A knowledge-driven intelligent moxibustion scheme generation method and system

Abstract

This invention discloses a knowledge-driven intelligent moxibustion scheme generation method and system, belonging to the interdisciplinary field of traditional Chinese medicine information engineering and artificial intelligence. First, this invention collects Zhou Meisheng's moxibustion monographs and medical records to construct a domain-specific knowledge graph containing moxibustion sensation features, completing the digital representation of the academic thoughts of this renowned traditional Chinese medicine practitioner. Second, it constructs a graph neural network reasoning model incorporating a multi-head attention mechanism to simulate Professor Zhou Meisheng's diagnostic and treatment thinking in moxibustion. Finally, it generates personalized moxibustion schemes, providing interpretability through knowledge graph path backtracking and achieving closed-loop dynamic control of the scheme based on real-time moxibustion sensation feedback. This invention solves the problems of difficulty in inheriting Zhou Meisheng's moxibustion experience and the lack of interpretability and dynamic control capabilities in intelligent moxibustion decision-making. It can be directly embedded into intelligent moxibustion devices and primary healthcare systems, promoting the standardization and large-scale application of high-quality traditional Chinese medicine moxibustion resources.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of traditional Chinese medicine information engineering and artificial intelligence decision - making, and particularly relates to a method and system for automatically generating, interpretably reasoning, and dynamically and closed - loop regulating clinical treatment plans for traditional Chinese medicine moxibustion based on natural language processing, knowledge graph, and graph neural network technologies, integrating the academic thoughts of the famous veteran traditional Chinese medicine doctor Zhou Meisheng's moxibustion method. Background Technique

[0002] Professor Zhou Meisheng is a famous acupuncture and moxibustion expert in China. His monograph "The Rope of Moxibustion" put forward the core theories of the three phases of moxibustion sensation (directional conduction, accumulation at the action point, and meridian - following sensation transmission) and the importance of moxibustion for heat syndromes, breaking through the taboos of traditional moxibustion methods and laying an important foundation for modern moxibustion clinical practice. At present, the core experience of Zhou Meisheng's moxibustion method is retained in unstructured forms such as paper works, handwritten medical records, and master - apprentice oral teachings, presenting problems in digital inheritance: First, the spatio - temporal logic and syndrome - differentiation thinking of "qi reaching the diseased site" in "The Rope of Moxibustion" cannot be analyzed by traditional digital means such as keyword retrieval. Second, the efficiency of experience inheritance is low and it is easy to be distorted, making it difficult to be reused on a large scale in primary medical care and intelligent devices.

[0003] Existing moxibustion - assisted diagnosis and treatment systems and intelligent devices have obvious technical defects: First, they rely on general acupoint databases and do not construct a domain ontology for Zhou Meisheng's moxibustion method, unable to handle complex diagnosis and treatment logics such as "treating the same disease with different methods and treating different diseases with the same method". Second, the AI model only recommends acupoints based on statistical probability, lacking causal reasoning ability and having no clinical interpretable basis, making it difficult to gain the trust of doctors. Third, there is no real - time moxibustion sensation feedback regulation mechanism, unable to dynamically optimize the moxibustion implementation plan. Summary of the Invention

[0004] The present invention aims to solve the technical problems existing in the existing moxibustion intelligent decision - making system, namely, the difficulty in digitizing the experience of famous veteran traditional Chinese medicine doctors, the lack of interpretability in the reasoning process, and the inability to dynamically adjust the plan according to real - time moxibustion sensation feedback. Specifically, it is to solve how to transform the implicit thinking of Professor Zhou Meisheng's "three phases of moxibustion sensation" into a computable model by a computer, and how to construct a closed - loop moxibustion decision - making system that not only has high accuracy but also has the ability of theoretical traceability.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A knowledge - driven intelligent moxibustion implementation plan generation method integrating Zhou Meisheng's thoughts, comprising the following steps:

[0007] S1 constructs a knowledge graph for Zhou Meisheng's moxibustion method: collect Professor Zhou Meisheng's "Moxibustion Rope" and "Zhou Meisheng's Clinical Essentials" as well as clinical medical records, and establish a domain ontology model containing entities with moxibustion sensation characteristics; use the BERT-BiLSTM-CRF model for named entity recognition, and use the CasRel layered pointer labeling network to extract relational triples between entities to solve the problem of overlapping entities in medical records, and store the cleaned triples data in the graph database;

