Graph convolution heart disease detection method fusing trust loss and decision reclassification

Abstract

The invention provides a graph convolution heart disease detection method fusing trust loss and decision reclassification. Belongs to the technical field of medical information intelligent processing and solves the problem that uncertain data affects classification precision in heart disease detection. According to the technical scheme, the method comprises the steps that firstly, a heart disease information data sample set is collected, then the data are preprocessed, and an adjacent matrix is constructed; inputting the processed heart disease information data into a graph convolutional neural network GCN for model training; and finally, the model enters a test stage, a preliminary classification result is obtained and then enters a decision-making layer, uncertain node information is screened out according to a threshold value, and a Gaussian kernel is used for mapping, so that the uncertain node information can be classified at a high dimension, and a re-classification effect is achieved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent processing of medical information, and in particular to a graph convolution heart disease detection method integrating trust loss and decision reclassification. Background Art

[0002] Accurate identification of heart disease types is of great significance for improving the early detection rate of the disease, improving treatment effectiveness and reducing misdiagnosis. Traditional heart disease classification methods mainly rely on doctors' experience and rule-based systems. These methods often rely on the judgment of medical experts and are easily affected by subjective factors of the tester, resulting in limitations in accuracy and generalization ability.

[0003] Currently, researchers are beginning to try to classify heart disease through deep learning methods. However, although the existing deep learning-based models have achieved good results in some tasks, there are still some problems.

[0004] In the medical field, data is often noisy, incomplete, and uncertain, which may cause traditional deep learning models to produce unstable prediction results when dealing with uncertain data. In particular, when the data quality is not high or there are large differences between samples, the prediction effect of the model is likely to decline. In addition, the classification of heart disease not only relies on the data of a single patient, but also should consider the similarities between patients. The existing model classification process often ignores the similarity information between patients and fails to fully utilize this information for model optimization.

[0005] How to solve the above problems is the subject faced by the present invention. Summary of the invention

[0006] The purpose of the present invention is to provide a graph convolutional heart disease detection method that integrates trust loss and decision reclassification. The method integrates trust adaptive weighted loss and negative log-likelihood loss, and has good effect in processing complex data. In addition, a re-decision mechanism is introduced to ensure good effect in processing uncertain data.

[0007] The present invention is achieved by the following measures: a graph convolution heart disease detection method integrating trust loss and decision reclassification comprises the following steps:

[0008] Step 1: Collect a sample set U of heart disease data, which includes the characteristic information and categories of heart disease. Give each piece of data a unique ID, select neighbors for each ID using similarity, and build an adjacency relationship.

[0009] Step 2: Preprocess the original sample set, convert the original categories into one-hot encoded labels, construct a sparse adjacency matrix based on the adjacency relationship, and divide the heart disease sample set data into a training sample set, a validation sample set, and a test sample set;

[0010] Step 3: Input the above training sample set data into the improved graph convolutional neural network, train the graph convolutional neural network according to the actual category of heart disease, integrate the trust-based adaptive loss function on the basis of the negative log-likelihood loss function, and perform back propagation according to the output value of the new loss function to update the model parameters;

[0011] Step 4: Input the validation sample set data into the model, and calculate the uncertain classification node threshold δ based on the classification accuracy of the last round of validation;

[0012] Step 5: Input the test sample set into the trained model to obtain the preliminary classification probability distribution Z = {z 1 ,z 2 ,z 3 ,…z n}, the probability distribution enters the decision layer, the information of uncertain nodes is filtered according to the uncertain classification node threshold δ, and the Gaussian kernel function is used to reclassify the uncertain nodes.

