Infectious disease risk assessment method and system based on multimodal graph coding

Abstract

The present invention relates to medical image processing technology, and discloses an infectious disease risk assessment method and system based on multimodal graph coding. The infectious disease risk assessment method based on multimodal graph coding is applied to a PACS system, and the method includes: extracting image features from input image data, and determining the image sample category based on the image features; splicing text features with image features, reducing the dimension, and obtaining the spliced ​​graph code through a two-layer graph neural network output, and inputting the spliced ​​graph code into a kmeans clustering model to determine the category of the inspection sample; after clustering, inputting the number of newly added samples of each category on the same day into an early warning model to determine its confidence; comparing the obtained confidence with a set threshold value to determine its risk coefficient. The present invention uses a graph neural network to convert simple pattern recognition features into graph codes, emphasizing the connection between similar samples, making the kmeans clustering effect more accurate, and thus more sensitive to risk assessment.

Description

Technical Field

[0001] The present invention relates to medical image processing technology, and in particular to an infectious disease risk assessment method and system based on multimodal graph coding applied in a NUMA architecture. Background Art

[0002] The existing PACS system of the hospital mainly stores images and reports. Since the hospital mainly relies on radiologists to read the images, when respiratory infectious diseases occur, there will be a run on medical resources. The backlog of images has no extra doctors to read the images and write reports, making it impossible to timely assess the risk of infectious diseases.

[0003] For example, the prior art with patent application number CN202011497563.8 generally conducts risk assessment of infectious diseases based on a single sample. It does not evaluate the correlation between samples and does not take advantage of the similar characteristics between infectious disease samples, because the assessment results lack sufficient evidence of infectious diseases. Summary of the invention

[0004] In view of the problem that the risk assessment of infectious diseases in the prior art is generally based on a single sample, the assessment has poor correlation and inaccurate results, the present invention provides an infectious disease risk assessment method and system based on multimodal graph coding.

[0005] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0006] The infectious disease risk assessment method based on multimodal graph coding is applied to the PACS system, and the method includes:

[0007] Determine the image sample category, determine the image features of the input image data, and determine the image sample category based on the image features;

[0008] To determine the category of the inspection sample, the text features of the input image report text are obtained, and the text features are spliced ​​with the image features, and the dimension is reduced and the splicing graph code is obtained through the output of the two-layer graph neural network. The splicing graph code is input into the kmeans clustering model to determine the category of the inspection sample;

[0009] Determination of the confidence level of the newly added samples: the number of newly added samples of each category on the day is input into the early warning model to determine its confidence level;

[0010] Risk assessment compares the obtained confidence level with the set threshold to determine its risk factor.

[0011] Preferably, the method further includes updating the kmeans clustering model, inputting the newly added sample splicing graph encoding into the kmeans clustering model for iteration, and updating the kmeans clustering.

[0012] Preferably, the method for determining the image sample category includes:

[0013] Acquisition of image features: inputting the acquired image data into the image pattern recognition model and outputting the image features;

[0014] Dimensionality reduction of feature vectors: input image features into the image feature PCA model for dimensionality reduction;

[0015] To obtain the encoding features, the feature vector after dimension reduction is input into the two-layer graph neural network, and the encoding features are obtained through the output of the two-layer graph neural network;

[0016] Determination of image sample categories, encoding features and three cluster centers c i，i∈{1，2，3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category.

[0017] Preferably, the method for determining the inspection sample category includes:

[0018] Acquisition of text feature vectors: inputting the image report text into the semantic analysis model to obtain the text feature vectors;

[0019] Feature vector concatenation: concatenating the image feature vector with the text feature vector to obtain a concatenated feature vector;

[0020] Dimensionality reduction of feature vectors: input the concatenated feature vectors into the text concatenation PCA model for dimensionality reduction;

[0021] The acquisition of spliced ​​graph encoding features, inputting the reduced-dimensional feature vector into a two-layer graph neural network, and obtaining the spliced ​​graph encoding features through the output of the two-layer graph neural network;

[0022] Determination of image sample categories, mosaic image encoding features and three cluster centers c i，i∈{1，2，3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category.

