An acupoint recommendation method based on multi-label text classification

Abstract

The application discloses a kind of based on multi-label text classification's matching point recommendation method, it is related to through information technology in the field of traditional Chinese medicine acupoint, including: S1, the preprocessing of traditional Chinese medicine symptom original text is carried out to convert traditional Chinese medicine symptom text into structured input new symptom text;S2, from acupoint label library, the acupoint prescription that is adapted to new symptom text description is obtained, and acupoint prescription is converted into multidimensional label vector by symptom coding layer;S3, construct including word embedding layer, attention layer, GNN feature interaction layer, multi-label prediction layer's matching point recommendation model, and based on new text symptom, multidimensional label vector is trained, verified;S4, the matching point recommendation model is based on the input traditional Chinese medicine symptom text and carries out matching point recommendation.The application fuses global symptom vector and acupoint node feature, utilizes adjacency matrix transmission synergistic effect, explicitly models the compatibility rule between acupoints, avoids isolated recommendation, and improves overall rationality of prescription.

Description

Technical Field

[0001] The present invention relates to the field of traditional Chinese medicine acupuncture points through information technology, and more specifically, to a method for recommending acupuncture points based on multi-label text classification. Background Art

[0002] Multi-label text classification (MLTC) is an important task in natural language processing (NLP). Its goal is to assign multiple labels or categories to each input text, rather than just a single label or category. Unlike traditional multi-classification tasks, multi-label prediction allows an instance to belong to multiple categories simultaneously. This is of great significance in many practical application scenarios, such as text classification, image annotation, and biomedical information processing. How to efficiently process this text data and assign it accurate labels has become a pressing issue. Multi-label text classification technology has been developed precisely to address this challenge.

[0003] Unlike traditional single-label text classification problems, multi-label text classification often faces additional challenges, such as high-dimensional sparsity, dependencies between labels, and missing datasets. Traditional multi-label classification approaches typically convert the problem into multiple binary classification problems or single-label multi-classification problems, or directly process multi-label data and refine algorithms to accommodate the characteristics of multi-label classification. These approaches are simple and easy to implement, but they ignore the correlation between labels.

[0004] With the continuous development of deep learning, methods based on deep neural networks have been proposed. These methods primarily improve classification performance by extracting enhanced text representations and modeling label relationships. Deep learning models, by learning semantic representations and contextual information from text, can better capture key information in the text, thereby improving classification performance. Deep learning models such as TextCNN, TextRNN, and BERT have achieved good results in processing text features and label relationships. However, these methods primarily rely on manually extracted features and then use machine learning algorithms for classification. With the advancement of deep learning technology, neural network-based methods have become a major research direction, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and the attention model BERT. Numerous deep learning models have been proposed for multi-label text classification, such as L-RCap, CRL4, and BR-CNN. These deep learning methods have greatly enriched the semantic representation of text and significantly improved the performance of multi-label annotation. However, these methods focus on extracting text representations and fail to consider the interactive relationships between labels and text. Furthermore, several specialized multi-label classification algorithms have also been proposed, such as those based on classification trees, binary relations, and matrix factorization. These algorithms are optimized for the characteristics of multi-label classification, improving the accuracy and efficiency of classification.

[0005] In acupuncture treatment of clinical diseases, multi-label classification of the relationship between symptoms and acupoints is a key issue in intelligent TCM research. The goal is to automatically identify multiple therapeutic acupoints corresponding to symptoms using artificial intelligence. However, traditional methods rely on expert experience, resulting in high subjectivity and inefficient knowledge transfer. Clinical practice has demonstrated that the mapping of symptoms to acupoints exhibits many-to-many relationships (e.g., "headache" can be mapped to six primary acupoints, such as Baihui and Fengchi) and context sensitivity (different concurrent symptoms lead to variations in acupoint selection). Traditional rule-based systems struggle to model these complex nonlinear relationships. In recent years, the integration of natural language processing (NLP) and graph neural networks (GNNs) has provided new insights in this area. Summary of the Invention

