Acupuncture large model dialectical method and conditioning optimization system fused with dynamic knowledge graph

Abstract

The invention discloses an acupuncture large model dialectical method and conditioning optimization system fusing a dynamic knowledge graph, and the method comprises the steps: obtaining first knowledge data of the acupuncture field, carrying out the entity extraction, relation extraction and attribute extraction of the first knowledge data, and constructing an initial knowledge graph of the acupuncture field; according to the dynamic change of acupuncture clinical practice, an initial knowledge graph is updated in real time to form a dynamic knowledge graph, clinical features of a patient at multiple different time points are extracted to form a clinical feature time sequence, and time sequence embedded features are generated according to the clinical feature time sequence. Performing semantic representation learning on the dynamic knowledge graph based on a graph attention network to obtain node feature information of the dynamic knowledge graph; according to the method, the time sequence embedded features are fused and enhanced to obtain fused features, and the acupuncture treatment scheme of the patient is differentiated by using the node feature information and the fused features based on a reinforcement learning algorithm, so that the flexibility and individuation level of acupuncture point selection are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information technology, and in particular to a large acupuncture model syndrome differentiation method fusing a dynamic knowledge graph and a regulation and optimization system. BACKGROUND

[0002] The core of acupuncture treatment is to determine the acupoint combination scheme through syndrome differentiation, and its curative effect directly depends on the accuracy and flexibility of acupoint selection. However, in the current clinical practice, the selection of acupuncture points still faces many technical bottlenecks, which seriously restricts the modernization development of acupuncture therapy.

[0003] With the development of information technology, traditional acupuncture gradually evolves towards digitization and intelligentization. However, the existing technical solutions still have obvious defects in actual application, and it is difficult to meet the clinical demand for flexibility of acupoint selection. SUMMARY

[0004] Therefore, the present application provides a large acupuncture model syndrome differentiation method fusing a dynamic knowledge graph and a regulation and optimization system, which solves the technical problem that it is difficult to meet the clinical demand for flexibility of acupoint selection.

[0005] The first aspect of the present application provides a large acupuncture model syndrome differentiation method fusing a dynamic knowledge graph, comprising: Obtaining first knowledge data in the field of acupuncture and acupunctur e, and performing entity extraction, relation extraction and attribute extraction on the first knowledge data to construct an initial knowledge graph of the field of acupuncture and acupunctur e; According to the dynamic changes of acupuncture clinical practice, the initial knowledge graph is updated in real time to form a dynamic knowledge graph; Extracting clinical features of a patient at multiple different time points to form a clinical feature time sequence, and generating a time sequence embedding feature according to the clinical feature time sequence; Based on the graph attention network, the dynamic knowledge graph is subjected to semantic representation learning to obtain node feature information of the dynamic knowledge graph; the time sequence embedding feature is fused and enhanced to obtain a fusion feature; Based on a reinforcement learning algorithm, the node feature information and the fusion feature are used to differentiate the acupuncture treatment scheme of the patient to obtain the acupuncture treatment scheme of the patient.

[0006] Preferably, the obtaining of the first knowledge data in the field of acupuncture and acupunctur e, and the entity extraction, relation extraction and attribute extraction on the first knowledge data to construct the initial knowledge graph of the field of acupuncture and acupunctur e, comprises: Collecting first knowledge data in the field of acupuncture and acupunctur e from multiple sources; Structural conversion is performed on the first knowledge data to obtain structured first knowledge data; data cleaning and standardization processing is performed on the structured first knowledge data to obtain standardized first knowledge data; entity extraction, relation extraction and attribute extraction are performed on the standardized first knowledge data to obtain triples of entity, relation and attribute; wherein the entity is acupoint and disease, the attribute includes acupoint name, belonging meridian, positioning and treated disease, and the relation is the treated relation between acupoint and disease; According to the triples, an initial knowledge graph of the acupuncture field is constructed.

[0007] Preferably, the initial knowledge graph is updated in real time according to the dynamic changes of acupuncture clinical practice to form a dynamic knowledge graph, comprising: According to the dynamic changes of acupuncture clinical practice, second knowledge data of the acupuncture field is generated; Cluster analysis is performed on the second knowledge data, and cluster center data of each cluster is obtained; Triple extraction is performed on the cluster center data of each cluster to obtain triples of the cluster center data; The time sequence similarity of the triples of the cluster center data and each triple in the initial knowledge graph of the acupuncture field is determined; In at least one case where the time sequence similarity is greater than a preset similarity threshold, the triples of the cluster center data are added to the initial knowledge graph to obtain the dynamic knowledge graph.

[0008] Preferably, the clinical features of the patient at multiple different time points are extracted to form a clinical feature time sequence, and a time sequence embedding feature is generated according to the clinical feature time sequence, comprising: Periodically acquire tongue image and pulse waveform data of the patient at multiple different time points; Feature extraction is performed on the tongue image and the pulse waveform data of each time point respectively to obtain clinical features corresponding to the tongue image and the pulse waveform data respectively, and the clinical features are arranged in time sequence to obtain clinical feature time sequences corresponding to the tongue image and the pulse waveform data respectively; Feature encoding is performed on the clinical feature time sequences corresponding to the tongue image and the pulse waveform data respectively based on a Transformer encoder through a self-attention mechanism to obtain time sequence embedding features corresponding to the tongue image and the pulse waveform data respectively.

[0009] Preferably, the clinical features corresponding to the tongue image include multi-dimensional tongue texture features and multi-dimensional tongue fur features, and the clinical features corresponding to the pulse waveform data include energy distribution features. The feature extraction is respectively performed on the tongue image and the pulse waveform data of each time point to obtain clinical features corresponding to the tongue image and the pulse waveform data respectively, including: The tongue image is segmented to separate a tongue body image region and a tongue fur image region; Multidimensional tongue body features and multidimensional tongue fur features are extracted from the tongue body image region and the tongue fur image region respectively; The pulse waveform data is preprocessed; the preprocessing includes removing baseline drift and filtering and denoising; Short-time Fourier transform is used to perform time-frequency analysis on the preprocessed pulse waveform data to extract energy distribution features in the 5-15 Hz frequency band.