[0008] S2 constructs and trains an intelligent reasoning model based on graph neural networks: it uses the GraphSAGE algorithm to vectorize and embed heterogeneous nodes in the knowledge graph, transforming discrete TCM concepts into continuous vector space representations; it constructs a graph attention network GAT that incorporates a multi-head attention mechanism, taking the patient symptom entity set as input, triggering multi-hop message passing in the graph and dynamically calculating the importance weights of nodes; during training, it uses FocalLoss as the loss function to solve the problem of imbalance between positive and negative samples in acupoint recommendations.

[0009] S3 generates personalized moxibustion plans and achieves closed-loop dynamic control: Based on the probability distribution score list output by the model, it generates the optimal main acupoints, acupoint combinations, and corresponding moxibustion types and parameters; using graph path backtracking technology, it retrieves the knowledge path of syndromes and recommended acupoints, generating interpretable evidence including original citations from "Moxibustion Rope"; during the moxibustion process, it receives real-time feedback from patients on their moxibustion sensations, and automatically triggers parameter correction commands when the feedback is unsatisfactory, thus achieving closed-loop optimization of the moxibustion plan.

[0010] As a further technical solution of the present invention: in step S1, the entity types of the domain ontology model include at least five categories: disease, syndrome, acupoint, moxibustion sensation, and meridian; the relationship types between entities include at least six categories: meridian tropism, inclusion, recommendation, triggering, treatment method association, and feedback regulation; the graph database adopts Neo4j graph database.

[0011] As a further technical solution of the present invention: In step S1, data acquisition and preprocessing specifically involves: digitizing paper documents using a 600DPI industrial scanner, extracting text using an OCR engine based on the ResNet-50 backbone network, setting a confidence threshold for manual proofreading of low-confidence characters, and cleaning the text using the Jieba word segmentation tool, which integrates more than 4,500 TCM terminology dictionaries, to obtain a standardized token sequence.

[0012] As a further technical solution of the present invention: In step S2, the GraphSAGE algorithm transforms discrete TCM concepts into 128-dimensional dense vectors; the graph attention network GAT uses 8 parallel attention heads to capture the features of different semantic subspaces of cold and heat, exterior and interior, and deficiency and excess, simulating Professor Zhou Meisheng's diagnostic and treatment thinking of "differentiating syndromes and applying moxibustion with clear distinction between primary and secondary aspects".

[0013] As a further technical solution of the present invention: In step S3, the BeamSearch algorithm is used to retrieve the optimal knowledge path. The path score function is the product of the weights of all edges on the path. The optimal path is converted into a natural language explanation and associated with the original text index of "Jiusheng".

[0014] As a further technical solution of the present invention: In step S3, a Markov decision process is constructed based on the Q-Learning reinforcement learning algorithm to process real-time moxibustion sensation feedback. The state space is the moxibustion temperature, duration and moxibustion sensation category, and the action space is the adjustment of moxibustion parameters and the change of technique. A reward function is designed based on Zhou Meisheng's theory of "effectiveness comes from qi" to learn the optimal control strategy.

[0015] A knowledge-driven intelligent moxibustion scheme generation system integrating Zhou Meisheng's ideas, used to execute the above method, the system comprising the following components connected in sequence via communication:

[0016] The data acquisition and preprocessing module is used for the acquisition, digital processing, and standardized corpus construction of multi-source data on Zhou Meisheng's moxibustion method.

[0017] The knowledge graph construction module is used for the construction and management of knowledge graphs in the fields of TCM entity recognition, relation extraction, and Zhou Meisheng moxibustion.

[0018] The intelligent reasoning module is used for node vectorization embedding, intelligent reasoning of moxibustion treatment schemes, and model training optimization.

[0019] The scheme generation and interpretability module is used for outputting personalized moxibustion schemes and generating diagnostic and treatment evidence;

[0020] The closed-loop feedback control module is used for real-time collection of moxibustion sensations and dynamic optimization of moxibustion treatment plans.