[0006] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0007] In order to achieve these objectives and other advantages of the present invention, a method for recommending acupuncture points based on multi-label text classification is provided, comprising:

[0008] S1. Preprocessing the original text of TCM symptoms to convert the TCM symptom text into a new symptom text of structured input;

[0009] S2. Obtaining acupoint prescriptions that match the new symptom text description from the acupoint label library, and converting the acupoint prescriptions into multidimensional label vectors through the symptom encoding layer;

[0010] S3. Construct an acupuncture point recommendation model that includes a word embedding layer, an attention layer, a GNN feature interaction layer, and a multi-label prediction layer, and train and verify it based on new text symptoms and multi-dimensional label vectors;

[0011] S4. The acupoint recommendation model recommends acupoints based on the input TCM symptom text.

[0012] Preferably, in S1, the preprocessing is to inject noise into the original text of the TCM symptoms using a random character deletion technique to obtain a new symptom text;

[0013] The new symptom text is segmented into subwords using the BERT tokenizer. The subwords are padded to a length of 128 tokens by adding a start marker [CLS] and a space marker [SEP] to generate a corresponding binary vector sequence as the input to the symptom encoding layer.

[0014] Among them, the noise factor in the random character deletion technique is set to 0.1.

[0015] Preferably, in S2, the symptom coding layer uses dynamic label coding to construct a multi-dimensional label vector for the input binary vector sequence, so as to map the acupoint label into a binary vector space to obtain a hidden state matrix H;

[0016] in, H =[ h 1, h 2,..., h T ]∈R T×768 , T is the sequence length.

[0017] Preferably, in S3, the word embedding layer uses each symptom in the new symptom text as an edge to connect the symptom grammatical structure in the acupoint prescription, and adjusts the edge weight according to the symptom to enhance the relevant acupoint compatibility;

[0018] Among them, the word embedding layer mines the topic of each symptom through Dirichlet distribution to obtain the corresponding edges.

[0019] Preferably, in S3, the attention layer performs a nonlinear transformation on the hidden state matrix H by learning the parameter matrix to obtain the corresponding score vector V, and V={ e 1, e 2,..., e T};

[0020] in, e i It is characterized by the following formula:

[0021] e i =tanh( W a × h i + b a )× v a

[0022] In the above formula, e i is the hidden state matrix H The attention score of the i-th position in , W a ∈R T×768 is the weight matrix, v a is the projection vector, b a is the bias term, and tanh() is the hyperbolic tangent activation function.

[0023] Preferably, in S3, the attention layer applies Softmax to generate probability distribution, and normalizes and weights the attention weights to generate a global semantic vector by the following formula: c :

[0024] .

[0025] Preferably, in S3, the GNN feature interaction layer constructs an acupoint relationship graph G, and G=(V,E), wherein V is a node of an acupoint label, and E is an edge constructed by the prior co-occurrence symptom information of the label. If there is a synergistic relationship between two labels, the corresponding nodes are connected by an edge;

[0026] Let A be the adjacency matrix of the acupoint relationship graph, Represents the initial node feature matrix, maps the global symptom semantics to the acupoint graph space, and uses the graph neural network to analyze the node features c proj Update is performed, and the update formula is as follows:

[0027] c proj = W p × c + b p , ( W p ∈R 256×768 , bp ∈R 256 )

[0028] In the above formula, c proj represents the global symptom semantic features after projection, W p represents the projection weight matrix, b p represents the bias vector;

[0029] The GNN feature interaction layer performs feature transformation and dimensionality reduction on the updated nodes through the following formula to obtain the output features of the GNN feature interaction layer: :

[0030]

[0031] In the above formula, Represents the added self-loop edge, is the degree matrix, Indicates that c proj Combined with the characteristics of each acupoint label, W g are learnable weights, σ for ReLU Activation function.