[0013] As a graph convolution heart disease detection method integrating trust loss and decision reclassification provided by the present invention, step 3 includes the following steps:

[0014] Step 3.1 Set the random seed seed = 42, the total number of training epochs = 200, the learning rate lr = 0.01, the dropout rate Dropout = 0.5, and set the model to enter the training mode. In the training mode, Dropout randomly drops a part of the neurons and inputs the sample data feature X and the adjacency matrix A into the graph convolutional neural network. The formula is as follows:

[0015]

[0016] in W (0) is the weight from the input layer to the hidden layer, W (1) is the weight from the hidden layer to the output layer, Z = {z 1 ,z 2 ,z 3 ,…z n} is the probability distribution of n nodes corresponding to each category, and the nonlinear activation function ReLU of the hidden layer is expressed as:

[0017] ReLU(x i )=max(0,x i ) (2)

[0018] The output layer softmax function uses logarithmic transformation to output logarithmic probability distribution. The formula is as follows:

[0019]

[0020] where x i is the i-th component in the input vector x;

[0021] Step 3.2 Use the exp function to convert the logarithmic probability distribution Z = {z 1 ,z 2 ,z 3 ,…z n}, and is reduced to the ordinary probability distribution P = {p 1 ,p 2 ,p 3 ,…p n}, retaining the logarithmic probability distribution, the calculation expression is as follows:

[0022] p i =σ exp (x i )=exp(σ(x i )) (4)

[0023] Step 3.3 According to the probability distribution p of the i-th node i ={p i1 ,p i2 ,p i3 ,…p im}The maximum probability v for each node i,1 =max(p i1 ,p i2 ,p i3 ,…p im ) and the second highest probability v i,2 =max({p i1 ,p i2 ,p i3 ,…p im}\{v i,1}) Make a difference, the operation is:

[0024] Δ i =v i,1 -v i,2 (5)

[0025] Get the maximum difference matrix Δ=[Δ 1 ,Δ 2 ,…,Δ n ], the difference is used as the confidence of node classification;

[0026] Step 3.4 Given the logarithmic probability distribution z of the i-th node i ={z i1 ,z i2 ,z i3 ,…z im}, training set sample index idxtrain , the number of training set samples N, the true category label y of the i-th node i , calculate the negative log-likelihood loss, the function calculation formula is:

[0027]

[0028] Step 3.5 Use the maximum difference matrix Δ = [Δ 1 ,Δ 2 ,…,Δ n ], the negative log-likelihood loss is given an adaptive weight through the Hadamard product to obtain the trust adaptive loss. The loss function formula is as follows:

[0029]

[0030] The trust adaptive loss function is combined with the negative log-likelihood loss function to obtain the final loss function, which is expressed as:

[0031]

[0032] The higher the node trust, the greater the impact on the model. The lower the node trust, the smaller the impact on the model. For nodes with uncertain properties, reduce their interference with the model and improve the model's classification ability for nodes with certain properties. Set λ=0.2, perform backpropagation according to the loss value, update the model parameters, and output the loss and classification accuracy of each round of training.

[0033] As a graph convolution heart disease detection method integrating trust loss and decision reclassification provided by the present invention, step 4 includes the following steps:

[0034] Step 4.1 The model enters the evaluation mode, Dropout is disabled, all neurons are involved in the calculation, the model is verified using the complete network structure, the verification sample set data is input into the model, and the number of correctly classified samples is calculated. The calculation formula is:

[0035]

[0036] in, is the indicator function, when 1 when it is, otherwise 0;

[0037] Step 4.2 calculates the classification accuracy based on the number of correctly classified samples. The output is the ratio of the number of correctly classified samples to the total number of input samples. The calculation formula is as follows:

[0038]

[0039] Step 4.3: The classification accuracy of the last round of verification Rounded to two decimal places, the formula is:

[0040]

[0041] According to classification accuracy Calculate the threshold value of uncertain classification nodes

[0042] As a graph convolution heart disease detection method integrating trust loss and decision reclassification provided by the present invention, step 5 includes the following steps:

[0043] Step 5.1 Input the test sample data into the trained model and retain the initial classification result y = [y 1 ,y 2, …,y n ], calculate the maximum difference matrix Δ = [Δ 1 ,Δ 2 ,…,Δ n ], calculate the number of selected elements k according to the threshold δ min , the calculation expression is as follows:

