Traditional Chinese medicine multi-label evidence element classification method and system based on graph attention network

Abstract

The invention discloses a traditional Chinese medicine multi-label evidence element classification method and system based on a graph attention network. The method comprises the following steps: performing bidirectional context semantic modeling on a clinical text by utilizing a pre-trained BERT model, generating a context-related clinical text embedded vector, and generating an initial embedded representation of a syndrome element tag through a pre-trained word vector model; and then fusing clinical text embedding and syndrome element label embedding by adopting a graph attention network to generate a clinical information-syndrome element embedding matrix, inputting the embedding matrix into a multi-layer perceptron classifier, performing multi-label classification training in combination with a Focal Loss loss function, and outputting a prediction result of a clinical information-syndrome element relationship. According to the method, the complex association between clinical information and syndrome elements can be accurately captured, high-precision and explainable decision support is provided for traditional Chinese medicine clinical diagnosis, and the method is suitable for scenes such as personalized treatment scheme recommendation, curative effect evaluation and knowledge base construction.

Description

Technical Field

[0001] The present invention belongs to the field of intersectional technology of traditional Chinese medicine informatization and artificial intelligence, and in particular relates to a method and system for traditional Chinese medicine multi-label syndrome classification based on graph attention network. Background Art

[0002] Traditional Chinese medicine diagnosis and treatment systems face many technical bottlenecks, such as static knowledge. Existing knowledge bases rely on manual data entry and are unable to integrate clinical data and the latest research results in real time, resulting in poor knowledge timeliness. Image data such as tongue and pulse patterns are separated from text descriptions, failing to achieve cross-modal feature fusion and incomplete diagnostic information. Plan generation relies on a fixed rule base and lacks a dynamic optimization mechanism based on efficacy feedback. Individual differences, such as patient constitution, are not considered, resulting in plans that are highly universal but less targeted. Currently, existing technologies, such as static knowledge graph methods, use the TransR model to train entity relationships, but the node weights are fixed and cannot reflect the latest clinical evidence. Single-modal analysis systems extract tongue image features through ResNet, but are not associated with textual symptoms, limiting diagnostic accuracy. Rule-driven recommendation engines generate plans based on a classifier chain, but do not introduce an efficacy feedback mechanism, resulting in low iteration efficiency. Summary of the Invention

[0003] The purpose of the present invention is to provide a TCM multi-label syndrome classification method and system based on graph attention network, aiming to solve the low-order interaction limitations, data redundancy and category imbalance problems existing in traditional TCM syndrome classification.

[0004] The present invention is implemented as follows: a method for multi-label syndrome classification in traditional Chinese medicine based on a graph attention network, the method comprising the following steps:

[0005] S1. Screen high-frequency syndrome tags from clinical diagnosis and treatment records and construct a standardized dataset containing disease-type syndrome tags and disease-location syndrome tags;

[0006] S2. Use the pre-trained BERT model to perform bidirectional contextual semantic modeling on clinical text and generate context-related clinical text embedding vectors;

[0007] S3, generate the initial embedding representation of the certificate element label through the pre-trained word vector model;

[0008] S4. Construct a clinical information-syndrome interaction graph, use a graph attention network to fuse clinical text embeddings and syndrome label embeddings, and generate a clinical information-syndrome embedding matrix;

[0009] S5. Input the embedding matrix into a multi-layer perceptron classifier, perform multi-label classification training in combination with a Focal Loss function, and output the prediction result of the clinical information-syndrome factor relationship.

[0010] Preferably, in step S1, the screening is to filter out syndrome factor labels with an occurrence frequency of less than 2%, and retain 40 pathological syndrome factors and 11 pathological location syndrome factors.

[0011] Preferably, in step S2, the bidirectional contextual semantic modeling of the clinical text using the pre-trained BERT model is specifically as follows: the global semantic information of the clinical text is extracted through the bidirectional Transformer architecture of the BERT model, and the contextual relationship between words is calculated using the self-attention mechanism.

[0012] Preferably, in step S4, the graph attention network includes 8 attention heads, updates the clinical text features by weighted aggregation of neighbor node information, and uses a LeakyReLU activation function with a negative slope of 0.01.