[0021] As a further technical solution of the present invention: the knowledge graph construction module has a built-in BERT-BiLSTM-CRF named entity recognition model and a CasRel relation extraction model; the intelligent reasoning module has a built-in GraphSAGE node embedding unit, a GAT reasoning unit incorporating a multi-head attention mechanism and a FocalLoss training optimization unit.

[0022] As a further technical solution of the present invention: the closed-loop feedback control module has a built-in moxibustion sensation interaction acquisition unit, a reinforcement learning decision unit based on Q-Learning algorithm, and a moxibustion parameter correction execution unit.

[0023] This technology proposes a knowledge-driven intelligent moxibustion plan generation method and system, which has the following advantages and beneficial effects:

[0024] 1. Digital inheritance: For the first time, Zhou Meisheng's implicit diagnostic and treatment thinking of "three phases of moxibustion sensation" and "moxibustion is valuable for heat syndromes" has been transformed into a machine-computable model, realizing the digital immortality of the moxibustion thought of the famous old Chinese medicine doctor;

[0025] 2. Precise and Explainable: Based on a dedicated knowledge graph and graph neural network reasoning, the solution is more accurate than traditional rule bases and can be traced back to the original text of "Jiu Sheng" to improve clinical credibility and compliance;

[0026] 3. Dynamic closed loop: Introducing real-time moxibustion sensation feedback and reinforcement learning regulation to achieve a leap from static schemes to dynamic adaptive optimization;

[0027] 4. Large-scale application: It can be adapted to moxibustion robots and primary care systems, promoting the standardization, intelligentization and large-scale implementation of Zhou Meisheng's moxibustion method. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall process for generating intelligent moxibustion plans.

[0029] Figure 2 This is a schematic diagram of the ontological structure of Zhou Meisheng's moxibustion method.

[0030] Figure 3 This is a diagram of the architecture of a multi-head attention graph neural network inference model.

[0031] Figure 4 This is a schematic diagram illustrating the module composition and interaction logic of an intelligent moxibustion system. Detailed Implementation

[0032] The present invention will be further described below with reference to the embodiments. It should be noted that these are merely examples and descriptions of the inventive concept. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in the claims, they should all be considered to fall within the protection scope of the present invention.

[0033] like Figures 1-4 As shown, this invention proposes a knowledge-driven intelligent moxibustion scheme generation method that integrates Zhou Meisheng's ideas;

[0034] First, a knowledge graph of Zhou Meisheng's moxibustion method was constructed. By collecting data from Professor Zhou Meisheng's *Moxibustion Rope*, *Zhou Meisheng's Clinical Essentials*, and clinical case records, a domain ontology model containing entities with "moxibustion sensation characteristics" was established. Named entity recognition was performed using the BERT-BiLSTM-CRF model, and a CasRel layered pointer annotation network was employed to extract relational triples between entities, resolving the entity overlap problem in the case records. Finally, the cleaned data was stored in a graph database.

[0035] Secondly, the present invention constructs and trains an intelligent reasoning model based on a graph neural network. The GraphSAGE algorithm is used to vectorize and embed heterogeneous nodes in the knowledge graph, converting discrete traditional Chinese medicine concepts into a continuous vector space. On this basis, a graph attention network (GAT) incorporating a multi-head attention mechanism (Multi-headAttention) is designed to simulate the diagnostic and treatment thinking of Professor Zhou Meisheng's "differentiating syndromes for moxibustion and distinguishing primary and secondary aspects". The model receives the set of symptom entities of the patient as input, triggers multi-hop message passing on the graph, and dynamically calculates the importance weights of different neighbor nodes for the central node. During the training process, FocalLoss is used as the loss function to solve the problem of imbalance between positive and negative samples, making the model pay more attention to samples that are difficult to classify.

[0036] Finally, the present invention generates a personalized moxibustion plan and conducts dynamic regulation. The system intercepts the optimal combination of main acupoints and supplementary acupoints according to the probability distribution score list output by the model, and outputs the prediction vectors for the moxibustion type and moxibustion parameters in parallel. At the same time, using the graph path backtracking technology, the knowledge path connecting "patient syndromes" and "recommended acupoints" is retrieved, and an explanatory basis including the original text citation of "Moxibustion Rope" is generated. During the moxibustion process, the system receives the real-time moxibustion sensation feedback of the patient (such as the arrival or non-arrival of qi) through the interaction interface. If the feedback is not ideal, the system automatically triggers a parameter correction instruction to achieve the closed-loop optimization of the plan.