[0010] Preferably, the semantic representation learning of the dynamic knowledge graph based on the graph attention network obtains node feature information of the dynamic knowledge graph, including: The information aggregation of neighbor nodes is performed on each node in the dynamic knowledge graph by using the graph attention network to obtain aggregated features of each node; The aggregated features are subjected to nonlinear transformation to obtain semantic representations of each node; The node feature information of the dynamic knowledge graph is obtained according to the semantic representations.

[0011] Preferably, the fusion and enhancement of the time series embedding features obtain fusion features, including: The time series embedding features corresponding to the tongue image and the time series embedding features corresponding to the pulse waveform data are spliced to obtain spliced features; The spliced features are subjected to attention mechanism weighting processing to obtain the fusion features.

[0012] Preferably, the acupuncture treatment scheme for the patient is diagnosed based on the node feature information and the fusion features by using the reinforcement learning algorithm to obtain the acupuncture treatment scheme for the patient, including: An intelligent agent for acupuncture treatment scheme diagnosis is designed, and the intelligent agent includes a state space, an action space and a reward function; The node feature information and the fusion features are input into the intelligent agent as states in the state space; According to the state, the reinforcement learning intelligent agent selects an acupoint combination action in the action space to generate an acupuncture treatment scheme; The acupuncture treatment scheme is executed, and a weighted sum of a therapeutic effect prediction value and a knowledge graph update gain is obtained as a reward value in the reward function; Adjust the policy of the agent according to the reward value until the policy of the agent converges, to obtain an optimal acupuncture treatment scheme; The efficacy prediction value is obtained by predicting the expected efficacy of the patient after receiving acupuncture treatment according to a prediction model trained according to historical data; and the knowledge graph update gain is obtained by evaluating the gain brought by updating the dynamic knowledge graph according to the knowledge data collected after the execution of the acupuncture treatment scheme.

[0013] In a second aspect, the present application provides a conditioning optimization system for syndrome differentiation of an acupuncture large model fused with a dynamic knowledge graph, comprising: A graph construction module is configured to acquire first knowledge data in the field of acupuncture and perform entity extraction, relation extraction and attribute extraction on the first knowledge data to construct an initial knowledge graph in the field of acupuncture. A graph updating module is configured to update the initial knowledge graph in real time according to the dynamic changes of acupuncture clinical practice to form a dynamic knowledge graph. A feature extraction module is configured to extract clinical features of a patient at multiple different time points to form a clinical feature time sequence and generate time sequence embedding features according to the clinical feature time sequence. A feature processing module is configured to perform semantic representation learning on the dynamic knowledge graph based on a graph attention network to obtain node feature information of the dynamic knowledge graph, and fuse and enhance the time sequence embedding features to obtain fused features. A scheme differentiation module is configured to perform syndrome differentiation on an acupuncture treatment scheme for the patient based on a reinforcement learning algorithm, using the node feature information and the fused features, to obtain the acupuncture treatment scheme for the patient.

[0014] In a third aspect, the present application further provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to cause the processor to perform the steps of the method for syndrome differentiation of an acupuncture large model fused with a dynamic knowledge graph according to the first aspect.

[0015] As can be seen from the above technical solutions, this invention acquires first knowledge data in the field of acupuncture and extracts entities, relationships, and attributes from the first knowledge data to construct an initial knowledge graph in the field of acupuncture. Based on the dynamic changes in acupuncture clinical practice, the initial knowledge graph is updated in real time to form a dynamic knowledge graph. This allows for more timely absorption of new research evidence and clinical experience compared to traditional static rule bases. The dynamic update mechanism ensures the timeliness and advancement of the system's knowledge. By extracting the clinical characteristics of patients at multiple different time points, a clinical feature time sequence is formed, and time-series embedding features are generated based on the clinical feature time sequence. Semantic representation learning is performed on the dynamic knowledge graph based on a graph attention network to obtain the node feature information of the dynamic knowledge graph. The time-series embedding features are fused and enhanced to obtain fused features, thereby making the formulation of acupuncture treatment plans more reasonable. Based on reinforcement learning algorithms, the acupuncture treatment plan for the patient is differentiated using node feature information and fused features, improving the flexibility and personalization of acupoint selection. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of a diagnostic method for acupuncture based on a dynamic knowledge graph provided in an embodiment of the present invention; Figure 2 A schematic diagram of the structure of a large-scale acupuncture model for syndrome differentiation and optimization, which integrates dynamic knowledge graphs, is provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown in the figure, this application provides a diagnostic method for acupuncture based on a large-scale model that integrates dynamic knowledge graphs, including the following steps S1 to S5. Wherein: Step S1, acquire first knowledge data in the field of acupuncture and moxibustion, and perform entity extraction, relation extraction and attribute extraction on the first knowledge data to construct an initial knowledge graph in the field of acupuncture and moxibustion.

[0020] Among them, the knowledge data in the field of acupuncture and moxibustion can be based on the information recorded in traditional Chinese medicine ancient books, such as acupoint compatibility rules, acupuncture techniques, and treatment of diseases, etc. At the same time, public acupuncture and moxibustion clinical case data from many top three hospitals in the country are collected. These case data contain detailed information such as patient's basic information, symptom performance, diagnosis result, acupoint selection scheme, treatment effect, etc.

[0021] Among them, entity extraction refers to identifying entities related to acupuncture and moxibustion from the first knowledge data in the field of acupuncture and moxibustion, such as acupoints, diseases, etc.; relation extraction refers to determining the relationship between these entities, such as the treatment relationship between acupoints and diseases; attribute extraction is to extract the specific attributes of the entity, such as the name of the acupoint, the meridian it belongs to, the positioning and the treatment of the disease, etc. Through the extraction and arrangement of these information, an initial knowledge graph in the field of acupuncture and moxibustion can be constructed, which presents the knowledge system in the field of acupuncture and moxibustion in a structured way.

[0022] Step S2, real-time update the initial knowledge graph according to the dynamic changes of acupuncture and moxibustion clinical practice, form a dynamic knowledge graph.

[0023] Among them, in order to ensure the accuracy and timeliness of the acupuncture and moxibustion big model syndrome differentiation method, the embodiment of the present application proposes a real-time updating mechanism. This mechanism can capture the latest dynamics of acupuncture and moxibustion clinical practice, including new acupoint compatibility rules, new acupuncture techniques, new treatment of diseases, etc., and integrate these information into the initial knowledge graph, thus forming a dynamic knowledge graph. This dynamic update not only improves the flexibility of the acupuncture and moxibustion big model syndrome differentiation method, but also makes it better adapt to the diversity and complexity of clinical practice.

