Traditional Chinese Medicine Intelligent Inquiry Method and System Based on AI Large Language Model

Abstract

The present invention discloses a traditional Chinese medicine intelligent consultation method and system based on an AI large language model. The system includes a data acquisition module, an image processing module, a semantic analysis module, a disease database management module, a multimodal fusion module, a neural network retrieval module, an interaction optimization module, and an output module. The intelligent consultation method includes obtaining facial images, body images, and tongue coating images of a patient; constructing a disease database; based on the image data, collecting the voice responses of the patient, extracting the semantic content and language breath features in the voice, and the language breath features include speech rate, pitch, and pause frequency; converting the semantic content of the patient into traditional Chinese medicine professional terms, associating with the synonym list in the disease database, and normalizing inconsistent terms; fusing the image data, normalized terms, and language breath features through a multimodal neural network model; based on the comprehensive feature vector, performing a difference analysis on the remaining symptoms of the candidate diseases; and outputting the final diagnosis result.

Description

Technical Field

[0001] The present invention relates to the technical field of traditional Chinese medicine interrogation, and particularly to an intelligent interrogation method and system for traditional Chinese medicine based on an AI large language model. Background Art

[0002] Traditional Chinese medicine emphasizes adapting treatment to individual patients, times, and local conditions. According to differences in a person's constitution, gender, age, etc., as well as seasons and geographical environments, appropriate treatment methods are formulated for treatment. Due to the existence of the three factors of individual, time, and place, and the different factions of traditional Chinese medicine in different regions, the same diseases or Chinese herbal medicines may have different names due to regional reasons. Moreover, each school of traditional Chinese medicine has its own way of understanding and practicing traditional Chinese medicine, operating independently and developing separately, making it difficult to standardize. As a result, the experience of traditional Chinese medicine can only be passed down orally. Therefore, the high-quality data input into the AI large language model is limited, leading to difficulties in traditional Chinese medicine AI interrogation and low accuracy when dealing with non-reference diseases. Summary of the Invention

[0003] Aiming at the shortcoming that it is difficult to standardize the results of traditional Chinese medicine in the prior art, which leads to difficulties in AI interrogation, the present invention provides an intelligent interrogation method and system for traditional Chinese medicine based on an AI large language model.

[0004] To solve the above technical problems, the present invention is solved through the following technical solutions. The intelligent interrogation method for traditional Chinese medicine based on an AI large language model includes the following steps.

[0005] Step S1: Construct a disease library, which stores a number of disease entries. Each disease entry is associated with the cause, symptoms, medical history, living habits, and corresponding traditional Chinese medicine professional terms, and data standardization is achieved by deconstructing the disease into four elements: cause, symptoms, medical history, and living habits. When there are terms with the same semantics but different names in the input data of different schools, normalization processing is performed through a synonym table, and they are merged into a unified semantic entry.

[0006] Step S2: Obtain the facial image, body image, and tongue coating image of the patient through an image collector, and perform denoising, standardization, and feature extraction on the images to generate image data.

[0007] Step S3: Based on the image data, match the symptom features of the disease library, generate an initial question set through the AI large language model, and dynamically adjust the questioning strategy according to the patient's answer. The question generation uses the BERT model combined with a graph convolutional network to optimize the interrogation logic to ensure that the questions cover key discrimination points.

[0008] Step S4: Collect the patient's voice answer through a voice collector, extract the semantic content and language breath features, including speech rate, pitch, pause frequency, and breathing rhythm, and use Mel-frequency cepstral coefficients and long short-term memory networks to analyze the emotional tendency.

[0009] Step S5: Convert the semantic content into traditional Chinese medicine (TCM) professional terms and perform normalization processing by associating with the synonym list in the disease database.

[0010] Step S6: Integrate the image data, normalized terms, and language breath features through a multi-modal neural network model to generate a comprehensive feature vector.

[0011] Step S7: Retrieve the matching disease entries in the disease database based on the comprehensive feature vector, screen the top N candidate diseases with confidence, and the sorting basis is the weighted score of cosine similarity and symptom coverage rate.

[0012] Step S8: Conduct a differential analysis on the remaining symptoms of the candidate diseases, generate a minimized set of quadratic problems through a sequence generation model, maximize the distinction of the remaining diseases with each question, and iteratively update until a unique disease is matched or the top N disease results are output.

[0013] Step S9: Output the diagnosis result and treatment plan, accompanied by the term explanations generated by the AI large language model for the patient to confirm.

