A medical data feature interpolation method based on a graph network

Abstract

The application relates to the medical field and in particular to a medical data feature interpolation method based on a graph network, which comprises the following steps: constructing and training a multi-task graph network interpolation model; constructing to-be-processed data into a medical bipartite graph and inputting the to-be-processed data 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, the information extraction module comprises three cascaded EMRHGRAPHSage modules; the application solves the shortcomings of traditional interpolation methods when medical data is processed, thereby significantly improving the performance of downstream tasks.

Description

TECHNICAL FIELD

[0001] The present application relates to the medical field, in particular to a medical data feature interpolation method based on a graph network. BACKGROUND

[0002] In reality, medical data often has missing values due to various factors during 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 techniques, as an important means to address data missing problems, have received extensive attention in recent years. However, medical data has high complexity and specificity, making the application of general data interpolation models in this field not ideal. Medical data usually has the following challenges: first, they usually have high dimensionality and heterogeneity, and the correlation between different features (such as physiological indicators, laboratory test results, clinical diagnoses, etc.) is complex and difficult to model; second, the distribution of medical data is often unbalanced, and some rare clinical states or test results are extremely rare in the data, but they may play a key role in disease diagnosis; third, 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, K-nearest neighbor interpolation, etc. often fail to effectively address the above challenges, resulting in interpolation results that lack sufficient robustness and accuracy.

[0003] Although existing GNN methods can capture complex dependency relationships between features, they still have shortcomings in fully utilizing edge information and handling heterogeneous features. Therefore, how to design an interpolation model that combines structured information representation and dynamic dependency modeling capability has become a key problem in current research. SUMMARY

[0004] To solve the shortcomings of existing traditional interpolation methods in handling medical data (such as high dimensionality, heterogeneity, and unbalanced data distribution), the present application proposes a medical data feature interpolation method based on a graph network, which includes constructing and training a multi-task graph network interpolation model; constructing the data to be processed into a medical bipartite graph and inputting it into the trained multi-task graph network interpolation model to obtain the interpolated data; the multi-task graph network interpolation model includes an information extraction module and an 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 reinforced medical bipartite graph into the MLP to obtain the interpolation accuracy and label prediction accuracy;

[0009] S4. construct a combined loss function according to the edge-level prediction task and the node-level prediction task, calculate the loss through the combined loss function, and train the model parameters through back propagation until the model converges.

[0010] The beneficial effects of the present application are:

[0011] The present application constructs a data set containing missing data into a bipartite graph with missing information, in which samples and features are respectively taken as two types of nodes, and the observed feature values are taken as attribute edges connecting the sample and feature nodes. Under this graph structure, the feature interpolation task can be naturally regarded as an edge-level prediction problem, and the label prediction task is regarded as a node-level prediction problem. The present application solves the problems of feature missing and label prediction in a heterogeneous graph structure through modular design. Through effective combination of edge perception multi-scale attention, heterogeneous edge update and residual network modules, the model can capture and fuse the feature information of nodes and edges from multiple perspectives. In addition, the present application adopts the concept of multi-task learning, takes feature interpolation as the main task, and introduces label prediction as an auxiliary task. Through the design of a combined loss function, the joint optimization of the two tasks is realized. The design of multi-task learning improves the model's perception ability of the graph structure, thereby generating higher quality feature representation and ultimately improving the performance of subsequent tasks. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The method flowchart of the present application is shown in the figure;

[0013] Figure 2 The model structure diagram of the present application is shown in the figure;

[0014] Figure 3 The edge perception multi-scale attention (MMA) module flowchart of the present application is shown in the figure. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] The application provides a medical data feature interpolation method based on a graph network, comprising constructing and training a multi-task graph network interpolation model; constructing to-be-processed data as a medical bipartite graph to input 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] As shown in Figure 1 , the multi-task graph network interpolation model training process comprises the following steps:

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

[0019] Specifically, step S1. Process the obtained medical data into a medical bipartite graph with missing values comprises:

[0020] Each sample in the medical data is taken as a sample node to obtain a sample node set V A ={u1,u2,…,u m}, m represents the number of samples; each medical feature is taken as a feature node to obtain a feature node set V B ={v1,v2,…,v m}, n represents the number of medical feature categories; wherein the medical feature refers to various data and indexes used to describe and analyze the health status of individuals in medical research or clinical practice, including demographic information, laboratory test results, lifestyle factors, clinical features, etc.

