Traditional Chinese medicine symptom association prediction method based on similarity and graph neural network

Abstract

The invention provides a traditional Chinese medicine symptom association prediction method based on similarity and a graph neural network. According to the method, multi-dimensional characteristics such as taste, nature, channel tropism and rising and falling of traditional Chinese medicines and characteristics such as exogenous causes, endogenous causes, qi, blood, body fluid and viscera of symptoms of the traditional Chinese medicines are extracted from a traditional Chinese medicine knowledge base, and a quantitative expression system is constructed. And constructing a three-dimensional similarity vector matrix of the traditional Chinese medicine and the symptoms by calculating the similarity of each first-level feature, and fusing the three-dimensional similarity vector matrix with the traditional Chinese medicine-symptom incidence matrix to form a GNN input graph structure. In the GNN, feature interaction between traditional Chinese medicine and symptom nodes is realized by using a message passing mechanism, node embedding is updated through three times of iteration, and a correlation score matrix is output in combination with a double-line decoder. And aiming at the problem of data imbalance, a weighted cross entropy loss function optimization model is adopted, a correlation threshold is automatically generated, and accurate prediction is realized. The method effectively excavates deep correlation between traditional Chinese medicine and symptoms, and provides a data-driven prediction tool for modern research of traditional Chinese medicine.

Description

Technical Field

[0001] The present invention relates to the field of traditional Chinese medicine, and particularly to a method for predicting potential associations between traditional Chinese medicines and symptoms through graph neural network technology. Background Art

[0002] Currently, the matching between traditional Chinese medicines and symptoms mainly relies on experience and rules. Traditional methods fail to effectively utilize the multi-level features of traditional Chinese medicines and symptoms, nor can they fully explore potential associations. With the increasing richness of traditional Chinese medicine knowledge bases, data-driven technologies such as graph neural networks are expected to improve the prediction accuracy of associations between traditional Chinese medicines and symptoms. However, the prior art has not proposed a solution that can construct symptom and traditional Chinese medicine features from the perspective of traditional Chinese medicine, nor construct a similarity vector matrix and integrate it into a graph neural network for efficient prediction. Summary of the Invention

[0003] The present invention provides a method for predicting associations between traditional Chinese medicines and symptoms based on similarity and graph neural networks. This method extracts multi-dimensional features of traditional Chinese medicines and symptoms, constructs a similarity vector matrix between traditional Chinese medicines and symptoms, and at the same time uses the existing association matrix of traditional Chinese medicines and symptoms to perform multi-level message passing in a graph neural network, and finally predicts the potential associations between traditional Chinese medicines and symptoms; This model is implemented based on the following steps: (1) Feature extraction step: Extract multi-dimensional features of traditional Chinese medicines and symptoms from a traditional Chinese medicine knowledge base, and construct a feature representation vector matrix of traditional Chinese medicines and symptoms. The primary features of traditional Chinese medicines are "taste", "nature", "meridian tropism", "ascending, descending, floating, and sinking", and the secondary features are specifically "cold, heat, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridian tropisms, ascending, descending, floating, and sinking"; The primary features of symptoms are "external pathogenic factors", "internal pathogenic factors", "qi, blood, and body fluids", "zang-fu organs", and the secondary features are "wind, cold, summer heat, dampness, dryness, fire, joy, anger, worry, pensiveness, sadness, fear, shock, diet, overwork and underwork, qi, blood, body fluids, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, triple energizer"; (2) Similarity matrix construction step: Each secondary feature belongs to a primary feature. By calculating the similarity under each primary feature, the primary similarities between traditional Chinese medicines and symptoms form similarity vectors between traditional Chinese medicines and between symptoms, and a three-dimensional similarity matrix between traditional Chinese medicines and symptoms is obtained; (3) Message passing step: Perform message passing in the GNN to update the features of each node, depending on the previously constructed three-dimensional similarity matrix. First, perform message passing between symptoms and traditional Chinese medicines, and then perform message passing between symptoms and between traditional Chinese medicines.