[0024] Step S3, extract the clinical features of the patient at multiple different time points to form a clinical feature time sequence, and generate a time sequence embedding feature according to the clinical feature time sequence.

[0025] Among them, clinical features are important basis for judging the health status of patients. In order to fully reflect the physical condition and disease development trend of patients, the embodiment of the present application proposes to extract the clinical features of patients from multiple different time points, and arrange these features in chronological order to form a clinical feature time sequence. This process not only considers the static features of patients at different time points, but also integrates the time factor, making the description of clinical features more three-dimensional and comprehensive.

[0026] Step S4, perform semantic representation learning on the dynamic knowledge graph based on graph attention network to obtain node feature information of the dynamic knowledge graph; and perform fusion enhancement on the time sequence embedding feature to obtain a fusion feature.

[0027] Among them, the graph attention network is an effective graph neural network model, which can aggregate the information of neighbor nodes of each node in the graph through the attention mechanism, so as to learn the semantic representation of each node. In the embodiment of the application, the semantic representation learning of each acupoint and disease node in the dynamic knowledge graph is carried out by using the graph attention network, so that the feature information of these nodes can be obtained, which can reflect the importance of the nodes in the knowledge graph and the association relationship with other nodes.

[0028] Meanwhile, the fusion of the time sequence embedding features can enhance the comprehensive and accurate fusion features, which fuse the information of the tongue image and the pulse waveform data, and can more comprehensively reflect the physical condition and disease development trend of the patient.

[0029] Step S5, based on the reinforcement learning algorithm, the node feature information and the fusion features are used to diagnose the acupuncture treatment scheme of the patient, and the acupuncture treatment scheme of the patient is obtained.

[0030] Among them, the reinforcement learning algorithm is a machine learning method that simulates the human decision-making process, which learns the optimal strategy through the interaction between the agent and the environment. In the embodiment of the application, an agent for diagnosing the acupuncture treatment scheme is designed, which can select the optimal acupoint combination action in the state space according to the node feature information and the fusion features of the dynamic knowledge graph of the patient, and generate the acupuncture treatment scheme. The strategy of the agent is optimized through continuous trial and error and adjustment. After each execution of the acupuncture treatment scheme, the performance of the agent is evaluated according to the weighted sum of the efficacy prediction value and the knowledge graph update gain as the reward value, and the strategy of the agent is adjusted according to the reward value, until the strategy converges, and the optimal acupuncture treatment scheme is obtained. This process not only improves the individualization and precision of the acupuncture treatment scheme, but also enables the acupuncture large model diagnosis method to continuously learn and progress, and better serves the clinical practice.

[0031] It should be noted that, in the embodiment of the present application, the first knowledge data in the field of acupuncture and moxibustion is acquired, and entity extraction, relation extraction and attribute extraction are performed on the first knowledge data to construct an initial knowledge graph in the field of acupuncture and moxibustion; according to the dynamic changes of acupuncture and moxibustion clinical practice, the initial knowledge graph is updated in real time to form a dynamic knowledge graph, so that the new research evidence and clinical experience can be more timely absorbed than the traditional static rule base, and the dynamic updating mechanism ensures the timeliness and advancement of the system knowledge. By extracting the clinical characteristics of the patient at multiple different time points, a clinical feature time sequence is formed, and a time sequence embedding feature is generated according to the clinical feature time sequence. The dynamic knowledge graph is semantically represented and learned based on a graph attention network to obtain node feature information of the dynamic knowledge graph. The time sequence embedding feature is fused and enhanced to obtain a fusion feature, so that the acupuncture treatment plan is more reasonable. Based on the reinforcement learning algorithm, the node feature information and the fusion feature are used to diagnose the acupuncture treatment plan for the patient, and the flexibility and individualization level of acupoint selection are improved.

[0032] In some embodiments, the first knowledge data in the field of acupuncture and moxibustion is acquired, and entity extraction, relation extraction and attribute extraction are performed on the first knowledge data to construct an initial knowledge graph in the field of acupuncture and moxibustion, comprising: Step S101, collecting first knowledge data in the field of acupuncture and moxibustion from multiple sources.

[0033] Step S102, structuring the first knowledge data to obtain structured first knowledge data.

[0034] The structured first knowledge data refers to the first knowledge data collected in the field of acupuncture and moxibustion after cleaning, sorting and standardization processing, which is converted into an information format with clear structure and semantics. This process includes converting unstructured text data into structured table or database form, so that the entities, relationships and attributes in the data are clearly presented, facilitating subsequent entity extraction, relation extraction and attribute extraction operations. Through structured conversion, it can ensure that the knowledge data in the field of acupuncture and moxibustion has high accuracy and consistency when constructing the initial knowledge graph.

[0035] For unstructured first knowledge data, a pre-trained model (such as RoBERTa-wwm) is used to structure the first knowledge data to obtain structured first knowledge data.

[0036] Step S103, data cleaning and standardization processing on the structured first knowledge data to obtain standardized first knowledge data.

[0037] The data cleaning is a process of quality inspection and correction of the structured first knowledge data, aiming to eliminate errors, redundancies and inconsistencies in the data. This process includes checking the integrity, accuracy and consistency of the data, repairing or deleting erroneous data, filling in missing values, and unifying the format and unit of the data, etc. Through data cleaning, it can ensure that the knowledge data in the field of acupuncture has high quality and reliability when constructing the initial knowledge graph.

[0038] The standardization processing is a normalization processing of the cleaned first knowledge data, so as to unify the description and comparison when constructing the initial knowledge graph. This process includes unifying the naming rules of acupoints, the classification standards of diseases, and the description methods of acupuncture techniques, so that data of different sources and formats can be compared and analyzed under the same framework. Through standardization processing, the comparability and usability of the knowledge data in the field of acupuncture can be improved.

[0039] In step S104, entity extraction, relation extraction and attribute extraction are performed on the standardized first knowledge data to obtain triples of entity, relation and attribute; wherein the entity is acupoint and disease, the attribute includes acupoint name, belonging meridian, positioning and treating disease; the relation is the treating relation between acupoint and disease.