[0044]

[0045] Select the smallest k from the maximum difference matrix min elements, record the difference value and the index in the original set, and the operation expression is:

[0046]

[0047] where |S| = k min ;

[0048] Step 5.2 Calculate the Euclidean distance matrix D between nodes based on the node feature X. The matrix form is as follows:

[0049]

[0050] Set σ = 1, input the Euclidean distance matrix D into the Gaussian kernel function, and obtain the similarity matrix S based on the Gaussian kernel. The matrix is ​​expressed as:

[0051]

[0052] Step 5.3 obtains the neighborhood similarity matrix N by element-by-element multiplication of the adjacency matrix A and the similarity matrix S, which is expressed as:

[0053] N=A⊙S (16)

[0054] If there is an edge connection between node i and node j, then N ij It is the similarity value S between node i and node jij , if there is no edge connection, then N ij =0;

[0055] Step 5.4 Get the similarity S between the current uncertain node i and other adjacent nodes i =[S i1 ,S i2 ,…,S ik ], get the label of the adjacent node, the operation expression is as follows:

[0056] Y i ={y j |j∈N(i)} (17)

[0057] Extract the probability value of the adjacent node label corresponding to node i in the probability distribution to obtain the weight vector W i =[W i1 ,W i2 ,…,W ik ], for the similarity S i =[S i1 ,S i2 ,…,S ik ] is weighted to obtain the weighted similarity The operation is as follows:

[0058]

[0059] Step 5.5 finds the node index with the largest similarity in the weighted similarity matrix, as described below:

[0060]

[0061] Similarity node index i max Get reclassification labels Update y = [y 1 ,y 2 ,…,y n ], loop step 5.4 to step 5.5 until all uncertain nodes are reclassified, and output the final classification result.

[0062] Compared with the prior art, the present invention has the following beneficial effects:

[0063] 1. The present invention integrates the trust adaptive loss with the negative log-likelihood loss, assigns a trust value to each node according to the uncertainty measurement, reduces the influence of nodes with low trust values ​​on the model, and enhances the influence of nodes with high trust values ​​on the model, which can effectively improve the classification accuracy and provide medical personnel with more accurate detection basis.

[0064] 2. The present invention introduces a reclassification mechanism to screen out uncertain nodes according to a threshold, and reclassifies the uncertain nodes through implicit dimensionality increase using a Gaussian kernel function, thereby improving the model's ability to handle uncertain data and further improving classification accuracy.

[0065] 3. Aiming at the characteristics of heart disease data such as fuzzy category boundaries and uncertainty, the present invention proposes a graph convolution heart disease detection method that integrates trust loss and decision reclassification. This method starts from the loss function and probability distribution, fully considers the uncertainty of data and probability, uses trust weighting to "denoise" the model, and uses Gaussian kernel function for reclassification, which effectively improves the classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0067] Figure 1 This is the overall framework diagram of the graph convolution heart disease detection method that integrates trust loss and decision reclassification in the present invention.

[0068] Figure 2 This is a flowchart of the algorithm improved based on graph convolution in the present invention.

[0069] Figure 3 This is a diagram of the decision reclassification architecture of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0071] Example 1

[0072] See also Figures 1 to 3 The technical solution provided in this embodiment is a graph convolution heart disease detection method integrating trust loss and decision reclassification, comprising the following steps:

[0073] Step 1: Collect 200 samples of heart disease data, including heart disease feature information and heart disease categories. Give each piece of data a unique ID, select neighbors for each ID using similarity, and build an adjacency relationship.