[0023] Preferably, the two-layer graph neural network that outputs the encoded features is trained by unsupervised clustering.

[0024] Preferably, the unsupervised clustering training method includes:

[0025] Step 1: Input two feature vectors, namely v i and v j , output the similarity coefficient c of the two sample features ij ; With two sample features v i and v jWhether all samples are verified is the label. If they are all verified samples, it is 1; otherwise, it is 0 and the gradient descent method is used for training;

[0026] Step 2, based on the similarity coefficient c between samples ij , set the similarity threshold θ, when c ij When >θ, two samples are connected to construct an undirected graph;

[0027] Step 3: Construct the graph topology matrix A and the h-order similarity matrix M.

[0028] M=(B+B 2 +...+B h ) / h Formula 1

[0029] Among them, B is the transfer matrix of the training set samples. If the sample feature v i and v j If there is an edge between ij =1, B ij =1 / d i , d i For sample v i otherwise, A ij =0, B ij =0;

[0030] Step 4: For the target sample v i and its neighbor node v j ∈N r , where N r v j Neighbor node set, calculate its attention coefficient α ij ;

[0031]

[0032] Step 5. Build a two-layer graph neural network W 0 and W 1 , with node v i For example, it is encoded as x through a two-layer graph neural network i , which can be expressed as

[0033]

[0034] Among them, x′ i is the intermediate conversion result, σ is a nonlinear function, using a two-layer neural network, based on each node v i Neighbor nodes reconstruct graph relationships;

[0035] Step 6: Based on the encoded feature x i , construct the decoding topology matrix A′,

[0036]

[0037] Construct the loss function L r ,

[0038]

[0039] W of the two-layer graph neural network in step 5 0 and W 1 To minimize the loss L r The goal is to use gradient descent method for training.

[0040] In order to solve the above technical problems, the present invention also provides an infectious disease risk assessment system based on multimodal graph coding, which is applied to a PACS system, and includes: an image sample category determination module, an inspection sample category determination module, a new sample confidence determination module and a risk assessment module;

[0041] An image sample category determination module, which determines the image features of the input image data and determines the image sample category based on the image features;

[0042] Inspection sample category determination module: The inspection sample category determination module obtains the text features of the input image report text, and splices the text features with the image features, reduces the dimension and obtains the splicing graph code through the two-layer graph neural network output. The splicing graph code is input into the kmeans clustering model to determine the category of the inspection sample;

[0043] A new sample confidence determination module is added. The new sample confidence determination module inputs the number of new samples of each category on the day into the early warning model to determine its confidence;

[0044] The risk assessment module compares the obtained confidence level with the set threshold to determine its risk factor.

[0045] Preferably, a kmeans clustering update module is further included, and the kmeans clustering update module inputs the newly added sample splicing graph encoding into the kmeans clustering for iteration, and updates the kmeans clustering.

[0046] In order to solve the above technical problems, the present invention further provides an electronic device, comprising:

[0047] at least one processor and memory;

[0048] The memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the infectious disease risk assessment method based on multimodal graph coding as described.

[0049] In order to solve the above technical problems, the present invention also provides a computer-readable storage medium, characterized in that computer execution instructions are stored in the computer-readable storage medium, and when the processor executes the computer execution instructions, the infectious disease risk assessment method based on multimodal graph coding as described is implemented.

[0050] The present invention has significant technical effects due to the adoption of the above technical solution:

[0051] The present invention conducts dual analysis of images and reports, and can identify images alone when reports are missing. When medical resources are sufficient and image reports are written and completed in a timely manner, images and reports can be comprehensively analyzed to timely assess the risk of infectious diseases.

[0052] The present invention uses graph neural networks to convert simple pattern recognition features into graph codes, emphasizing the connection between similar samples, so that the subsequent kmeans clustering will be more accurate and the risk assessment will be more sensitive. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of the present invention.

[0054] Figure 2 This is a flow chart of Example 1 of the present invention.

