Traditional Chinese medicine-symptom association prediction method fusing traditional Chinese medicine theory and causal reasoning

Abstract

The invention provides a traditional Chinese medicine-symptom association prediction method fusing traditional Chinese medicine theories and causal reasoning. The method comprises the following steps: firstly, extracting multi-dimensional characteristics of'taste, nature, channel tropism, rising and falling 'and the like of traditional Chinese medicines and multi-dimensional characteristics of'exogenous pathogenesis, endogenous pathogenesis, qi, blood, body fluid, viscera' and the like of symptoms from a traditional Chinese medicine knowledge base, constructing a quantitative system, then constructing a hierarchical traditional Chinese medicine causal diagram, and determining a bidirectional or unidirectional causal connection relationship among nodes; designing an edge relationship construction rule, and introducing a causal weight in message transmission; constructing a dual-channel heterogeneous graph neural network architecture, and updating node embedding through three rounds of iteration; and finally, through a combined loss function optimization model, outputting a scoring matrix with a confidence coefficient threshold value by a double-branch prediction head so as to accurately predict. The method excavates deep correlation between traditional Chinese medicines and symptoms, and provides a prediction tool for modernization of traditional Chinese medicines.

Description

Technical Field

[0001] This invention relates to the field of traditional Chinese medicine, and in particular to a method for predicting the potential association between traditional Chinese medicine and symptoms using graph neural network technology and causal reasoning. Background Technology

[0002] Currently, the matching of traditional Chinese medicine (TCM) herbs with symptoms mainly relies on experience and rules. Traditional methods have failed to effectively utilize the multi-level features of TCM herbs and symptoms, nor have they fully explored potential associations. With the increasing richness of TCM knowledge bases, data-driven technologies, such as graph neural networks, hold promise for improving the predictive accuracy of the association between TCM herbs and symptoms. However, existing technologies have not yet proposed a method to construct symptom and herb features from a TCM perspective, nor have they proposed a solution for constructing similarity vector matrices and causal graphs, and then integrating them into graph neural networks for efficient prediction. Summary of the Invention

[0003] This invention provides a method for predicting the association between traditional Chinese medicine (TCM) herbs and symptoms by integrating TCM theory and causal reasoning. This method extracts multi-dimensional features from TCM herbs and symptoms to construct similarity vector matrices and causal graphs between TCM herbs and symptoms. Simultaneously, it utilizes existing association matrices between TCM herbs and symptoms to perform multi-level message passing within a graph neural network. Finally, it predicts the potential association between TCM herbs and symptoms based on our newly created dual-branch prediction formula. This model is implemented based on the following steps: (1) Feature extraction steps: Extract multidimensional features of Chinese medicine and symptoms from the TCM knowledge base, and construct feature representation vector matrix of Chinese medicine and symptoms. The first-level features of Chinese medicine are "taste", "nature", "meridian tropism" and "ascending, descending and sinking". The second-level features are "cold, hot, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians, ascending, descending, sinking and floating". The first-level features of symptoms are "external pathogenic factors", "internal pathogenic factors", "qi, blood and body fluids" and "viscera". The second-level features are "wind, cold, summer heat, dampness, dryness, fire, joy, anger, worry, thought, grief, fear, fright, 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 belongs to a primary feature. By calculating the similarity under each primary feature, the primary similarity between Chinese medicine and symptoms constitutes the similarity vector between Chinese medicine and symptoms, thus obtaining the three-dimensional similarity matrix between Chinese medicine and symptoms. (3) Causal graph construction steps: Extract Chinese medicine, symptoms, symptom features and Chinese medicine features from the Chinese medicine knowledge base, construct ternary groups <Chinese medicine features-symptom features-symptom> and <symptom features-Chinese medicine features-Chinese medicine>, and customize four types of nodes: symptoms, Chinese medicine, symptom features. Symptom features include "cold", "heat", "warm", "cool", "sour", "bitter", "sweet", "pungent", "salty", "twelve meridians", "ascending", "descending", "sinking", "floating". Chinese medicine features include "wind", "cold", "summer heat", "dampness", "dryness", "fire", "joy", "anger", "worry", "thinking", "sadness", "fear", "fright", "diet", "work and rest", "qi", "blood", "body fluids", "heart", "liver", "spleen", "lung", "kidney", "gallbladder", "stomach", "small intestine", "large intestine", "bladder", "three jiaos". Symptom nodes and symptom feature nodes perform bidirectional causal connections, Chinese medicine nodes and Chinese medicine feature nodes perform bidirectional causal connections, and Chinese medicine features unidirectionally treat symptom nodes. Then, assign weights to the edges of the graph according to the treatment frequency. (4) Message passing steps: Message passing is performed in GNN to update the features of each node. Based on the previously constructed three-dimensional similarity matrix, message passing is first performed between symptoms and Chinese medicine, and then message passing is performed between symptoms, between Chinese medicine, between symptoms and symptom features, and between Chinese medicine and Chinese medicine features.