[0013] The present invention further discloses a TCM multi-label syndrome classification system based on a graph attention network, the system comprising:

[0014] The data preprocessing module is used to screen high-frequency syndrome factor labels from clinical diagnosis and treatment records and construct a standardized data set containing disease-type syndrome factors and disease-location syndrome factors;

[0015] The clinical text embedding module is used to perform bidirectional contextual semantic modeling on clinical text using the pre-trained BERT model to generate context-related clinical text embedding vectors;

[0016] The certificate element label embedding module is used to generate the initial embedding representation of the certificate element label through the pre-trained word vector model;

[0017] The graph attention network embedding module is used to construct a clinical information-syndrome interaction graph and use the graph attention network to fuse clinical text embedding and syndrome label embedding to generate a clinical information-syndrome embedding matrix;

[0018] The multi-label classification training module is used to input the embedding matrix into the multi-layer perceptron classifier, perform multi-label classification training in combination with the FocalLoss loss function, and output the prediction result of the clinical information-syndrome factor relationship.

[0019] Preferably, in the data preprocessing module, the screening is to filter out syndrome factor labels with an occurrence frequency of less than 2%, and retain 40 disease-related syndrome factors and 11 disease-site syndrome factors.

[0020] Preferably, in the clinical text embedding module, the bidirectional contextual semantic modeling of the clinical text using the pre-trained BERT model is specifically as follows: the global semantic information of the clinical text is extracted through the bidirectional Transformer architecture of the BERT model, and the contextual relationship between words is calculated using the self-attention mechanism.

[0021] Preferably, in the graph attention network module, the graph attention network includes 8 attention heads, updates the clinical text features by weighted aggregation of neighbor node information, and uses a LeakyReLU activation function with a negative slope of 0.01.

[0022] The present invention further discloses a TCM auxiliary diagnosis and treatment device, which includes a data acquisition device, the above-mentioned TCM multi-label syndrome classification system based on the graph attention network, and a visual interface for outputting diagnostic suggestions.

[0023] Preferably, the data acquisition device includes a sensor that supports collecting tongue images and extracts tongue coating texture features through a ResNet-50 model.

[0024] The present invention overcomes the shortcomings of the prior art and provides a method and system for TCM multi-label syndrome classification based on a graph attention network. The present invention constructs a dynamic knowledge fusion framework that, through the coordinated optimization of a pre-trained language model and a graph attention network, enables efficient modeling of the complex relationship between clinical information and syndromes (i.e., the TCM multi-label syndrome classification model based on a graph attention network in the method or system of the present invention, referred to as the ML-SCGAN model).

[0025] In the data preprocessing stage, the present invention extracts symptom descriptions and syndrome labels from clinical diagnosis and treatment records, removes low-frequency labels with a frequency of less than 2%, and constructs a standardized data set containing 40 disease syndromes and 11 disease location syndromes. In order to solve the problem of heteromorphic characters in ancient books and documents, a heteromorphic character mapping database is constructed for unified replacement. For example, the various historical expressions of "chest pain" are normalized into modern standard terms, and paper case records are digitized in combination with OCR technology. The information of uncommon characters is supplemented through the human-computer interaction interface to ensure the integrity and consistency of the data. For unstructured image data such as tongue and pulse, an improved ResNet-50 model is used to extract HSV spatial features, in which the calculation of the redness of the tongue is not only based on the ratio of the RGB channels, but also introduces a texture analysis algorithm to quantify the roughness of the tongue coating, and is deeply integrated with the text features through a cross-modal alignment loss function. The formula is:

[0026]

[0027] in, is the image feature vector, is the text feature vector, p img is the image modality feature distribution, p text is the text modal feature distribution, and the coefficient λ is the weight for balancing the two losses, which needs to be tuned through experiments.

