CN118161169A - Intelligent ECG detection method for coronary heart disease based on graph neural network - Google Patents

Abstract

The present invention discloses a method for intelligent detection of ECG for coronary heart disease based on graph neural network, comprising the following steps: obtaining a clinical data set, desensitizing it and then having a professional doctor mark it on the clinical data set; preprocessing the marked clinical data set, extracting the original signal in the clinical data, obtaining a 12-lead ECG electrocardiogram signal, and then performing noise reduction and standardization processing; constructing a diagnostic model based on an improved GCN network, taking each lead of the 12-lead ECG electrocardiogram signal as a node, adding topological connections according to physiological spatial relationships to construct an ECG graph, inputting the ECG graph into the diagnostic model for training, and using GNNExplain for interpretability analysis; constructing a coronary heart disease detection system based on graph neural network and 12-lead ECG electrocardiogram according to the diagnostic model. This method can simultaneously focus on the time characteristics in the electrocardiogram and the physiological spatial connection between different leads, which helps to better capture the overall state of heart activity, reduce the pressure of manual reading of the graph, and make screening for coronary heart disease more efficient.

Description

Intelligent ECG detection method for coronary heart disease based on graph neural network

Technical Field

The present invention relates to the field of smart medical technology, and in particular to a method for intelligent detection of ECG of coronary heart disease based on graph neural network.

Background technique

Coronary heart disease refers to a heart disease caused by myocardial ischemia and hypoxia due to coronary atherosclerosis. The root cause of the disease is the continuous accumulation of fat or unhealthy cholesterol in the arterial wall, which eventually causes the arterial wall to narrow and block. Arrhythmia, angina pectoris and myocardial infarction are the most common clinical manifestations of coronary heart disease. Coronary heart disease not only has a high mortality rate, but also has a risk of recurrence after the patient is cured and discharged from the hospital. It often causes acute myocardial infarction or even sudden death due to delayed treatment due to untimely diagnosis. Therefore, early detection, early diagnosis and early intervention treatment of coronary heart disease are crucial to improving patient prognosis and survival rate.

At present, the gold standard for identifying coronary artery lesions is coronary angiography, but because it is an invasive technique that requires arterial puncture, the examination cost is high and there is a risk of radiation exposure. In clinical practice, non-invasive, convenient and economical electrocardiogram diagnosis is more commonly used to screen for coronary heart disease. The electrocardiogram-based diagnosis method of coronary heart disease is made by examining the electrocardiogram waveform and relying on the long-term experience and subjective judgment of experts. However, these decisions can be supported and replaced to a certain extent by automated data-driven methods.

Research on ECG diagnosis of coronary heart disease based on deep learning is still in the exploratory stage. When applying traditional deep learning models, ECG signals are usually only regarded as synchronous signals arranged in Euclidean space. When analyzing ECGs, these methods usually treat ECG signals as simple channel-level arrangements, focusing only on the temporal characteristics of the signals, while ignoring the physiological spatial connections between different leads in the ECG. The spatial relationships between these leads are physiologically crucial and have practical value for the diagnosis of coronary heart disease.

Summary of the invention

The purpose of the present invention is to propose an intelligent ECG detection method for coronary heart disease based on graph neural network. By adding topological connections in non-Euclidean data and using graph neural network form to model, physiological spatial relationships are obtained, which can better capture the overall state of heart activity and thus improve the accuracy of diagnosis.

To achieve the above object, the present invention provides a method for intelligent detection of ECG for coronary heart disease based on graph neural network, comprising the following steps:

Obtain clinical data sets, desensitize them, and have professional doctors annotate them on the clinical data sets, dividing them into abnormal coronary arteries and normal coronary arteries;

Preprocess the annotated clinical data set, extract the original signal from the clinical data, obtain the 12-lead ECG signal, and then perform noise reduction and standardization processing;

A diagnostic model was constructed based on the improved GCN network. Each lead of the 12-lead ECG signal was used as a node. The ECG graph was constructed by adding topological connections according to the physiological spatial relationship. The ECG graph was input into the diagnostic model for training and GNNExplain was used for interpretability analysis.

According to the diagnostic model, a coronary heart disease detection system based on graph neural network and 12-lead ECG signals was constructed.