[0040] The entity extraction, relation extraction and attribute extraction are performed by natural language processing technology. The entity extraction refers to accurately identifying entities related to acupuncture from the standardized first knowledge data, such as acupoints and diseases, etc. The relation extraction refers to determining the association between these entities, such as the treating relation between acupoints and diseases. The attribute extraction refers to extracting specific attribute information of the entity, such as the name of acupoint, the belonging meridian, the positioning location and the specific disease treated, etc. Through detailed extraction and arrangement of these information, a series of triples composed of entity, relation and attribute can be obtained, which are the basic units for constructing the initial knowledge graph in the field of acupuncture.

[0041] For example, "acupoint A can treat symptom B", "disease mechanism C corresponds to treatment method D, common acupoints include E, F", etc. The extracted entity and relation are integrated with the structured data to construct the initial knowledge graph.

[0042] To ensure the accuracy of the knowledge, in the data fusion process, domain experts are introduced for review and correction, and the knowledge with conflicts or uncertainties is marked and verified.

[0043] In step S105, according to the triples, the initial knowledge graph in the field of acupuncture is constructed.

[0044] In some embodiments, according to the dynamic changes of acupuncture clinical practice, the initial knowledge graph is updated in real time to form a dynamic knowledge graph, including: Step S201, generate the second knowledge data in the field of acupuncture and moxibustion according to the dynamic changes of acupuncture clinical practice.

[0045] Among them, the second knowledge data can come from the latest acupuncture research results, clinical practice experience summary, expert consensus or guidelines, etc. These data reflect the latest progress in the field of acupuncture and new findings in clinical practice.

[0046] For example: based on the DeepSeek-R1 streaming computing framework, automatically crawl the latest acupuncture research literature, clinical guidelines and expert consensus from academic platforms such as PubMed, CNKI and Wanfang every 15 minutes.

[0047] Step S202, cluster analysis is performed on the second knowledge data, and the cluster center data of each cluster is obtained.

[0048] Among them, the BERTopic topic model is used to cluster the second knowledge data obtained to determine the theme and core content of different research or practice experience.

[0049] BERTopic is a text clustering method based on BERT pre-training model and LDA topic model, which combines the semantic understanding ability of deep learning and the clustering characteristics of traditional topic model, and can more accurately capture the topic structure of text data.

[0050] Among them, the cluster center data of each cluster is the representative data obtained after processing by the BERTopic topic model, which can summarize and reflect the main content and characteristics of each cluster. By clustering analysis of the second knowledge data, similar or related research and practice experience can be classified into a category, which is convenient for subsequent knowledge update and integration.

[0051] Step S203, re-extract triplets from the cluster center data of each cluster to obtain the triplets of the cluster center data.

[0052] Among them, re-extracting triplets from the cluster center data of each cluster means extracting the head entity h (such as acupoint), the relationship r (such as treatment), and the tail entity t (such as disease).

[0053] Step S204, determine the time sequence similarity of the triplets of the cluster center data and each triplet in the initial knowledge graph in the field of acupuncture and moxibustion.

[0054] Among them, the time sequence similarity refers to the similarity of the triplets of the cluster center data and the triplets in the initial knowledge graph in terms of time sequence and logical relationship. By calculating the time sequence similarity, the relevance and consistency between new knowledge data and existing knowledge graph can be evaluated, which provides a basis for subsequent knowledge update. Specifically, the calculation of time sequence similarity is:

[0055] In the formula, For temporal similarity, n is the number of data points within the window. As weight, Angle The cosine of .

[0056] in,

[0057] In the formula, The relation vectors corresponding to the triples of the cluster center data (mapped using the TransE model). The relation vectors corresponding to the triples in the initial knowledge graph (mapped using the TransE model). Let be the magnitude of the relation vector corresponding to the triples of the cluster center data. This is the magnitude of the relation vector corresponding to the triples in the initial knowledge graph.

[0058] Among them, weight The weighting is determined based on factors such as the credibility of the source, sample size, and research quality. For example, randomized controlled trials published in authoritative journals have higher weightings, while case reports have lower weightings. The smaller the angle, the more similar the relationship between the old and new studies.

[0059] Step S205: If at least one temporal similarity is greater than a preset similarity threshold, the triples of the cluster center data are added to the initial knowledge graph to obtain a dynamic knowledge graph.

[0060] Specifically, if at least one temporal similarity is greater than a preset similarity threshold (e.g., 0.85), it indicates that the new evidence has a high degree of consistency and credibility with existing knowledge, triggering a knowledge graph update and adding the new triple to the knowledge graph; otherwise, the new knowledge is marked and further reviewed and evaluated by domain experts before deciding whether to include it in the knowledge graph.

[0061] In some embodiments, clinical features of a patient at multiple different time points are extracted to form a clinical feature time series, and time series embedding features are generated based on the clinical feature time series, including: Step S301: Periodically acquire tongue images and pulse waveform data of the patient at multiple different time points.

[0062] Periodic acquisition refers to the regular collection of patients' tongue images and pulse waveform data at certain time intervals. This process ensures the continuity and timeliness of the acquired clinical characteristic data, such as acquiring patients' tongue images and pulse waveform data at multiple different time points every 60 seconds.

[0063] Tongue imaging is an important component of traditional Chinese medicine's diagnostic method of observation. By observing the tongue's color, shape, and coating, one can determine the patient's internal condition, such as whether it is hot or cold, deficient or excessive, and the state of qi and blood. Pulse waveform data, obtained through pulse diagnosis, reflects the flow of qi and blood and the functional status of the internal organs, serving as a crucial basis for TCM diagnosis and treatment.

[0064] In acquiring pulse waveform data, pulse waveform data of the patient's radial artery were collected using a pressure sensor. During the acquisition process, the pressure and position of the sensor were strictly controlled to ensure the stability and consistency of the data.

[0065] Step S302: Extract features from the tongue image and pulse waveform data at each time point to obtain the clinical features corresponding to the tongue image and pulse waveform data respectively, and arrange the clinical features in chronological order to obtain the chronological order of the clinical features corresponding to the tongue image and pulse waveform data respectively.

[0066] This study processes tongue images and pulse waveforms that possess visual or waveform characteristics to extract effective feature information. Specifically, the clinical features corresponding to tongue images include multidimensional tongue texture features and multidimensional tongue coating features; the clinical features corresponding to pulse waveforms include energy distribution features.