[0074] Step 2: Preprocess the original sample set, convert the original categories into one-hot encoded labels, construct a sparse adjacency matrix based on the adjacency relationship, and divide the heart disease sample set data into a training sample set, a validation sample set, and a test sample set;

[0075] Step 3: Input the above training sample set data into the improved graph convolutional neural network, train the graph convolutional neural network according to the actual category of heart disease, integrate the trust-based adaptive loss function on the basis of the negative log-likelihood loss function, and perform back propagation according to the output value of the new loss function to update the model parameters;

[0076] Step 4: Input the validation sample set data into the model, and calculate the uncertain classification node threshold δ based on the classification accuracy of the last round of validation;

[0077] Step 5: Input the test sample set into the trained model to obtain the preliminary classification probability distribution Z = {z 1 ,z 2 ,z 3 ,…z n}, the probability distribution enters the decision layer, the information of uncertain nodes is filtered according to the uncertain classification node threshold δ, and the Gaussian kernel function is used to reclassify the uncertain nodes.

[0078] Preferably, step 3 comprises the following steps:

[0079] Step 3.1 Set the random seed seed = 42, the total number of training epochs = 200, the learning rate lr = 0.01, the dropout rate Dropout = 0.5, and set the model to enter the training mode. In the training mode, Dropout randomly drops a part of the neurons and inputs the sample data feature X and the adjacency matrix A into the graph convolutional neural network. The formula is as follows:

[0080]

[0081] in W (0) is the weight from the input layer to the hidden layer, W (1) is the weight from the hidden layer to the output layer, Z = {z 1 ,z 2 ,z 3 ,…z n} is the probability distribution of n nodes corresponding to each category, and the nonlinear activation function ReLU of the hidden layer is expressed as:

[0082] ReLU(x i )=max(0,x i ) (4)

[0083] The output layer softmax function uses logarithmic transformation to output logarithmic probability distribution. The formula is as follows:

[0084]

[0085] where x iis the i-th component in the input vector x;

[0086] Step 3.2 Use the exp function to convert the logarithmic probability distribution Z = {z 1 ,z 2 ,z 3 ,…z n}, and is reduced to the ordinary probability distribution P = {p 1 ,p 2 ,p 3 ,…p n}, retaining the logarithmic probability distribution, the calculation expression is as follows:

[0087] p i =σ exp (x i )=exp(σ(x i )) (6)

[0088]

[0089] Step 3.3 According to the probability distribution p of the i-th node i ={p i1 ,p i2 ,p i3 ,…p im}The maximum probability v for each node i,1 =max(p i1 ,p i2 ,p i3 ,…p im ) and the second highest probability v i,2 =max({p i1 ,p i2 ,p i3 ,…p im}\{v i,1}) Make a difference, the operation is:

[0090] Δ i =v i,1 -v i,2 (8)

[0091]

[0092] Get the maximum difference matrix Δ=[Δ 1 ,Δ 2 ,…,Δ n ], the difference is used as the trust value of node classification;

[0093] Step 3.4 Given the logarithmic probability distribution z of the i-th node i ={z i1 ,z i2 ,z i3 ,…zim}, training set sample index idx train , the number of training set samples N, the true category label y of the i-th node i , calculate the negative log-likelihood loss, the function calculation formula is:

[0094]

[0095] Step 3.5 Use the maximum difference matrix Δ = [Δ 1 ,Δ 2 ,…,Δ n ], the negative log-likelihood loss is given an adaptive weight through the Hadamard product to obtain the trust adaptive loss. The loss function formula is as follows:

[0096]

[0097] The trust adaptive loss function is combined with the negative log-likelihood loss function to obtain the final loss function, which is expressed as:

[0098]

[0099] The higher the node trust value, the greater the impact on the model. The lower the node trust value, the smaller the impact on the model. For nodes with uncertain properties, reduce their interference with the model and improve the model's classification ability for nodes with certain properties. Set λ=0.2, perform backpropagation according to the loss value, update the model parameters, and output the loss and classification accuracy of each round of training.

