Heterogeneous graph and graph neural network-based emphysema risk prediction method and equipment

Through the method based on heterogeneous graph and graph neural network, multi-source medical data is integrated, and the subjectivity and low sensitivity of emphysema risk prediction in the prior art are solved, achieving more efficient and accurate emphysema risk prediction .

CN118942702AInactive Publication Date: 2024-11-12CENT SOUTH UNIV
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411041793.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-31
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art has subjectivity and low sensitivity of imaging examinations in the prediction of emphysema risk, and multi-source medical data has not been fully utilized, resulting in low prediction accuracy.

Method used

Using a method based on heterogeneous graph and graph neural network, the heterogeneous graph is constructed and the graph neural network model is captured by acquiring multi-source data (CT images, physiological data, biochemical marker data) for preprocessing and fusion, and the heterogeneous graph is constructed and the graph neural network model is trained to capture the data between the data. complex relationships identify subtle pathological features of emphysema.

Benefits of technology

It improves the accuracy and efficiency of emphysema risk prediction, can identify the complex pathological characteristics of emphysema, and provide more accurate prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118942702A_ABST
    Figure CN118942702A_ABST
Patent Text Reader

Abstract

The invention provides an emphysema risk prediction method and device based on a heterogeneous graph and a graph neural network, and the method integrates emphysema medical data from different sources, extracts key features from multi-source data through the construction of the heterogeneous graph, and comprehensively analyzes a plurality of data sources through the graph neural network. Complex relations and interaction among different data types can be captured, subtle and complex pathological features of emphysema can be identified, and emphysema risk prediction accuracy and efficiency are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of medical information processing and statistical analysis, and in particular relates to a method and device for predicting emphysema risk based on heterogeneous graphs and graph neural networks. Background Art

[0002] Emphysema is a common chronic obstructive pulmonary disease (COPD), which is characterized by permanent expansion of the alveoli and destruction of their walls, leading to a significant decrease in respiratory function. Patients often experience symptoms such as dyspnea, cough, and sputum, which seriously affect their quality of life. The causes of emphysema are complex, including smoking, air pollution, occupational exposure, and genetic factors. Currently, the number of patients suffering from emphysema is increasing worldwide, and it has become a serious public health problem.

[0003] The impact of emphysema on the human body is mainly reflected in two aspects: first, damage to the respiratory system. The destruction of alveoli and obstruction of airways seriously impair the patient's ventilation and gas exchange functions, oxygen cannot fully enter the blood, and carbon dioxide cannot be effectively discharged, resulting in hypoxemia and hypercapnia in patients. Secondly, it affects the cardiovascular system. Due to long-term hypoxia, patients often experience pulmonary hypertension, which in turn leads to right heart failure, further aggravating the condition.

[0004] At present, the risk prediction of emphysema mainly relies on imaging examinations and pulmonary function tests. CT (computed tomography) is the most commonly used imaging examination method, which can clearly show the lung structure and pathological conditions. By observing the changes in the alveoli and the obstruction of the airways, doctors can preliminarily determine whether the patient has emphysema. However, the interpretation of CT examinations depends on the doctor's experience and has a certain degree of subjectivity and uncertainty. Pulmonary function tests mainly evaluate the lung function status by measuring the patient's vital capacity, forced expiratory volume and other indicators. Although it has a certain objectivity, the sensitivity of the test is low for patients with early emphysema.

[0005] With the development of artificial intelligence and machine learning technologies, more and more studies have tried to apply these technologies to assist in the diagnosis of emphysema. However, imaging, physiological and biochemical data are often processed in a scattered manner, and the information is not fully utilized for accurate judgment, resulting in low accuracy in emphysema risk prediction. Summary of the invention

[0006] The purpose of the present invention is to address the shortcomings of the existing technology and provide a method and device for predicting emphysema risk based on heterogeneous graphs and graph neural networks, which improves the accuracy and efficiency of emphysema risk prediction by fusing and analyzing multi-source medical data.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A method for predicting emphysema risk based on heterogeneous graphs and graph neural networks, comprising the following steps:

[0009] S1: Acquire multi-source data of emphysema patients, wherein the multi-source data includes CT images, physiological data, and biochemical marker data;

[0010] S2: preprocessing the multi-source data;

[0011] S3: Construct a heterogeneous graph based on the preprocessed multi-source data to obtain the node feature matrix and adjacency matrix;