[0067] Feature extraction was performed on the tongue image and pulse waveform data at each time point to obtain the corresponding clinical features, including: Step S3021: Segment the tongue image to separate the tongue body image region and the tongue coating image region.

[0068] Among these methods, an improved U-Net model can be used to segment the acquired tongue images, accurately separating the tongue body and tongue coating areas.

[0069] The improved U-Net model is based on the traditional U-Net, introducing an attention mechanism and residual connections to improve the accuracy and robustness of segmentation.

[0070] Step S3022: Extract multidimensional tongue texture features and multidimensional tongue coating features from the tongue texture image region and the tongue coating image region, respectively.

[0071] Among them, the multidimensional tongue body characteristics include the RGB mean value, texture features, and color saturation of the tongue body, while the multidimensional tongue coating characteristics include the thickness, color, distribution uniformity, crack density, and number of spots of the tongue coating. These characteristics can effectively reflect the patient's pathological state, such as cold or heat, deficiency or excess, and the state of qi and blood, providing an important basis for diagnosis.

[0072] Step S3023: Preprocess the pulse waveform data; preprocessing includes removing baseline drift and filtering for noise reduction.

[0073] Baseline drift removal aims to eliminate baseline shifts in pulse waveform data caused by the instrument itself or patient physiological factors. This is achieved using digital filters to ensure data accuracy. Noise reduction filtering eliminates high-frequency noise or interference that may be introduced during acquisition, improving the signal-to-noise ratio. This process can be implemented using bandpass or low-pass filters, preserving the effective information in the pulse waveform while removing unnecessary noise components.

[0074] Step S3024: Use short-time Fourier transform to perform time-frequency analysis on the preprocessed pulse waveform data and extract the energy distribution characteristics of the 5-15Hz frequency band.

[0075] The 5-15Hz frequency band is a crucial band in pulse waveform data, reflecting the function of internal organs and the state of qi and blood flow. Time-frequency analysis using short-time Fourier transform can reveal the energy distribution within this band, thereby extracting characteristic information relevant to traditional Chinese medicine diagnosis. These energy distribution characteristics can reveal the strength and rhythm of pulse patterns (such as floating pulse, deep pulse, slow pulse, and rapid pulse) as well as pathological changes in different internal organs.

[0076] Energy distribution characteristics include peak energy, mean energy, and energy fluctuation range. These characteristics can quantitatively describe different aspects of the pulse waveform, such as strength, smoothness, and possible pathological changes. For example, the level of the peak energy reflects the strength of the pulse, the mean energy reflects the overall intensity level of the pulse, and the energy fluctuation range may reveal rhythmic changes or instability in the pulse.

[0077] By meticulously extracting features from tongue images and pulse waveform data, a series of clinical characteristics reflecting the patient's internal condition can be obtained. These features are arranged in chronological order, forming a temporal sequence of clinical characteristics corresponding to each tongue image and pulse waveform data. These temporal features not only record the changes in the patient's physiological state at different time points but also provide rich data support for subsequent diagnosis and treatment.

[0078] Step S303: Based on the Transformer encoder, the clinical features corresponding to the tongue image and pulse waveform data are encoded in time sequence using a self-attention mechanism to obtain the time sequence embedding features corresponding to the tongue image and pulse waveform data respectively.

[0079] After obtaining the temporal sequence of clinical features, a deep learning model is used to process these temporal features to generate temporal embedding features. In this embodiment, based on the Transformer encoder, the long-distance dependencies of the temporal sequences of clinical features corresponding to the tongue image and pulse waveform data are analyzed through a self-attention mechanism. Combined with positional encoding, this accurately reflects the temporal order and development trend of symptoms, thus obtaining temporal embedding features.

[0080] For example, a patient with a cold may have symptoms that progress from initial nasal congestion and runny nose to cough and sore throat in the middle stage, and then to fatigue and loss of appetite in the later stage. The Transformer encoder can encode the dynamic changes of these symptoms into a temporal embedding, providing more comprehensive information for diagnosis.

[0081] In some embodiments, semantic representation learning of a dynamic knowledge graph is performed based on a graph attention network to obtain node feature information of the dynamic knowledge graph, including: Step S401: Use a graph attention network to aggregate the information of the neighboring nodes of each node in the dynamic knowledge graph to obtain the aggregated features of each node.

[0082] Among them, graph attention networks are graph neural networks based on attention mechanisms. They can automatically learn the strength of associations between different nodes in a graph and aggregate the information of neighboring nodes based on these association strengths, thereby obtaining the aggregated features of each node. This process ensures that each node not only contains its own information but also incorporates the information of its neighboring nodes, enhancing the expressive power of node features.

[0083] When aggregating neighbor node information for each node in a dynamic knowledge graph, the graph attention network calculates attention coefficients between nodes based on the edges (i.e., relationships) and node attributes. These attention coefficients reflect the strength and importance of the associations between nodes, enabling more accurate capture of key information in the graph during information aggregation. Through weighted aggregation of neighbor node information, each node obtains an aggregated feature containing richer information, which more comprehensively reflects the node's position and role in the knowledge graph. Specifically, the aggregated features are:

[0084] In the formula, For the first l+ The aggregated feature vector of node v in layer 1. For activation functions (such as the ReLU function). Let v be the set of neighboring nodes. , Let v and u be the degrees of nodes v and u, respectively. For the first l The learnable weight matrix of the layer, For the first l The aggregated feature vector of node v in the layer. This is a bias term.

[0085] By stacking multiple layers of graph attention network, the local neighbor information of nodes in the knowledge graph can be aggregated into the node's own features, enhancing the expressive power of node features and providing richer knowledge features for subsequent dialectical reasoning.

[0086] Step S402: Perform a nonlinear transformation on the aggregated features to obtain the semantic representation of each node.

[0087] Semantic representation is the result of nonlinear transformation of aggregated features, which can further capture the complex relationships and inherent patterns of node features.

[0088] In this embodiment, a multilayer perceptron (MLP) is used to perform nonlinear transformations on the aggregated features. By introducing nonlinear activation functions (such as ReLU or Tanh), the model can learn higher-order interactions and complex patterns between node features. This process results in each node receiving a fixed-dimensional semantic representation vector, which comprehensively and accurately reflects the semantic information and role of the node in the knowledge graph, providing strong support for subsequent dialectical reasoning. By learning the semantic representation of each node in the dynamic knowledge graph, we obtain a node feature matrix containing rich semantic information, which provides a solid foundation for subsequent dialectical reasoning and optimization.