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

[0005] S101: The Chinese herbal entities we extracted from the Traditional Chinese Medicine knowledge base are... ,symptom Where n is the total number of Chinese medicines, and m is the number of symptoms. , NER is a method for extracting keywords from text, for each Chinese medicine. Corresponding to a set of characteristics, the primary characteristics of traditional Chinese medicine include "taste," "nature," "meridian tropism," and "ascending, descending, sinking, floating," while the secondary characteristics specifically include "cold, hot, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridian tropisms, ascending, descending, sinking, and floating"; each symptom Corresponding to a set of characteristics, the primary characteristics of symptoms include "external causes", "internal causes", "qi, blood and body fluids", and "internal organs", while the secondary characteristics include "wind, cold, summer heat, dampness, dryness, fire, joy, anger, worry, pensiveness, grief, fear, fright, diet, work and rest, qi, blood, body fluids, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, and triple burner".

[0006] S102: For each Chinese medicine and symptoms Its characteristics include multiple primary and secondary features. Let the set of primary features of traditional Chinese medicine be denoted as . Each first-level feature Composed of several secondary features For example, traditional Chinese medicine The characteristic expression is Each dimension represents a different characteristic; a value of 0 indicates the absence of this characteristic, while 1, 2, and 3 represent the degree of this characteristic. Taking the "cold" nature of a traditional Chinese medicine as an example, severe cold is represented by 3, moderate cold by 2, and minor cold by 1, etc. This set of characteristics expresses that this traditional Chinese medicine is sweet in taste, moderately cold in nature, enters the stomach meridian, and ascends in terms of "ascending, descending, sinking, floating." The characteristic expression of symptoms is similar to the above. Similarly, we define the first-level feature set of symptoms as follows. , z indicates that it consists of z secondary features.

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

[0008] S201: Based on the above-mentioned secondary feature expressions of traditional Chinese medicine and symptoms, this 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. For the traditional Chinese medicine node... and Their similarity can be calculated by examining the similarity under each primary feature. The similarity of traditional Chinese medicine under the k-th primary feature is... The similarity matrix between Chinese medicines can be represented as: ,in, It is a node of traditional Chinese medicine and The similarity at the k-th first-level feature can be calculated using the following formula, where... Let d represent the first-level feature, and d represent the dimension of that first-level feature. Through calculation, we will obtain an n×k×n three-dimensional similarity matrix of Chinese medicines, where k represents the dimension of the similarity vector between Chinese medicines.

[0009] S202: For symptom nodes and Their similarity can be calculated by examining the similarity under each primary feature. The similarity of the symptoms under the z-th primary feature is... Then the similarity matrix between symptoms can be represented as: ,in, Symptom nodes and The similarity under the z-th first-level feature can be calculated using the following formula, where... Let d represent the first-level feature, and d represent the dimension of that first-level feature. By 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: Data source integration: Extract causal relationship triplets <TCM characteristics-symptom characteristics-symptom> and <symptom characteristics-TCM characteristics-TCM> from TCM databases and modern TCM pharmacology research.