[0004] Further, the feature extraction step includes the following specific steps:

[0005] S101: The traditional Chinese medicine entities we extracted from the traditional Chinese medicine knowledge base are , symptoms where n is the total number of traditional Chinese medicines and m is the number of symptoms, , , NER is a method for extracting keywords from text, and each traditional Chinese medicine corresponds to a set of features. The first-level features of traditional Chinese medicines are "taste", "nature", "meridian tropism", and "ascending, descending, floating, and sinking". The specific second-level features are "cold, heat, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians tropism, ascending, descending, floating, and sinking"; each symptom corresponds to a set of features. The first-level features of symptoms are "exogenous etiological factors", "endogenous etiological factors", "qi, blood, and body fluids", and "zang-fu organs". The second-level features are "wind, cold, summerheat, dampness, dryness, fire, joy, anger, worry, pensiveness, grief, fear, shock, diet, overwork and underwork, qi, blood, body fluids, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, and triple energizer";

[0006] S102: For each traditional Chinese medicine and symptom its features include multiple first-level features and second-level features. Let the set of first-level features of traditional Chinese medicines be , and each first-level feature consists of several second-level features . For example, the feature expression of traditional Chinese medicine is , and each dimension represents its different features. If it is 0, then this feature does not exist. 1, 2, and 3 represent the degree of this feature. Taking the "cold" of the "nature" of traditional Chinese medicine as an example, severe cold is represented by 3, moderate cold by 2, and mild cold by 1, etc. This string of features represents that this traditional Chinese medicine has a sweet "taste", a moderate cold "nature", a stomach "meridian tropism", and an ascending "ascending, descending, floating, and sinking". The feature expression of symptoms is also as shown above. Similarly, we set the set of first-level features of symptoms as , and z represents that it consists of z second-level features.

[0007] Furthermore, the steps for constructing the similarity matrix include the following specific steps:

[0008] S201: According to the above second-level feature expressions of traditional Chinese medicines and symptoms, the present invention calculates the similarity between traditional Chinese medicines and symptoms through the similarity matrix of each first-level feature; the calculation of each similarity is based on their respective first-level and second-level features. For the traditional Chinese medicine nodes and , their similarity can be obtained through the similarity calculation under each first-level feature. The similarity under the k-th first-level feature of traditional Chinese medicine is , and the similarity matrix between traditional Chinese medicines can be expressed as: , where, is the traditional Chinese medicine node and The similarity under the k-th first-level feature, which can be calculated by the following formula, where represents the first-level feature and d represents the dimension of this first-level feature: , through calculation, we will obtain a three-dimensional traditional Chinese medicine similarity matrix of n×k×n, where k represents the dimension of the similarity vector between traditional Chinese medicines.

[0009] S202: For symptom nodes and , their similarity can be obtained by calculating the similarity under each first-level feature. The similarity under the z-th first-level feature of the symptom is , then the similarity matrix between symptoms can be expressed as: , where is the similarity between symptom nodes and under the z-th first-level feature, which can be calculated by the following formula, where represents the first-level feature and d represents the dimension of this first-level feature: , through calculation, a three-dimensional symptom similarity matrix of m×z×m is obtained, where z represents the dimension of the similarity vector between symptoms.

[0010] 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; thus, we expand it into a matrix G of (m + n)×(m + n) to construct a GNN (Graph Convolutional Network).

[0011] Furthermore, the message passing step includes the following specific steps:

[0012] S302: We perform message passing between symptoms and traditional Chinese medicines through G, and its message passing method is as follows: , where and represent the node embeddings of the (l + 1)-th layer and the l-th layer respectively, D is the degree matrix of graph G, and W(l) is the learnable weight matrix of the l-th layer.