[0021] A feature matrix A of m*n size is constructed, the element A ij in the i-th row and the j-th column of the matrix feature A represents the feature value of the j-th feature of the i-th sample; due to the problem of missing feature values caused by missing data, a mask matrix M of m*n size is constructed, the element M ij in the i-th row and the j-th column of the mask feature M is a binary variable, if M ij =0, it indicates that the j-th feature of the i-th sample is missing and the value of A ij cannot be observed, if M ij =1, it indicates that the j-th feature of the i-th sample is not missing and the value of A ij can be observed;

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

[0023] Typically, datasets include labels for downstream tasks. Let Y be the label of the i-th sample. i Set a binary variable W for each sample. i ∈{0,1}, through the binary variable W i To split the training / test set, only in W i Y can only be observed in training and testing when Y = 1. i Furthermore, due to the inherent gaps in medical data, this invention, after generating the mask matrix M, will... ij The training and test sets are constructed using the part equal to 1. After iterating and optimizing to find the optimal model, M is then used. ij The model is partially input with a value of 0 to obtain the final interpolation result, and the accuracy of the interpolation is verified using a downstream task (label prediction).

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

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

[0026] S21. The input graph structure is processed by 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 the edge-aware multi-scale attention module to obtain the enhanced feature representation of each node;

[0028] S23. Merge the input graph structure updated in step S21 and the input graph structure updated in step S22 to obtain the effective feature representation of each node;

[0029] S24. The input graph structure updated in step S24 is passed through a 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, and the processing procedure includes:

[0031] S211. The embedding vector of each node is initialized, denoted as

[0032]

[0033] wherein, represents the initialized embedding vector of node i, x i represents the feature vector of node i;

[0034] S212. In the l = 1, 2, …, L layers, the edge information is supplemented as the node information, and the message passing function takes the concatenation of the node embedding vector and the edge feature as the input to update the embedding vector of each node, denoted as

[0035]

[0036]

[0037] wherein, 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 aggregated vector of node i in the l-th layer, and represents the embedding vector of node i in the l-th layer, and σ() represents the activation function, which is ReLU here; N(i, ε drop ) represents the neighbor node set of node i, and ε drop represents the edge set with dropout processing;

[0038] S213. The update result of the last layer is output 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 understanding of graph structure information by the graph neural network. Specifically, the edge-aware multi-scale attention module includes three parallel convolution branches, the first convolution branch adopts a 1x1 convolution, the second convolution branch adopts a 3x3 convolution, and the third convolution branch adopts a 5x5 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 the feature information in a larger range. The output of each branch generates a corresponding feature matrix for further attention weighting.

[0040] Specifically, as shown in Figure 3 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 needs to be particularly noted that the node feature matrix obtained here is obtained by processing the last 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 obtained in step S1).

[0042] S222. The node feature matrix is obtained by the first convolution branch to obtain the local node feature matrix, the node feature matrix is obtained by the second convolution branch to obtain the neighborhood node feature matrix, and the node feature matrix is obtained by the third convolution branch to obtain the large-range node feature matrix.

[0043] S223. Different linear transformations are performed on the edge matrix to obtain the first edge matrix, the second edge matrix, and the third edge matrix, which maps the feature dimension of each edge to the feature dimension of the corresponding node, so as to operate with the node feature when calculating the attention weight subsequently.

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

[0045] Specifically, the edge-aware attention mechanism performs weighted calculation on the features of each node and its adjacent nodes and edge information, which is completed through Q, K, and V matrices (i.e., query, key, and value). One of the input node feature matrices is used as the Q matrix, the adjacent node edge representation matrix is calculated as the K matrix, and the neighbor node feature matrix is used as the V matrix, which can be represented as a whole

[0046]

[0047] d k denotes the number of feature dimensions. It is described in detail that the local node feature matrix and the first edge matrix pass through the edge-aware attention mechanism to obtain the first attention weight matrix, the local node feature matrix is 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, and the first attention weight matrix is calculated according to the Q matrix, the K matrix and the V matrix; wherein, the acquisition process of the adjacent node edge representation matrix and the neighbor node feature matrix is: for each node, all neighbor nodes are obtained, and the feature representation of all neighbor nodes is extracted and fused in the local node feature to obtain the neighbor aggregated feature representation, the edge feature representation between the node and all neighbor nodes is extracted and fused in the first edge matrix to obtain the edge aggregated feature representation, the neighbor aggregated feature representation of all nodes is combined to form the neighbor node feature matrix, and the edge aggregated feature representation of all nodes is combined to form the adjacent node edge representation matrix. The neighbor aggregated feature representation, the edge aggregated feature representation and the feature representation of the node in the local node feature matrix are of the same size.