[0032] Preferably, the multi-label prediction layer performs binary classification for each acupoint node to output an unnormalized logit vector logits i :

[0033]

[0034] The multi-label prediction layer independently calculates the activation probability of each acupoint by the following formula: p i :

[0035]

[0036] In the above formula, 、 Respectively represent i The weight matrix and bias term of each acupoint, For the i GNN output features of acupoints.

[0037] Preferably, in S3, the verification method of the acupoint recommendation model includes:

[0038] The following formula is used to compare the prescription P predicted by the acupoint recommendation model with the standard prescription S kThe similarity S between im ( P,S k ) to calculate:

[0039]

[0040] In the above formula, P 主 To predict the main acupoints in prescription P, P 配 To predict the acupoints in prescription P, S k,主 Standard prescription S k The main point, S k,配 Standard prescription S k The paired acupoints;

[0041] like Sim ( P,S k )≥0.75, it is considered that the predicted prescription P is consistent with the standard prescription S k High consistency;

[0042] If 0.5≤ Sim ( P,S k )<0.75, it is considered that the predicted prescription P is consistent with the standard prescription S k Partially consistent, otherwise it is judged as inconsistent and the acupoint recommendation model continues to be trained.

[0043] The present invention has at least the following beneficial effects:

[0044] 1. The global symptom vector is integrated with the acupoint node features, and the synergistic effect is transferred using the adjacency matrix. The compatibility rules between acupoints are explicitly modeled to avoid isolated recommendations, improve the overall rationality of prescriptions, and comply with the TCM compatibility principle of "monarch, minister, assistant and envoy". This provides clinical assistance, provides standardized acupoint recommendations for primary care physicians, and reduces reliance on experience.

[0045] 2. The preprocessing stage uses random character deletion to simulate real-world text irregularities. The symptom encoding layer maps acupoint labels into adaptive binary vectors through dynamic label encoding. This provides strong tolerance to spelling errors and discrepancies in the input text, enhancing the model's robustness and generalization. The invented framework can be applied to scenarios such as Traditional Chinese Medicine prescription recommendations and disease diagnosis.

[0046] 3. Through the unified learning of symptom semantics and acupoint coordination through GNN and attention mechanism, the adaptability of each acupoint is independently evaluated using multi-label binary classification, taking into account the overall compatibility. The recommendation results have both individual adaptability and overall rationality of the prescription, which is close to the expert level.

[0047] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flow chart of acupoint recommendation by the acupoint recommendation model of the present invention;

[0049] Figure 2 This is a flowchart for recommending acupoints based on the TCM text symptoms of "abdominal distension with loss of appetite" and "epigastric pain" in one embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0051] The present invention proposes an acupuncture point recommendation method based on multi-label text classification. The acupuncture point recommendation model in this method mainly consists of a word embedding layer, an attention layer, a GNN feature interaction layer, and a multi-label prediction layer. The processing flow of the acupuncture point recommendation model is as follows: Figure 1 As shown, it mainly includes:

[0052] 1. Dataset Construction. Using 68 classic acupuncture texts, including the Huangdi Neijing and the Zhenjiu Jiayijing, as the primary data source, we established a dataset on acupuncture treatment for functional gastrointestinal disorders (FGIDs). We identified keywords that define and characterize FGIDs, including "hiccups," "reflux," "abdominal pain," "abdominal distension," "vomiting," "loss of appetite," "stomach bloating," and "stomach pain." Based on these criteria, we collected 1,086 relevant data points. Information on symptoms and acupoints was extracted and classified using a two-person double-entry process. Data including book titles, publication years, all nomenclature, clinical manifestations, acupuncture prescriptions, and categorized acupuncture prescriptions and methods is available upon request. This double-review process ensured the rigor and reliability of our data collection.