[0012] S4: using the node feature matrix and the adjacency matrix as inputs of a graph neural network model to train the graph neural network model;

[0013] The graph neural network model includes a plurality of cascaded convolutional networks, and the convolutional network includes a convolutional layer, an activation layer, and a regularization layer connected in sequence;

[0014] The graph neural network model training process includes:

[0015] The adjacency matrix A is standardized to obtain a standardized adjacency matrix

[0016] Through convolution and activation operations, the node feature matrix is ​​updated. The expression of the updated node feature matrix is ​​as follows:

[0017] Normalize the node features output by the last layer of convolutional network to obtain the probability that each node belongs to a different category

[0018] The graph neural network model was iteratively trained using the cross entropy loss function to obtain an emphysema risk prediction model.

[0019] Where D is the degree matrix, H (l) is the node feature matrix of the l-th layer convolutional network, H (l+1) is the node feature matrix of the l+1th layer convolutional network, is the normalized adjacency matrix, W (l) is the weight matrix of the l-th layer convolutional network, σ is the ReLU activation function, Softmax is the activation function, H i is the feature vector of node i in the last layer of convolutional network.

[0020] Emphysema is predicted using the emphysema risk prediction model.

[0021] The present invention integrates emphysema medical data from different sources and comprehensively analyzes multiple data sources through heterogeneous graphs and graph neural networks. It can capture the complex relationships and interactions between different data types, identify the subtle and complex pathological characteristics of emphysema, and improve the accuracy and efficiency of emphysema risk prediction.

[0022] Furthermore, the implementation process of S2 includes:

[0023] Adaptive histogram equalization technology is used to enhance the contrast of CT images, and Laplacian filter is used to sharpen the edges of CT images after histogram equalization.

[0024] Convert physiological data into standardized scores;

[0025] The principal component analysis method was used to reduce the dimensionality of biochemical marker data.

[0026] By preprocessing CT images to improve image visibility and highlight edge details, the boundaries of the lesion area are made clearer, facilitating subsequent feature extraction; by standardizing physiological data, the effects of different measurement units can be eliminated and data consistency can be ensured for comparison; by using principal component analysis to reduce the dimensionality of biochemical marker data, the main features can be extracted.

[0027] Furthermore, the implementation process of S3 includes:

[0028] The preprocessed multi-source data is fused using a fully connected network to obtain the node feature matrix H = FCM (F CT ,F Physioi ,F Physio ,F Bio );

[0029] Establish edges between each node feature;

[0030] The adjacency matrix of the heterogeneous graph is constructed through the features of each node and the edges between the node features. The expression of the adjacency matrix A is as follows:

[0031]

[0032] Among them, FCN represents a fully connected network, F CT represents the CT image node feature vector, F Physio represents the physiological data node feature vector, F Bio represents the biochemical marker node feature vector, w ij represents the edge weight between node i and node j.

[0033] By constructing heterogeneous graphs, we can capture the complex relationships between features and extract key features from these multi-source data.

[0034] Furthermore, the correlation between nodes with different features is identified by calculating the similarity between node features of the last layer of convolutional network;

[0035] The node features of the last layer of convolutional network are clustered by clustering method to identify node feature groups.

[0036] Being able to predict possible but unobserved edges in heterogeneous graphs and identify closely related groups of nodes in heterogeneous graphs can help reveal potential biomarkers or pathological processes of emphysema.

[0037] Based on the same inventive concept, the present invention also provides an electronic device, including:

[0038] one or more processors;

[0039] A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of a method for predicting emphysema risk based on a heterogeneous graph and a graph neural network.

[0040] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for predicting emphysema risk based on heterogeneous graphs and graph neural networks.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] The present invention integrates and analyzes emphysema medical data from different sources, and comprehensively analyzes multiple data sources through heterogeneous graphs and graph neural networks, so as to identify the subtle and complex pathological characteristics of emphysema and improve the accuracy and efficiency of emphysema risk prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the emphysema risk prediction method based on heterogeneous graph and graph neural network of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be described in detail below in conjunction with the embodiments. It should be noted that the embodiments and features of the embodiments of the present invention can be combined with each other without conflict. For the convenience of description, if the words "upper", "lower", "left" and "right" appear in the following, they only indicate that the upper, lower, left and right directions are consistent with the drawings themselves, and do not limit the structure.