[0013] S401: For message passing between symptoms and between traditional Chinese medicines, we rely on the three-dimensional similarity matrix created by S2: we first expand the dimensions of the symptom and traditional Chinese medicine node features to the dimensions of their corresponding similarity vectors, and then through and Message passing is performed, and {X}^{'}=[{X}^{'}_{1},{X}^{'}_{2,}...,{X}^{'}_{k}]\in {R}^{n×64×k} For the nodes with the original 2D embeddings, we expand their second dimensions by k times (symptoms) and z times (traditional Chinese medicine) respectively, and then convert them into 3D embeddings through the reshape() function. represents the expanded traditional Chinese medicine node embeddings. represents the dimension of the traditional Chinese medicine nodes. represents the dimension of the symptom nodes.

[0014] S402: Then perform message passing on the graph neural network: Here, is the feature representation of symptom i after the k-th 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.

[0015] S501: After each round of message passing, the features of each node will be updated. At this time, the graph neural network needs to compress the dimension of the features to ensure that the dimension of the final output features is 64. We perform dimension compression on the features of each node through a linear transformation (such as a fully connected layer). Assume that after the k-th round of message passing, the new symptom feature representation is We then compress it into 2D features through the reshape() function, and then compress the features to 64 dimensions through a fully connected layer. The herb features are also passed through the same method for message passing and dimension conversion: where W is a trainable weight matrix.

[0016] S502: Then in each round, use the normalization function and the activation function to activate the embeddings after message passing.

[0017] S601: After three layers of GNN, we obtain the final embeddings of traditional Chinese medicine and symptoms. We obtain the final association score matrix through a bilinear decoder as follows: represents the final association score. represents the final traditional Chinese medicine embedding. represents the final symptom embedding. is a trainable weight matrix.

[0018] S602: Through the final score matrix we obtained, we automatically generate a threshold through the score matrix, and select the associations higher than the threshold as positive associations, and the rest as negative associations.

[0019] S701: We select the traditional Chinese medicine - symptom association pairs as positive instances and all other combinations as negative instances. We represent the positive and negative instances of traditional Chinese medicine with and respectively. Since the number of observed associations is significantly less than the number of unobserved associations, we choose weighted cross - entropy as the loss function, , where , and represent the quantities of and respectively, and (i, j) represents the traditional Chinese medicine - symptom pair with traditional Chinese medicine and symptom . The parameter acts as a weight factor to mitigate the impact of data imbalance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 FIG. is a schematic structural diagram of a method for predicting the association between traditional Chinese medicine and symptoms based on similarity and graph neural network according to the present invention.

[0021] Figure 2 FIG. is a schematic flow diagram of steps S301 - S302 of a method for predicting the association between traditional Chinese medicine and symptoms based on similarity and graph neural network according to the present invention.

[0022] Figure 3 FIG. is a schematic flow diagram of step S602 of a method for predicting the association between traditional Chinese medicine and symptoms based on similarity and graph neural network according to the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0023] Example 1

[0024] As Figure 1 shown, the present application provides a method for predicting the association between traditional Chinese medicine and symptoms based on similarity and graph neural network, which specifically includes the following steps.

[0025] S101: Feature extraction: The traditional Chinese medicine entities we extracted from the traditional Chinese medicine knowledge base are , which includes "Eclipta prostrata", and the symptoms which includes "coma", where n is the total number of traditional Chinese medicines and m is the number of symptoms, , .

[0026] 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 tropism", "ascending, descending, floating and sinking". These features are further divided into primary features and secondary features. For example, "nature" is a primary feature, while "cold" (divided into severe cold, moderate cold, mild cold, etc.) is the corresponding secondary feature. The expression of features uses digital coding. If a certain feature does not exist, it is 0, and the degree of the feature is represented by 1, 2, 3, etc.

[0027] 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, blood and body fluids", "zang-fu organs", etc. These features are also detailedly divided and coded.