[0055] Figure 3 This is a flow chart of Example 5 of the present invention. DETAILED DESCRIPTION

[0056] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0057] Example 1

[0058] The infectious disease risk assessment method based on multimodal graph coding is applied to the PACS system, and the method includes:

[0059] Determine the image sample category, determine the image features of the input image data, and determine the image sample category based on the image features;

[0060] To determine the category of the inspection sample, the text features of the input image report text are obtained, and the text features are spliced ​​with the image features, and the dimension is reduced and the splicing graph code is obtained through the output of the two-layer graph neural network. The splicing graph code is input into the kmeans clustering model to determine the category of the inspection sample;

[0061] Determination of the confidence level of the newly added samples: the number of newly added samples of each category on the day is input into the early warning model to determine its confidence level;

[0062] Risk assessment compares the obtained confidence level with the set threshold to determine its risk factor.

[0063] It also includes the updating of the kmeans clustering model, inputting the newly added sample splicing graph encoding into the kmeans clustering model for iteration, and updating the kmeans clustering.

[0064] Methods for determining image sample categories include:

[0065] Acquisition of image features: inputting the acquired image data into the image pattern recognition model and outputting the image features;

[0066] Dimensionality reduction of feature vectors: input image features into the image feature PCA model for dimensionality reduction;

[0067] To obtain the encoding features, the feature vector after dimension reduction is input into the two-layer graph neural network, and the encoding features are obtained through the output of the two-layer graph neural network;

[0068] Determination of image sample categories, encoding features and three cluster centers c i，i∈{1，2，3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category.

[0069] Methods for determining the inspection sample category include:

[0070] Acquisition of text feature vectors: inputting the image report text into the semantic analysis model to obtain the text feature vectors;

[0071] Feature vector concatenation: concatenating the image feature vector with the text feature vector to obtain a concatenated feature vector;

[0072] Dimensionality reduction of feature vectors: input the concatenated feature vectors into the text concatenation PCA model for dimensionality reduction;

[0073] The acquisition of spliced ​​graph encoding features, inputting the reduced-dimensional feature vector into a two-layer graph neural network, and obtaining the spliced ​​graph encoding features through the output of the two-layer graph neural network;

[0074] Determination of image sample categories, mosaic image encoding features and three cluster centers c i，i∈{1，2，3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category.

[0075] The two-layer graph neural network that outputs the encoded features is trained via unsupervised clustering.

[0076] Unsupervised clustering training methods include:

[0077] Step 1: Input two feature vectors, namely v i and v j, output the similarity coefficient c of the two sample features ij ; With two sample features v i and v j Whether all samples are verified is the label. If they are all verified samples, it is 1; otherwise, it is 0 and the gradient descent method is used for training;

[0078] Step 2, based on the similarity coefficient c between samples ij , set the similarity threshold θ, when c ij When >θ, two samples are connected to construct an undirected graph;

[0079] Step 3: Construct the graph topology matrix A and the h-order similarity matrix M.

[0080] M=(B+B 2 +…+B h ) / h Formula 1

[0081] Among them, B is the transfer matrix of the training set samples. If the sample feature v i and v j If there is an edge between ij =1, B ij =1 / d i , d i For sample v i otherwise, A ij =0, B ij =0;

[0082] Step 4: For the target sample v i and its neighbor node v j ∈N r , where N r v j Neighbor node set, calculate its attention coefficient α ij ;

[0083]

[0084] Step 5. Build a two-layer graph neural network W 0 and W 1 , with node v i For example, it is encoded as x through a two-layer graph neural network i , which can be expressed as

[0085]

[0086] Among them, x′ i is the intermediate conversion result, σ is a nonlinear function, using a two-layer neural network, based on each node v i Neighbor nodes reconstruct graph relationships;

[0087] Step 6: Based on the encoded feature x i , construct the decoding topology matrix A′,

[0088]

[0089] Construct the loss function L r ,

[0090]

[0091] W of the two-layer graph neural network in step 5 0 and W 1 To minimize the loss L r The goal is to use gradient descent method for training.

[0092] The early warning model is a Gaussian kernel density estimation. It inputs the daily number of new cases in each of the three categories and outputs a floating point confidence level of 0-100%. The confidence level indicates the percentile of the current value in the total historical value, that is, how many historical values ​​it exceeds. The higher the confidence level, the greater the risk, which is used as the infectious disease risk assessment value.