[0028] In the clinical text embedding stage, this paper adopts the 12-layer Transformer structure of the BERT-base model to perform bidirectional encoding on the symptom text, inputs the symptom sequence, and dynamically captures the contextual dependencies between words through the self-attention mechanism. Specifically, the clinical text is first represented as a word sequence {Syn1, Syn2, ..., Syn m}, through the embedding layer E∈R d*|V| Convert these words into word vector matrices {S1, S2, ..., S m}, where the i-th word in the dictionary V is represented by Syn i , d is the word vector dimension, |V| is the dictionary length, S i is the word vector for the i-th word. Next, these word vectors are processed by the BERT model. BERT uses a bidirectional Transformer framework to capture the relationship between words in context. The representation of each word in the model is calculated using the following formula:

[0029]

[0030] in, Represents the representation of the i-th word in the layer, and Transformer(*) is the calculation function of the Transformer model. BERT uses the self-attention mechanism to adjust the representation of each word. The self-attention mechanism is calculated as:

[0031]

[0032] Among them, Q, K, V are the matrices of query, key and value respectively, d k is the dimension of the key. Through this mechanism, the model is able to capture the contextual relationship between words.

[0033] For example, the symptom "chest pain with back pain" is encoded into a 768-dimensional vector, capturing the correlation between "chest pain" and "back pain." Positional encoding is also used to distinguish the semantic differences between "chest pain → back pain" and "back pain → chest pain." Embedded representations of syndrome labels are generated using the GloVe.6B.300d pre-trained model. For example, the vector for the syndrome "liver" captures its Traditional Chinese Medicine (TCM) properties of "dredging" and "storing blood," and is aligned with BERT features through a linear transformation. This alignment process enables semantic interaction between symptoms and syndrome labels within a unified high-dimensional space, laying the foundation for subsequent graph structure modeling.

[0034] After BERT processing, word vectors are aggregated to form fixed-dimensional sentence vectors. These feature vectors integrate the global semantic information of the clinical text. Ultimately, these generated clinical text feature vectors, along with the embeddings of syndrome factor labels, are input into a graph neural network for relationship learning and prediction, thereby accurately predicting the relationship between clinical information and syndrome factors.

[0035] During the certificate element label embedding phase, each certificate element label needs to be effectively embedded into a continuous vector space so that the model can learn the semantic characteristics of the label and make accurate relationship predictions. The model first uses the pre-trained word vector model GloVe to generate initial word vectors for each certificate element label. These word vectors provide basic semantic information of the label, expressed as follows:

[0036] Z i =E d (label i )

[0037] Among them, Z i is the word vector of the i-th label, E d (*) is the pre-trained word vector model, label i is the i-th certificate label.

[0038] In the graph attention network embedding stage, the present invention constructs a clinical information-symptom factor interaction graph and dynamically aggregates node information through a multi-head graph attention network (GAT). The nodes of the interaction graph include clinical text embedding vectors and syndrome factor label vectors, and the edge weights are calculated by co-occurrence frequency. Specifically, the embedding vector of the clinical text generated based on BERT is represented as:

[0039] S i =BERT(Syn i )

[0040] Among them, Syn i Represents the i-th clinical text. Next, generate the initial word vector Z for each syndrome label j At this point, the ML-SCGAN model of the present invention needs to fuse and optimize the relationship between clinical information and syndrome factors. Graph convolution constructs a clinical information-syndrome graph to represent the relationship between clinical text and syndrome factor labels. Each clinical text S i Will verify the element label Z with its neighbors j Specifically, GAT updates the feature vector of a node by weighted aggregation of information from neighboring nodes. The update process can be expressed as:

[0041] S′ i =GAT(S i ,{Z j} j∈N(i) )

[0042] where S′ i is the updated clinical text feature vector, N(i) is the set of neighbor syndrome elements of clinical information i, {Z j} j∈N(i)is the embedding of these neighbor syndromes. GAT assigns different weights to each neighbor node through the attention mechanism to capture the relationship between clinical information and syndromes.

[0043] During the graph attention network embedding phase, the graph attention network generates an optimized clinical information-syndrome embedding matrix. This matrix captures the high-level semantic relationships between clinical text features and syndrome labels. This embedding matrix not only reflects the individual characteristics of symptoms and syndromes but also integrates the complex relationships between them, resulting in a more comprehensive and accurate representation of each symptom-syndrome pair.