[0014] By adopting the above technical solutions, data standardization and term normalization are achieved. By constructing a disease database and deconstructing diseases into four types of elements: cause, symptom, medical history, and living habits, the structured storage of TCM diagnosis data is realized. Combining with the synonym list, semantic normalization processing is performed on terms from different schools or regions, solving the problem of semantic confusion caused by regional or school differences in TCM terms, and improving the unity and retrievability of data. Under the action of multi-modal data fusion and dynamic interaction, through the multi-modal data collection and analysis of images, voices, and texts, combined with the dynamic adjustment of the questioning strategy by the AI model, the physiological and psychological states of patients can be captured more comprehensively. For example, the BERT model is used to optimize the interrogation logic to ensure that the questions cover key differential points, and at the same time, through voice emotion analysis, such as speech rate and breathing rhythm, to assist in judging the condition, significantly improving the accuracy of diagnosis. Under the action of intelligent diagnosis and iterative optimization, a comprehensive feature vector is generated through the fusion of a multi-modal neural network, and the weight is dynamically allocated by combining the attention mechanism, enabling the system to adapt to the feature priorities of different cases. The screening and differential analysis of candidate diseases further optimize the questioning path through the sequence generation model, gradually narrowing the diagnosis scope until a unique result is output, reducing the misdiagnosis rate and improving the efficiency, increasing the interpretability and patient participation. Finally, the output includes the treatment plan and the term explanations generated by the AI, helping patients understand the diagnostic basis. The overall method promotes the intelligence and modernization of TCM interrogation through the combination of standardized processes and AI technologies.

[0015] The present invention is further configured that in step S1, the construction method of the synonym list includes

[0016] Step S101: Collect traditional Chinese medicine terms and dialect expressions from different regions, and invite experts from regional schools to annotate the interpretations of the terms.

[0017] Step S102: Calculate the semantic similarity of terms through the Word2Vec model, and establish a synonymous relationship graph by combining expert annotations.

[0018] Step S103: Use knowledge graph technology to store term nodes and relationships, supporting dynamic updates and fuzzy queries.

[0019] By adopting the above technical solutions, focusing on the construction method of the synonym table, the technical effects include solving the problem of regional differences in terms. By collecting traditional Chinese medicine terms and dialect expressions from different regions and inviting experts to annotate the interpretations, a term knowledge base covering multiple regions is established. For example, "excessive liver fire" may have different expressions in different regions, such as "hyperactivity of liver yang". Through expert annotation, semantic consistency is ensured, avoiding diagnostic deviations caused by term differences. In addition, for semantic similarity calculation and knowledge graph storage, the Word2Vec model is used to calculate the semantic similarity of terms, and a synonymous relationship graph is constructed by combining expert annotations, realizing the dynamic association and efficient retrieval of terms. The application of knowledge graph technology supports fuzzy queries, such as automatically matching normalized terms when inputting dialect words, improving the flexibility and practicality of the system. There are also effects on dynamic updates and clinical adaptability. The knowledge graph supports dynamic updates and can continuously optimize the synonym table as new terms or clinical cases accumulate. For example, when a new disease description appears in a certain region, the system can quickly associate it with existing terms, ensuring that the diagnostic model is always synchronized with clinical practice and enhancing the long-term applicability of the system.

[0020] The present invention is further configured such that in step S8, the sequence generation model is a Transformer model architecture, the input is the remaining symptom feature vectors of the candidate disease, and the output is the minimized set of questions; the objective function is designed as

[0021]

[0022] where is the objective function for model optimization; T is the termination index, indicating the total number of rounds of questions; t = 1 is the starting index, indicating the start of the first round of questions; The conditional probability represents the probability of generating question q t at the t-th round of questioning, given all previous questions q <t and the comprehensive feature vector F; λ is the expert knowledge distillation weight, is used to measure the difference between two probability distributions; is the question priority distribution generated by the model; The priority distribution of questions marked by TCM experts; the base of the logarithmic symbol log is the natural constant e, and the logarithmic operation in KL divergence is the same.