[0053] 2. Data preprocessing: Based on data related to acupuncture treatment for functional gastrointestinal disorders, including symptom descriptions and corresponding acupoint information, the symptom descriptions were used as input text, and the acupoint information was output as multi-labels. The original symptom-acupoint data was injected with noise (noise factor 0.1) using Random Token Deletion (RTD) to generate new symptom text. Next, a multidimensional label vector was constructed using a dynamic label encoding (MultiLabelBinarizer), mapping the acupoint labels to a binary vector space. The symptom text was segmented into subwords using the BERT tokenizer (bert-base-uncased), padded to a uniform length of 128 tokens, and generated to concatenate with the original text sequence.

[0054] 2. Constructing a word embedding layer: To capture the sequence, structure, and semantic information of symptom text, it is necessary to model irregular text data. Using each symptom in the text as an edge, we connect the symptom grammatical structure in the acupoint prescriptions. Edge weights are adjusted based on the symptom, enhancing relevant acupoint combinations and constructing a word embedding layer.

[0055] Specifically, in order to further capture its semantic information, Dirichlet distribution is used to mine the topic of each symptom and use it as an edge. (N is the total number of acupoints) is the acupoint label set, and the initial adjacency matrix A is defined, where A ij represents the adjacency matrix A i Rank j Elements of the column:

[0056] (1)

[0057] 3. Use BERT's tokenizer, BertTokenizer, to tokenize the symptom description text and convert it into corresponding token IDs. BertTokenizer is the tokenizer used in the BERT model for text processing. It is primarily responsible for breaking the input text into subword units (tokens) that the model can process and performing normalization. Specifically, it uses a bidirectional Transformer encoder to generate context-aware token embeddings to better understand the words in the symptoms. Special tokens [CLS] and [SEP] are added, and corresponding attention_masks are generated. The original text sequence is concatenated with the symptoms by constructing auxiliary sentence pairs and input into the BERT model. The input method is as follows:

[0058] Input=[CLS]+ S +[SEP]+ a i +[SEP] (2)

[0059] Among them, [CLS] is a special marker representing the beginning of the symptom in the BERT model, [SEP] is a special marker used to separate different symptoms, and S is the core symptom text. a i It is the text content of the second sequence that matches the symptom description S.

[0060] 4. In the multi-label classification task of Traditional Chinese Medicine text, the attention layer uses dynamic weight allocation to help the model focus on key terms in symptom descriptions (such as "abdominal distension" and "dull pain") and establish associations between these terms and specific acupoints (such as Zusanli and Zhongwan). The association system is established as follows:

[0061] Suppose the symptom description is encoded by the BERT model to obtain the hidden state matrix H , H =[ h 1, h 2,..., h T ]∈R T×768 , where each token corresponds to a 768-dimensional vector, T is the sequence length.

[0062] Nonlinear transformation via learnable parameter matrix:

[0063] e i =tanh( W a × h i + b a )× v a (3)

[0064] in, W a ∈R T×768 is the weight matrix, v a is the projection vector, b a As the bias term, we get the score vector V={ e 1, e 2,..., e T By simulating the clinical acupoint selection process of acupuncture, nonlinear transformation is used to extract the dialectical value of symptom terms.

[0065] Apply Softmax to generate probability distribution, normalize attention weights, and weighted aggregate key information to synthesize context vectors based on symptoms. c ,and :

[0066] (4)

[0067] 5. In the TCM acupuncture knowledge system, acupoints have complex compatibility relationships and meridian connections. Define an acupoint relationship graph G, where G = (V, E), where V represents the nodes corresponding to the acupoint labels, and E represents the edges constructed from the labels' prior co-occurrence symptom information. If two labels have a synergistic relationship, the corresponding nodes are connected by an edge. Initialize the node features as 256-dimensional random embedding vectors.

[0068] Let A be the adjacency matrix of the acupoint relationship graph, H （0） Represent the initial node feature matrix, map the global symptom semantics to the acupoint graph space, and use graph neural network to update the node features.