[0044] During the multi-label classification training phase, the optimized clinical text-symptom embedding matrix is ​​fed into the classifier for relationship prediction. The classifier's core task is to convert these high-dimensional embedding vectors into specific relationship predictions. A classifier typically consists of a multi-layer perceptron, consisting of several fully connected layers. Each layer processes and nonlinearly transforms the input features using a weight matrix and activation function to extract underlying patterns and features from the data. Specifically, the classifier is calculated using the following formula:

[0045] y=σ(W2*(ReLu(W1*h+b1))+b2)

[0046] Here, h represents the clinical text-symptom embedding matrix, W1 and W2 are weight matrices, b1 and b2 are bias values, ReLu is the activation function, and y is the final predicted value. Through these calculations, the classifier can convert high-dimensional embedding vectors into probability values, indicating the strength of the relationship between each pair of clinical text-symptom factors.

[0047] During training, the classifier optimizes the model parameters by minimizing the error between the predicted results and the true labels. A common loss function for multi-label classification tasks is the cross entropy loss function, which measures the gap between the predicted probability and the true label. The formula is:

[0048]

[0049] Where N is the number of samples, C is the number of labels, and y ij is the true label of the jth label of the i-th sample, p ij is the predicted probability.

[0050] In addition, in order to deal with the problem of class imbalance, Focal Loss can be used. Focal Loss is an improvement to the cross entropy loss, which gives more weight to samples that are difficult to classify, thereby reducing the attention paid to simple samples.

[0051]

[0052] Among them, α ijis the weight of sample i on label j, and γ is a modulating factor that controls the attention paid to difficult and easy samples. Ultimately, the classifier can predict the relationship between newly input symptom-syndrome pairs, providing accurate diagnostic support and helping to accurately predict the relationship between clinical information and syndrome factors.

[0053] Compared with the shortcomings and deficiencies of the existing technology, the present invention has the following beneficial effects: the present invention can accurately capture the complex relationship between clinical information and syndrome factors, provide high-precision and explainable decision support for clinical diagnosis of traditional Chinese medicine, and is suitable for scenarios such as personalized treatment plan recommendation, efficacy evaluation and knowledge base construction. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 This is a schematic diagram of the structure of the TCM multi-label syndrome classification system of the present invention;

[0055] Figure 2 This is the architecture diagram of the TCM multi-label syndrome classification system of the present invention;

[0056] Figure 3 is a bar chart comparing model performance in an embodiment of the present invention;

[0057] Figure 4 This is a performance comparison chart of GAT and GCN models in multi-label classification tasks;

[0058] Figure 5 It is a flow chart of the steps of the method of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0060] The present invention discloses a TCM multi-label syndrome classification system (ML-SCGAN model) based on graph attention network, such as Figure 1 、 2 As shown, the system includes:

[0061] The data preprocessing module is used to screen high-frequency syndrome factor labels from clinical diagnosis and treatment records and construct a standardized data set containing disease-type syndrome factors and disease-location syndrome factors;

[0062] The clinical text embedding module is used to perform bidirectional contextual semantic modeling on clinical text using the pre-trained BERT model to generate context-related clinical text embedding vectors;

[0063] The certificate element label embedding module is used to generate the initial embedding representation of the certificate element label through the pre-trained word vector model;

[0064] The graph attention network embedding module is used to construct a clinical information - syndrome element interaction graph, and adopts a graph attention network to fuse clinical text embeddings and syndrome element label embeddings to generate a clinical information - syndrome element embedding matrix;

[0065] The multi-label classification training module is used to input the embedding matrix into a multi-layer perceptron classifier, and perform multi-label classification training in combination with the FocalLoss loss function to output the prediction result of the clinical information - syndrome element relationship.

[0066] In the embodiment of the present invention, the data preprocessing module belongs to the input layer in the system architecture and can receive clinical texts (such as "dry mouth, constipation") and syndrome element labels (such as "lung, yin deficiency"). In the data preprocessing module, the data source extracts clinical information and syndrome element labels from real clinical diagnosis and treatment records, covering image data such as tongue images and pulse images and text descriptions. Among them, the data preprocessing includes the following contents:

[0067] (1) Label screening, eliminating low-frequency syndrome element labels with a frequency lower than 2%, and constructing a standardized data set, including 40 disease nature syndrome elements (such as qi deficiency, blood stasis, damp-heat) and 11 disease location syndrome elements (such as liver, spleen, lung);

[0068] (2) Variant character processing, unifying ancient Chinese terms through a variant character mapping database (such as mapping "chài" to "yù"), digitizing paper cases in combination with OCR technology, and supplementing rare character information through a human-computer interaction interface;

[0069] (3) Image feature extraction, collecting tongue image and pulse image, using an improved ResNet-50 model (Reference: Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. Deep Residual Learning for Image Recognition arXiv: 1512.03385) to extract HSV space features (such as the redness degree of the tongue body is calculated based on the RGB ratio, and the roughness of the tongue coating is quantified through texture analysis), and fusing with text features through a cross-modal alignment loss function.