[0011] S302: Node Definition: Construct four types of nodes—①Symptom Nodes ②Symptom Feature Nodes, secondary symptom features: cold, heat, warm, cool, sour, bitter, sweet, pungent, salty, twelve meridians, ascending, descending, sinking, floating ③Traditional Chinese Medicine Feature Nodes (secondary features of traditional Chinese medicine: wind, cold, summer heat, dampness, dryness, fire, joy, anger, worry, thought, grief, fear, fright, diet, work and rest, qi, blood, body fluids, heart, liver, spleen, lung, kidney, gallbladder, stomach, small intestine, large intestine, bladder, triple burner) ④Traditional Chinese Medicine Nodes.

[0012] S303: Edge relationship construction: Through the TCM features and TCM prescription statistics in step (1), establish causal edges, perform bidirectional causal connections between symptom nodes and symptom feature nodes, perform bidirectional causal connections between TCM nodes and TCM feature nodes, and perform unidirectional treatment connections between TCM features and symptom nodes. The weight formula is: α = (frequency of occurrence of feature pairs in total number of documents) × 100. Then, weights are assigned according to the treatment frequency. The weighting coefficient formula is: α = (frequency of occurrence / total number of documents) × 100.

[0013] S401: We construct a Chinese medicine-symptom association matrix using a traditional Chinese medicine knowledge base, as follows: ,when When =1, it indicates that there is a correlation between the Chinese medicine h and the symptom s. In addition, we construct the matrix between the Chinese medicine h and the Chinese medicine feature node hc, and the matrix between the symptom s and the symptom feature hs based on the constructed causal graph. We then concatenate these with the Chinese medicine similarity matrix and the symptom similarity matrix to expand them into a (m+n+i+j)×(m+n+i+j) matrix G to construct a GNN (Graph Convolutional Network), where m+n represents the sum of the number of Chinese medicine symptoms and i+j represents the sum of the number of Chinese medicine features and symptom feature nodes.

[0014] Furthermore, the message passing steps include the following specific steps:

[0015] S402: We use G to communicate between symptoms and traditional Chinese medicine, and the message transmission method is as follows: ,in and Let represent the node embeddings of the (l+1)th layer and the lth layer, respectively. D is the degree matrix of graph G, and W(l) is the learnable weight matrix of the lth layer.

[0016] S501: Regarding message passing between symptoms and traditional Chinese medicines, we rely on the three-dimensional similarity matrix created in 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... and To perform message passing, {X}^{'}=[{X}^{'}_{1},{X}^{'}_{2,}...,{X}^{'}_{k}]\in {R}^{n×64×k} The original node was a 2D embedding. We expanded its second dimension by a factor of k (symptoms) and a factor of z (traditional Chinese medicine) respectively, and then used the reshape() function to transform it into a 3D embedding. This represents the extended embedding of the traditional Chinese medicine node. Indicates the node dimension of traditional Chinese medicine. This represents the symptom node dimension.

[0017] S502: Message passing is based on a causal graph: Symptom nodes and symptom feature nodes form bidirectional causal connections, as do traditional Chinese medicine (TCM) nodes and TCM feature nodes. TCM feature nodes form unidirectional therapeutic connections to symptom nodes. The aggregation coefficient utilizes the edge weights of the causal graph, α = (frequency of occurrence / total number of documents) × 100. The message passing formula is expressed as follows: ,in This represents the node that points to node i in the causal graph.

[0018] S503: Then perform message passing on the graph neural network: ,here, This is the feature representation of symptom i after the k-th iteration. This represents the set of symptoms adjacent to herb i. This is a characteristic of herbal medicine. It represents the similarity between symptom i and symptom j across f dimensions.

[0019] S601: Characteristics of each node after each round of message passing All will be updated. At this point, the graph neural network needs to compress the dimensionality of the features to ensure that the final output feature dimension is 64. We compress the dimensionality of the features of each node through a linear transformation (such as a fully connected layer). Suppose that after the k-th round of message passing, the new symptom features are represented as follows: We then compress it into 2D features using the reshape() function, and then compress the features to 64 dimensions using a fully connected layer. The herbal features are also processed using the same message passing and dimensionality transformation method. W is the trainable weight matrix.

