edge-aware multi-scale attention (MMA) module and the heterogeneous edge update (HEU) module. The processing process of each EMRHGRAPHSage module on the input graph structure includes:

[0026] S21. The input graph structure passes through the EGSAGE module to obtain the initial representation of each node in the input graph structure;

[0027] S22. The input graph structure updated in step S21 is passed through an edge-aware multi-scale attention module to obtain an enhanced feature representation of each node;

[0028] S23. Fusing the input graph structure updated in step S21 with the input graph structure updated in step S22 to obtain a valid feature representation of each node;

[0029] S24. Pass the input graph structure updated in step S24 through the heterogeneous edge update module to update the edge features of each edge.

[0030] Specifically, the EGSAGE module adopts a graph neural network architecture inspired by GraphSAGE. The processing process includes:

[0031] S211. Initialize the embedding vector of each node, expressed as

[0032]

[0033] in, represents the embedding vector initialized by node i, x i Represents the feature vector of node i;

[0034] S212. In the l=1, 2, …, Lth layer, the edge information is used as a supplement to the node information. The message passing function takes the concatenation of the node embedding vector and the edge feature as input to update the embedding vector of each node, which is expressed as

[0035]

[0036]

[0037] Among them, AGG l (·) represents the aggregation function, e ij represents the edge feature between node i and node j, P (l) , Q (l) represents the weight matrix, CONCAT(·) represents the concatenation function, represents the aggregate vector representation of node i at layer l, represents the embedding vector of node i in layer l, σ() represents the activation function, ReLU is used here; N(i,ε drop ) represents the set of neighbor nodes of node i, ε drop Represents the edge set after dropout processing;

[0038] S213. Output the updated result of the last layer to obtain the initial representation of each node.

[0039] Specifically, the core goal of the edge-aware multi-scale attention module is to extract the features of nodes and edges by combining multi-scale convolution operations and edge-aware attention mechanisms, thereby enhancing the graph neural network's understanding of graph structure information. Specifically, the edge-aware multi-scale attention module includes three parallel convolution branches. The first convolution branch uses a 1×1 convolution, the second convolution branch uses a 3×3 convolution, and the third convolution branch uses a 5×5 convolution. The first convolution branch is mainly used to extract the local features of each node, the second convolution branch is mainly used to capture the features of the node and its neighborhood, and the third convolution branch is mainly used to obtain feature information in a larger range. The output of each branch will generate a corresponding feature matrix for further attention weighting.

[0040] Specifically, if Figure 3 As shown in Figure 2, the specific processing process of the edge-aware multi-scale attention module includes:

[0041] S221. Extract the node feature matrix and edge feature matrix of the input graph structure; it should be noted that the node feature matrix obtained here is obtained through processing by the previous module (EGSAGE module), and each row in the node feature matrix represents the initial representation of a node; each row in the edge feature matrix represents the edge feature of an edge (i.e., the edge feature constructed and obtained in step S1).

[0042] S222. The node feature matrix is ​​passed through the first convolution branch to obtain a local node feature matrix, the node feature matrix is ​​passed through the second convolution branch to obtain a neighborhood node feature matrix, and the node feature matrix is ​​passed through the third convolution branch to obtain a large-scale node feature matrix.

[0043] S223. Perform different linear transformations on the edge matrix to obtain the first edge matrix, the second edge matrix and the third edge matrix. This operation enables the feature dimension of each edge to be mapped to the feature dimension of the corresponding node so as to operate with the node features when subsequently calculating the attention weight.

[0044] S224. The local node feature matrix and the first edge matrix are combined through the edge-aware attention mechanism to obtain the first attention weight matrix; the neighborhood node feature matrix and the second edge matrix are combined through the edge-aware attention mechanism to obtain the second attention weight matrix; the large-scale node feature matrix and the third edge matrix are combined through the edge-aware attention mechanism to obtain the third attention weight matrix.