[0070] In the embodiment of the present invention, semantic embedding includes clinical text representation learning and syndrome element label representation learning. Among them, in the clinical text embedding module, in clinical text representation learning, a 12-layer Transformer structure of the BERT-base model is adopted, and the symptom sequence (such as "chest pain radiating to the back") is input, and a 768-dimensional context semantic vector is extracted through the self-attention mechanism. The formula is as follows:

[0071]

[0072] Position coding is also performed to distinguish the order of symptoms, such as the semantic difference between "chest pain → back pain" and "back pain → chest pain".

[0073] In the system of the present invention, the certificate label embedding module belongs to the BERT encoder in terms of architecture and is used to extract contextual semantic features of symptom text.

[0074] In the syndrome label embedding module, the GloVe.6B.300d pre-trained model (GloVe.6B.300d is a pre-trained model from the 2014 paper "GloVe: Global Vectors for Word Representation" by Jeffrey Pennington, Richard Socher, and Christopher D. Manning. The paper link is: https: / / nlp.stanford.edu / pubs / glove.pdf) is used to generate initial vectors for syndrome labels (such as "liver"), which incorporate Traditional Chinese Medicine (TCM) characteristics such as "dredging" and "storing blood." GloVe vectors are aligned with BERT semantic vectors through a linear transformation to ensure a unified semantic space.

[0075] In the system of the present invention, the graph attention network embedding module is architecturally part of the GAT fusion layer, using an 8-head graph attention network to dynamically optimize the relationship between clinical information and syndrome factors. In the graph attention network embedding module, graph network optimization includes the construction of the clinical information-syndrome interaction graph and GAT dynamic aggregation.

[0076] Among them, in the construction of the clinical information-syndrome interaction graph, advanced node definition, clinical text embedding (generated by BERT) and syndrome label embedding (generated by GloVe) are used as nodes, and then the edge weights are calculated, and the edge weights are dynamically allocated based on the co-occurrence frequency, such as the number of co-occurrences of "chest pain" and "qi deficiency".

[0077] In GAT dynamic aggregation, the multi-head attention mechanism adopts an 8-head GAT layer to update clinical information features by weighted aggregation of neighbor node information; the activation function uses LeakyReLU with a negative slope of 0.01 and a Dropout ratio of 0.3 to prevent overfitting.

[0078] In this embodiment of the present invention, the multi-label classification training module is used to output the prediction results and confidence of the clinical information-syndrome relationship in combination with Focal Loss. In the multi-label classification training module, the classifier design adopts a network structure, using a multi-layer perceptron classifier (MLP) with 2 layers of full connection, and the formula is:

[0079] y=σ(W2*(ReLu(W1*h+b1))+b2)

[0080] Where h represents the clinical information-syndrome embedding matrix, W1 and W2 are weight matrices, b1 and b2 are bias values, ReLu(*) is the activation function, and y is the final predicted value.

[0081] In addition, the loss function Focal Loss is used to solve the problem of category imbalance. The formula is:

[0082]

[0083] Among them, α ij is the weight of sample i on label j, and γ is an adjustment factor that controls the attention of difficult and easy samples. In this embodiment of the present invention, the parameter γ is set to 2, α ij Dynamically adjust according to label distribution.

[0084] Based on the above content, the present invention further trains and optimizes the system. In the parameter setting, the input dimension is 768, which is consistent with the BERT output; the training round is 50 epochs, the batch size is 32; the optimizer Adam learning rate is 1e-4; the hardware configuration is: NVIDIA V100 GPU, 32GB video memory.