[0020] S602: Then, in each round, the normalization function and the activation function are used to activate the embedding after message passing.

[0021] S701: After passing through three layers of GNN, the final embeddings of traditional Chinese medicine and symptoms are obtained. We use a two-branch prediction head to obtain the final scoring matrix: , Representing the weighting factors, we obtain the causal inference prediction score matrix through a dual-line decoder, as follows: , Represents the causal reasoning prediction score matrix. This indicates the final embedding of traditional Chinese medicine. This indicates the final symptom embedding. For trainable weight matrix, This represents the embedding of feature nodes of traditional Chinese medicine. This represents the embedding of symptom feature nodes. We obtain the association prediction score matrix through a bilinear decoder, as follows: , Represents the correlation prediction score matrix, This indicates the final embedding of traditional Chinese medicine. This indicates the final symptom embedding. This is a trainable weight matrix.

[0022] S702: Through The final rating matrix we obtained was used to automatically generate a threshold. Associations above the threshold were selected as positive associations, and the rest were negative associations.

[0023] S801: We selected the Traditional Chinese Medicine (TCM)-symptom association pair as positive instances and all other combinations as negative instances. We then used separate methods to distinguish between the positive and negative TCM instances. and This indicates that, since the number of observed associations is significantly less than the number of unobserved associations, we choose weighted cross-entropy as the mean. and The predicted loss function ,in , and They represent and The amount, (i, j) represents the effect of traditional Chinese medicine on symptoms and the effect of traditional Chinese medicine. and symptoms .parameter As a weighting factor to mitigate the impact of data imbalance. The total loss is expressed as: in Indicates the associated predicted loss. This indicates the loss predicted by causal reasoning. This represents a weighting factor that adjusts the losses of both factors. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of a Chinese medicine-symptom association prediction method that integrates traditional Chinese medicine theory and causal reasoning according to the present invention.

[0025] Figure 2 This is a flowchart of the Chinese medicine-symptom association prediction method S301-S302 of the present invention, which integrates traditional Chinese medicine theory and causal reasoning.

[0026] Figure 3 This is a flowchart illustrating S602 of the present invention, which is a method for predicting the association between traditional Chinese medicine and symptoms by integrating traditional Chinese medicine theory and causal reasoning. Detailed Implementation

[0027] Example 1

[0028] like Figure 1 As shown, this application provides a method for predicting the association between Chinese medicine and symptoms by integrating traditional Chinese medicine theory and causal reasoning, which specifically includes the following steps.

[0029] S101: Feature Extraction: The Chinese herbal entities we extracted from the Traditional Chinese Medicine knowledge base are... This includes "Eclipta prostrata", symptoms This includes "confusion," where n is the total number of Chinese medicinal herbs and m is the number of symptoms. , .

[0030] S102: Traditional Chinese Medicine Feature Extraction: Extract traditional Chinese medicine entities from the TCM knowledge base and assign a set of features to each herb, such as "taste," "nature," "meridian tropism," and "ascending / descending / floating / sinking" characteristics. These features are further subdivided into primary and secondary features. For example, "nature" is a primary feature, while "cold" (divided into severe cold, moderate cold, mild cold, etc.) is a corresponding secondary feature. Features are represented using numerical codes; if a feature is absent, it is represented by 0, and the degree of the feature is represented by 1, 2, 3, etc.

[0031] S103: Symptom Feature Extraction: Similar to traditional Chinese medicine, symptom entities are extracted from the TCM knowledge base, and a set of features is assigned to each symptom, such as "external causes", "internal causes", "qi, blood and body fluids", and "viscera". These features are also divided and coded in detail.

[0032] S201: Similarity Matrix Construction: Based on the secondary feature representations of the above-mentioned traditional Chinese medicines and symptoms, this invention calculates the similarity between traditional Chinese medicines and symptoms using the similarity matrix of each primary feature. The calculation of each similarity is based on its respective primary and secondary features.