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

[0049] Wherein, the multiplication of the local node feature matrix and the first attention weight matrix to obtain the first representation matrix can be represented as

[0050]

[0051] Wherein, W1 represents the weight matrix corresponding to 1x1 convolution, h j denotes the initial representation of node j in the node feature matrix, denotes the first attention weight between node i and node j, denotes 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 fused to obtain a node enhanced feature representation matrix, so as to obtain the enhanced feature representation of each node.

[0053] Specifically, as Figure 2To effectively reserve and utilize the input information, the present application adopts a 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 and the output of the MMA module are weighted and summed. Through the residual connection, the original input features can be directly transmitted 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 expression ability 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 the edge features by using the features of adjacent nodes, edge features and type information of the check item (i.e. the category of medical features). This module combines various sources of context information to strengthen the representation ability of the edge, so that the model can better capture the complex relationship 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, obtains the enhanced feature representation of the two nodes, maps the feature node connected by the edge through a linear encoder to obtain the check item category feature containing context information; the edge features, the enhanced feature representation of the sample node, the enhanced feature representation of the feature node, and the check item category feature are concatenated and then nonlinearly converted through an MLP to obtain the updated edge features, which can be represented as

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

[0057] Where e i ' j represents the updated edge features, x i represents the enhanced feature representation of node i, check_types represents the check item category feature, || represents the concatenation operation, W e , W c represent weight matrices, b e , b c represent bias terms, and σ() represents an activation function.

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

[0059] Specifically, the present application includes two tasks of 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 is an edge between any two nodes in the graph and the potential weight of the edge; in the present application, the missing feature imputation is used as the edge-level prediction. The missing feature imputation can be represented as learning an edge value prediction mapping:

[0060]

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

[0062] The present application takes label prediction as the node-level prediction. The prediction of the downstream node label can be represented as learning a mapping:

[0063]

[0064] wherein, represents the label prediction result of sample i, g i (·) represents a function for predicting the node label, that is, a node-level prediction mapping. In the training process, by minimizing the difference between the label prediction result and the true label Y i , the model gradually learns to accurately predict the node label in the case of incomplete features.

[0065] S4. Construct a combined loss function according to the edge-level prediction task and the node-level prediction task, calculate the loss through the combined loss function, and train the model parameters through back propagation 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 an adaptive coefficient a, 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. The combined loss can adaptively balance the contributions of the two tasks, ensuring that the model can consider both feature missing and label prediction when processing, thereby improving the performance of the overall model. Loss can be represented as:

[0067] L = a · L edge + (1-a) · L node

[0068] wherein, 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 determines the proportion of importance of the two tasks in the total loss.

[0069] In order to verify the effectiveness of the method of the present application, based on the United States National Health and Nutrition Examination Survey (NHANES) and the National Lung Screening Trial (NLST), the method proposed in the present application is compared with the k-nearest neighbors algorithm proposed in the literature "A benchmark for data imputation methods", the MICE algorithm proposed in the literature "Multiple imputation using chained equations: issues and guidance for practice", the SVD algorithm proposed in the literature "Missing value estimation methods for DNA microarrays", the GRAPE algorithm proposed in the literature "Handling missing data with graph representation learning", the Spectral algorithm proposed in the literature "Spectral regularization algorithms for learning large incomplete matrices", the AE algorithm proposed in the literature "Autoencoder-based multi-task learning for imputation and classification of incomplete data", and the DIMV algorithm proposed in the literature "Conditional expectation with regularization for missing data imputation".

[0070] In the benchmark experiments, the present application uses the National Health and Nutrition Examination Survey (NHANES) and National Lung Screening Trial (NLST) datasets. For NHANES, the present application selects the data from 17-20 years, which contains various types of medical features such as demographic information, laboratory test results, lifestyle factors, etc. The present application screens for diabetes as a label, removes duplicate fields (such as mg / dl and mmol / L) and useless fields (such as annotation codes), and finally the present application obtains two datasets NH-Diabetes1 (including demographic data and examination data, with 14673 samples and 38 types of medical features, the positive to negative sample ratio is 1:9, and the overall missing rate is 30%) and NH-Diabetes2 (mainly including laboratory test data, with 14673 samples and 93 types of medical features, the positive to negative sample ratio is 1:9, and the overall missing rate is 21%). For NLST, the present application selects medical features including participant characteristics, screening results, and mortality, etc. comprehensive data, and finally obtains a dataset-NLST containing 11598 samples, 129 types of medical features, and an overall missing rate of 28%. In order to reflect the accuracy and stability of the model in handling missing value imputation, the RMSE (Root Mean Square Error) index is used to evaluate the performance of the algorithm. The results are shown in Table 1 below, and the best results are highlighted in bold in the table.