[0069] c proj = W p × c + b p ( W p ∈R 256×768 , b p ∈R 256 ) (5)

[0070] In the above formula, c proj represents the global symptom semantic features after projection, W p represents the projection weight matrix, b p represents the bias vector;

[0071] The acupoint relationship graph G is constructed through the GNN feature interaction layer, encoding this prior symptom knowledge and enhancing the rationality of feature representation. The 768-dimensional global semantic vector (key symptom information) is integrated with the acupoint graph structure, and a graph neural network (GNN) is used for feature transformation and dimensionality reduction.

[0072] (6)

[0073] In the above formula, Represents the added self-loop edge, is the degree matrix, Indicates that c proj Combined with the characteristics of each acupoint label, W g are learnable weights, σ for ReLUActivation function.

[0074] 6. Mapping the 256-dimensional acupoint features to N acupoint label spaces supports multi-label independent prediction, which is in line with the clinical prescription principle of acupuncture "main acupoint + auxiliary acupoint". If the output feature of GNN is , is the acupoint label set, performs binary classification for each acupoint node, and outputs the unnormalized logit vector logits i , the activation probability of each acupoint is calculated independently through the sigmoid function p i , the sum of probabilities is not forced to be 1, and multiple acupoints can be activated simultaneously.

[0075] (7)

[0076] (8)

[0077] In the above formula, 、 Respectively represent i The weight matrix and bias term of each acupoint, For the i GNN output features of acupoints.

[0078] After obtaining the acupoint activation probability p i , generate personalized prescriptions that comply with the "main acupoints + auxiliary acupoints" principle, and realize end-to-end mapping from symptom semantics to personalized acupuncture prescriptions.

[0079] 7. Further structure the original prescription in step 1 to form a "standard prescription library". Compare the predicted prescription P = {main acupoint} ∪ {auxiliary acupoints} generated in step 6 with the standard prescriptions in the library. S k Calculate weighted Jaccard similarity Sim ( P,S k ).

[0080]

[0081] In the above formula, It means that the recommended main acupoints are exactly the same as the main acupoints in the ancient standard prescriptions. Indicates that the recommended acupoints are exactly the same as those in the standard prescriptions in ancient books. Indicates all the non-repeated main acupoints that appear in the model and the ancient standard prescriptions. It represents all non-repetitive acupoint combinations that appear in the model and standard prescriptions in ancient books.

[0082] like Sim ( P,S k ) ≥ 0.75 is considered "highly consistent"; 0.5 ≤ Sim ( P,S k ) < 0.75 is "partially consistent"; otherwise "inconsistent".

[0083] For the "inconsistent" prescriptions, locate the most similar standard prescription, highlight the different acupoints, and pop up a list of differences and the original ancient medical texts for the acupuncture experts to make a secondary judgment.

[0084] Example:

[0085] Taking "abdominal distension accompanied by loss of appetite" and "epigastric pain" as the input of traditional Chinese medicine text symptoms, the acupoint recommendation process is as Figure 2 shown, including the following steps:

[0086] 1. Data preprocessing

[0087] Input the traditional Chinese medicine symptom texts: "abdominal distension accompanied by loss of appetite", "epigastric pain", noise injection (random character deletion, noise factor 0.1), delete "accompanied by" → generate new text symptoms: "abdominal distension", "loss of appetite", "epigastric pain". Using the true prescription in the multi-label encoding (assuming the acupoint label library: [Zusanli, Zhongwan, Neiguan, Pishu, Tianshu]) as "Zusanli, Zhongwan", its binary vector is [1, 1, 0, 0, 0]. Taking each symptom in the text as an edge, connect the symptom syntax structures in the acupoint prescription, and adjust the edge weights according to the symptoms to enhance the relevant acupoint compatibility.