[0023] By adopting the above technical solution, the Transformer model is used to generate a minimized set of questions. By analyzing the remaining symptom characteristics of the candidate diseases, the questions with the highest discrimination are dynamically generated. For example, for the identification of headaches and migraines, the model can give priority to asking whether there is photophobia or nausea, thereby quickly narrowing the scope of diagnosis and reducing unnecessary rounds of questions. The KL divergence constraint is introduced in the objective function to align the question priority generated by the model with the annotation distribution of Chinese medicine experts. For example, experts may pay attention to the color of the tongue coating rather than lifestyle habits. By learning this priority, the model ensures that the question logic is consistent with clinical experience, avoids logical errors caused by data bias in the AI ​​model, and integrates expert knowledge to improve reliability. Through the design of the termination index T, the system can automatically decide whether to continue asking questions based on the current diagnostic confidence. For example, when the similarity difference of the candidate diseases is large enough, the remaining questions are terminated in advance to shorten the consultation time and balance efficiency and accuracy.

[0024] The present invention is further configured such that, in step S9, the method for generating term explanations is to input TCM professional terms and classic annotations into a large language model to generate popular explanations, and optimize the explanation content through patient feedback.

[0025] By adopting the above technical solutions, the patient's cognitive threshold is lowered. Through the large language model, TCM professional terms such as qi stagnation and blood stasis are converted into popular explanations, such as pain caused by poor circulation of qi and blood, to help patients understand the nature of the disease and reduce anxiety or misunderstanding caused by obscure terminology. The explanation content is dynamically optimized and the generation logic is adjusted based on patient feedback. For example, if most patients are confused about the explanation of dampness blocking the spleen, the system can automatically optimize it to a description that is closer to daily life, such as excessive moisture in the body affects digestive function, thereby improving communication effectiveness, enhancing doctor-patient trust, and providing sources of annotations in classics, such as relevant discussions in the "Yellow Emperor's Internal Classic", to make the explanation more authoritative. Patients can enhance their sense of identity with the diagnosis results and promote treatment compliance by comparing their own symptoms with descriptions in classics.

[0026] The present invention is further configured such that, in step S4, the language tone characteristics further include sentiment analysis and voiceprint recognition, the sentiment analysis identifies positive emotional states or negative emotional states through a pre-trained model RoBERTa, and the voiceprint recognition uses an i-vector algorithm to distinguish patient identities.

[0027] By adopting the above technical solutions, for the auxiliary diagnosis of emotional states, the RoBERTa model is used to identify positive or negative emotional states, where negative emotional states include emotions such as anxiety and exhaustion. Combining with the theory of traditional Chinese medicine that emotions cause diseases, it provides additional basis for diagnosis. For example, anxiety may be related to stagnation of liver qi. The system can adjust the key points of questioning or prescription suggestions accordingly. Voiceprint recognition prevents data confusion. The i-vector algorithm is used to distinguish the identities of patients to ensure that the consultation data is bound to specific patients. For example, in the scenario of shared household devices, it prevents the cross-contamination of voice data of different users and guarantees data privacy and diagnostic accuracy. Multi-dimensional feature fusion integrates emotional analysis and voiceprint recognition into the voice feature vector, enabling the system to not only analyze semantic content but also capture the physiological state of patients. For example, rapid breathing indicates cardiopulmonary problems, improving the comprehensiveness of diagnosis.

[0028] The present invention is further configured to further include step S10, where the patient uploads the curative effect feedback after taking the medicine. The curative effect feedback obtains the facial image, body image, and tongue coating image of the patient after taking the medicine through an image acquisition device, and analyzes the relevance between the prescription and the curative effect through the LIME algorithm to dynamically adjust the dosage or replace the medicinal materials.

[0029] By adopting the above technical solutions, the treatment plan is optimized in a closed loop. After the patient takes the medicine, the curative effect feedback is uploaded, such as the image of the change in the tongue coating. The system analyzes the relevance between the prescription ingredients and the curative effect through the LIME algorithm. For example, if it is found that a certain medicinal material, such as Coptis chinensis, has a significant effect on a specific patient group, the dosage can be dynamically adjusted or the medicinal material can be replaced, such as Scutellaria baicalensis, to achieve personalized treatment. Data-driven clinical decision-making. By continuously collecting curative effect data, a prescription-curative effect association database is constructed to provide reference for subsequent diagnoses. For example, if a certain prescription is generally effective in patients with damp-heat constitution, the system can give priority to recommending this plan to improve the treatment efficiency. Enhance the self-learning ability of the system. The curative effect feedback data is fed back to the disease library and model training, enabling the system to adapt to changes in clinical practice, such as newly emerging disease variants, and maintaining the timeliness of the diagnostic model.