[0045] Example

[0046] like Figure 1As shown, this embodiment proposes a pulmonary emphysema risk prediction system based on heterogeneous graphs and graph neural networks (GNNs), which integrates multimodal data analysis technology and includes the following steps:

[0047] Step 1: Data Collection

[0048] The data mainly includes lung CT image data, physiological parameter data, and biochemical marker data.

[0049] A1: Lung CT image data

[0050] Lung CT images are used to identify lesion areas and assess lung tissue damage and emphysema areas. All images conform to the DICOM standard and contain necessary technical parameters.

[0051] A2: Physiological data

[0052] Physiological parameter data include lung function test results, blood oxygen saturation, etc. These data files are usually stored in CSV format, and each file contains all parameter values ​​at multiple timestamps.

[0053] A3: Biochemical marker data

[0054] Biochemical marker data include C-reactive protein (CRP), white blood cell count, etc. These data help assess the patient's inflammatory status and infection. These data files are usually stored in JSON format.

[0055] Step 2: Data Preprocessing

[0056] Data preprocessing is a key step to ensure the accuracy and efficiency of subsequent analysis. In this embodiment, all kinds of data are subjected to a special preprocessing process to ensure data quality and consistency and optimize the subsequent deep learning processing effect.

[0057] B1: Image data preprocessing

[0058] For lung CT image data, image enhancement algorithms are applied to improve prediction accuracy. Specifically, image contrast is enhanced using adaptive histogram equalization technology, which adjusts the histogram distribution of the image to increase the visibility of low-contrast areas that are considered lesions. In addition, an edge detection-based sharpening algorithm is applied, which uses the Laplacian filter to highlight the edge details of the image, making the boundaries of the lesion area clearer and facilitating subsequent feature extraction.

[0059] The application of image enhancement algorithm is described in detail as follows:

[0060] Histogram equalization process: In order to enhance the contrast of lung CT images, adaptive histogram equalization is used. The specific operation can be expressed as:

[0061]

[0062] Among them, O(x,y) is the pixel value of the output image, I(x,y) is the pixel value of the original image, and p(r) is the histogram distribution of the original image. This process adjusts the histogram to improve the visibility of the image, especially in low-contrast areas.

[0063] Edge sharpening process: In order to more clearly display the edge of the lesion area in the CT image, a sharpening algorithm based on edge detection is applied. The specific operation can be expressed as follows:

[0064] I sharp =I+λ·(L*I)

[0065] Among them, I is the image after histogram equalization, L is the Laplacian filter, λ is the intensity adjustment coefficient, and * represents the convolution operation, which enhances the edge by highlighting the high-frequency part of the image.

[0066] B2: Physiological data preprocessing

[0067] For physiological data, standardized score technology is used: various physiological data are converted into standardized scores for comparison. The operation formula is as follows:

[0068]

[0069] Among them, X is the original score, μ1 is the mean value of the physiological data sample, and δ is the sample standard deviation. This can eliminate the influence of different measurement units and ensure the consistency of the data.

[0070] B3: Biochemical marker data preprocessing

[0071] Use principal component analysis (PCA) to reduce the data dimension and extract the main features:

[0072]

[0073] PCA(x1)=W T (x1-μ)

[0074] Among them, x i is the data vector of each sample, μ is the mean of the biochemical marker data, ∑ is the covariance matrix, W is the eigenvector matrix of the covariance matrix, W T is the transpose of the eigenvector matrix of the covariance matrix, and x1 is the biochemical marker data.

[0075] Step 3: Constructing a heterogeneous graph

[0076] In this embodiment, the construction of the heterogeneous graph includes the following contents:

[0077] C1: Multimodal Fusion Algorithm

[0078] The pre-processed multi-source data are integrated into a unified feature representation using a multi-modal fusion algorithm. The node feature matrix H is formed by concatenation, weighted summation or neural network fusion.

[0079] Use neural networks to nonlinearly fuse features from different data sources. Neural networks can capture the complex relationships between features and generate richer and more abstract comprehensive feature vectors.

[0080] H=FCM(F CT ,F Physio ,F Bio )

[0081] Among them, FCN represents a fully connected network, which is used to fuse features.

[0082] C2: Define Node

[0083] Define nodes of different data types:

[0084] Lung CT image feature node: the node feature vector is represented by F CT .

[0085] Physiological data node: The node feature vector is represented by F Physio .

[0086] Biochemical marker node: The node feature vector is represented by F Bio .