[0045] Specifically, the edge-aware attention mechanism performs weighted calculations on the features of each node with its neighboring nodes and edge information. This process is completed through the Q, K, and V matrices (i.e., query, key, and value). Using an input node feature matrix as the Q matrix, the adjacent node edge representation matrix as the K matrix, and the neighbor node feature matrix as the V matrix, the whole can be expressed as

[0046]

[0047] d k Indicates the number of feature dimensions. The first attention weight matrix is ​​obtained by the edge-aware attention mechanism using the local node feature matrix and the first edge matrix. The local node feature matrix is ​​taken as the Q matrix. The adjacent node edge representation matrix and the neighbor node feature matrix are calculated according to the local node feature matrix and the first edge feature matrix as the K matrix and the V matrix respectively. The first attention weight matrix is ​​calculated according to the Q matrix, the K matrix and the V matrix. Among them, the acquisition process of the adjacent node edge representation matrix and the neighbor node feature matrix is ​​as follows: for each node, all its neighbor nodes are obtained, and the feature representations of all neighbor nodes are extracted from the local node features and fused to obtain the neighbor aggregation feature representation, the edge feature representations between the node and all its neighbor nodes are extracted from the first edge matrix and fused to obtain the edge aggregation feature representation, the neighbor aggregation feature representations of all nodes are composed of the neighbor node feature matrix, and the edge aggregation feature representations of all nodes are composed of the adjacent node edge representation matrix. The neighbor aggregation feature representation and the edge aggregation feature representation are the same size as the feature representation of the nodes in the local node feature matrix.

[0048] S225. Multiply the local node feature matrix by the first attention weight matrix to obtain a first representation matrix, multiply the neighborhood node feature matrix by the second attention weight matrix to obtain a second representation matrix, and multiply the large-scale node feature matrix by the third attention weight matrix to obtain a third representation matrix;

[0049] Among them, the first representation matrix obtained by multiplying the local node feature matrix with the first attention weight matrix can be expressed as

[0050]

[0051] Among them, W1 represents the weight matrix corresponding to the 1×1 convolution, h j represents the initial representation of node j in the node feature matrix, represents the first attention weight between node i and node j, represents the first representation of node i in the first representation matrix; the operations of obtaining the second representation matrix and the third representation matrix are the same as those of the first representation matrix.

[0052] S226. The first representation matrix, the second representation matrix and the third representation matrix are merged to obtain a node enhanced feature representation matrix, thereby obtaining an enhanced feature representation of each node.

[0053] Specifically, Figure 2As shown, in order to effectively retain and utilize the input information, the present invention adopts residual connection in the EMRHGRAPHSage module, that is, an additional branch is drawn at the output of the EGSAGE module, and the output of the EGSAGE module is weighted and summed with the output of the MMA module. Through the residual connection, the original input features can be directly transferred to the deep layer of the network, which helps to alleviate the gradient vanishing problem in the deep network, and at the same time enables the model to better maintain the integrity of the input features, thereby improving the expressiveness of the features and accelerating the convergence of the model.

[0054] Specifically, the core idea of ​​the heterogeneous edge update module is to dynamically update edge features using the features of adjacent nodes, edge features, and the type information of the examination items (i.e., the category of medical features). This module strengthens the representation ability of edges by combining contextual information from multiple sources, so that the model can better capture the complex relationships between nodes. Specifically, the heterogeneous edge update module updates the edge features of each edge of the input graph structure, including:

[0055] Each edge connects a sample node and a feature node, and the enhanced feature representation of the two nodes is obtained. The feature nodes connected by the edge are mapped through a linear encoder to obtain the inspection item category features containing context information; the edge features, the enhanced feature representation of the sample node, the enhanced feature representation of the feature node, and the inspection item category features are concatenated and nonlinearly transformed through MLP to obtain the updated edge features, which can be expressed as

[0056] e i ' j =σ(W e ·[x i ||x j ||e ij ||(W c check_types+b c )]+b e )

[0057] Among them, e i ' j represents the updated edge feature, x i represents the enhanced feature representation of node i, check_types represents the check item category feature, || represents the concatenation operation, and W e , W c represents the weight matrix, b e , b c represents the bias term, and σ() represents the activation function.

[0058] S3. Input the enhanced medical bipartite graph into MLP to obtain edge prediction results and label prediction results.

[0059] Specifically, the present invention includes two tasks: edge-level prediction and node-level prediction. The purpose of the edge-level prediction task in the graph neural network is to determine whether there may be an edge between any two nodes in the graph, as well as the potential weights of these edges; in the present invention, missing feature interpolation is used as edge-level prediction. Interpolating missing features can be expressed as learning edge value prediction mapping:

[0060]

[0061] represents the edge prediction result between node i and node j, which also refers to the prediction result of the element in the i-th row and j-th column in the feature matrix A; f ij (·) represents the edge-level prediction mapping function, which is used to predict missing values ​​in the feature matrix. Its input is the graph G (including the structural information of the graph, node features, and edge features). During the training process, With A ij By calculating the difference between the two (cross entropy), the model can gradually learn how to fill in the missing data more accurately.