Furthermore, the specific method of constructing the ECG graph is:

The 12-lead ECG electrocardiogram signal is divided into a limb lead area and a chest lead area, wherein the limb lead area includes limb lead I, limb lead II, limb lead III, limb enhanced lead aVR, limb enhanced lead aVF, and limb enhanced lead aVL; the chest lead area includes chest lead V1, chest lead V2, chest lead V3, chest lead V4, chest lead V5, and chest lead V6, and each lead represents an activity record of a specific heart area;

The leads in each area are connected according to the physiological spatial relationship, and the leads in different areas are connected together using limb lead I, limb lead III, chest lead V1, and chest lead V6 to achieve communication between areas.

Furthermore, the diagnostic model includes a feature extraction module and a graph neural network module;

The feature extraction module includes a long short-term memory network, which takes each original signal as input, obtains the lead-level time domain feature fb, and uses the output feature vector to initialize the nodes in the entire graph network;

The graph neural network module is implemented through graph convolution and graph pooling operations, which contains a 3-layer GCN model that updates the representation of each node by aggregating information from neighboring nodes.

Furthermore, for the i-th layer in the GCN model, its next layer output is:

Among them, _Hi represents the node matrix of the i-th layer, A represents the adjacency matrix between nodes, and D is the degree matrix. is the adjacency matrix with self-connection, I is the identity matrix, /> For/> The diagonal matrix of , _Wi represents the weight matrix of the i-th layer, σ represents the activation function, and this module uses the Relu function as the activation function.

Furthermore, the graph neural network module uses a self-attention-based TopK graph pooling mechanism to learn a score representing the importance of each node in the graph, and discards some nodes with low scores based on the ranking of this score, and downsamples the N nodes in the entire graph to kN nodes; the entire pooling process is expressed as follows:

idx = top-rank(score,kN)

A _i+1 =A _i (idx,idx)

Among them, _Hi (idx,:) means that the feature matrix of the new subgraph is sliced in rows according to the value of the vector idx, and _Ai (idx,idx) means that the adjacency matrix of the new subgraph is sliced in rows and columns at the same time according to the value of the vector idx;

The node importance score based on self-attention is as follows:

Where X is the input feature of the graph and θ _att is the only parameter based on the features and topology of the graph.

Furthermore, the graph neural network module iteratively performs graph convolution and graph pooling operations to generate multiple new subgraphs, aggregates all node representations in the subgraphs and sums them to obtain a fixed-size graph-level feature fg.

Furthermore, the diagnostic model aggregates the lead-level time-domain features fb and the image-level features fg through a classifier, which uses the Sigmoid function to predict the diagnostic label of the entire image, and defines the loss function as the cross entropy of the prediction on the label. The output diagnostic results are divided into two categories, including normal coronary artery and abnormal coronary artery;

The sigmoid function is:

The loss function is:

Where N is the size of the training sample, M is the number of categories, _Yij is the ground truth, and _Pij is the predicted probability of belonging to class j.

Furthermore, the labeled clinical data set is divided according to the patient ID to obtain a training set, a validation set and a test set; the normal coronary artery data in the clinical data set is marked as (1,0), and the abnormal coronary artery data is marked as (0,1); the training set is input into the diagnostic network for training.

As a further step, GNNExplain is used to visualize and explain the diagnostic results. Specifically, GNNExplain takes the trained diagnostic model and its prediction results as input, then selects a node v, monitors the predicted value of a subgraph on node v and the predicted value of fewer features on the subgraph, and provides a subgraph or feature-based explanation through the difference between the two, so as to maximize the mutual information MI between the prediction of the original image and the prediction of the subgraph, which can be expressed as:

Where _Gs and _Xs are the features of the subgraph and its nodes, Y is the predicted label distribution, and its entropy H(Y) is a constant.

As a further step, the coronary heart disease detection system is built as follows:

The diagnostic model is deployed on the server side, the electronic medical records on the web page are obtained, the original 12-lead ECG signal is extracted, and the signal is denoised and normalized;

The pre-processed clinical data is converted into ECG graphs and then input into the trained diagnostic model for coronary heart disease detection to obtain the current patient's diagnosis results;

The diagnosis results are transmitted back to the front-end system, and each diagnosis result is visually explained on the user interface, showing the feature basis of the current classification result.

Compared with the prior art, the above technical solution adopted by the present invention has the following advantages: the method can simultaneously focus on the time characteristics in the electrocardiogram and the physiological spatial connection between different leads, which helps to better capture the overall state of cardiac activity, reduce the pressure of manual reading of the image, make screening for coronary heart disease more efficient, improve diagnostic confidence, and provide clinicians with more accurate diagnostic tools, thereby saving hospital resources and improving patients' cardiovascular health.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings in the specification, which constitute a part of the present application, are used to provide further understanding of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

FIG1 is a flow chart of a method for intelligent detection of ECG for coronary heart disease based on a graph neural network;

Fig. 2 is a schematic diagram of the structure of the diagnosis model;

Fig. 3 is a schematic diagram of the structure of an ECG graph;

FIG4 is a schematic diagram of the structure of GNNExplainer.