[0087] C3: Create edges

[0088] The edges between nodes are established based on the biological relevance, functional association or statistical relevance between data. Specifically, by consulting medical literature and databases, combined with expert knowledge, known biological relationships are determined and fixed weights are assigned to these relationships. Through functional network analysis and experimental research, methods such as Granger causal analysis are used to identify the functional relationships between different features, and weights are calculated based on the analysis results.

[0089] The edge between the lung CT image feature node and the biochemical marker node represents the relationship between the lung CT image feature and the biochemical marker.

[0090] The edge between the physiological data node and the biochemical marker node indicates the association between the physiological parameter change and the biochemical marker node.

[0091] C4: Constructing the adjacency matrix

[0092] By defining nodes and edges, we construct the adjacency matrix A of the heterogeneous graph. The element w in the adjacency matrix is ijrepresents the edge weight between node i and node j. If there is an edge between i and j, w ij The value of is the weight of the edge; otherwise, the value is 0.

[0093]

[0094] Step 4: Graph Neural Network Analysis

[0095] The graph neural network model is trained with the node feature matrix of the heterogeneous graph and its adjacency matrix A as input. The graph neural network is used to extract and identify pathological features, and features related to emphysema are identified by classifying node embeddings.

[0096] D1: Training of graph neural network models

[0097] Graph Neural Network (GNN) is a deep learning model for processing graph structured data. GNN learns the feature representation of nodes and edges by transferring and aggregating information between nodes in the graph. In this embodiment, a heterogeneous graph and an adjacency matrix A are used as input to train and analyze the model.

[0098] 1. Graph Convolutional Layer (GCN Layer):

[0099] In each layer of graph convolution, the feature representation of a node is updated by the features of its neighboring nodes.

[0100]

[0101] Among them, H (l) is the node feature matrix of the l-th layer convolutional network, H (l+1) is the node feature matrix of the l+1th layer convolutional network, is the normalized adjacency matrix, W (l) is the weight matrix of the l-th layer graph convolution, and σ is the ReLU activation function.

[0102] 2. Standardization of adjacency matrix:

[0103] To ensure numerical stability and model convergence, the adjacency matrix needs to be normalized.

[0104] The normalization formula is as follows:

[0105]

[0106] Where D is the degree matrix and A is the original adjacency matrix.

[0107] 3. Model training process:

[0108] Definition of loss function: Use cross entropy loss function for training node classification tasks.

[0109]

[0110] Among them, y i is the true label, is the predicted probability.

[0111] Backpropagation and optimization: Update the model parameters using a gradient descent algorithm such as the Adam optimizer.

[0112]

[0113] Among them, θ t is the value of the model parameter at the tth iteration, η is the learning rate, is the gradient of the loss function with respect to the parameters θ.

[0114] D2: Application of Graph Neural Networks

[0115] 1. Node classification:

[0116] GNN is used to classify nodes and identify nodes of different categories, for example, distinguishing healthy and diseased nodes. The classification results are normalized by the Softmax function to obtain the probability that each node belongs to a different category.

[0117]

[0118] Among them, H i is the feature vector of node i in the last convolutional network layer.

[0119] 2. Link Prediction:

[0120] Predict edges that may exist in the graph but have not yet been observed, such as predicting which biochemical indicators may be associated with specific CT features. Link prediction is achieved by calculating the similarity between node features, using dot products here.

[0121] Similarity(u,v)=H u ·H v

[0122] Among them, H u and H v is the feature representation of node u and node v, and Similarity represents the similarity between node u and node v.

[0123] 3. Cluster analysis:

[0124] Identify closely related groups of nodes in the graph, which can help reveal potential biomarkers or pathological processes of emphysema. Cluster the node features using unsupervised learning methods such as Kmeans clustering.

[0125]

[0126] Where k is the number of clusters, C i is the i-th cluster, x is the feature vector of the node, μ i is the center of the ith cluster.

[0127] Through the above steps, the graph neural network model can effectively extract and identify the pathological characteristics of emphysema, provide accurate risk prediction results, and improve the efficiency and accuracy of risk prediction of emphysema.

[0128] Step 5: Automatic prediction

[0129] A risk prediction report is generated based on the results of graph neural network analysis, which includes the identified pathological features and abnormal nodes.

[0130] After the GNN analysis, the generated risk prediction report includes the following:

[0131] Identified pathological features: GNN is used to classify node embeddings and identify features related to emphysema, for example, which CT image features and biochemical markers are significantly abnormal.