[0062] The present invention regards label prediction as node-level prediction. The predicted downstream node label can be expressed as a learning mapping:

[0063]

[0064] in, represents the label prediction result of sample i, g i (·) represents the function used to predict node labels, that is, the node-level prediction mapping. During the training process, the label prediction result is minimized. and the true label Y i By comparing the differences between the two, the model gradually learns to accurately predict node labels even when the features are incomplete.

[0065] S4. Construct a combined loss function based on the edge-level prediction task and the node-level prediction task, calculate the loss through the combined loss function, and back-propagate the training model parameters until the model converges.

[0066] Specifically, based on the definition of S3, the model combines the losses of the two tasks by sharing parameters and information. By using the adaptive coefficient α, the loss of the edge-level prediction task and the loss of the node-level prediction task are weighted and summed to form a combined loss function. This combined loss can adaptively balance the contribution of the two tasks, ensuring that the model can take both into account when dealing with feature missing and label prediction, thereby improving the performance of the overall model. Loss can be expressed as:

[0067] L=α·L edge +(1-α)·L node

[0068] Among them, L edge is the loss of the edge-level prediction task; L node is the loss of the node-level prediction task; the adaptive coefficient α determines the importance ratio of the two tasks in the total loss.

[0069] In order to verify the effectiveness of the method of the present invention, based on the National Health and Nutrition Examination Survey (NHANES) and the National Lung Screening Trial (NLST), the method proposed in the present invention was compared with the k-nearest neighbors algorithm proposed in the document “A benchmark for data imputation methods”, the MICE algorithm proposed in the document “Multiple imputation using chained equations: issues and guidance for practice”, the SVD algorithm proposed in the document “Missing value estimation methods for dnamicroarrays”, the GRAPE algorithm proposed in the document “Handling missing data with graphrepresentation learning”, the Spectral proposed in the document “Spectral regularization algorithms for learning large incomplete matrices”, the AE algorithm proposed in the document “Autoencoder-based multi-task learning for imputation and classification of incomplete data”, and the DIMV algorithm proposed in the document “Conditional expectation withregularization for missing data imputation” for comparative experiments.

[0070] In the benchmark experiment, the present invention uses the National Health and Nutrition Examination Survey (NHANES) and the National Lung Screening Trial (NLST) datasets. For NHANES, the present invention selected data from 17-20 years, which contains various types of medical characteristics such as demographic information, laboratory test results, and lifestyle factors. The present invention uses diabetes as a label for screening, removes repeated fields (such as mg / dl and mmol / L) and useless fields (such as annotation codes), and finally obtains two sets of data sets NH-Diabetes1 (including demographic data and examination data, including 14,673 samples and 38 types of medical characteristics, the positive and negative sample ratio is 1:9, and the overall missing rate is 30%) and NH-Diabetes2 (mainly including laboratory test data, with 14,673 samples and 93 types of medical characteristics, the positive and negative sample ratio is 1:9, and the overall missing rate is 21%). For NLST, the present invention selects comprehensive data such as medical characteristics including participant characteristics, screening results and mortality, and finally obtains a data set containing 11,598 samples, 129 types of medical characteristics, and an overall missing rate of 28% - NLST. In order to reflect the accuracy and stability of the model when dealing with missing value interpolation, the RMSE (root mean square error) indicator is used to evaluate the performance of the algorithm. The results are in Table 1 below, and the best results are highlighted in bold in the table.

[0071] Table 1 RMSE comparison of different interpolation methods on test data, laboratory data and NLST datasets

[0072]

[0073] Experimental results on three datasets show that traditional feature interpolation methods, such as mean interpolation (mean), KNN and SVD, show certain limitations when dealing with the problem of missing medical data, especially on complex data sets, where the RMSE values ​​are generally high and it is difficult to fully recover the information of missing features. Most of these methods rely on global statistics or simple strategies based on sample similarity. Although they are computationally efficient, they are generally ineffective when dealing with high-dimensional and nonlinear data. In comparison, methods such as MICE and Spectral introduce more complex inference mechanisms. Although they have improved on some data sets, their RMSE values ​​are still high, especially on laboratory data sets. The errors of almost all methods are close to 0.7, indicating that the problem of missing features in this data set is more complex and difficult to be effectively solved by traditional interpolation techniques.

