A traditional chinese medicine symptom association prediction method based on similarity and graph neural network

Abstract

The present application provides a kind of similarity and graph neural network-based traditional Chinese medicine symptom association prediction method.Method extracts the multi-dimensional characteristics of traditional Chinese medicine such as "taste, nature, meridian, ascending and descending" and the characteristics of symptoms such as "external cause of disease, internal cause of disease, qi and blood, viscera" from traditional Chinese medicine knowledge base, and constructs a quantitative expression system.By calculating the similarity of each first-level feature, a three-dimensional similarity vector matrix of traditional Chinese medicine and symptoms is constructed, and is fused with traditional Chinese medicine-symptom association matrix to form GNN input graph structure.In GNN, feature interaction between traditional Chinese medicine and symptom nodes is realized by using message passing mechanism, and node embedding is updated after three rounds of iteration, and an association score matrix is output by combining double-line decoder.For the problem of data imbalance, a weighted cross-entropy loss function is used to optimize the model, and an association threshold is automatically generated to realize accurate prediction.This method effectively mines the deep association between traditional Chinese medicine and symptoms, and provides a data-driven prediction tool for modernization research of traditional Chinese medicine.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of traditional Chinese medicine, and in particular to a method for predicting potential associations between traditional Chinese medicine and symptoms through graph neural network technology. BACKGROUND

[0002] Currently, the matching between traditional Chinese medicine and symptoms mainly relies on experience and rules. Traditional methods have failed to effectively utilize the multi-level characteristics of traditional Chinese medicine and symptoms, and have not fully explored potential associations. With the increasing richness of traditional Chinese medicine knowledge base, data-driven technologies such as graph neural networks are expected to improve the prediction accuracy of associations between traditional Chinese medicine and symptoms. However, existing technologies have not proposed a solution that can construct symptom and traditional Chinese medicine features from a traditional Chinese medicine perspective, and has not constructed a similarity vector matrix to integrate into a graph neural network for efficient prediction. SUMMARY

[0003] The present application provides a method for predicting associations between traditional Chinese medicine and symptoms based on similarity and graph neural networks. The method extracts multi-dimensional features of traditional Chinese medicine and symptoms, constructs a similarity vector matrix between traditional Chinese medicine and symptoms, and uses existing association matrices of traditional Chinese medicine and symptoms to perform multi-level message passing in a graph neural network, ultimately predicting the potential associations between traditional Chinese medicine and symptoms.

[0004] The model is implemented based on the following steps:

[0005] (1) Feature extraction step: Extract multi-dimensional features of traditional Chinese medicine and symptoms from a traditional Chinese medicine knowledge base, construct a feature representation vector matrix of traditional Chinese medicine and symptoms, and the first-level features of traditional Chinese medicine include "taste", "nature", "meridian", "ascending and descending", and the second-level features include "cold", "hot", "warm", "cool", "sour", "bitter", "sweet", "pungent", "salty", "twelve meridians", "ascending", "descending", "sinking", "floating"; The first-level features of symptoms include "external pathogenic factors", "internal pathogenic factors", "Qi and blood", "viscera", and the second-level features include "wind", "cold", "humidity", "wetness", "dryness", "fire", "joy", "anger", "sorrow", "thought", "grief", "fear", "surprise", "diet", "labor and rest", "Qi", "blood", "body fluid", "heart", "liver", "spleen", "lung", "kidney", "gallbladder", "stomach", "large intestine", "small intestine", "bladder", "sanjiao";

[0006] (2) Similarity matrix construction step: Each second-level feature belongs to a first-level feature. By calculating the similarity of each first-level feature, the first-level similarity of traditional Chinese medicine and symptoms constitutes the similarity vector between traditional Chinese medicine and symptoms, and a three-dimensional similarity matrix between traditional Chinese medicine and symptoms is obtained.

[0007] (3) Message passing step: Perform message passing in GNN, update the features of each node, and rely on the three-dimensional similarity matrix constructed previously. First, perform message passing between symptoms and traditional Chinese medicine, and then perform message passing between symptoms and traditional Chinese medicine.