[0132] Abnormal node marking: Abnormal nodes are marked in the heterogeneous graph, which may correspond to lesion areas or abnormal physiological and biochemical indicators.

[0133] 1. Node embedding representation:

[0134] GNN generates embedded representations of nodes through layer-by-layer convolution and aggregation operations:

[0135]

[0136] The node embedding representation in the final layer is used to classify and label abnormal nodes.

[0137] 2. Classification result processing:

[0138] Use the Softmax function to normalize the node embedding to get the probability that each node belongs to a different category:

[0139]

[0140] Based on the classification results, pathological features and abnormal nodes related to emphysema were identified.

[0141] This embodiment provides an electronic device, including:

[0142] one or more processors;

[0143] A memory having one or more programs stored thereon, which, when executed by one or more processors, enables the one or more processors to implement the steps of a method for predicting emphysema risk based on heterogeneous graphs and graph neural networks.

[0144] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0145] In some other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors of various types, which are not limited herein.

[0146] This embodiment provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of a method for predicting emphysema risk based on a heterogeneous graph and a graph neural network.

[0147] The contents explained in the above embodiments should be understood as these embodiments are only used to more clearly illustrate the present invention, and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

Claims

1. A method for predicting emphysema risk based on heterogeneous graphs and graph neural networks, characterized in that: The following steps are involved: S1: Acquire multi-source data of emphysema patients, wherein the multi-source data includes CT images, physiological data, and biochemical marker data; S2: preprocessing the multi-source data; S3: Construct a heterogeneous graph based on the preprocessed multi-source data to obtain the node feature matrix and adjacency matrix; S4: using the node feature matrix and the adjacency matrix as inputs of a graph neural network model, and using a cross entropy loss function to iteratively optimize and train the graph neural network model to obtain an emphysema risk prediction model; The graph neural network model includes multiple cascaded convolutional networks, and the convolutional network includes a convolutional layer, an activation layer and a regularization layer connected in sequence.

2. The method for predicting emphysema risk based on heterogeneous graph and graph neural network according to claim 1, characterized in that: The implementation process of S2 includes: Adaptive histogram equalization technology is used to enhance the contrast of CT images, and Laplacian filter is used to sharpen the edges of CT images after histogram equalization. Convert physiological data into standardized scores; The principal component analysis method was used to reduce the dimensionality of biochemical marker data.

3. The method for predicting emphysema risk based on heterogeneous graph and graph neural network according to claim 1, characterized in that: The implementation process of S3 includes: The preprocessed multi-source data is fused using a fully connected network to obtain the node feature matrix H = FCM (F CT , F Physio , F Physio , F Bio ); Establish edges between each node feature; The adjacency matrix of the heterogeneous graph is constructed through the features of each node and the edges between the node features. The expression of the adjacency matrix A is as follows: Among them, FCN represents a fully connected network, F CT represents the CT image node feature vector, F Physio represents the physiological data node feature vector, F Bio represents the biochemical marker node feature vector, w ij represents the edge weight between node i and node j.

4. The method for predicting emphysema risk based on heterogeneous graph and graph neural network according to claim 1, characterized in that: The correlation between nodes with different features is identified by calculating the similarity between the node features of the last layer of the convolutional network.

5. The method for predicting emphysema risk based on heterogeneous graph and graph neural network according to claim 1, characterized in that: The node features of the last layer of convolutional network are clustered by clustering method to identify node feature groups.

6. The method for predicting emphysema risk based on heterogeneous graph and graph neural network according to claim 1, characterized in that: Emphysema is predicted using the emphysema risk prediction model.

7. An electronic device, characterized in that: include: one or more processors; A memory having one or more programs stored thereon, which, when the one or more programs are executed by the one or more processors, enables the one or more processors to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that: The computer program is stored therein, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Citation Information

Patent Citations

  • Epilepsy detection system based on linear graph convolutional neural network

    CN112294337A

  • Heterogeneous graph classification method based on double-layer attention mechanism

    CN112381179A

  • Lung cancer lymph node metastasis auxiliary diagnosis system based on multi-modal diagram forest

    CN115424724A

  • Disease prediction system based on mixed multi-modal data and graph convolutional neural network

    CN117524460A

  • Subpopulation based patient risk prediction using graph attention networks

    US20230238141A1