[0100] Preferably, step 4 comprises the following steps:

[0101] Step 4.1 The model enters the evaluation mode, Dropout is disabled, all neurons are involved in the calculation, the model is verified using the complete network structure, the verification sample set data is input into the model, and the number of correctly classified samples is calculated. The calculation formula is:

[0102]

[0103] in, is the indicator function, when 1 when it is, otherwise 0;

[0104] Step 4.2 calculates the classification accuracy based on the number of correctly classified samples. The output is the ratio of the number of correctly classified samples to the total number of input samples. The calculation formula is as follows:

[0105]

[0106] Step 4.3: The classification accuracy of the last round of verification Rounded to two decimal places, the formula is:

[0107]

[0108] According to classification accuracy Calculate the threshold value of uncertain classification nodes

[0109] Preferably, step 5 comprises the following steps:

[0110] Step 5.1 Input the test sample data into the trained model and retain the initial classification result y = [y 1 ,y 2 ,…,y n ], calculate the maximum difference matrix Δ = [Δ 1 ,Δ 2 ,…,Δ n ], calculate the number of selected elements k according to the threshold δ min , the calculation expression is as follows:

[0111]

[0112] Select the smallest k from the maximum difference matrix min elements, record the difference value and the index in the original set, and the operation expression is:

[0113]

[0114] where |S| = k min ;

[0115] Step 5.2 Calculate the Euclidean distance matrix D between nodes based on the node feature X. The matrix form is as follows:

[0116]

[0117] Set σ = 1, input the Euclidean distance matrix D into the Gaussian kernel function, and obtain the similarity matrix S based on the Gaussian kernel. The matrix is ​​expressed as:

[0118]

[0119] Step 5.3 obtains the neighborhood similarity matrix N by element-by-element multiplication of the adjacency matrix A and the similarity matrix S, which is expressed as:

[0120]

[0121] If there is an edge connection between node i and node j, then N ij It is the similarity value S between node i and node j ij , if there is no edge connection, then Nij =0;

[0122] Step 5.4 Get the similarity S between the current uncertain node i and other adjacent nodes i =[S i1 ,S i2 ,…,S ik ], get the label of the adjacent node, the operation expression is as follows:

[0123] S i =[S i1 ,S i2 ,…,S ik ]=[0.1977 0.1982 … 0.1978] (21)

[0124] Y i ={y j |j∈N(i)}=[0 3 … 3] (22)

[0125] Extract the probability value of the adjacent node label corresponding to node i in the probability distribution to obtain the weight vector W i =[W i1 ,W i2 ,…,W ik ], for the similarity S i =[S i1 ,S i2 ,…,S ik ] is weighted to obtain the weighted similarity The operation is as follows:

[0126]

[0127] Step 5.5 finds the node index with the largest similarity in the weighted similarity matrix, as described below:

[0128]

[0129] Similarity node index i max Get reclassification labels Update y = [y 1 ,y 2 ,…,y n ], loop step 5.4 to step 5.5 until all uncertain nodes are reclassified, and output the final classification result.

[0130] Example 2

[0131] Referring to Example 1, this example will use the parameters and results calculated in Example 1 to compare with the traditional method to prove the superiority of this example. In the specific comparison, we used the comparison of different models in the accuracy index. The final result shows that this example is preferred over the traditional method.

[0132] 1. Traditional Model

[0133] With reference to relevant research at home and abroad, this embodiment selects the following comparison models:

[0134] 1) Deep Walk Model:

[0135] Deep Walk is a graph embedding method based on random walks. It can map nodes in a graph into a low-dimensional vector space, thereby achieving representation learning of nodes. It traverses the entire graph in a random walk manner, and then uses these random walk sequences to learn the vector representation of nodes. Specifically, Deep Walk first performs several random walks on each node to obtain some random walk sequences. Then, these sequences are regarded as sentences, and the word vector model is used to learn the vector representation of the nodes.

[0136] 2) GCN model:

[0137] GCN works on graph data, and its core idea is to propagate information through adjacency relationships in the graph structure. Through graph convolution, nodes can not only retain their own information, but also obtain information about neighboring nodes, thereby updating the feature representation of nodes at each layer. Increasing the number of GCN layers allows nodes to obtain information about neighbors at a greater distance. The graph convolution operation aggregates the information of node neighbors through the adjacency matrix A, and the node features of each layer depend on the features of the previous layer.