[0028] S201: Similarity Matrix Construction: According to the above-mentioned secondary feature expressions of traditional Chinese medicine and symptoms, the present invention 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 its respective primary and secondary features.

[0029] S202: Traditional Chinese Medicine Similarity Matrix Construction: Based on the secondary feature expression of traditional Chinese medicine, calculate the similarity under each primary feature. For any two traditional Chinese medicines, their similarity is comprehensively obtained from the similarities under each primary feature. The calculation of similarity takes into account the dimension of the primary feature and is calculated using a specific formula. Finally, a three-dimensional traditional Chinese medicine similarity matrix of n×k×n is obtained, where n is the total number of traditional Chinese medicines and k is the number of primary features.

[0030] S203: Symptom Similarity Matrix Construction: Similar to the construction of the traditional Chinese medicine similarity matrix, the similarity between symptoms is also calculated based on their features. A three-dimensional symptom similarity matrix of m×z×m is obtained, where m is the number of symptoms and z is the number of primary features of symptoms.

[0031] S301: Construct an Association Matrix: Use the traditional Chinese medicine knowledge base to construct a traditional Chinese medicine - symptom association matrix to represent the known association relationships between traditional Chinese medicine and symptoms. , when = 1, it means there is an association between traditional Chinese medicine i and symptom j. There is no association between "Ecliptae Herba" and "coma". Based on this, we expand it into a matrix G of (m + n)×(m + n) to construct a GNN (Graph Convolutional Network).

[0032] S302: Perform message passing on the graph neural network to update the features of each node. The message passing method between traditional Chinese medicine and symptoms is: , which depends on and the 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 in each iteration.

[0033] S401: For message passing between symptoms and traditional Chinese medicines, we rely on the three-dimensional similarity matrix created by S2. First, we expand the dimensions of the symptom and traditional Chinese medicine node features to the dimensions of their corresponding similarity vectors, {X}^{'}=[{X}^{'}_{1},{X}^{'}_{2,}...,{X}^{'}_{k}]\in {R}^{n×64×k} The nodes were originally two-dimensional embeddings. We expand their second dimensions by k times (for symptoms) and z times (for traditional Chinese medicines) respectively, and then convert them into three-dimensional embeddings through the reshape() function. represents the expanded traditional Chinese medicine node embedding. represents the dimension of the traditional Chinese medicine node. represents the dimension of the symptom node.

[0034] S402: Then, through and perform message passing and conduct message passing on the graph neural network: Here, is the feature representation of symptom i after the k-th 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.

[0035] S501: After each round of message passing, the features of each node will be updated. At this time, the graph neural network needs to compress the dimension of the features to ensure that the dimension of the final output features is 64. We perform dimension compression on the features of each node through a linear transformation (such as a fully connected layer). Suppose after the k-th round of message passing, the new symptom feature representation is , and we then compress it into a two-dimensional feature through the reshape() function, and then compress the feature to 64 dimensions through a fully connected layer. The herb features are also passed and dimensionally transformed in the same way: .

[0036] S502: Then, in each round, use the normalization function and the activation function to activate the embeddings after message passing.

[0037] S601: After three layers of GNN, we obtain the final embeddings of traditional Chinese medicines and symptoms. We obtain the final association score matrix through a bilinear decoder as follows: , represents the final association score. represents the final traditional Chinese medicine embedding. represents the final symptom embedding. is a trainable weight matrix.

[0038] S602: Through the scoring matrix we finally obtained, we automatically generate a threshold through the scoring matrix, select the associations higher than the threshold as positive associations, and the rest as negative associations.

[0039] S701: We select the traditional Chinese medicine - symptom association pairs as positive instances, and all other combinations as negative instances. We use and to represent the positive instances and negative instances of traditional Chinese medicine respectively. Since the number of observed associations is significantly less than the number of unobserved associations, we choose weighted cross-entropy as the loss function, , where , and represent the amounts of and respectively, and (i, j) represents the traditional Chinese medicine - symptom pair and traditional Chinese medicine and symptom . The parameter is used as a weight factor to mitigate the impact of data imbalance.