[0071] Table 1 RMSE comparison of different imputation methods on detection data, laboratory data and NLST dataset

[0072]

[0073] The experimental results on the three datasets show that traditional feature imputation methods such as mean, KNN and SVD have certain limitations in handling missing medical data problems, especially on complex structured datasets, the RMSE value is generally high, and it is difficult to fully recover the information of the missing features. These methods mostly rely on global statistics or simple strategies based on sample similarity, although the calculation efficiency is high, but the effect is general when dealing with high-dimensional, nonlinear data. In comparison, methods such as MICE and Spectral introduce more complex inference mechanisms, although there is some improvement on some datasets, but the RMSE value is still high, especially on the laboratory dataset, the error of almost all methods is close to 0.7, indicating that the feature missing problem of this dataset is more complex, and it is difficult to be effectively solved by traditional imputation techniques.

[0074] The emerging GRAPE method based on graph neural network and the DIMV method combined with conditional expectation alleviate this problem to some extent compared with traditional methods, and show certain advantages on the detection dataset and the NLST dataset, and the RMSE is reduced. However, the performance of the method of the present application on all datasets is particularly outstanding, significantly lower than other methods, especially on the NLST dataset, the RMSE value of the method of the present application is the lowest, which shows its excellent ability in dealing with complex, nonlinear data missing problems. In addition, even in the laboratory dataset, the method of the present application still achieves good results, proving its robustness and wide adaptability in a variety of different data scenarios.

[0075] In summary, the method of the present application can better capture the potential structure of the data through the innovative interpolation strategy, and effectively deal with the feature missing problem in different datasets, showing strong generalization ability and robustness. This makes the method of the present application 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 application, unless otherwise explicitly specified and limited, the terms "mounting", "setting", "connecting", "fixing", "rotating" and the like should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium; it can be the internal communication of two elements or the interaction relationship between two elements, unless otherwise explicitly limited, the above-mentioned terms in the present application can be understood according to the specific meaning in the present application by the person skilled in the art according to the specific circumstances.

[0077] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for medical data feature imputation based on graph network, characterized in that, The method comprises the following steps: The multi-task graph network imputation 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 imputation model comprises the following steps: S1. processing the obtained medical data into a medical bipartite graph with missing values; S2. inputting the medical bipartite graph into the information extraction module to obtain a strengthened medical bipartite graph; The EMRHGRAPHSage module comprises an EGSAGE module, an edge-aware multi-scale attention module and a heterogeneous edge updating module, and the processing process of each EMRHGRAPHSage module on the input graph structure comprises: S21. inputting the graph structure through the EGSAGE module to obtain the initial representation of each node in the input graph structure; S22. inputting the updated input graph structure of step S21 through the edge-aware multi-scale attention module to obtain the enhanced feature representation of each node; S23. fusing the updated input graph structure of step S21 and the updated input graph structure of step S22 to obtain the effective feature representation of each node; S24. inputting the updated input graph structure of step S24 through the heterogeneous edge updating module to update the edge feature of each edge; The heterogeneous edge updating module updates the edge feature of each edge of the input graph structure, comprising: obtaining the enhanced feature representation of the two nodes connected by the edge, and mapping the feature nodes connected by the edge through a linear encoder to obtain the inspection item category feature containing context information; e′ ij = σ(W e · [x i || x j || e ij || (W c · check_types + b c )] + b e ) where e' = e + Δe ij 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 splicing operation, W e , W c represents the weight matrix, b e , b c represents the bias term, and σ() represents the activation function. splicing the edge feature, the enhanced feature representation of the sample node, the enhanced feature representation of the feature node and the inspection item category feature, and then performing nonlinear conversion through the MLP to obtain the updated edge feature, which can be represented as S3. inputting the strengthened medical bipartite graph into the MLP to obtain the edge prediction result and the label prediction result; 2. The medical data feature imputation method based on graph network according to claim 1, characterized in that, S4. constructing a combined loss function according to the edge-level prediction task and the node-level prediction task, calculating the loss through the combined loss function, and training the model parameters through back propagation until the model converges. Each sample in the medical data is treated as a sample node, resulting in a sample node set V. A ={u1,u2,…,u m }, where m represents the number of samples; each medical feature is treated as a feature node, resulting in a feature node set V. B ={v1,v2,…,v m }, where n represents the number of medical feature categories; A feature matrix A of size m x n is constructed, where the element A ij represents the value of the jth feature of the ith sample; a mask matrix M of size m x n is constructed, where the element M ij is a binary variable, M ij = 0 indicates that the jth feature of the ith sample is missing and the value of A ij cannot be observed, M ij = 1 indicates that the jth feature of the ith sample is not missing and the value of A ij can be observed; According to the feature matrix A and the mask matrix M, a medical bipartite graph G = {V, E} is constructed, 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 an edge set, and e ij represents an edge feature between the sample node u i and the feature node v j . If M ij = 1, there is an edge between the sample node u i and the feature node v j , and e ij = A ij .