[0008] Further, the feature extraction step includes the following specific steps:

[0009] S101: We extract the traditional Chinese medicine entities from the traditional Chinese medicine knowledge base , symptoms where n is the total number of traditional Chinese medicines, m is the number of symptoms, , NER is a method of extracting keywords from text, each traditional Chinese medicine corresponds to a set of features, and the first-level features of traditional Chinese medicines are "taste", "nature", "meridian", "ascending, descending and floating", and the second-level features are "cold, heat, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians, ascending, descending, sinking, floating"; each symptom corresponds to a set of features, and the first-level features of symptoms are "external pathogenic factors", "internal pathogenic factors", "qi, blood and body fluid", and "viscera", and the second-level features are "wind, cold, damp, heat, dryness, fire, joy, anger, sadness, thought, grief, fear, surprise, diet, rest, qi, blood, body fluid, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, and Sanjiao";

[0010] S102: For each traditional Chinese medicine and symptom , the features include multiple first-level features and second-level features. Let the first-level feature set of traditional Chinese medicines be , and each first-level feature is composed of several second-level features . For example, the feature expression of traditional Chinese medicine is , each dimension represents a different feature, and 0 means that the feature does not exist, and 1, 2, and 3 represent the degree of the feature. For example, the "cold" of the "nature" of traditional Chinese medicine is represented by 3 for severe cold, 2 for moderate cold, and 1 for mild cold. This string of features represents that the taste of the traditional Chinese medicine is sweet, the nature is moderate cold, the meridian is stomach, and the ascending, descending and floating is ascending. The feature expression of symptoms is also as shown above. Similarly, let the first-level feature set of symptoms be , and z represents that it is composed of z second-level features.

[0011] Further, the similarity matrix construction step includes the following specific steps:

[0012] S201: According to the second-level feature expression of the traditional Chinese medicine and the symptom, the similarity between the traditional Chinese medicine and the symptom is calculated by the similarity matrix of each first-level feature; the calculation of each similarity is based on the first-level and second-level features. For traditional Chinese medicine nodes and , their similarity can be obtained by calculating the similarity under each first-level feature. The similarity of the kth first-level feature of traditional Chinese medicine is , the similarity matrix between traditional Chinese medicines can be expressed as: wherein, is the node of traditional Chinese medicine and the similarity under the kth primary feature, which can be calculated by the following formula, wherein represents the primary feature, and d represents the dimension of the primary feature: , through calculation, we will get a three-dimensional traditional Chinese medicine similarity matrix of n x k x n, and k represents the dimension of the similarity vector between traditional Chinese medicines.

[0013] S202: For symptom nodes and , their similarity can be calculated by the similarity under each primary feature. The similarity of the zth primary feature of the symptom is , then the similarity matrix between symptoms can be expressed as: , wherein, is the node of symptom and the similarity under the zth primary feature, which can be calculated by the following formula, wherein represents the primary feature, and d represents the dimension of the primary feature: , through calculation, we will get a three-dimensional symptom similarity matrix of m x z x m, and z represents the dimension of the similarity vector between symptoms.

[0014] S301: We construct a traditional Chinese medicine-symptom association matrix through the traditional Chinese medicine knowledge base, as follows: , when =1, it indicates that there is an association between traditional Chinese medicine h and symptom s; in this way, we expand it into a (m+n) x (m+n) matrix G to construct GNN (graph convolutional network).

[0015] Further, the message passing step includes the following specific steps:

[0016] S302: We pass messages between symptoms and traditional Chinese medicines through G, and the message passing method is as follows: , wherein and respectively represent the node embedding of the (l+1)th layer and the lth layer, D is the degree matrix of graph G, and W(l) is the learnable weight matrix of the lth layer.