[0040] The technical solutions in the above embodiments of the present application have at least the following technical effects or advantages: When we extract the features of traditional Chinese medicine and symptoms, the symptom contains the traditional Chinese medicine symptom s, and this symptom is also a typical clinical symptom of the disease d. There is a traditional Chinese medicine m in the traditional Chinese medicine. In the final scoring matrix, they are not associated in the input association matrix. In the prediction, the correlation coefficient between m and s is higher than the threshold and is used as a positive correlation traditional Chinese medicine. In the current medical field, m has become a drug worthy of research and has potential therapeutic effects in the treatment of d and is being studied by researchers. When we lack any one of the feature extraction step, similarity matrix construction step, and message passing step, we cannot successfully predict that there is a positive correlation between m and s. By applying traditional Chinese medicine knowledge to the association prediction of traditional Chinese medicine symptoms in the graph neural network, the potential relationship between traditional Chinese medicine and symptoms can be deeply explored, 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 makes full use of the interaction between traditional Chinese medicine and symptoms through the message passing mechanism, making the prediction result more comprehensive and accurate.

[0041] Embodiment 2

[0042] On the basis of Embodiment 1, the message passing mechanism can also be optimized, such as Figure 2 . Specifically, during the message passing process.

[0043] S301: In addition, the graph attention network (GAT) can be used to replace the traditional graph convolutional network (GCN) to more flexibly capture the interaction between TCM and symptoms.

[0044] 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 connections and graph attention networks, the model can be trained more stably and the interaction between Chinese medicine and symptoms can be captured more flexibly, improving the accuracy and generalization ability of association prediction. x represents the input of the l+1 layer, Represents the output of layer l.

[0045] Embodiment 3

[0046] Based on the first embodiment, the association prediction method is optimized, such as Figure 3 .

[0047] S602: Ensemble learning methods can also be used to combine the outputs of multiple association prediction models to improve the accuracy and robustness of predictions. We design sequence 1: first perform message passing between symptoms and Chinese medicines, then perform message passing between symptoms and Chinese medicines. Sequence 2: first perform message passing between symptoms and Chinese medicines, then perform message passing between symptoms and Chinese medicines. The final prediction is obtained by integrating the prediction results of the two ensemble learning. By introducing multi-task learning and ensemble learning methods, the information of Chinese medicines and symptoms can be more comprehensively utilized to improve the accuracy and generalization ability of association predictions. By applying graph neural networks to the association prediction of Chinese medicines and symptoms, the potential connection between Chinese medicines and symptoms can be deeply explored to improve the accuracy and reliability of predictions. This method not only takes into account the characteristic information of Chinese medicines and symptoms, but also makes full use of the interaction between Chinese medicines and symptoms through the message passing mechanism, making the prediction results more comprehensive and accurate.

[0048] This invention does not have the function of directly guiding, determining or replacing the doctor's diagnosis and treatment behavior, nor does it provide specific medication recommendations for the public. The final medication selection should be judged and decided by a qualified doctor based on the patient's physical signs, medical history and other medical examination data.