3. The medical data feature imputation method based on graph network according to claim 1, characterized in that, Step S1 of processing the obtained medical data into a medical bipartite graph with missing values comprises: The EGSAGE module adopts a graph neural network architecture based on GraphSAGE heuristic, and the processing process comprises: wherein, xi represents an initialized embedding vector of node i, x i xi represents a feature vector of node i; S211. initializing the embedding vector of each node, represented as where AGG l denotes an aggregation function, e ij denotes the edge feature between node i and node j, P (l) , Q (l) denotes a weight matrix, CONCAT(·) denotes a concatenation function, denotes the aggregated vector representation of node i at the l-th layer, denotes the embedding vector of node i at the l-th layer, σ() denotes an activation function; N(i,ε drop ) denotes the neighbor node set of node i, ε drop denotes the edge set with dropout processing; S212. updating the embedding vector of each node at the l = 1, 2, …, L layer, represented as 4. The medical data feature imputation method based on graph network according to claim 1, characterized in that, S213. outputting the update result of the last layer to obtain the initial representation of each node.

5. The medical data feature imputation method based on graph network according to claim 4, characterized in that, The edge-aware multi-scale attention module comprises three parallel convolution branches, the first convolution branch adopts 1 1×1 convolution, the second convolution branch adopts 1 3×3 convolution, and the third convolution branch adopts 1 5×5 convolution. The specific processing process of the edge-aware multi-scale attention module comprises: S221. Extract the node feature matrix and the edge feature matrix of the input graph structure; 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; S222. Obtain the local node feature matrix by passing the node feature matrix through the first convolution branch, obtain the neighborhood node feature matrix by passing the node feature matrix through the second convolution branch, and obtain the wide-range node feature matrix by passing the node feature matrix through the third convolution branch; S223. Perform different linear transformations on the edge matrix to obtain the first edge matrix, the second edge matrix, and the third edge matrix; S224. Obtain the first attention weight matrix by passing the local node feature matrix and the first edge matrix through the edge-aware attention mechanism; obtain the second attention weight matrix by passing the neighborhood node feature matrix and the second edge matrix through the edge-aware attention mechanism; and obtain the third attention weight matrix by passing the wide-range node feature matrix and the third edge matrix through the edge-aware attention mechanism; S225. Multiply the local node feature matrix and the first attention weight matrix to obtain the first representation matrix, multiply the neighborhood node feature matrix and the second attention weight matrix to obtain the second representation matrix, and multiply the wide-range node feature matrix and the third attention weight matrix to obtain the third representation matrix; S226. Fuse the first representation matrix, the second representation matrix, and the third representation matrix to obtain the node enhanced feature representation matrix, thereby obtaining the enhanced feature representation of each node.

6. The medical data feature imputation method based on graph network according to claim 5, characterized in that, In step S224, the local node feature matrix and the first edge matrix are passed through the edge-aware attention mechanism, 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, and the first attention weight matrix is calculated according to the Q matrix, the K matrix, and the V matrix; wherein the acquisition process of the adjacent node edge representation matrix and the neighbor node feature matrix is as follows: for each node, all neighbor nodes are obtained, the feature representation of all neighbor nodes is extracted from the local node feature and fused to obtain the neighbor aggregated feature representation, the edge feature representation between the node and all neighbor nodes is extracted from the first edge matrix and fused to obtain the edge aggregated feature representation, the neighbor aggregated feature representations of all nodes are combined to obtain the neighbor node feature matrix, and the edge aggregated feature representations of all nodes are combined to obtain the adjacent node edge representation matrix.

Images