[0138] 2. Comparison indicators

[0139] This embodiment is used for classification tasks. In classification tasks, we use accuracy (denoted as ACC) as an evaluation indicator. ACC is the ratio of the number of correctly classified samples to the total number of samples. It reflects the accuracy of the overall classification of the model.

[0140] This embodiment selects the accuracy evaluation method to objectively compare this embodiment with other models, aiming to highlight the superiority of this embodiment.

[0141] 3. Comparison results

[0142] As can be seen from the table, this embodiment performs well in accuracy evaluation indicators and is significantly better than other models. The classification accuracy of this embodiment reaches 82.5%, which is higher than 80.3% of the GCN model and much higher than 67.9% of the Deep Walk model. This shows that this embodiment has higher accuracy in overall classification and can better distinguish multiple types of samples.

[0143] Table 1 Experimental comparison of the present invention, Deep Walk and GCN models

[0144] method Dataset ACC(%) This article UCI 82.5 GCN UCI 80.3 Deep Walk UCI 67.9

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A graph convolutional heart disease detection method integrating trust loss and decision reclassification, characterized by The following steps are involved: Step 1: Collect a sample set U of heart disease data, which includes the characteristic information and categories of heart disease. Give each piece of data a unique ID, select neighbors for each ID using similarity, and build an adjacency relationship. Step 2: Preprocess the original sample set, convert the original categories into one-hot encoded labels, construct a sparse adjacency matrix based on the adjacency relationship, and divide the heart disease sample set data into a training sample set, a validation sample set, and a test sample set; Step 3: Input the above training sample set data into the improved graph convolutional neural network, train the graph convolutional neural network according to the actual category of heart disease, integrate the trust-based adaptive loss function on the basis of the negative log-likelihood loss function, and perform back propagation according to the output value of the new loss function to update the model parameters; Step 4: Input the validation sample set data into the model, and calculate the uncertain classification node threshold δ based on the classification accuracy of the last round of validation; Step 5: Input the test sample set into the trained model to obtain the preliminary classification probability distribution Z = {z1,z2,z3,…z n }, the probability distribution enters the decision layer, the information of uncertain nodes is filtered according to the uncertain classification node threshold δ, and the Gaussian kernel function is used to reclassify the uncertain nodes.

2. The method for detecting heart disease by graph convolution integrating trust loss and decision reclassification according to claim 1, characterized in that: The step 3 comprises the following steps: Step 3.1 Set the random seed seed = 42, the total number of training epochs = 200, the learning rate lr = 0.01, the dropout rate Dropout = 0.5, and set the model to enter the training mode. In the training mode, Dropout randomly drops a part of the neurons and inputs the sample data feature X and the adjacency matrix A into the graph convolutional neural network. The formula is as follows: in W (0) is the weight from the input layer to the hidden layer, W (1) is the weight from the hidden layer to the output layer, Z = {z1,z2,z3,…z n } is the probability distribution of n nodes corresponding to each category, and the nonlinear activation function ReLU of the hidden layer is expressed as: ReLU(x i )=max(0,x i ) (2) The output layer softmax function uses logarithmic transformation to output logarithmic probability distribution. The formula is as follows: where x i is the i-th component in the input vector x; Step 3.2 Use the exp function to convert the logarithmic probability distribution Z = {z1, z2, z3, … z n }, and is reduced to the ordinary probability distribution P = {p1, p2, p3, ... p n }, retaining the logarithmic probability distribution, the calculation expression is as follows: p i =s exp (x i )=exp(σ(x i )) (4) Step 3.3 According to the probability distribution p of the i-th node i ={p i1 ,p i2 ,p i3 ,…p im }The maximum probability v for each node i,1 =max(p i1 ,p i2 ,p i3 ,…p im ) and the second highest probability v i,2 =max({p i1 ,p i2 ,p i3 ,…p im }\{v i,1 }) Make a difference, the operation is: Δ i =v i,1 -v i,2 (5) Get the maximum difference matrix Δ=[Δ1,Δ2,…,Δ n ], the difference is used as the trust value of node classification; Step 3.4 Given the logarithmic probability distribution z of the i-th node i ={z i1 ,z i2 ,z i3 ,…z im }, training set sample index idx train , the number of training set samples N, the true category label y of the i-th node i , calculate the negative log-likelihood loss, the function calculation formula is: Step 3.5 Use the maximum difference matrix Δ=[Δ1,Δ2,…,Δ n ], the negative log-likelihood loss is given an adaptive weight through the Hadamard product to obtain the trust adaptive loss. The loss function formula is as follows: The trust adaptive loss function is combined with the negative log-likelihood loss function to obtain the final loss function, which is expressed as: The higher the node trust, the greater the impact on the model. The lower the node trust, the smaller the impact on the model. For nodes with uncertain properties, reduce their interference with the model and improve the model's classification ability for nodes with certain properties. Set λ=0.2, perform backpropagation according to the loss value, update the model parameters, and output the loss and classification accuracy of each round of training.