[0088] Use the BERT tokenizer BertTokenizer to tokenize the symptom description text, add special tokens [CLS] and [SEP], the tokenization result: ["[CLS]", "abdominal", "distension", "loss of", "appetite", "[SEP]"], pad to 128 tokens (only 8 actual valid tokens), generate attention_mask = [1, 1, 1, 1, 1, 1, 1, 1, 0,..., 0].

[0089] 2. Word embedding and attention layer

[0090] BERT encoding: input the binary vector sequence of symptoms "abdominal distension", "loss of appetite", "epigastric pain" → the BERT output hidden state matrix H ∈ R^{128×768}. Calculate the attention weights: calculate the importance score of each token through formula (3) , Such as: e_abdominal distension = 0.9, e_loss of appetite = 0.8, e_epigastric pain = 1.5 (the core symptom terms have higher scores).

[0091] Apply Softmax to generate probability distribution, normalize Softmax and perform weighted aggregation (Formula 4) to generate the global semantic vector c∈R^{768}:

[0092] c=0.15*h_abdominal+0.14*h_bloating+0.35*h_loss of appetite+...

[0093] 3. GNN feature interaction layer

[0094] The adjacency matrix A encodes the meridian association and compatibility experience (e.g., Zusanli + Zhongwan are classic acupoint combinations for gastrointestinal diseases).

[0095] Symptom feature fusion and update: The projected global symptom semantic features c proj Projection to acupoint space (Formula 5):

[0096] c proj = W p c + b p ( c proj ∈R^{256})

[0097] Splicing to each acupuncture point node feature:

[0098] H^{(0)}=[random vector_{Zusanli}⊕ c proj ; Random vector _{Zhongwan}⊕ c proj ...]

[0099] Graph convolution update (Formula 6):

[0100]

[0101] The feature interaction between Zusanli and Zhongwan nodes is significantly enhanced due to the strong edge connection (weight 0.8).

[0102] 4. Multi-label prediction layer

[0103] Calculation of independent acupoint probability:

[0104] Perform binary classification on the acupoint features output by GNN:

[0105] logits_{Zusanli}=W_{fc}^{(1)}·h_{Zusanli}^{(1)}+b_{fc}^{(1)}

[0106] Sigmoid output p_{Zusanli}=0.92

[0107] logits_{Zhongwan}=W_{fc}^{(2)}·h_{Zhongwan}^{(1)}+b_{fc}^{(2)}

[0108] p_{Zhongwan}=0.87

[0109] Probability of other acupoints: p_{Neiguan}=0.21, p_{Pishu}=0.09, p_{Tianshu}=0.33

[0110] Prediction results:

[0111] The threshold is set to 0.5, and the recommended acupoints are: [Zusanli, Zhongwan], with a binary vector of [1,1,0,0,0], which is consistent with the true label.

[0112] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacements and / or modifications can be made according to user needs.

[0113] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and exemplary embodiments. They can be applied to a variety of fields suitable for the present invention. Further modifications will be readily apparent to those skilled in the art. Therefore, the present invention is not limited to the specific details and illustrations shown and described herein without departing from the general concept defined by the claims and their equivalents.