[0089] Step S403: Obtain node feature information of the dynamic knowledge graph based on semantic representation.

[0090] Node feature information is determined through semantic representation, encompassing the node's core semantic information and role positioning within the knowledge graph. This node feature information not only reflects the node's own attributes and relationships but also incorporates information from its neighboring nodes and the global structural information of the graph, providing comprehensive knowledge support for subsequent dialectical reasoning and optimization. Specifically, node feature information can include the node's semantic vector, category label, and importance score. The node's semantic vector is a direct result of semantic representation, comprehensively and accurately reflecting the node's semantic information and role in the knowledge graph; the node's category label is obtained based on the semantic vector and predefined category classification rules, used to identify the type or category to which the node belongs; and the node's importance score is based on the node's position and connections within the knowledge graph.

[0091] In some embodiments, the temporal embedding features are fused and enhanced to obtain fused features, including: Step S411: Concatenate the temporal embedding features corresponding to the tongue image and the temporal embedding features corresponding to the pulse waveform data to obtain the concatenated features.

[0092] The splicing feature involves merging the temporal embedding features corresponding to the tongue image and pulse waveform data respectively, forming a feature vector containing richer information. This process is similar to integrating data from different sources in data fusion to obtain a more comprehensive perspective. Through splicing, the features of the tongue image and pulse are combined, providing a richer data foundation for subsequent analysis and reasoning.

[0093] Step S412: Apply attention mechanism weighting to the spliced ​​features to obtain fused features.

[0094] Specifically, the weights of different modalities are dynamically adjusted based on their relevance to the current dialectical task. This is achieved by calculating the attention weights for each modality using the Softmax function.

[0095] In the formula, Let be the attention weight for the i-th concatenated feature. Here is the attention weight matrix. Temporal embedding features corresponding to tongue image Temporal embedding features corresponding to pulse waveform data The splicing features after splicing.

[0096] Among them, attention weight This reflects the importance of each modality of data in the current diagnostic task; a higher weight indicates a greater impact of that modality on the diagnostic result. Through attention-based fusion, multimodal data can be organically combined to improve the accuracy of diagnosis. For example, for diseases primarily characterized by tongue appearance changes, the attention weight of tongue appearance features will be increased accordingly; while for diseases with significant symptom changes, the weight of symptom temporal features will be greater.

[0097] In some embodiments, based on a reinforcement learning algorithm, the acupuncture treatment plan for the patient is analyzed using node feature information and fused features to obtain the patient's acupuncture treatment plan, including: Step S501: Design an intelligent agent that dialectically applies to acupuncture treatment plans. The intelligent agent includes a state space, an action space, and a reward function.

[0098] The state space includes node feature information and fusion features, comprehensively reflecting the patient's condition and the knowledge background in the knowledge graph. The action space consists of acupoint combinations, generally containing 361 regular acupoints and 128 extra acupoints. Each action represents a specific acupoint combination scheme, i.e., an acupuncture treatment plan.

[0099] Step S502: Input the node feature information and fused features as state inputs into the agent in the state space.

[0100] In this process, the agent selects an acupoint combination scheme as an action from the action space based on the current state (i.e., the patient's specific condition and relevant knowledge in the knowledge graph). This selection process is based on a policy network, which outputs a probability distribution representing the probability of each action being selected based on the input state information. The agent then randomly selects an action to execute according to this probability distribution.

[0101] Step S503: Based on the state, the reinforcement learning agent selects acupoint combination actions in the action space to generate an acupuncture treatment plan.

[0102] When selecting acupoint combinations, the agent considers the characteristics of the current condition, relevant knowledge in the knowledge graph, and the potential therapeutic effects of different acupoint combinations. Through a policy network, the agent can learn the optimal or near-optimal acupoint combinations under different conditions, thereby generating personalized acupuncture treatment plans for patients.

[0103] The policy network, a deep neural network, takes states from the state space as input and outputs a probability distribution in the action space. This probability distribution represents the probability of different acupoint combinations being selected. The agent randomly selects an acupoint combination as the current acupuncture treatment plan based on this probability distribution. By continuously interacting with the environment (i.e., the patient's condition and treatment effect) and adjusting the parameters of the policy network based on the reward signal from the reward function, the agent gradually learns the ability to select the optimal or near-optimal acupoint combination under different states.

[0104] Step S504: Execute the acupuncture treatment plan and obtain the weighted sum of the efficacy prediction value and the knowledge graph update gain as the reward value in the reward function.

[0105] Among them, the efficacy prediction value is obtained by predicting the expected efficacy of patients after receiving acupuncture treatment based on a prediction model trained on historical data; the knowledge graph update gain is obtained by evaluating the gain brought about by updating the dynamic knowledge graph based on the knowledge data collected after the acupuncture treatment plan is implemented.

[0106] Specifically, this embodiment considers both the efficacy prediction value and the knowledge graph update gain. The efficacy prediction value is obtained by quantitatively evaluating the expected treatment effect after the acupuncture treatment plan is implemented, reflecting the degree to which the treatment plan improves the patient's condition. The knowledge graph update gain is the information gain obtained by dynamically updating the knowledge graph based on the actual changes in the patient's condition during the implementation of the treatment plan, reflecting the degree to which the treatment plan improves and optimizes the knowledge graph. By weighting and calculating the efficacy prediction value and the knowledge graph update gain, a comprehensive reward value is obtained, which is used to evaluate the effect of the acupuncture treatment plan selected by the agent. The higher the reward value, the better the effect of the treatment plan, and the better the agent's strategy when selecting acupoint combinations. The agent continuously adjusts the parameters of its strategy network based on the reward value to optimize the selection process of the acupuncture treatment plan, ultimately achieving personalized and precise acupuncture treatment. The reward function is:

[0107] In the formula, As a reward value, The weight of the predictive value of therapeutic effect, This is a predictive value for therapeutic efficacy. Update the weights of the gain for the knowledge graph. Update the gain for the knowledge graph.

[0108] Among them, the weight of the predictive value of efficacy Weights of knowledge graph update gain The sum of these values ​​is 1. Generally, the weight of the predictive value for therapeutic efficacy is... =0.7, =0.3.