[0037] Example 1: Detailed implementation of the construction of the knowledge graph in the field of Zhou Meisheng's moxibustion method (S1);

[0038] This step aims to transform unstructured medical literature into a machine-readable structured graph. The specific implementation process is divided into the following three sub-steps;

[0039] 1. Data collection and preprocessing;

[0040] First, a dedicated corpus for Zhou Meisheng's moxibustion method is established. The data sources include "Moxibustion Rope", "Compilation of Zhou Meisheng's Clinical Practice" and more than 500 handwritten / printed clinical medical records. For paper-based literature, a high-resolution industrial scanner (resolution set to 600 DPI) is used for digitization, and an OCR (optical character recognition) engine based on the ResNet-50 backbone network is used for text extraction. In order to reduce the recognition error rate, a confidence threshold is set , for character fragments with a confidence lower than , the system automatically marks and introduces a manual proofreading link. In the preprocessing stage, the Jieba word segmentation tool combined with a special dictionary in the field of traditional Chinese medicine (including 4,500 proper nouns such as meridians, acupoints, and pulses) is used to clean the text, remove stop words (such as "de", "le"), and obtain a standardized Token sequence .

[0041] 2. Named Entity Recognition (NER) Model Construction;

[0042] To accurately identify special entities such as "moxibustion sensation features", this embodiment adopts a joint architecture of BERT-BiLSTM-CRF.

[0043] (1) Embedding layer: Input the token sequence into the pre-trained BERT-Base-Chinese model to obtain character vector representations containing contextual semantics. .

[0044] (2) Coding layer: Input a bidirectional long short-term memory (BiLSTM) network to capture long-range dependency features. For time... Forward hidden state and backward hidden state Complete features are obtained by splicing. .

[0045] (3) Decoding layer: The label sequence is constrained using a Conditional Random Field (CRF) (e.g., B-LOC cannot be directly followed by I-PER). The objective function of the CRF layer is to maximize the target label sequence. conditional probability

[0046] in, The scoring function is defined as the sum of the emission score (output by the BiLSTM) and the transition score (determined by the CRF transition matrix):

[0047]

[0048] In the formula, Indicates from the label Transferred to The probability, Indicates the first The character was predicted as a label. The probability of.

[0049] 3. Relation Extraction (RE) and Triple Construction

[0050] To address the common issue of entity overlap in medical records (such as one acupoint corresponding to multiple therapeutic effects), CasRel (a cascaded pointer annotation network) is employed. This model first identifies the head entity (Subject), and then, for each identified Subject, predicts its corresponding Object and Relation. Its loss function... It consists of two parts:

[0051]

[0052] in, Binary cross-entropy loss for head entity labeling, Given a header entity, this represents the loss of object and relation tags.

[0053] Ultimately, the system extracts forms such as The triples (e.g., <Zusanli, trigger, meridian sensation>) are generated and all triples are imported into the Neo4j graph database to complete the graph construction.

[0054] Example 2: Detailed implementation of the intelligent reasoning model (S2) based on graph neural networks;

[0055] This step is the core invention point, which uses deep learning to simulate the diagnostic thinking of traditional Chinese medicine, and consists of the following three steps;

[0056] 1. Heterogeneous graph definition and node embedding;

[0057] Define a heterogeneous graph G = (V, E), where V contains nodes of four categories: disease, syndrome, acupoint, and moxibustion sensation. First, initialize the node embeddings using the GraphSAGE algorithm. For any node V, its K-th layer embedding vector... The calculation formula is:

[0058]

[0059]

[0060] in, Represents a node The neighborhood group, It is the mean aggregation function. The weight matrix is ​​a learnable matrix. The ReLU is a nonlinear activation function. This step transforms the discrete TCM concepts into a dense 128-dimensional vector.

[0061] 2. Inference mechanism based on multi-head attention (GAT);

[0062] To achieve "differentiated moxibustion" (i.e., dynamically adjusting the weight of acupoints according to the severity of the condition), a graph attention network (GAT) is introduced.