[0033] S202: Construction of the Traditional Chinese Medicine (TCM) Similarity Matrix: Based on the secondary feature representation of TCM, the similarity under each primary feature is calculated. For any two TCMs, their similarity is derived by combining the similarities under each primary feature. The similarity calculation considers the dimension of the primary features and uses a specific formula. The final result is an n×k×n three-dimensional TCM similarity matrix, where n is the total number of TCMs and k is the number of primary features.

[0034] S203: Symptom Similarity Matrix Construction: Similar to the construction of the Chinese medicine similarity matrix, the similarity between symptoms is also calculated based on their features. This results in an m×z×m three-dimensional symptom similarity matrix, where m is the number of symptoms and z is the number of first-level features of each symptom.

[0035] S204: Cause-effect graph construction: includes four types of nodes: symptoms, traditional Chinese medicine, symptom features, and traditional Chinese medicine features, and assigns weights to the edges of the cause-effect graph according to their frequency of occurrence.

[0036] S301: Constructing an association matrix: Using a traditional Chinese medicine knowledge base, construct a Chinese medicine-symptom association matrix to represent the known associations between Chinese medicines and symptoms. ,when When =1, it indicates that there is a correlation between Chinese medicine i and symptom j, while "Mohanlian" and "shenhun" are not related. In addition, we construct matrices between Chinese medicine h and Chinese medicine feature nodes hc, and between symptom s and symptom features hs, based on the constructed causal graph. We then expand these matrices into a (m+n+i+j)×(m+n+i+j) matrix G to construct a GNN (Graph Convolutional Network), where m+n represents the sum of the number of Chinese medicine symptoms, and i+j represents the sum of the number of Chinese medicine features and symptom feature nodes.

[0037] S302: Message passing is performed on the graph neural network to update the features of each node. The message passing method between traditional Chinese medicine and symptoms is as follows: This relies on the traditional Chinese medicine-symptom association matrix. The message passing process is completed through multiple iterations, with each iteration updating the feature representation of the node.

[0038] S401: For message passing between symptoms and traditional Chinese medicines, we rely on the three-dimensional similarity matrix created in 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} The original node was a 2D embedding. We expanded its second dimension by a factor of k (symptoms) and a factor of z (traditional Chinese medicine) respectively, and then used the reshape() function to transform it into a 3D embedding. This represents the extended embedding of the traditional Chinese medicine node. Indicates the node dimension of traditional Chinese medicine. This represents the symptom node dimension.

[0039] S402: Message passing is based on the causal graph: Symptom node S ↔ Symptom feature node SC forms a bidirectional causal connection; Traditional Chinese medicine node H ↔ Traditional Chinese medicine feature node HC forms a bidirectional causal connection; and Symptom feature node SC → Traditional Chinese medicine feature node HC forms a unidirectional therapeutic connection. The aggregation coefficient utilizes the edge weights of the causal graph, α = (frequency of occurrence / total number of documents) × 100. The message passing formula is expressed as follows: ,in This indicates a node that points to node i.

[0040] S403: Pass through again and Message passing is performed on a graph neural network. ,here, It is the feature representation of symptom i after the k-th iteration. This represents the set of symptoms adjacent to herb i. This is a characteristic of herbal medicine. It represents the similarity between symptom i and symptom j across f dimensions.

[0041] S501: Characteristics of each node after each round of message passing All will be updated. At this point, the graph neural network needs to compress the dimensionality of the features to ensure that the final output feature dimension is 64. We compress the dimensionality of the features of each node through a linear transformation (such as a fully connected layer). Suppose that after the k-th round of message passing, the new symptom features are represented as follows: We then compress it into 2D features using the reshape() function, and then compress the features to 64 dimensions using a fully connected layer. The herbal features are also processed using the same message passing and dimensionality transformation method. .

[0042] S502: Then, in each round, the normalization function and the activation function are used to activate the embedding after message passing.

[0043] S602: Using the final rating matrix we obtain, we automatically generate a threshold and select associations that are above the threshold as positive associations, and the rest as negative associations.