Specific implementation methods

The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and/or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and/or combinations thereof.

It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the methods and systems according to various embodiments of the present disclosure. It should be noted that each box in the flowchart or block diagram can represent a module, a program segment, or a part of a code, and the module, the program segment, or a part of the code may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, or they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the flowchart and/or block diagram, and the combination of boxes in the flowchart and/or block diagram can be implemented using a dedicated hardware-based system that performs a specified function or operation, or can be implemented using a combination of dedicated hardware and computer instructions.

Example 1

As shown in FIG1 , this embodiment provides a method for intelligent detection of ECG for coronary heart disease based on a graph neural network, comprising the following steps:

Step 1. Obtain clinical data sets, desensitize them, and have professional doctors annotate them on the clinical data sets, dividing them into abnormal coronary arteries and normal coronary arteries;

Specifically, the clinical data consisted of 2756 electronic diagnostic reports in XML format, including 1901 12-lead ECG records of 1842 patients with abnormal coronary arteries and 855 12-lead ECG records of 847 patients with normal coronary arteries at a sampling frequency of 500 Hz.

Step 2: Preprocess the annotated clinical data set, extract the original signal in the clinical data, obtain the 12-lead ECG signal, and then perform noise reduction and standardization;

Specifically, the annotated clinical data set uses wavelet threshold denoising to filter out noise such as line drift; the db4 wavelet function with a decomposition level of 6 is selected for wavelet decomposition, and the coefficients of the wavelet decomposition are soft-thresholded, and finally the denoised signal is obtained through wavelet reconstruction. The clinical data is standardized using maximum and minimum normalization, and the original signal data is converted to the [0,1] range by a linear method. The normalization formula is as follows:

Where x _max is the maximum value of the clinical sample data, and x _min is the minimum value of the clinical sample data.

The musexmlexport package was used to extract the raw signals from the clinical data and convert the clinical data into csv format for future use.

Step 3. Build a diagnostic model based on the improved GCN network, as shown in Figure 2; take each lead of the 12-lead ECG signal as a node, add topological connections according to the physiological spatial relationship to build an ECG graph, input the ECG graph into the diagnostic model for training and use GNNExplain for interpretability analysis;

Specifically, the 12-lead ECG signal is divided into 12 single-channel segments according to the leads, and these segments are converted into nodes. The ECG data of each patient is modeled as an ECG graph. The characteristics of the nodes in the graph are the time series of the segments, and the spatial connection between nodes is achieved through edges. The model learns and processes the characteristics and connection relationships of each node to provide a classification label for the entire graph, indicating the diagnosis result of the current patient. The specific implementation method is:

Step 31: Construct an ECG graph. The ECG signal can comprehensively display the heart activity from different angles by recording the potential difference between electrodes placed at specific locations on the body surface. The 12-lead ECG signal includes 3 limb leads (I, II, III), 3 limb enhancement leads (aVR, aVF, aVL) and 6 chest leads (V1-V6), each of which represents the activity record of a specific heart region. Therefore, according to the spatial distribution of the electrodes, the 12 leads can be divided into 2 lead regions: the limb lead region and the chest lead region. The leads in the same lead region are connected according to the physiological spatial relationship. In order to exchange information between different lead regions, four leads reflecting different heart regions (limb lead I, limb lead III, chest lead V1, chest lead V6) are connected together to realize communication between lead regions. The connection method is shown in Figure 3.

Step 32: construct a diagnostic model, which includes a feature extraction module and a graph neural network module;

The feature extraction module is mainly composed of a long short-term memory network (LSTM), which is a variant of a recurrent neural network (RNN). It effectively captures and memorizes long sequence dependencies by introducing a structure called a "gate" to control the flow of information. Due to its ability to effectively process long sequence data and capture time dependencies, LSTM has been widely used in ECG signal classification tasks. This module takes each lead raw signal as input, obtains the lead-level time domain feature fb, and uses the output feature vector to initialize the nodes in the entire graph network.