[0109] Among them, the efficacy prediction value is the value output by the efficacy prediction model trained based on historical clinical data. After inputting the state, it outputs the expected treatment effect of a certain acupoint combination on the current patient, with a value range of [0,10] (10 points is "excellent efficacy", 0 points is "ineffective").

[0110] Knowledge graph update gain quantifies the "value" of a specific acupoint combination scheme in supplementing new evidence to a dynamic knowledge graph, with a value range of [0, 10] (10 points signifies "extremely high value," and 0 points signifies "no new value"). The formula for calculating knowledge graph update gain is as follows:

[0111] In the formula, , , , All are weighted values, and their sum is 1. Score the credibility of the evidence. For scarcity scoring, Scoring based on knowledge relevance. Score the knowledge completion rate.

[0112] Each scoring dimension has been standardized, with values ​​ranging from [0, 10]. This formula, through multi-dimensional collaborative evaluation, considers scarcity and completeness, as well as the quality of evidence, logical connections between knowledge, and clinical application value, making it more aligned with the actual needs of dynamic updates to the acupuncture knowledge graph.

[0113] in,

[0114] In the formula, For the type of research, randomized controlled trials (RCT) = 10 points, cohort studies = 8 points, case-control studies = 6 points, case reports = 3 points, and theoretical derivations = 1 point. To score sample quality, a sample size > 1000 cases = 10 points, 500-1000 cases = 8 points, 100-500 cases = 6 points, and < 100 cases = 3 points; The number of validations is scored as follows: ≥3 multicenter validations = 10 points, 2 validations = 8 points, 1 validation = 6 points, and no validation = 2 points.

[0115]

[0116] In the formula, This represents the number of times the disease entity within the triple appears in the knowledge graph. The number of all diseases in the knowledge graph.

[0117]

[0118] In the formula, m is the number of existing similar triples. Let be the cosine of the angle between the vector of the new triplet and the j-th existing triplet (values ​​[-1, 1]).

[0119]

[0120] In the formula, The gaps are scored according to priority: core gaps (such as the association between "rare syndrome types and acupoints") = 10 points, important gaps (such as the association between "common syndrome types and new acupoints") = 8 points, and general gaps (such as the supplementation of "common syndrome types and conventional acupoints") = 5 points. To complete the score, fully filling the gap = 10 points, partially filling (in conjunction with other evidence) = 6 points, and providing only supplementary information = 3 points.

[0121] Step S505: Adjust the agent's strategy according to the reward value until the agent's strategy converges to obtain the optimal acupuncture treatment plan.

[0122] In this process, the agent continuously optimizes its strategy for selecting acupuncture treatment plans through trial and error and learning. The agent tries different combinations of acupoints and evaluates the effectiveness of these plans based on reward values. Higher reward values ​​indicate better results, and the agent is more likely to choose similar plans. Through multiple iterations and learning, the agent eventually converges to a stable strategy that can select the optimal or near-optimal acupuncture treatment plan given a specific condition.

[0123] The policy convergence condition is that the acupuncture treatment plan selected by the agent tends to be consistent in multiple consecutive iterations, and the corresponding reward value reaches or exceeds a preset threshold.

[0124] This application's embodiments not only consider the patient's condition and relevant knowledge in the knowledge graph, but also incorporate historical clinical data and dynamically updated knowledge graph information, thereby achieving personalized and precise acupuncture treatment. The optimal acupuncture treatment plan can provide targeted treatment for patients, improve treatment outcomes, and also help to improve and optimize the acupuncture knowledge graph, providing more accurate and reliable knowledge support for future acupuncture treatment.

[0125] like Figure 2 As shown, this application provides a conditioning and optimization system for acupuncture based on a dynamic knowledge graph and a large-scale model of syndrome differentiation, including: The graph construction module 100 is used to acquire first knowledge data in the field of acupuncture and extract entities, relationships and attributes from the first knowledge data to construct an initial knowledge graph in the field of acupuncture. The knowledge graph update module 200 is used to update the initial knowledge graph in real time according to the dynamic changes in acupuncture clinical practice, forming a dynamic knowledge graph. The feature extraction module 300 is used to extract the clinical features of patients at multiple different time points, form a clinical feature time series, and generate time series embedded features based on the clinical feature time series; The feature processing module 400 is used to learn semantic representations of dynamic knowledge graphs based on graph attention networks to obtain node feature information of dynamic knowledge graphs; and to fuse and enhance temporal embedded features to obtain fused features. The scheme dialectics module 500 is used to dialectically analyze the patient's acupuncture treatment plan based on reinforcement learning algorithm, using node feature information and fusion features, to obtain the patient's acupuncture treatment plan.

[0126] like Figure 3As shown, this application provides an electronic device. The electronic device 10 includes a memory 20 and a processor 30. The memory 20 stores a computer program. When the computer program is executed by the processor 30, the processor 30 performs the steps of the acupuncture big model diagnostic method that integrates dynamic knowledge graphs as described in the above embodiment.

[0127] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system and electronic equipment described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0128] It should be noted that the user information (including but not limited to user images, user tongue images, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0129] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0130] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0131] In the embodiments provided by this invention, it should be understood that the disclosed systems, electronic devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0133] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0134] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A diagnostic method for acupuncture using a large-scale model integrating dynamic knowledge graphs, characterized in that, include: Acquire first knowledge data in the field of acupuncture, and extract entity, relation and attribute data from the first knowledge data to construct an initial knowledge graph in the field of acupuncture. The initial knowledge graph is updated in real time based on the dynamic changes in acupuncture clinical practice to form a dynamic knowledge graph; Extract the patient's clinical features at multiple different time points to form a clinical feature time series, and generate time series embedding features based on the clinical feature time series; Semantic representation learning is performed on the dynamic knowledge graph based on graph attention network to obtain the node feature information of the dynamic knowledge graph; The temporal embedding features are fused and enhanced to obtain fused features; Based on the reinforcement learning algorithm, the acupuncture treatment plan for the patient is analyzed using the node feature information and the fusion feature to obtain the patient's acupuncture treatment plan.