[0017] S401: In the message passing between symptoms and traditional Chinese medicines, we rely on the three-dimensional similarity matrix created by S2: we first expand the dimension of the symptom and traditional Chinese medicine node features to the dimension of their corresponding similarity vectors, and then pass them through and Message passing, {X}^{'}=[{X}^{'}_{1},{X}^{'}_{2,}...,{X}^{'}_{k}]\in {R}^{n×64×k} , where the original node is a 2-dimensional embedding, we expand its second dimension by k times (symptoms) and z times (TCM), respectively, and then convert it to a three-dimensional embedding through the reshape() function, represents the expanded TCM node embedding, represents the TCM node dimension, represents the symptom node dimension.

[0018] S402: Message passing on the graph neural network again: , here, is the feature representation of symptom i after the kth iteration, represents the set of symptoms adjacent to herb i, is the feature of herb j, is the similarity between symptom i and symptom j in f dimensions.

[0019] S501: After each round of message passing, the feature of each node is updated. At this time, the graph neural network needs to compress the dimension of the feature to ensure that the final output feature dimension is 64. We compress the dimension of the feature of each node through linear transformation (such as fully connected layer). Assuming that after the kth round of message passing, the new symptom feature representation is , we compress it into 2-dimensional feature through the reshape() function, and then pass it through a fully connected layer to compress it to 64-dimensional, and the herb feature also passes through the same method for message passing and dimension conversion: , W is a trainable weight matrix.

[0020] S502: Then use the normalization function in each round, use the activation function to activate the embedding after message passing.

[0021] S601: After three layers of GNN, we get the final embedding of TCM and symptoms, and we get the final association score matrix through the bi-linear decoder as follows: , represents the final association score, represents the final TCM embedding, represents the final symptom embedding, is a trainable weight matrix.

[0022] S602: Through the score matrix we finally obtain, we automatically generate a threshold value through the score matrix, and select the associations higher than the threshold value as positive associations, and the rest as negative associations.

[0023] S701: We select the traditional Chinese medicine-symptom pair as a positive instance, and all other combinations as negative instances. We denote the traditional Chinese medicine positive instances and negative instances as and respectively. Since the number of observed associations is significantly less than the number of unobserved associations, we choose the weighted cross-entropy as the loss function, where , and represent the quantities of and , respectively, (i, j) represents the traditional Chinese medicine-symptom pair and traditional Chinese medicine and symptom . The parameter acts as a weighting factor to mitigate the effects of data imbalance. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 It is a structural schematic diagram of a traditional Chinese medicine symptom association prediction method based on similarity and graph neural network.

[0025] Figure 2 It is a flowchart of the traditional Chinese medicine symptom association prediction method S301-S302 based on similarity and graph neural network.

[0026] Figure 3 It is a flowchart of the traditional Chinese medicine symptom association prediction method S602 based on similarity and graph neural network. DETAILED DESCRIPTION

[0027] Example 1

[0028] As shown in Figure 1 , the present application provides a traditional Chinese medicine symptom association prediction method based on similarity and graph neural network, which specifically includes the following steps.

[0029] S101: Feature extraction: The traditional Chinese medicine entities extracted from the traditional Chinese medicine knowledge base are , which contains "Eclipta prostrata", and the symptoms contains "delirium", where n is the total number of traditional Chinese medicines, m is the number of symptoms, , .

[0030] S102: Traditional Chinese medicine feature extraction: Extract traditional Chinese medicine entities from the traditional Chinese medicine knowledge base, and assign a set of features to each traditional Chinese medicine, such as "taste", "nature", "meridian", "ascending and descending". These features are further divided into primary and secondary features, for example, "nature" is a primary feature, and "cold" (divided into large cold, medium cold, small cold, etc.) is a corresponding secondary feature. The expression of the feature uses digital coding, such as 0 for no certain feature, and 1, 2, 3, etc. for the degree of the feature.

[0031] S103: Symptom feature extraction: Similar to traditional Chinese medicine, extract symptom entities from the traditional Chinese medicine knowledge base, and assign a set of features to each symptom, such as "external pathogenic factors", "internal pathogenic factors", "qi and blood", "zang-fu organs" and so on. These features are also divided in detail and coded.