The graph neural network module is implemented through graph convolution and graph pooling operations, which contains a 3-layer GCN model. GCN updates the representation of each node by aggregating the information of neighboring nodes. This module uses a self-attention-based TopK graph pooling mechanism to downsample nodes while retaining more key information of the original graph. It learns a score representing the importance of each node in the graph, discards some low-score nodes based on the ranking of this score, and downsamples the N nodes in the full graph to kN nodes; finally, it iteratively performs graph convolution and graph pooling operations to generate multiple new subgraphs, aggregates all node representations in the subgraphs and sums them to obtain a fixed-size (128) graph-level feature fg.

The diagnostic model aggregates the fb and fg features and classifies them through a classifier. The classifier uses the Sigmoid function to predict the diagnostic label of the entire image, defines the loss function as the cross entropy of the prediction on the label, and outputs two diagnostic results, including normal coronary artery and abnormal coronary artery.

Step 33: Diagnostic model. Divide the labeled clinical data set by patient ID into a training set, a validation set, and a test set at a ratio of 7:2:1. The normal coronary data in the clinical data set is marked as (1,0), and the abnormal coronary data is marked as (0,1); input the processed data into the above diagnostic model for training. The specific steps are as follows:

Step 331. Send the training set data to the diagnostic model for reasoning;

Step 332: Calculate the loss function based on the inference result of the diagnostic model and the actual label;

Step 333. Perform gradient back propagation through the loss function to update the diagnostic model weights. If the training is not finished, return to step 1. If the training is finished, go to step 4.

Step 334: Apply the trained diagnostic model to the test set and obtain classification indicators.

Step 34: Interpretability analysis. The present invention uses the visualize_subgraph method of GNNExplain to visualize and explain the diagnostic results. GNNExplainer captures important input features by generating masks that convey key semantics, thereby producing predictions similar to the original predictions. It learns soft masks of edge and node features and explains predictions through mask optimization. As shown in Figure 4. The explanation process of GNNExplainer can be mainly divided into selecting target nodes, building perturbation graphs, calculating neighborhood importance to generate explanations and visualizations. In the task of the present invention, GNNExplainer takes the trained diagnostic model and its prediction results as input, then selects a node v, monitors the predicted value of a subgraph on the node v and the predicted value of fewer features on the subgraph, and provides a subgraph or feature-based explanation through the difference between the two, so as to maximize the mutual information MI between the prediction of the original image and the prediction of the subgraph.

Step 4. Construct a coronary heart disease detection system based on graph neural network and 12-lead ECG electrocardiogram signals according to the diagnostic model;

Step 41: Build a web-based detection platform: The diagnostic model is deployed on the server. After receiving the XML-formatted electronic medical record from the web-based end, extract the original 12-lead ECG signal, perform noise reduction and normalization on the signal to ensure that it is consistent with the data format used during training;

Step 42: Diagnostic classification: The pre-processed clinical data is converted into ECG graphs and then input into the trained diagnostic model for coronary heart disease detection to obtain the current patient's diagnosis result.

Step 43: Result visualization: The diagnosis results are transmitted back to the front-end system, and each diagnosis result is visually explained on the user interface, showing the feature basis of the current classification result.

Those skilled in the art should understand that the modules or steps of the present disclosure can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present disclosure is not limited to any specific combination of hardware and software.

The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims (10)

1. The coronary heart disease ECG intelligent detection method based on the graph neural network is characterized by comprising the following steps of:

Acquiring a clinical data set, and marking the clinical data set by a professional doctor after desensitizing the clinical data set, wherein the clinical data set is divided into abnormal coronary arteries and normal coronary arteries;

preprocessing the marked clinical data set, extracting original signals in the clinical data, obtaining 12-lead ECG electrocardiosignals, and then carrying out noise reduction and standardization treatment;

Constructing a diagnosis model based on an improved GCN network, taking each lead of a 12-lead ECG electrocardiograph signal as a node, adding topological connection according to a physiological space relation to construct an ECG graph, inputting the ECG graph into the diagnosis model for training and using GNNExplain for interpretation analysis;

and constructing a coronary heart disease detection system based on the graphic neural network and the 12-lead ECG electrocardiosignal according to the diagnosis model.

2. The intelligent detection method for coronary heart disease ECG based on graph neural network according to claim 1, wherein the specific mode of constructing the ECG graph is as follows:

Dividing the 12-lead ECG electrocardiograph signal into a limb lead area and a chest lead area, wherein the limb lead area comprises a limb lead I, a limb lead II, a limb lead III, a limb enhancement lead aVR, a limb enhancement lead aVF and a limb enhancement lead aVL; the chest lead region comprises chest leads V1, V2, V3, V4, V5 and V6, and each lead represents an activity record of a specific heart region;

The leads in each region are connected according to the physiological space relation, and the leads in different regions are connected together by using the limb lead I, the limb lead III, the chest lead V1 and the chest lead V6 to realize the communication between the regions.