[0074] The emerging GRAPE method based on graph neural network and the DIMV method combined with conditional expectation have alleviated this problem to a certain extent compared with traditional methods, showing certain advantages on the detection data set and NLST data set, and the RMSE has decreased. However, the performance of the method of the present invention on all data sets is particularly outstanding, significantly lower than other methods, especially on the NLST data set, where the RMSE value of the method of the present invention is the lowest, showing its excellent ability to deal with complex, nonlinear data missing problems. In addition, even in laboratory data sets, the method of the present invention still achieved good results, proving its robustness and wide adaptability in a variety of different data scenarios.

[0075] In summary, the method of the present invention can better capture the potential structure of the data through innovative interpolation strategies, and effectively deal with the problem of missing features in different data sets, showing strong generalization ability and robustness. This makes the method of the present invention have obvious advantages in complex application scenarios, and can provide higher quality completion results for medical data analysis in practical applications.

[0076] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "rotation" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral one; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0077] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A medical data feature interpolation method based on graph network, characterized in that: The method comprises constructing and training a multi-task graph network interpolation model; constructing the data to be processed as a medical bipartite graph and inputting the trained multi-task graph network interpolation model to obtain interpolated data; the multi-task graph network interpolation model comprises an information extraction module and an MLP, and the information extraction module comprises three cascaded EMRHGRAPHSage modules; The training process of the multi-task graph network interpolation model includes the following steps: S1. Process the acquired medical data into a medical bipartite graph with missing values; S2. Input the medical bipartite graph into the information extraction module to obtain an enhanced medical bipartite graph; S3. Input the enhanced medical bipartite graph into the MLP to obtain edge prediction results and label prediction results; S4. Construct a combined loss function based on the edge-level prediction task and the node-level prediction task, calculate the loss through the combined loss function, and back-propagate the training model parameters until the model converges.

2. A method for medical data feature interpolation based on graph network according to claim 1, characterized in that: Step S1 processes the acquired medical data into a medical bipartite graph with missing values, including: Take each sample in the medical data as a sample node and get the sample node set V A ={u1,u2,…,u m }, m represents the number of samples; each medical feature is taken as a feature node, and the feature node set V is obtained B ={v1,v2,…,v m }, n represents the number of medical feature categories; Construct a feature matrix A of size m×n. The element A in the i-th row and j-th column of the matrix feature A is ij Represents the eigenvalue of the jth feature of the i-th sample; constructs a mask matrix M of size m×n, and the element M in the i-th row and j-th column of the mask feature M ij is a binary variable, if M ij = 0, it means that the jth feature of the ith sample is missing and A cannot be observed ij If the value of M ij =1, it means that the jth feature of the ith sample is not missing, and A can be observed ij The value of According to the feature matrix A and the mask matrix M, a medical bipartite graph G = {V, E}, V = V A ∪V B represents a node set, E={(u i ,v j ,e ij )|u i ∈V A ,v j ∈V B ,M ij =1} represents the edge set, e ij Represents the sample node u i With feature node v j The edge features between ij =1, then the sample node u i With feature node v j There is an edge between them, and e ij =A ij .

3. The method for medical data feature interpolation based on graph network according to claim 1, characterized in that: The EMRHGRAPHSage module includes the EGSAGE module, the edge-aware multi-scale attention module and the heterogeneous edge update module. The processing process of each EMRHGRAPHSage module on the input graph structure includes: S21. The input graph structure passes through the EGSAGE module to obtain the initial representation of each node in the input graph structure; S22. The input graph structure updated in step S21 is passed through an edge-aware multi-scale attention module to obtain an enhanced feature representation of each node; S23. Fusing the input graph structure updated in step S21 with the input graph structure updated in step S22 to obtain a valid feature representation of each node; S24. Pass the input graph structure updated in step S24 through the heterogeneous edge update module to update the edge features of each edge.