3. The graph convolution heart disease detection method integrating trust loss and decision reclassification according to claim 1, characterized in that: The step 4 comprises the following steps: Step 4.1 The model enters the evaluation mode, Dropout is disabled, all neurons are involved in the calculation, the model is verified using the complete network structure, the verification sample set data is input into the model, and the number of correctly classified samples is calculated. The calculation formula is: in, is the indicator function, when 1 when it is, otherwise 0; Step 4.2 calculates the classification accuracy based on the number of correctly classified samples. The output is the ratio of the number of correctly classified samples to the total number of input samples. The calculation formula is as follows: Step 4.3: The classification accuracy of the last round of verification Rounded to two decimal places, the formula is: According to classification accuracy Calculate the threshold value of uncertain classification nodes 4. The graph convolution heart disease detection method integrating trust loss and decision reclassification according to claim 1, characterized in that: The step 5 comprises the following steps: Step 5.1 Input the test sample data into the trained model and retain the initial classification result y = [y1,y 2, …,y n ], calculate the maximum difference matrix Δ=[Δ1,Δ2,…,Δ n ], calculate the number of selected elements k according to the threshold δ min , the calculation expression is as follows: Select the smallest k from the maximum difference matrix min elements, record the difference value and the index in the original set, and the operation expression is: where |S| = k min ; Step 5.2 Calculate the Euclidean distance matrix D between nodes based on the node feature X. The matrix form is as follows: Set σ = 1, input the Euclidean distance matrix D into the Gaussian kernel function, and obtain the similarity matrix S based on the Gaussian kernel. The matrix is ​​expressed as: Step 5.3 obtains the neighborhood similarity matrix N by element-by-element multiplication of the adjacency matrix A and the similarity matrix S, which is expressed as: N=A⊙S (16) If there is an edge connection between node i and node j, then N ij It is the similarity value S between node i and node j ij , if there is no edge connection, then N ij =0; Step 5.4 Get the similarity S between the current uncertain node i and other adjacent nodes i =[S i1 ,S i2 ,…,S ik ], get the label of the adjacent node, the operation expression is as follows: AND i ={and j |j∈N(i)} (17) Extract the probability value of the adjacent node label corresponding to node i in the probability distribution to obtain the weight vector W i =[W i1 ,W i2 ,…,W ik ], for the similarity S i =[S i1 ,S i2 ,…,S ik ] is weighted to obtain the weighted similarity The operation is as follows: Step 5.5 finds the node index with the largest similarity in the weighted similarity matrix, as described below: Similarity node index i max Get reclassification labels Update y = [y1,y 2, …,y n ], loop step 5.4 to step 5.5 until all uncertain nodes are reclassified, and output the final classification result.

Images