3. The coronary heart disease ECG intelligent detection method based on the graph neural network according to claim 1, wherein the diagnosis model comprises a feature extraction module and a graph neural network module;

The feature extraction module comprises a long-term and short-term memory network, takes each original signal as input, obtains lead-level time domain features fb, and uses the output feature vectors to initialize nodes in the whole graph network;

the graph neural network module is implemented by graph rolling and graph pooling operations, which includes a layer 3 GCN model that updates the representation of each node by aggregating the information of neighboring nodes.

4. The intelligent detection method for coronary heart disease ECG based on graphic neural network according to claim 3, wherein for the ith layer in the GCN model, the output of the next layer is:

Wherein H _i represents a node matrix of the i-th layer, A represents an adjacency matrix between nodes, D is a degree matrix, For the adjacency matrix with self-connection, I is the identity matrix,/>For/>W _i represents the weight matrix of the i-th layer, σ represents the activation function, and the module uses Relu function as the activation function.

5. The method for intelligent detection of coronary heart disease ECG based on a graph neural network according to claim 3, wherein the graph neural network module learns a score representing node importance for each node in the graph by using a TopK graph pooling mechanism based on self-attention, discards a part of low-score nodes based on the ordering of the score, and downsamples N nodes in the whole graph to kN nodes; the entire pooling process is represented as follows:

idx＝top-rank(score,kN)

A_i+1＝A_i(idx，idx)

Wherein H _i (idx, i) represents that the new sub-graph feature matrix is subjected to row slicing according to the value of the vector idx, and A _i (idx ) represents that the adjacent matrix of the new sub-graph is subjected to row slicing and column slicing simultaneously according to the value of the vector idx;

the node importance score based on self-attention is as follows:

Where X is the input feature of the graph and θ _att is the only parameter based on the feature and topology of the graph.

6. A coronary heart disease ECG intelligent detection method based on a graph neural network according to claim 3, wherein the graph neural network module generates a plurality of new subgraphs by iteratively performing graph convolution and graph pool operations, and all nodes in the aggregated subgraphs represent and sum to obtain a graph-level feature fg of a fixed size.

7. The coronary heart disease ECG intelligent detection method based on the graphic neural network according to claim 1, wherein the diagnostic model aggregates lead level time domain features fb and graphic level features fg to be classified by a classifier, the classifier predicts the diagnostic label of the whole graph by using a Sigmoid function, the loss function is defined as the cross entropy predicted on the label, and the output diagnostic results are divided into two types including coronary artery normal and coronary artery abnormal;

the sigmoid function is:

The loss function is:

Where N is the size of the training sample, M is the number of classes, Y _ij is the base true value, and P _ij is the prediction probability belonging to class j.

8. The intelligent detection method for coronary heart disease ECG based on the graph neural network according to claim 1, wherein a training set, a verification set and a test set are obtained by dividing a clinical data set after labeling according to a patient PATIENTID; coronary normal data in the clinical dataset is marked as (1, 0), coronary abnormal data is marked as (0, 1); the training set is input into a diagnostic network for training.

9. The intelligent detection method for coronary heart disease ECG based on the graphic neural network according to claim 1, wherein GNNExplain is used for visually explaining the diagnosis result, specifically: GNNExplain takes the trained diagnostic model and the predicted result thereof as input, then selects a node v, monitors the predicted value of a sub-graph on the node v and the predicted value of fewer features on the sub-graph, provides interpretation based on the sub-graph or the features through the difference change of the two, maximizes the mutual information MI between the prediction of the original graph and the prediction of the sub-graph, and is expressed as:

Wherein G _s and X _s are characteristics of the subgraph and its nodes, Y is the predictive label distribution, and its entropy H (Y) is a constant.

10. The intelligent coronary heart disease ECG detection method based on the graph neural network according to claim 1, wherein the coronary heart disease detection system is constructed in the following way:

The diagnosis model is deployed at a server end, a webpage end electronic medical record is obtained, an original 12-lead ECG electrocardiosignal is extracted, and noise reduction and normalization processing are carried out on the signal;

The preprocessed clinical data are converted into an ECG (electrocardiogram) chart and then input into a trained diagnosis model for coronary heart disease detection, so that a diagnosis result of a current patient is obtained;

the diagnosis results are transmitted back to the front-end system, each diagnosis result is visually interpreted on the user interface, and the characteristic basis of the current classification result is displayed.