[0032] S201: Similarity matrix construction: Based on the secondary feature expression of traditional Chinese medicine and symptoms, the present application calculates the similarity between traditional Chinese medicine and symptoms through the similarity matrix of each primary feature. The calculation of each similarity is based on the primary and secondary features.

[0033] S202: Traditional Chinese medicine similarity matrix construction: Based on the secondary feature expression of traditional Chinese medicine, the similarity under each primary feature is calculated. For any two traditional Chinese medicines, their similarity is obtained by synthesizing the similarity under each primary feature. The calculation of similarity considers the dimension of primary feature and uses a specific formula for calculation. Finally, a three-dimensional traditional Chinese medicine similarity matrix of n x k x n is obtained, where n is the total number of traditional Chinese medicines and k is the number of primary features.

[0034] S203: Symptom similarity matrix construction: Similar to the construction of traditional Chinese medicine similarity matrix, the similarity between symptoms is also calculated based on their features. A three-dimensional symptom similarity matrix of m x z x m is obtained, where m is the number of symptoms and z is the number of primary features of symptoms.

[0035] S301: Construction of association matrix: Use traditional Chinese medicine knowledge base to construct a traditional Chinese medicine-symptom association matrix, which represents the known association between traditional Chinese medicine and symptoms, , when =1, it means that there is an association between traditional Chinese medicine i and symptom j, and "Eclipta prostrata" and "delirium" are not associated. In this way, we expand it into a (m+n) x (m+n) matrix G to construct GNN (graph convolutional network).

[0036] S302: Message passing on graph neural network, updating the feature of each node, the message passing method between traditional Chinese medicine and symptoms is: , which depends on and traditional Chinese medicine-symptom association matrix. The process of message passing is completed through multiple iterations, and the feature representation of the node is updated each time.

[0037] S401: In the message passing between symptoms and between traditional Chinese medicines, we rely on the three-dimensional similarity matrix created by S2. We first expand the dimension of the symptom and traditional Chinese medicine node features to the dimension of their corresponding similarity vectors, {X}^{'}=[{X}^{'}_{1},{X}^{'}_{2,}...,{X}^{'}_{k}]\in {R}^{n×64×k} , where the original node embedding is two-dimensional, we expand the second dimension by k times (symptoms) and z times (traditional Chinese medicines) respectively, and then convert it to a three-dimensional embedding through the reshape() function, represents the expanded traditional Chinese medicine node embedding, represents the traditional Chinese medicine node dimension, represents the symptom node dimension.

[0038] S402: Then pass messages through and on the graph neural network: , where, is the feature representation of symptom i after the kth iteration, represents the set of symptoms adjacent to herb i, is the feature of herb j, is the similarity between symptom i and symptom j in f dimensions.

[0039] S501: After each round of message passing, the features of each node are updated. At this time, the graph neural network needs to compress the dimension of the features to ensure that the final output feature dimension is 64. We compress the dimension of the features of each node through linear transformation (such as a fully connected layer). Assuming that after the kth round of message passing, the new symptom feature representation is , we compress it into a 2-dimensional feature through the reshape() function, and then pass it through a fully connected layer to compress it to 64 dimensions. The herb feature also passes through the same method for message passing and dimension conversion: .

[0040] S502: Then use the normalization function in each round to activate the embedding after message passing.

[0041] S601: After three layers of GNN, we get the final embedding of traditional Chinese medicines and symptoms. We get the final association score matrix through the bi-linear decoder as follows: , represents the final association score, represents the final traditional Chinese medicine embedding, represents the final symptom embedding, Trainable weight matrix.

[0042] S602: Through the score matrix we finally obtained, we automatically generate a threshold value through the score matrix, and select the correlation higher than the threshold value as a positive correlation, and the rest as a negative correlation.

[0043] S701: We select traditional Chinese medicine-symptom correlation pairs as positive instances, and all other combinations as negative instances. We represent traditional Chinese medicine positive instances and negative instances as and respectively. Since the number of observed correlations is significantly less than the number of unobserved correlations, we choose the weighted cross-entropy as the loss function, where , and represent the quantities of and , respectively, (i, j) represents the traditional Chinese medicine-symptom pair and traditional Chinese medicine and symptom . The parameter acts as a weight factor to mitigate the effects of data imbalance.