[0093] Example 2

[0094] Based on Example 1, the present invention further provides an infectious disease risk assessment system based on multimodal graph coding, which is applied to a PACS system, and includes: an image sample category determination module, an inspection sample category determination module, a new sample confidence determination module, and a risk assessment module;

[0095] An image sample category determination module, which determines the image features of the input image data and determines the image sample category based on the image features;

[0096] Inspection sample category determination module: The inspection sample category determination module obtains the text features of the input image report text, and splices the text features with the image features, reduces the dimension and obtains the splicing graph code through the two-layer graph neural network output. The splicing graph code is input into the kmeans clustering model to determine the category of the inspection sample;

[0097] A new sample confidence determination module is added. The new sample confidence determination module inputs the number of new samples of each category on the day into the early warning model to determine its confidence;

[0098] The risk assessment module compares the obtained confidence level with the set threshold to determine its risk factor.

[0099] Preferably, a kmeans clustering update module is further included, and the kmeans clustering update module inputs the newly added sample splicing graph encoding into the kmeans clustering for iteration, and updates the kmeans clustering.

[0100] Example 3

[0101] Based on Embodiment 1, the present invention further provides an electronic device, comprising: at least one processor and a memory;

[0102] The memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the infectious disease risk assessment method based on multimodal graph coding as described.

[0103] Example 4

[0104] Based on Example 1, the present invention also provides a computer-readable storage medium, characterized in that computer execution instructions are stored in the computer-readable storage medium. When the processor executes the computer execution instructions, the infectious disease risk assessment method based on multimodal graph coding as described is implemented.

[0105] Example 5

[0106] On the basis of the above-mentioned embodiment, the image report semantic analysis model of this embodiment uses node2vec as the preprocessing method and LSTM neural network as the basic architecture. When training the model, whether the sample contains descriptions related to pulmonary infectious diseases is the target, and the gradient descent method is used for training, and the final LSTM cell state is used as the text feature vector w1;

[0107] The image pattern recognition model is a three-dimensional ResNet model. When training the model, the goal is to determine whether the image sample contains pulmonary infectious disease lesions. The model is trained using the gradient descent method, and the input of the first fully connected layer of the model is used as the image feature vector w2.

[0108] The graph model may receive image features of length m or image-text concatenation features of length m+n. To ensure that the feature length of the input graph neural network is fixed, the features need to be first reduced to t (t<m<m+n) by PCA.

[0109] The PCA dimensionality reduction of the graph model includes two PCA models, which respectively reduce the image features and the image-text splicing features to t.

[0110] Assume that the three cluster centers of the kmeans clustering result of the current graph encoding feature are c i，i∈{1，2，3} On that day, 1,000 new chest CT scan examinations were added, including 1,000 image data and corresponding report text data.

[0111] S1. Read image data from the image database, input the image pattern recognition model, and output a floating-point image feature vector w2 with a length of m;

[0112] S2. Input the image feature vector w2 into the image feature PCA model to reduce it to a t-dimensional vector v2;

[0113] S3. Input the vector v into the two-layer graph neural network to obtain the encoded feature x2;

[0114] S4. Combine the encoded feature x with three cluster centers c i，i∈{1，2，3} Calculate the Euclidean distance respectively, and take the category of the cluster center with the smallest distance as the category of the current sample;

[0115] S5. After the image report of the sample is written, the image report text is input into the report semantic analysis model to obtain the n-dimensional text feature vector w1;

[0116] S6. Concatenate w1 and w2 to form an m+n-dimensional feature w, and input the text concatenation feature PCA model to reduce the dimension to a t-dimensional feature v, and then repeat S3 to S4 to obtain the final classification of the inspection sample;

[0117] S7. After the classification of all 1000 chest CT scans on the day is completed, the number of newly added samples on the day for the three categories of no infection, low probability of infection, and high probability of infection is calculated respectively, and input into the early warning model to output the confidence of the three categories;

[0118] S8. Compare the confidences of the three categories with the threshold θ. If the confidences of either or both of the two categories, low probability of infection or high probability of infection, are higher than the threshold θ, it indicates that there is an abnormal situation of pulmonary infectious disease on that day and an early warning is required;

[0119] S9. Incorporate the image text mosaic encoding features of all new samples added on that day into the iteration of kmeans clustering, and update the three cluster centers of kmeans clustering.