[0044] S701: We selected the Traditional Chinese Medicine (TCM)-symptom association pair as positive instances and all other combinations as negative instances. We then used separate methods to represent the positive and negative TCM instances. and This indicates that, since the number of observed associations is significantly less than the number of unobserved associations, we choose weighted cross-entropy as the mean. and The predicted loss function ,in , and They represent and The amount, (i, j) represents the effect of traditional Chinese medicine on symptoms and the effect of traditional Chinese medicine. and symptoms .parameter As a weighting factor, it mitigates the impact of data imbalance. The total loss is expressed as: ,in Indicates the associated predicted loss. This indicates the loss predicted by causal reasoning. This represents a weighting factor that adjusts the losses of both factors.

[0045] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: When extracting features from traditional Chinese medicine (TCM) herbs and symptoms, the symptoms include a TCM symptom 's', which is also a typical clinical symptom of disease 'd'. There is a TCM herb 'm'. In the final scoring matrix, the input association matrix shows no correlation between the two. However, in the prediction, the correlation coefficient between 'm' and 's' is higher than a threshold, indicating a positive correlation. In the current medical field, 'm' has become a drug worthy of research in treating disease 'd', possessing potential therapeutic effects and currently under investigation. However, when any one of the feature extraction, similarity matrix construction, or message passing steps is missing, the positive correlation between 'm' and 's' cannot be successfully predicted. By applying TCM knowledge to the association prediction of TCM symptoms using graph neural networks, we can deeply explore the potential connection between TCM herbs and symptoms, improving the accuracy and reliability of the prediction. This method not only considers the feature information of TCM herbs and symptoms but also fully utilizes the interaction between TCM herbs and symptoms through a message passing mechanism, making the prediction results more comprehensive and accurate.

[0046] Example 2

[0047] Based on Example 1, the message passing mechanism can be further optimized, such as... Figure 2 Specifically, during the message transmission process.

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

[0049] S302: Introduce residual connections during message passing to alleviate the vanishing gradient 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 capture the interaction between traditional Chinese medicine and symptoms more flexibly, thereby improving the accuracy and generalization ability of association prediction. x represents the input of the (l+1)th layer. This represents the output of the l-th layer.

[0050] Example 3

[0051] Based on Example 1, the association prediction method is optimized, such as... Figure 3 .

[0052] S602: Ensemble learning can also be used to combine the outputs of multiple association prediction models to improve the accuracy and robustness of predictions. We designed Model 1: trained using a GCN model. Model 2: trained using a GAT model. The final prediction is obtained by fusing the prediction results of the two ensemble learning models. By introducing multi-task learning and ensemble learning methods, we can more comprehensively utilize the information of traditional Chinese medicine and symptoms, improving the accuracy and generalization ability of association prediction. By applying graph neural networks to the association prediction of traditional Chinese medicine and symptoms, we can deeply explore the potential connections between traditional Chinese medicine and symptoms, improving the accuracy and reliability of predictions. 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 a message passing mechanism, making the prediction results more comprehensive and accurate.

[0053] This invention does not directly guide, determine, or replace physicians' diagnostic and treatment practices, nor does it provide specific medication advice to the general public. The final medication choice should be determined by a qualified physician based on the patient's physical signs, medical history, and other medical test data.

Claims

1. A method for predicting the association between traditional Chinese medicine and symptoms by integrating traditional Chinese medicine theory and causal reasoning, characterized in that, The following core steps are used to predict the association between traditional Chinese medicine and symptoms: (1) Steps for constructing a causal graph model: Construct a causal graph model containing four types of nodes: symptom nodes, traditional Chinese medicine nodes, symptom feature nodes, and traditional Chinese medicine feature nodes; (2) Design a multi-channel message passing mechanism based on causal graph and three-dimensional similarity matrix: There are four types of nodes in the causal graph. Symptom nodes and symptom feature nodes perform bidirectional causal connection message passing, traditional Chinese medicine nodes and traditional Chinese medicine feature nodes perform bidirectional causal connection message passing, traditional Chinese medicine feature unidirectional treatment connection symptom node message passing, and symptom nodes and traditional Chinese medicine nodes perform bidirectional message passing through the three-dimensional similarity matrix. (3) Causal constraint and dynamic scoring association prediction steps: The final association prediction result is generated by using the loss function with causal constraint and the dynamic scoring strategy.