[0044] The technical solutions in the embodiments of the application have at least the following technical effects or advantages:

[0045] When we extract traditional Chinese medicine features and symptom features, the symptom contains s, a TCM symptom, and this symptom is also a typical clinical symptom of disease d. There is m, a traditional Chinese medicine, in the traditional Chinese medicine. In the final score matrix, the two are not associated in the input correlation matrix. In the prediction, the correlation between m and s is higher than the threshold value, which is a kind of positive correlation traditional Chinese medicine. In the current medical field, m has become a drug worthy of research and has potential therapeutic effect in the treatment of d, and is being studied by researchers. When we lack any one of the feature extraction step, the similarity matrix construction step, and the message passing step, we cannot successfully predict that m and s have a positive correlation. By applying TCM knowledge to the correlation prediction of traditional Chinese medicine and symptoms in graph neural networks, the potential relationship between traditional Chinese medicine and symptoms can be deeply mined, and the accuracy and reliability of the prediction can be improved. This method not only considers the feature information of traditional Chinese medicine and symptoms, but also fully utilizes the interaction between traditional Chinese medicine and symptoms through the message passing mechanism, so that the prediction result is more comprehensive and accurate.

[0046] Embodiment Two

[0047] On the basis of embodiment one, the message passing mechanism can also be optimized, such as Figure 2 . Specifically, in the message passing process.

[0048] S301: In addition, the graph attention network (GAT) can be used instead of the traditional graph convolution network (GCN) to capture the interaction between traditional Chinese medicine and symptoms more flexibly.

[0049] S302: Introduce residual connection in the message passing process to alleviate the gradient vanishing problem and improve the training stability of the model: By introducing residual connection and graph attention network, the model can be trained more stably, and the interaction between traditional Chinese medicine and symptoms can be captured more flexibly, improving the accuracy and generalization ability of the association prediction. x represents the input of the l+1 layer,

[0050] Example Three

[0051] On the basis of example one, the association prediction method is optimized, such as Figure 3 .

[0052] S602: Ensemble learning method can also be used to combine the outputs of multiple association prediction models to improve the accuracy and robustness of the prediction. We design sequence 1: first perform message passing between symptoms and traditional Chinese medicine, and then perform message passing between symptoms and traditional Chinese medicine. Sequence 2: first perform message passing between symptoms and traditional Chinese medicine, and then perform message passing between symptoms and traditional Chinese medicine. The final prediction is obtained by fusing the prediction results of the ensemble learning. By introducing multi-task learning and ensemble learning methods, the information of traditional Chinese medicine and symptoms can be more fully utilized, and the accuracy and generalization ability of the association prediction can be improved. By applying graph neural network to the association prediction of traditional Chinese medicine and symptoms, the potential relationship between traditional Chinese medicine and symptoms can be deeply mined, and the accuracy and reliability of the prediction can be improved. This method not only considers the feature information of traditional Chinese medicine and symptoms, but also fully utilizes the interaction between traditional Chinese medicine and symptoms through the message passing mechanism, making the prediction result more comprehensive and accurate.

[0053] The present application does not have the function of directly guiding, determining or replacing the doctor's diagnosis and treatment behavior, and does not provide specific medication advice to the public individuals. The final medication choice should be judged and decided by the doctor with corresponding qualifications according to the patient's physical signs, medical history and other medical test data.