Claims

1. An infectious disease risk assessment method based on multimodal graph coding, applied in PACS system, characterized in that: Methods include: Determine the image sample category, extract image features from the input image data, and determine the image sample category based on the image features; To determine the category of the inspection sample, the text features of the input image report text are obtained, and the text features are spliced ​​with the image features, and the dimension is reduced and the splicing graph code is obtained through the output of the two-layer graph neural network. The splicing graph code is input into the kmeans clustering model to determine the category of the inspection sample; Determination of the confidence of new samples: input the number of new samples of each category on the day into the early warning model to determine its confidence; risk assessment: compare the obtained confidence with the set threshold to determine its risk factor; The confidence level indicates the percentile of the current value in the total historical value, that is, how many historical values ​​it exceeds. The higher the confidence level, the greater the risk, which is used as an infectious disease risk assessment value. Methods for determining image sample categories include: Acquisition of image features: inputting the acquired image data into the image pattern recognition model and outputting the image features; Dimensionality reduction of feature vectors: input image features into the image feature PCA model for dimensionality reduction; To obtain the encoding features, the feature vector after dimension reduction is input into the two-layer graph neural network, and the encoding features are obtained through the output of the two-layer graph neural network; Determination of image sample categories, encoding features and three cluster centers c i,i∈{1,2,3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category; Methods for determining the inspection sample category include: Acquisition of text feature vectors: inputting the image report text into the semantic analysis model to obtain the text feature vectors; Feature vector concatenation: concatenating the image feature vector with the text feature vector to obtain a concatenated feature vector; Dimensionality reduction of feature vectors: input the concatenated feature vectors into the text concatenation PCA model for dimensionality reduction; The acquisition of spliced ​​graph encoding features, inputting the reduced-dimensional feature vector into a two-layer graph neural network, and obtaining the spliced ​​graph encoding features through the output of the two-layer graph neural network; Determination of image sample categories, mosaic image encoding features and three cluster centers c i,i∈{1,2,3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category; The two-layer graph neural network that outputs the encoded features is trained via unsupervised clustering; Unsupervised clustering training methods include: Step 1: Input two feature vectors, namely v i and v j , output the similarity coefficient c of the two sample features ij ; With two sample features v i and v j Whether all samples are verified is the label. If they are all verified samples, it is 1; otherwise, it is 0 and the gradient descent method is used for training; Step 2, based on the similarity coefficient c between samples ij , set the similarity threshold θ, when c ij When >θ, two samples are connected to construct an undirected graph; Step 3: Construct the graph topology matrix A and the h-order similarity matrix M. M=(B+B 2 +…+B h ) / h Formula 1 Among them, B is the transfer matrix of the training set samples. If the sample feature v i and v j If there is an edge between ij =1,B ij =1 / d i , d i For sample v i otherwise, A ij =0,B ij =0; Step 4: For the target sample v i and its neighbor node v j ∈N r , where N r v j Neighbor node set, calculate its attention coefficient α ij ; Step 5. Build a two-layer graph neural network W 0 and W 1 , with node v i For example, it is encoded as x through a two-layer graph neural network i , expressed as: Among them, x' i is the intermediate conversion result, σ is a nonlinear function, using a two-layer neural network, based on each node v i Neighbor nodes reconstruct graph relationships; Step 6: Based on the encoded feature x i , construct the decoding topology matrix A', Construct the loss function L r , Step 5: W of the two-layer graph neural network 0 and W 1 To minimize the loss L r The goal is to use gradient descent method for training.

2. The infectious disease risk assessment method based on multimodal graph coding according to claim 1, characterized in that: It also includes the updating of the kmeans clustering model, inputting the newly added sample splicing graph encoding into the kmeans clustering model for iteration, and updating the kmeans clustering.