2. The method for predicting the association between traditional Chinese medicine and symptoms by integrating traditional Chinese medicine theory and causal reasoning as described in claim 1, characterized in that, The steps for constructing the cause-effect graph model include: (1) Node types: Symptom nodes, Chinese medicine nodes, symptom feature nodes, Chinese medicine feature nodes. Symptom feature nodes include "cold", "heat", "warm", "cool", "sour", "bitter", "sweet", "pungent", "salty", "twelve meridians", "ascending", "descending", "sinking", "floating". Chinese medicine feature nodes include "wind", "cold", "summer heat", "dampness", "dryness", "fire", "joy", "anger", "worry", "thinking", "sadness", "fear", "fright", "diet", "work and rest", "qi", "blood", "body fluids", "heart", "liver", "spleen", "lung", "kidney", "gallbladder", "stomach", "small intestine", "large intestine", "bladder", "triple burner". (2) Edge relationship construction: Based on TCM classics and clinical data, establish the ternary causal relationship of <TCM characteristics-symptom characteristics-symptom> and <symptom characteristics-TCM characteristics-TCM>. Symptom nodes and symptom characteristic nodes perform bidirectional causal connection, TCM nodes and TCM characteristic nodes perform bidirectional causal connection, and TCM characteristics unidirectionally treat symptom nodes. The weight formula is: α = frequency of occurrence of feature pairs in total number of documents × 100; (3) Graph structure extension: Integrate the TCM-symptom association matrix, TCM-TCM feature matrix, symptom-symptom feature matrix, TCM similarity matrix and symptom similarity matrix to form a (m+n+i+j)×(m+n+i+j) dimensional graph convolutional network input matrix.

3. The method for predicting the association between traditional Chinese medicine and symptoms by integrating traditional Chinese medicine theory and causal reasoning as described in claim 1, characterized in that, The multi-channel message passing mechanism designed based on causal graphs and a three-dimensional similarity matrix is ​​as follows: (1) Causal graph channel: Based on edge weight α, symptom nodes and symptom feature nodes perform bidirectional causal connections, traditional Chinese medicine nodes and traditional Chinese medicine feature nodes perform bidirectional causal connections, and traditional Chinese medicine feature nodes perform unidirectional treatment connections to symptom nodes. The transfer formula is: , represents the node pointing to node i in the causal graph, and α = the frequency of occurrence of feature pairs in the total number of documents × 100; (2) Three-dimensional similarity channel: Based on the three-dimensional similarity matrix, the three-dimensional similarity matrix of traditional Chinese medicine has a dimension of n×k×n, and the three-dimensional similarity matrix of symptoms has a dimension of m×z×m. After expanding the node dimension, it is reshaped into a three-dimensional tensor aggregation feature.

4. The method for predicting the association between traditional Chinese medicine and symptoms by integrating traditional Chinese medicine theory and causal reasoning as described in claim 1, characterized in that, The steps for causal constraints and dynamic rating association prediction include: (1) Combination loss function: in This represents the weighted cross-entropy loss for association prediction. This represents the weighted cross-entropy loss for causal reasoning prediction. This represents a weighting factor that adjusts the losses between the two factors. (2) Final scoring matrix: We use a two-branch prediction head: , Representing the weighting factors, we obtain the final causal inference prediction score matrix through a dual-line decoder, as follows: , Represents the causal reasoning prediction score matrix. This indicates the final embedding of traditional Chinese medicine. This represents the embedding of feature nodes of traditional Chinese medicine. For trainable weight matrix, This represents the embedding of symptom feature nodes. Representing the final symptom embeddings, we obtain the association score matrix through a bilinear decoder, as follows: , Represents the correlation score matrix. This indicates the final embedding of traditional Chinese medicine. For trainable weight matrix, This indicates the final symptom embedding.

Images