Claims

1. A method for predicting the association between traditional Chinese medicine symptoms based on similarity and graph neural network, characterized in that: include: (1) Feature extraction step: Extract the multidimensional features of Chinese medicine and symptoms from the TCM knowledge base, and construct the feature representation vector matrix of Chinese medicine and symptoms. The first-level features of Chinese medicine include "flavor", "nature", "meridian", and "rising and falling", and the second-level features include "cold, hot, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians, rising, falling, sinking, and floating"; the first-level features of symptoms include "exogenous causes", "internal causes", "qi, blood, and body fluids", and "viscera", and the second-level features include "wind, cold, heat, dampness, dryness, fire, joy, anger, worry, thinking, sadness, fear, shock, diet, work and rest, qi, blood, body fluids, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, and triple burner"; (2) Similarity matrix construction steps: Each secondary feature is subordinate to a primary feature. By calculating the similarity under each primary feature, the primary similarity between Chinese medicines and symptoms constitutes the similarity vectors between Chinese medicines and symptoms, and obtains a three-dimensional similarity matrix between Chinese medicines and symptoms; (3) Message passing step: Message passing is performed in GNN to update the features of each node. Relying on the three-dimensional similarity matrix constructed previously, message passing is first performed between symptoms and Chinese medicines, and then message passing is performed between symptoms and Chinese medicines.

2. The method for predicting the association between traditional Chinese medicine symptoms based on similarity and graph neural network according to claim 1, characterized in that: The feature extraction step comprises: (1) Extraction of TCM features: Extract TCM entities from the TCM knowledge base and assign a set of features to each TCM, including "flavor", "nature", "meridian", and "rising, falling, sinking and floating". The features are further divided into primary and secondary features. The primary features of TCM include "flavor", "nature", "meridian", and "rising, falling, sinking and floating". The secondary features include "cold, hot, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians, rising, falling, sinking, and floating", which are expressed in digital codes. (2) Symptom feature extraction: Symptom entities are extracted from the TCM knowledge base, and a set of features is assigned to each symptom, including "exogenous causes", "internal causes", "qi, blood, body fluids", and "viscera". The features are further divided into primary and secondary features. The primary features of symptoms include "exogenous causes", "internal causes", "qi, blood, body fluids", and "viscera". The secondary features include "wind, cold, heat, dampness, dryness, fire, joy, anger, worry, thinking, sadness, fear, shock, diet, work and rest, qi, blood, body fluids, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, and triple burner". They are expressed in digital codes.

3. The method for predicting the association between traditional Chinese medicine and symptoms based on similarity and graph neural network according to claim 1, characterized in that: The similarity matrix construction step comprises: (1) Construction of the similarity matrix of traditional Chinese medicine: Based on the secondary feature expression of traditional Chinese medicine, the similarity of each primary feature is calculated, and the similarity vector between any two traditional Chinese medicines is comprehensively obtained. Finally, a three-dimensional similarity matrix of traditional Chinese medicine of n×k×n is obtained, where n is the total number of traditional Chinese medicines and k is the number of primary features; (2) Construction of symptom similarity matrix: Based on the secondary feature expression of the symptoms, the similarity of each primary feature is calculated, and the similarity vector between any two symptoms is comprehensively obtained. Finally, a three-dimensional symptom similarity matrix of m×z×m is obtained, where m is the number of symptoms and z is the number of primary features of the symptoms.

4. The method for predicting the association between traditional Chinese medicine and symptoms based on similarity and graph neural network according to claim 1, characterized in that: The message transmission step includes: Message passing is performed in the graph neural network to update the features of each node. Message passing depends on the three-dimensional similarity matrix constructed previously. For the Chinese medicine and symptom nodes, the dimension of their features is first expanded to the dimension of the corresponding similarity vector, and then the features are updated by message passing with the similarity matrix. The nodes were originally 2-dimensional embeddings, and the second dimensions of the symptom and Chinese medicine nodes were expanded k times and z times respectively, and then converted into 3-dimensional embeddings through the reshape() function. represents the expanded TCM node embedding, Represents the dimension of Chinese medicine node, Represents the symptom node dimension, feature dimension compression and activation: After each round of message passing, the node feature dimension is compressed by linear transformation to ensure that the feature dimension of the final output is 64, and the activation function is used to activate the embedding after message passing. ,here, is the characteristic representation of symptom i after the kth iteration, represents the set of symptoms adjacent to herb i, is a characteristic of herbal medicine j, is the similarity between symptom i and symptom j in f dimensions.

Images