3. An infectious disease risk assessment system based on multimodal graph coding, applied in PACS system, characterized in that: include: Image sample category determination module, inspection sample category determination module, newly added sample confidence determination module and risk assessment module An image sample category determination module, which determines the image features of the input image data and determines the image sample category based on the image features; Inspection sample category determination module: The inspection sample category determination module obtains the text features of the input image report text, and splices the text features with the image features, reduces the dimension and obtains the splicing graph code through the two-layer graph neural network output. The splicing graph code is input into the kmeans clustering model to determine the category of the inspection sample; A new sample confidence determination module is added. The new sample confidence determination module inputs the number of new samples of each category on the day into the early warning model to determine its confidence; Risk assessment module: The risk assessment module compares the obtained confidence with the set threshold to determine its risk coefficient. The confidence indicates the percentile of the current value in the total historical value, that is, how many historical values ​​it exceeds. The higher the confidence, the greater the risk, which is used as the infectious disease risk assessment value; Methods for determining image sample categories include: Acquisition of image features: inputting the acquired image data into the image pattern recognition model and outputting the image features; Dimensionality reduction of feature vectors: input image features into the image feature PCA model for dimensionality reduction; To obtain the encoding features, the feature vector after dimension reduction is input into the two-layer graph neural network, and the encoding features are obtained through the output of the two-layer graph neural network; Determination of image sample categories, encoding features and three cluster centers c i,i∈{1,2,3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category; Methods for determining the inspection sample category include: Acquisition of text feature vectors: inputting the image report text into the semantic analysis model to obtain the text feature vectors; Feature vector concatenation: concatenating the image feature vector with the text feature vector to obtain a concatenated feature vector; Dimensionality reduction of feature vectors: input the concatenated feature vectors into the text concatenation PCA model for dimensionality reduction; The acquisition of spliced ​​graph encoding features, inputting the reduced-dimensional feature vector into a two-layer graph neural network, and obtaining the spliced ​​graph encoding features through the output of the two-layer graph neural network; Determination of image sample categories, mosaic image encoding features and three cluster centers c i,i∈{1,2,3} The kmeans clustering calculates the Euclidean distance respectively, and the category to which the cluster center with the smallest distance belongs is the image sample category; The two-layer graph neural network that outputs the encoded features is trained via unsupervised clustering; Unsupervised clustering training methods include: Step 1: Input two feature vectors, namely v i and v j , output the similarity coefficient c of the two sample features ij ; With two sample features v i and v j Whether all samples are verified is the label. If they are all verified samples, it is 1; otherwise, it is 0 and the gradient descent method is used for training; Step 2, based on the similarity coefficient c between samples ij , set the similarity threshold θ, when c ij When >θ, two samples are connected to construct an undirected graph; Step 3: Construct the graph topology matrix A and the h-order similarity matrix M. M=(B+B 2 +…+B h ) / h Formula 1 Among them, B is the transfer matrix of the training set samples. If the sample feature v i and v j If there is an edge between ij =1,B ij =1 / d i , d i For sample v i otherwise, A ij =0,B ij =0; Step 4: For the target sample v i and its neighbor node v j ∈N r , where N r v j Neighbor node set, calculate its attention coefficient α ij ; Step 5. Build a two-layer graph neural network W 0 and W 1 , with node v i For example, it is encoded as x through a two-layer graph neural network i , expressed as: Among them, x' i is the intermediate conversion result, σ is a nonlinear function, using a two-layer neural network, based on each node v i Neighbor nodes reconstruct graph relationships; Step 6: Based on the encoded feature x i , construct the decoding topology matrix A', Construct the loss function L r , Step 5: W of the two-layer graph neural network 0 and W 1 To minimize the loss L r The goal is to use gradient descent method for training.

4. The infectious disease risk assessment system based on multimodal graph coding according to claim 3, characterized in that: The method also includes a kmeans clustering update module, which inputs the newly added sample splicing graph encoding into the kmeans clustering for iteration and updates the kmeans clustering.

5. An electronic device, characterized in that: include: at least one processor and memory; The memory stores computer-executable instructions; The at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the infectious disease risk assessment method based on multimodal graph coding as described in any one of claims 1 to 2.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions. When the processor executes the computer-executable instructions, the infectious disease risk assessment method based on multimodal graph coding as described in any one of claims 1 to 2 is implemented.

Images