[0063] (1) Attention coefficient calculation: Calculate the correlation coefficient between node $i$ and its neighbor $j$. :

[0064]

[0065] in For attention vectors, This indicates vector concatenation.

[0066] (2) Normalized weights: The coefficients are normalized using the Softmax function to obtain the final attention weights. :

[0067] Physical meaning: If the input node For "fever", neighboring nodes If it is a "heat-clearing acupoint", then It will automatically learn larger values.

[0068] (3) Multi-head aggregation: K=8 attention heads are used for parallel computation to capture semantic features of different subspaces (such as hot and cold, exterior and interior, and virtual and real).

[0069]

[0070] 3. Moxibustion treatment plan prediction and loss function optimization;

[0071] The model's output layer scores all candidate acupoints through a fully connected layer (MLP). To address the extreme imbalance problem in traditional Chinese medicine prescriptions—characterized by a small number of positive samples (approximately 3-5 effective acupoints) and a large number of negative samples (over 360 acupoints)—FocalLoss is used instead of the traditional cross-entropy loss function.

[0072]

[0073] in, This is the model's predicted probability of the correct label. Setting the focus parameter... Balance factor This formula can significantly reduce the weight of simple negative samples (easily classifiable irrelevant acupoints), forcing the model to focus on training those difficult-to-distinguish "hidden acupoints" that conform to Zhou Meisheng's unique experience.

[0074] Example 3: Detailed implementation of personalized solution generation and closed-loop feedback (S3);

[0075] This step achieves closed-loop control from model inference to clinical execution, and consists of the following two steps;

[0076] 1. Interpretable generation based on path backtracking;

[0077] While outputting recommendation results, the system uses the BeamSearch algorithm to find the optimal path connecting "patient syndrome nodes" and "recommended acupoint nodes" in the knowledge graph.

[0078] Path scoring function Defined as the product of the weights of all edges on the path:

[0079]

[0080] The system selects the path with the highest score and converts the entities and relationships on the path into natural language text (e.g., "Because [stomach pain] belongs to [stomach meridian disease], [Zusanli] is selected to [regulate qi and relieve pain]"), and automatically associates it with the corresponding original text index of "Jiusheng" in the database.

[0081] 2. Dynamic feedback regulation based on reinforcement learning (RL);

[0082] A Markov decision process (MDP) is constructed to handle real-time moxibustion sensation feedback.

[0083] (1) State Current moxibustion parameters (temperature) Duration ) and the types of moxibustion sensations reported by patients ( ).

[0084] (2) Action Adjusting moxibustion parameters ( , Alternatively, you can change your approach.

[0085] (3) Reward function The reward is designed based on Mr. Zhou's theory of "effectiveness comes from the arrival of Qi":

[0086]

[0087] (4) Strategy update: The action value function is updated using the Q-Learning algorithm. In order to learn the optimal control strategy:

[0088]

[0089] in For learning rate, This is the discount factor.

[0090] When the system detects that the patient's feedback is "no sensation," the RL agent will automatically output an action based on the learned strategy. (For example, increasing the temperature by 2°C or changing to sparrow-pecking moxibustion), thereby achieving dynamic closed-loop optimization of the moxibustion treatment plan.

[0091] The above is an exemplary description of the invention. Obviously, the specific implementation of the invention is not limited to the above-described manner. Any non-substantial improvement made using the inventive concept and technical solution of the invention, or the direct application of the inventive concept and technical solution to other situations without modification, is within the protection scope of the invention.