[0085] In order to further verify the effect of the above-mentioned system of the present invention, the embodiment of the present invention carried out the following experimental results verification. Among them, Micro-F1 and Hamming Loss were selected as evaluation indicators because these two indicators can comprehensively reflect the performance of the model from different dimensions, covering precision, error rate and overall accuracy. Micro-F1 is particularly suitable for solving the problem of label category imbalance by comprehensively evaluating the precision and recall rate of all labels. Hamming Loss calculates the proportion of wrong labels of the model in all instances, which can clearly show the accuracy of the model in the label prediction of each instance. In multi-label classification tasks, Hamming Loss is particularly important because it can reveal the specific errors of the model in each label prediction, providing a clear direction for further improving the model. Through these two indicators, the classification effect of the model can be examined from both the overall and detailed aspects to ensure its robustness and reliability under different conditions.

[0086] The evaluation indicators are defined as follows:

[0087]

[0088] According to the experience of previous studies, a higher Micro-F1 value indicates a more accurate classification result, while a lower Hamming Loss value means a higher classification accuracy.

[0089] In the experimental comparison, HSVM performed best among all traditional machine learning models, achieving a Micro-F1 score of 0.508. However, the proposed ML-SCGAN model (i.e., the Multi-Label Syndrome Classification with GraphAttention Networks model for TCM) demonstrated significant advantages, achieving a 12.3% improvement in Micro-F1 score compared to HSVM. The specific results are shown in Table 1.

[0090] Table 1

[0091]

[0092] Among the neural network models, the convolutional neural network (CNN) performed in the middle, with a Micro-F1 score of 0.423, but its Hamming Loss was relatively high, reaching 0.054. This suggests that the CNN model may be overfitting or poorly handling sequential dependencies in multi-label tasks. In contrast, the bidirectional long short-term memory (Bi-LSTM) network performed the worst, with a Micro-F1 score of only 0.273, possibly due to its difficulty effectively learning long-term dependencies in multi-label tasks.

[0093] In addition, to more intuitively compare the performance of different models, Figure 3 The comparison results of Micro-F1 score and Hamming Loss are shown in the form of a bar graph. Based on the above analysis, the present invention further explores the performance difference between GAT (graph attention network) and GCN model (graph convolutional network). Understanding this difference is of great significance for optimizing the "symptom-syndrome" interaction matrix. It is worth noting that the GAT model improves the Micro-F1 score by 12.5% ​​compared with GCN. The specific performance comparison is shown in Table 2 (ML-SCGAN_GAT in the table is the system model of the present invention). The detailed visualization results are shown in Figure 4 .

[0094] Table 2

[0095]

[0096] In the actual application process of the embodiment of the present invention, taking the symptoms of "chest pain and back pain, shortness of breath, and difficulty breathing" as an example:

[0097] Input in the data preprocessing module: The symptom text is encoded into a 768-dimensional vector by BERT to capture the contextual association between "chest pain" and "back pain".

[0098] Graph aggregation is performed in the graph attention network embedding module: GAT dynamically assigns high attention weights to syndrome factors such as "lung, qi deficiency".

[0099] In the multi-label classification training module, the classification output is: predicted disease location "lung, liver", disease nature "Qi deficiency, Yin deficiency", mapped to the syndrome "lung Qi and Yin deficiency".

[0100] An embodiment of the present invention further discloses a TCM-assisted diagnosis and treatment device, which includes a data acquisition device, the above-mentioned TCM multi-label syndrome classification system based on graph attention network, and a visual interface for outputting diagnosis suggestions.

[0101] In an embodiment of the present invention, the data acquisition device includes a sensor that supports capturing tongue images and extracts tongue coating texture features using a ResNet-50 model. The visualization interface is preferably used to generate a three-level mapping relationship table of "clinical information - disease location and nature - syndrome" and provide diagnostic recommendations.

[0102] The embodiment of the present invention further discloses a TCM multi-label syndrome classification method based on graph attention network, such as Figure 5 As shown, the method includes the following steps:

[0103] S1. Screen high-frequency syndrome tags from clinical diagnosis and treatment records and construct a standardized dataset containing disease-type syndrome tags and disease-location syndrome tags;

[0104] S2. Use the pre-trained BERT model to perform bidirectional contextual semantic modeling on clinical text and generate context-related clinical text embedding vectors;

[0105] S3, generate the initial embedding representation of the certificate element label through the pre-trained word vector model;

[0106] S4. Construct a clinical information-syndrome interaction graph, use a graph attention network to fuse clinical text embeddings and syndrome label embeddings, and generate a clinical information-syndrome embedding matrix;

[0107] S5. Input the embedding matrix into a multi-layer perceptron classifier, perform multi-label classification training in combination with a Focal Loss function, and output the prediction result of the clinical information-syndrome factor relationship.