Claims

1. A method for recommending acupuncture points based on multi-label text classification, characterized in that: include: S1. Preprocessing the original text of TCM symptoms to convert the TCM symptom text into a new symptom text of structured input; S2. Obtaining acupoint prescriptions that match the new symptom text description from the acupoint label library, and converting the acupoint prescriptions into multidimensional label vectors through the symptom encoding layer; S3. Construct an acupuncture point recommendation model that includes a word embedding layer, an attention layer, a GNN feature interaction layer, and a multi-label prediction layer, and train and verify it based on new symptom text and multi-dimensional label vectors; S4, the acupoint recommendation model recommends acupoints based on the input TCM symptom text; In S3, the word embedding layer uses each symptom in the new symptom text as an edge to connect the symptom grammatical structure in the acupoint prescription and adjusts the edge weight according to the symptom to enhance the relevant acupoint compatibility; Among them, the word embedding layer mines the topic of each symptom through Dirichlet distribution to obtain the corresponding edge; In S3, assume that the symptom description is encoded by the BERT model to obtain the hidden state matrix H , H =[ h 1, h 2,..., h T ]∈R T×768 , the attention layer performs a nonlinear transformation on the hidden state matrix H by learning the parameter matrix to obtain the corresponding score vector V, and V={ e 1, e 2,..., e T }; in, e i It is characterized by the following formula: e i =tanh( W a × h i + b a )× v a In the above formula, e i is the first i The attention score of each position, W a ∈R T×768 is the weight matrix, v a is the projection vector, b a is the bias term, tanh() is the hyperbolic tangent activation function; In S3, the attention layer applies Softmax to generate probability distribution and normalizes the attention weights and performs weighted aggregation to generate a global semantic vector by the following formula: c : ； In S3, the GNN feature interaction layer constructs an acupoint relationship graph G, where G = (V, E), where V is the node of the acupoint label and E is the edge constructed by the prior co-occurrence symptom information of the label. If there is a synergistic relationship between two labels, the corresponding nodes are connected by an edge. Let A be the adjacency matrix of the acupoint relationship graph, Represents the initial node feature matrix, maps the global symptom semantics to the acupoint graph space, and uses the graph neural network to analyze the node features c proj Update is performed, and the update formula is as follows: c proj = W p × c + b p ,( W p ∈R 256×768 , b p ∈R 256 ) In the above formula, c proj represents the global symptom semantic features after projection, W p represents the projection weight matrix, b p represents the bias vector; The GNN feature interaction layer performs feature transformation and dimensionality reduction on the updated nodes through the following formula to obtain the output features of the GNN feature interaction layer: : In the above formula, Represents the added self-loop edge, is the degree matrix, Indicates that c proj Combined with the characteristics of each acupoint label, W g are learnable weights, σ for ReLU Activation function; The multi-label prediction layer performs binary classification for each acupoint node to output the unnormalized logit vector logits i : The multi-label prediction layer independently calculates the activation probability of each acupoint by the following formula: p i : In the above formula, 、 Respectively represent i The weight matrix and bias term of each acupoint, For the i GNN output features of acupoints.

2. The acupoint recommendation method based on multi-label text classification according to claim 1, characterized in that: In S1, the preprocessing is to inject noise into the original text of TCM symptoms using random character deletion technology to obtain a new symptom text; The new symptom text is segmented into subwords using the BERT tokenizer. The subwords are padded to a length of 128 tokens by adding a start marker [CLS] and a space marker [SEP] to generate a corresponding binary vector sequence as the input to the symptom encoding layer. Among them, the noise factor in the random character deletion technique is set to 0.

1.

3. The acupuncture point recommendation method based on multi-label text classification according to claim 2, characterized in that: In S2, the symptom coding layer uses dynamic label coding to construct a multi-dimensional label vector for the input binary vector sequence, so as to map the acupoint label into a binary vector space to obtain a hidden state matrix H; in, H =[ h 1, h 2,..., h T ]∈R T×768 , T is the sequence length.

4. The acupuncture point recommendation method based on multi-label text classification according to claim 1, characterized in that: In S3, the verification methods for the acupoint recommendation model include: The prescription is predicted by the acupoint recommendation model through the following formula P With standard prescription S k The similarity between Sim ( P,S k ) to calculate: In the above formula, P 主 To predict the main acupoints in prescription P, P 配 To predict the acupoints in prescription P, S k,主 Standard prescription S k The main point, S k,配 Standard prescription S k The paired acupoints; like Sim ( P,S k )≥0.75, it is considered that the predicted prescription P is consistent with the standard prescription S k High consistency; If 0.5≤ Sim ( P,S k )<0.75, it is considered that the predicted prescription P is consistent with the standard prescription S k Partially consistent, otherwise it is judged as inconsistent and the acupoint recommendation model continues to be trained.

Images