Claims

1. A knowledge-driven intelligent moxibustion treatment plan generation method, characterized in that, Includes the following steps: Step S1: Construct a knowledge graph for Zhou Meisheng's moxibustion method: Collect moxibustion ropes, Zhou Meisheng's Clinical Essentials, and clinical medical records to establish a domain ontology model containing entities with moxibustion sensation characteristics; use the BERT-BiLSTM-CRF model for named entity recognition, and use the CasRel layered pointer labeling network to extract relational triples between entities to solve the problem of overlapping entities in medical records. Store the cleaned triples data in the graph database. Step S2: Construct and train an intelligent reasoning model based on graph neural networks: Use the GraphSAGE algorithm to embed heterogeneous nodes in the knowledge graph into vectors, transforming discrete TCM concepts into continuous vector space representations; Construct a graph attention network GAT that incorporates a multi-head attention mechanism, taking the patient symptom entity set as input, triggering multi-hop message passing in the graph and dynamically calculating the importance weight of nodes. FocalLoss is used as the loss function during training to solve the problem of imbalance between positive and negative samples in acupoint recommendations; Step S3: Generate personalized moxibustion plans and achieve closed-loop dynamic control: Based on the probability distribution score list output by the model, generate the optimal main acupoints, acupoint combinations, and corresponding moxibustion types and parameters; use graph path backtracking technology to retrieve the knowledge path of syndromes and recommended acupoints, and generate interpretable evidence including the original text reference of the moxibustion rope; receive real-time feedback from patients on moxibustion sensations during the moxibustion process, and automatically trigger parameter correction instructions when the feedback is not ideal, thereby achieving closed-loop optimization of the moxibustion plan.

2. The method according to claim 1, characterized in that, In step S1, the entity types of the domain ontology model include at least five categories: disease, syndrome, acupoint, moxibustion sensation, and meridian. The relationship types between entities include at least six categories: meridian affiliation, inclusion, recommendation, triggering, treatment method association, and feedback regulation. The graph database used is Neo4j graph database.

3. The method according to claim 1, characterized in that, In step S1, the data acquisition and preprocessing are as follows: the paper documents are digitized using a 600DPI industrial scanner, the text is extracted using an OCR engine based on the ResNet-50 backbone network, and low-confidence characters are manually proofread by setting a confidence threshold; the text is cleaned using the Jieba word segmentation tool, which integrates more than 4,500 TCM terminology dictionaries, to obtain a standardized token sequence.

4. The method according to claim 1, characterized in that, In step S2, the GraphSAGE algorithm transforms discrete TCM concepts into 128-dimensional dense vectors; the graph attention network GAT uses 8 parallel attention heads to capture the features of different semantic subspaces such as cold and heat, exterior and interior, and deficiency and excess, simulating the diagnostic thinking of syndrome differentiation and moxibustion with clear distinction between primary and secondary aspects.

5. The method according to claim 1, characterized in that, In step S3, the BeamSearch algorithm is used to retrieve the optimal knowledge path. The path score function is the product of the weights of all edges on the path. The optimal path is then converted into a natural language interpretation and associated with the original text index of the moxibustion rope.

6. The method according to claim 1, characterized in that, In step S3, a Markov decision process is constructed based on the Q-Learning reinforcement learning algorithm to process real-time moxibustion sensation feedback. The state space is based on the moxibustion temperature, duration, and moxibustion sensation category, and the action space is based on the adjustment of moxibustion parameters and the change of techniques. A reward function is designed based on the theory that qi arrives and is effective, and the optimal control strategy is learned.

7. A knowledge-driven intelligent moxibustion plan generation system, characterized in that, For performing the method according to any one of claims 1-6, the system comprises the following components that are sequentially communicatively connected: The data acquisition and preprocessing module is used for the acquisition of multi-source data, digital processing, and construction of a standardized corpus. The knowledge graph construction module is used for the construction and management of knowledge graphs in the fields of TCM entity recognition, relation extraction, and Zhou Meisheng moxibustion. The intelligent reasoning module is used for node vectorization embedding, intelligent reasoning of moxibustion treatment schemes, and model training optimization. The scheme generation and interpretability module is used for outputting personalized moxibustion schemes and generating diagnostic and treatment evidence; The closed-loop feedback control module is used for real-time collection of moxibustion sensations and dynamic optimization of moxibustion treatment plans.

8. The system according to claim 7, characterized in that, The knowledge graph construction module incorporates the BERT-BiLSTM-CRF named entity recognition model and the CasRel relation extraction model; the intelligent inference module incorporates the GraphSAGE node embedding unit, the GAT inference unit with multi-head attention mechanism, and the FocalLoss training optimization unit.

9. The system according to claim 7, characterized in that, The closed-loop feedback control module incorporates a moxibustion sensation interaction acquisition unit, a reinforcement learning decision-making unit based on the Q-Learning algorithm, and a moxibustion parameter correction execution unit.

Images