[0108] The specific implementation content of the method embodiment of the present invention is the same as the specific implementation content of the above-mentioned system embodiment. The specific implementation content of the above-mentioned system embodiment is also used to explain the method embodiment of the present invention, and will not be repeated here.

[0109] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-label syndrome classification method for traditional Chinese medicine based on graph attention network, characterized by: The method comprises the following steps: S1. Screen high-frequency syndrome tags from clinical diagnosis and treatment records and construct a standardized dataset containing disease-type syndrome tags and disease-location syndrome tags; S2. Use the pre-trained BERT model to perform bidirectional contextual semantic modeling on clinical text and generate context-related clinical text embedding vectors; S3, generate the initial embedding representation of the certificate element label through the pre-trained word vector model; S4. Construct a clinical information-syndrome interaction graph, use a graph attention network to fuse clinical text embeddings and syndrome label embeddings, and generate a clinical information-syndrome embedding matrix; S5. Input the embedding matrix into a multi-layer perceptron classifier, perform multi-label classification training in combination with the FocalLoss loss function, and output the prediction result of the clinical information-syndrome factor relationship.

2. The method according to claim 1, wherein In step S1 , the screening is to filter out syndrome factor labels with an occurrence frequency of less than 2%, and retain 40 disease-related syndrome factors and 11 disease-location syndrome factors.

3. The method according to claim 1, wherein In step S2, the bidirectional contextual semantic modeling of clinical text using the pre-trained BERT model is specifically as follows: the global semantic information of the clinical text is extracted through the bidirectional Transformer architecture of the BERT model, and the contextual relationship between words is calculated using the self-attention mechanism.

4. The method according to claim 1, wherein In step S4, the graph attention network contains 8 attention heads, which updates the clinical text features by weighted aggregation of neighbor node information and uses the LeakyReLU activation function with a negative slope of 0.

01.

5. A multi-label syndrome classification system for traditional Chinese medicine based on graph attention network, characterized by: The system includes: The data preprocessing module is used to screen high-frequency syndrome factor labels from clinical diagnosis and treatment records and construct a standardized data set containing disease-type syndrome factors and disease-location syndrome factors; The clinical text embedding module is used to perform bidirectional contextual semantic modeling on clinical text using the pre-trained BERT model to generate context-related clinical text embedding vectors; The certificate element label embedding module is used to generate the initial embedding representation of the certificate element label through the pre-trained word vector model; The graph attention network embedding module is used to construct a clinical information-syndrome interaction graph and use the graph attention network to fuse clinical text embedding and syndrome label embedding to generate a clinical information-syndrome embedding matrix; The multi-label classification training module is used to input the embedding matrix into the multi-layer perceptron classifier, perform multi-label classification training in combination with the FocalLoss loss function, and output the prediction result of the clinical information-syndrome factor relationship.

6. The system according to claim 5, wherein: In the data preprocessing module, the screening is to filter out syndrome factor labels with an occurrence frequency of less than 2%, and retain 40 disease-related syndrome factors and 11 disease-site syndrome factors.

7. The system according to claim 5, wherein: In the clinical text embedding module, the bidirectional contextual semantic modeling of clinical text using the pre-trained BERT model is specifically as follows: the global semantic information of the clinical text is extracted through the bidirectional Transformer architecture of the BERT model, and the contextual relationship between words is calculated using the self-attention mechanism.

8. The system according to claim 5, wherein: In the graph attention network module, the graph attention network contains 8 attention heads, which updates the clinical text features by weighted aggregation of neighbor node information and uses the LeakyReLU activation function with a negative slope of 0.

01.

9. A TCM auxiliary diagnosis and treatment device, characterized in that: The device includes a data acquisition device, the system according to any one of claims 5 to 8, and a visual interface for outputting diagnostic suggestions.

10. The device according to claim 9, characterized in that The data acquisition device includes a sensor that supports collecting tongue images and extracts tongue coating texture features through a ResNet-50 model.

Images