2. The acupuncture large-scale model diagnostic method integrating dynamic knowledge graphs according to claim 1, characterized in that, The process of acquiring first knowledge data in the field of acupuncture and extracting entities, relationships, and attributes from the first knowledge data to construct an initial knowledge graph in the field of acupuncture includes: Collect first-hand knowledge data in the field of acupuncture from multiple sources; The first knowledge data is transformed into structured data to obtain structured first knowledge data. The structured first knowledge data is cleaned and standardized to obtain standardized first knowledge data; Entity extraction, relation extraction, and attribute extraction are performed on the standardized first knowledge data to obtain triples of entities, relations, and attributes; wherein, the entities are acupoints and diseases, and the attributes include the name of the acupoint, the meridian to which it belongs, the location, and the diseases it treats; the relations are the therapeutic relationships between the acupoints and the diseases. Based on the triples, an initial knowledge graph for the acupuncture field is constructed.

3. The acupuncture large-scale model diagnostic method integrating dynamic knowledge graphs according to claim 2, characterized in that, The process of updating the initial knowledge graph in real time based on the dynamic changes in acupuncture clinical practice to form a dynamic knowledge graph includes: Based on the dynamic changes in acupuncture clinical practice, second knowledge data in the field of acupuncture is generated; Cluster analysis was performed on the second knowledge data to obtain the cluster center data for each cluster; The cluster center data for each cluster is re-extracted into triples to obtain the triples of the cluster center data; Determine the temporal similarity between the triples of the cluster center data and each triple in the initial knowledge graph of the acupuncture domain; If at least one of the temporal similarities is greater than a preset similarity threshold, then the triples of the cluster center data are added to the initial knowledge graph to obtain the dynamic knowledge graph.

4. The acupuncture large-scale model diagnostic method integrating dynamic knowledge graphs according to claim 1, characterized in that, The process of extracting clinical features of patients at multiple different time points to form a clinical feature time series, and generating time-series embedded features based on the clinical feature time series, includes: The patient's tongue images and pulse waveform data at multiple different time points are periodically acquired; Feature extraction is performed on the tongue image and pulse waveform data at each time point to obtain the clinical features corresponding to the tongue image and pulse waveform data respectively. The clinical features are then arranged in chronological order to obtain the chronological order of the clinical features corresponding to the tongue image and pulse waveform data respectively. Based on the Transformer encoder, the clinical features corresponding to the tongue image and the pulse waveform data are encoded temporally using a self-attention mechanism to obtain the temporal embedding features corresponding to the tongue image and the pulse waveform data respectively.

5. The acupuncture large-scale model diagnostic method integrating dynamic knowledge graphs according to claim 4, characterized in that, The clinical features corresponding to the tongue image include multidimensional tongue body features and multidimensional tongue coating features; the clinical features corresponding to the pulse waveform data include energy distribution features. The step of extracting features from the tongue image and pulse waveform data at each time point to obtain the clinical features corresponding to the tongue image and pulse waveform data respectively includes: The tongue image is segmented to separate the tongue body image region and the tongue coating image region; Multidimensional tongue texture features and multidimensional tongue coating features are extracted from the tongue texture image region and the tongue coating image region, respectively. The pulse waveform data is preprocessed; the preprocessing includes baseline drift removal and noise reduction filtering. Short-time Fourier transform was used to perform time-frequency analysis on the preprocessed pulse waveform data to extract the energy distribution characteristics of the 5-15Hz frequency band.

6. The acupuncture large-scale model diagnostic method integrating dynamic knowledge graphs according to claim 5, characterized in that, The semantic representation learning of the dynamic knowledge graph based on the graph attention network is used to obtain the node feature information of the dynamic knowledge graph, including: A graph attention network is used to aggregate information from neighboring nodes for each node in the dynamic knowledge graph, resulting in aggregated features for each node. The aggregated features are subjected to a nonlinear transformation to obtain the semantic representation of each node; Based on the semantic representation, the node feature information of the dynamic knowledge graph is obtained.

7. The acupuncture large-scale model diagnostic method integrating dynamic knowledge graphs according to claim 5, characterized in that, The process of fusing and enhancing the temporal embedding features to obtain fused features includes: The temporal embedding features corresponding to the tongue image and the temporal embedding features corresponding to the pulse waveform data are concatenated to obtain the concatenated features. The spliced ​​features are weighted by an attention mechanism to obtain the fused features.

8. The acupuncture large-scale model diagnostic method integrating dynamic knowledge graphs according to any one of claims 1 to 7, characterized in that, The method based on reinforcement learning algorithm utilizes the node feature information and the fused features to differentiate the acupuncture treatment plan for the patient, thereby obtaining the patient's acupuncture treatment plan, including: Design an intelligent agent that dialectically addresses acupuncture treatment plans, the intelligent agent comprising a state space, an action space, and a reward function; The node feature information and the fused feature are used as state inputs in the state space to the agent; Based on the state, the reinforcement learning agent selects acupoints and combines actions in the action space to generate an acupuncture treatment plan; The acupuncture treatment plan is executed, and the weighted sum of the efficacy prediction value and the knowledge graph update gain is obtained as the reward value in the reward function; The agent's strategy is adjusted based on the reward value until the agent's strategy converges, thus obtaining the optimal acupuncture treatment plan. The efficacy prediction value is obtained by predicting the expected efficacy of a patient after receiving acupuncture treatment based on a prediction model trained on historical data; the knowledge graph update gain is obtained by evaluating the gain brought about by updating the dynamic knowledge graph based on the knowledge data collected after the acupuncture treatment plan is executed.

9. A conditioning and optimization system for acupuncture based on a dynamic knowledge graph-integrated large-scale model, characterized in that, include: The graph construction module is used to acquire first knowledge data in the field of acupuncture, and to extract entities, relationships and attributes from the first knowledge data to construct an initial knowledge graph in the field of acupuncture. The knowledge graph update module is used to update the initial knowledge graph in real time according to the dynamic changes in acupuncture clinical practice, forming a dynamic knowledge graph; The feature extraction module is used to extract the patient's clinical features at multiple different time points, form a clinical feature time series, and generate time series embedding features based on the clinical feature time series; The feature processing module is used to perform semantic representation learning on the dynamic knowledge graph based on a graph attention network to obtain the node feature information of the dynamic knowledge graph; The temporal embedding features are fused and enhanced to obtain fused features; The scheme dialectical module is used to dialectically analyze the patient's acupuncture treatment plan based on the node feature information and the fusion feature using a reinforcement learning algorithm, thereby obtaining the patient's acupuncture treatment plan.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor performs the steps of the acupuncture large model diagnostic method that integrates dynamic knowledge graphs as described in any one of claims 1-8.

Images