4. The method for medical data feature interpolation based on graph network according to claim 3, characterized in that: The EGSAGE module uses a graph neural network architecture inspired by GraphSAGE. The processing process includes: S211. Initialize the embedding vector of each node, expressed as in, represents the embedding vector initialized by node i, x i Represents the feature vector of node i; S212. In the l=1,2,…,Lth layer, update the embedding vector of each node, expressed as Among them, AGG l (·) represents the aggregation function, e ij represents the edge feature between node i and node j, P (l) , Q (l) represents the weight matrix, CONCAT(·) represents the concatenation function, represents the aggregate vector representation of node i at layer l, represents the embedding vector of node i in layer l, σ() represents the activation function; N(i,ε drop ) represents the set of neighbor nodes of node i, ε drop Represents the edge set after dropout processing; S213. Output the updated result of the last layer to obtain the initial representation of each node.

5. The method for medical data feature interpolation based on graph network according to claim 3, characterized in that: The edge-aware multi-scale attention module consists of three parallel convolution branches. The first convolution branch uses a 1×1 convolution, the second convolution branch uses a 3×3 convolution, and the third convolution branch uses a 5×5 convolution.

6. A method for medical data feature interpolation based on graph network according to claim 5, characterized in that: The specific processing process of the edge-aware multi-scale attention module includes: S221. Extracting a node feature matrix and an edge feature matrix of the input graph structure; each row in the node feature matrix represents an initial representation of a node; each row in the edge feature matrix represents an edge feature of an edge; S222. The node feature matrix is ​​passed through the first convolution branch to obtain a local node feature matrix, the node feature matrix is ​​passed through the second convolution branch to obtain a neighborhood node feature matrix, and the node feature matrix is ​​passed through the third convolution branch to obtain a large-scale node feature matrix; S223. Perform different linear transformations on the edge matrix to obtain a first edge matrix, a second edge matrix, and a third edge matrix; S224. The local node feature matrix and the first edge matrix are combined through the edge-aware attention mechanism to obtain a first attention weight matrix; the neighborhood node feature matrix and the second edge matrix are combined through the edge-aware attention mechanism to obtain a second attention weight matrix; the large-scale node feature matrix and the third edge matrix are combined through the edge-aware attention mechanism to obtain a third attention weight matrix; S225. Multiply the local node feature matrix by the first attention weight matrix to obtain a first representation matrix, multiply the neighborhood node feature matrix by the second attention weight matrix to obtain a second representation matrix, and multiply the large-scale node feature matrix by the third attention weight matrix to obtain a third representation matrix; S226. The first representation matrix, the second representation matrix and the third representation matrix are merged to obtain a node enhanced feature representation matrix, thereby obtaining an enhanced feature representation of each node.

7. The method for medical data feature interpolation based on graph network according to claim 6, characterized in that: In step S224, the local node feature matrix and the first edge matrix are connected through the edge-aware attention mechanism, and the local node feature matrix is ​​used as the Q matrix. The adjacent node edge representation matrix and the neighbor node feature matrix are calculated according to the local node feature matrix and the first edge feature matrix as the K matrix and the V matrix respectively, and the first attention weight matrix is ​​calculated according to the Q matrix, the K matrix and the V matrix; wherein, the process of obtaining the adjacent node edge representation matrix and the neighbor node feature matrix is ​​as follows: for each node, all its neighbor nodes are obtained, and the feature representations of all neighbor nodes are extracted from the local node features and fused to obtain the neighbor aggregation feature representation, the edge feature representations between the node and all its neighbor nodes are extracted from the first edge matrix and fused to obtain the edge aggregation feature representation, the neighbor aggregation feature representations of all nodes are composed into a neighbor node feature matrix, and the edge aggregation feature representations of all nodes are composed into an adjacent node edge representation matrix.

8. The method for medical data feature interpolation based on graph network according to claim 3, characterized in that: The heterogeneous edge update module updates the edge features of each edge of the input graph structure, including: Obtain enhanced feature representations of the two nodes connected by the edge, and map the feature nodes connected by the edge through a linear encoder to obtain the inspection item category features containing context information; The edge features, enhanced feature representation of sample nodes, enhanced feature representation of feature nodes, and inspection item category features are concatenated and nonlinearly transformed through MLP to obtain the updated edge features, which can be expressed as e i ' j =σ(W e ·[x i ||x j ||e ij ||(W c ·check_types+b c )]+b e ) Among them, e i ' j represents the updated edge feature, x i represents the enhanced feature representation of node i, check_types represents the check item category feature, || represents the concatenation operation, and W e , W c represents the weight matrix, b e , b c represents the bias term, and σ() represents the activation function.

Images