Claims

1. A traditional Chinese medicine symptom association prediction method based on similarity and graph neural network, characterized in that, Comprise: (1) feature extraction step: extract multi-dimensional features of traditional Chinese medicine and symptoms from traditional Chinese medicine knowledge base, construct feature representation vector matrix of traditional Chinese medicine and symptoms, the first level features of traditional Chinese medicine are "taste", "nature", "meridian", "ascending, descending and floating", the second level features are "cold, heat, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians, ascending, descending, sinking, floating"; the first level features of symptoms are "exogenous pathogen", "endogenous pathogen", "qi, blood and body fluid", "viscera", and the second level features are "wind, cold, heat, damp, dry, fire, joy, anger, worry, thought, sadness, fear, surprise, diet, rest, qi, blood, body fluid, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, triple energizer"; (2) similarity matrix construction step: each second level feature belongs to a first level feature, the similarity of each first level feature is calculated, the first level similarity of traditional Chinese medicine and symptoms constitutes the similarity vector between traditional Chinese medicine and symptoms, and a three-dimensional similarity matrix between traditional Chinese medicine and symptoms is obtained; (3) message passing step: message passing is carried out in GNN, and the features of each node are updated, which depends on the three-dimensional similarity matrix constructed before, the message passing between symptoms and traditional Chinese medicine is carried out first, and then the message passing between symptoms and traditional Chinese medicine is carried out.

2. The similarity and graph neural network-based traditional Chinese medicine symptom association prediction method according to claim 1, wherein, The feature extraction step comprises: (1) traditional Chinese medicine feature extraction: extract traditional Chinese medicine entities from traditional Chinese medicine knowledge base, and assign a set of features to each traditional Chinese medicine, including "taste", "nature", "meridian", "ascending, descending and floating", and further divide the features into first level features and second level features, the first level features of traditional Chinese medicine are "taste", "nature", "meridian", "ascending, descending and floating", and the second level features are "cold, heat, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians, ascending, descending, sinking, floating", which are expressed by digital coding; (2) symptom feature extraction: extract symptom entities from traditional Chinese medicine knowledge base, and assign a set of features to each symptom, including "exogenous pathogen", "endogenous pathogen", "qi, blood and body fluid", "viscera", and further divide the features into first level features and second level features, the first level features of symptoms are "exogenous pathogen", "endogenous pathogen", "qi, blood and body fluid", "viscera", and the second level features are "wind, cold, heat, damp, dry, fire, joy, anger, worry, thought, sadness, fear, surprise, diet, rest, qi, blood, body fluid, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, triple energizer", which are expressed by digital coding.

3. The similarity and graph neural network-based traditional Chinese medicine symptom association prediction method of claim 1, wherein, The similarity matrix construction step comprises: (1) traditional Chinese medicine similarity matrix construction: based on the expression of second level features of traditional Chinese medicine, the similarity of each first level feature is calculated, and the similarity vector between any two traditional Chinese medicines is obtained, and finally a three-dimensional traditional Chinese medicine similarity matrix of n*k*n is obtained, wherein n is the total number of traditional Chinese medicines, and k is the number of first level features; (2) symptom similarity matrix construction: based on the expression of second level features of symptoms, the similarity of each first level feature is calculated, and the similarity vector between any two symptoms is obtained, and finally a three-dimensional symptom similarity matrix of m*z*m is obtained, wherein m is the number of symptoms, and z is the number of first level features of symptoms.

4. The similarity and graph neural network-based traditional Chinese medicine symptom association prediction method according to claim 1, wherein, The message passing step comprises: Message passing is performed in the graph neural network to update the features of each node, and message passing relies on the three-dimensional similarity matrix constructed in advance. For the nodes of traditional Chinese medicine and symptoms, the dimension of the features is first expanded to the dimension of the corresponding similarity vector, and then the features are updated through message passing with the similarity matrix; , the original 2-dimensional embedding of the node, the second dimension of the symptom and traditional Chinese medicine nodes is expanded k times and z times respectively, and then reshaped into a three-dimensional embedding through the reshape() function, represents the expanded embedding of the traditional Chinese medicine node, represents the dimension of the traditional Chinese medicine node, represents the dimension of the symptom node, feature dimension compression and activation: after each round of message passing, the features of the nodes are compressed in dimension through linear transformation to ensure that the final output feature dimension is 64, and the embedding after message passing is activated using the activation function, , here, is the feature representation of symptom i after the kth iteration, represents the set of symptoms adjacent to herb i, is the feature of herb j, is the similarity between symptom i and symptom j in f dimensions.

Images