[0030] The present invention is further configured to be a traditional Chinese medicine intelligent consultation system based on the AI large language model, including

[0031] Data acquisition devices, including an image collector and a sound collector, for obtaining the facial image, body image, tongue coating image, and voice answers of the patient;

[0032] The image processing module denoises, standardizes the image, and extracts structured features;

[0033] The semantic analysis module extracts semantic content and language breath features according to the voice answer, converts the answer into traditional Chinese medicine professional terms through natural language processing technology, and associates with the synonym table. The language breath features include emotional analysis and voiceprint recognition;

[0034] A disease library that stores disease entries, synonym mappings, and dynamically updated clinical weights;

[0035] A multi-modal fusion module that fuses image, text, and speech features to generate a comprehensive feature vector;

[0036] A neural network retrieval module that screens candidate diseases through a deep matching model and sorts them by confidence;

[0037] An interaction optimization module that generates targeted questions and optimizes term explanations;

[0038] An output module that provides a diagnosis result, a treatment plan, and a visual explanation interface.

[0039] By adopting the above technical solutions, modular design improves scalability. Through the collaborative work of independent modules, such as image processing, semantic analysis, and multi-modal fusion, the system can flexibly connect to new devices or algorithms. For example, when a pulse sensor is added in the future, only the data acquisition module needs to be extended, and there is no need to reconstruct the overall architecture. The whole process is fully automated for diagnosis, from data acquisition, image and voice to feature extraction, term normalization, multi-modal fusion, and finally generating a diagnosis result, realizing the intelligence of the whole process of medical consultation, reducing the need for manual intervention, and ensuring standardization and traceability. The disease library stores clinical weights and synonym mappings to ensure the transparency of the diagnosis basis. For example, doctors can trace the matching disease entries and their weight distributions for a certain diagnosis, which is convenient for review or teaching applications.

[0040] The present invention is further configured to further include a collaborative medical consultation module, and the collaborative medical consultation module includes a data collaborative acquisition unit and a collaborative medical consultation unit.

[0041] The collaborative acquisition unit is connected to the data acquisition module, and obtains the facial image, body image, and tongue coating image of the patient through an image collector, and synchronously records the conversation content between the doctor and the patient through a voice collector.

[0042] The collaborative medical consultation unit is used to obtain an initial question set and a secondary question set, obtain the doctor's question set through voiceprint recognition, compare the initial question set and the secondary question set with the doctor's question set respectively, and generate a set of questions not asked by the doctor in real time.

[0043] Each disease entry in the disease library is associated with a case library. The collaborative medical consultation module stores the doctor's question set, the patient's answer set, the doctor's diagnosis result, the doctor's treatment plan, and the disease information obtained by the data acquisition module in the case library corresponding to the disease entry in the disease library, and establishes a traceable electronic medical record card.

[0044] By adopting the above technical solutions, doctor-patient collaboration enhances the reliability of diagnosis. By synchronously recording the conversations between doctors and patients, the system compares the set of questions generated by AI with the actual questions asked by doctors to generate a collection of unasked questions. For example, if a doctor does not ask about night sweats, the system can prompt for supplementation to reduce the risk of missed diagnosis. The traceability of the electronic medical record card correlates and stores the doctor's diagnosis process, the patient's answers, and the data collection results in the case library to form a complete electronic medical record. For example, during future follow-up consultations, doctors can quickly access historical consultation records to improve the quality of continuous medical care. Knowledge accumulation and sharing: The continuous accumulation of the case library provides real clinical data for the AI model, promoting system iteration. For example, the diagnostic experience of rare diseases can be shared through the case library to other medical institutions to promote the standardized dissemination of traditional Chinese medicine knowledge.

[0045] The present invention is further configured such that the collaborative consultation module further includes a system follow-up unit. The system follow-up unit semantically compares the doctor's diagnosis result with the system's diagnosis result. When the semantic comparison result shows different semantics, the system follow-up unit asks the doctor differential questions and records the reasons described by the doctor for the differences to generate a difference report.

[0046] By adopting the above technical solutions, a difference analysis and error correction mechanism: When the system's diagnosis result is inconsistent with the doctor's, it automatically generates differential questions, such as why qi stagnation and blood stasis are excluded, and records the reasons explained by the doctor. For example, a doctor may discover special symptoms based on clinical experience. The system enriches the exception rules of the disease library by recording these cases, improving the system's learning ability. The difference report is fed back to the case library for optimizing model parameters. For example, if multiple doctors all point out that the associated features of a certain disease are ignored by the system, the model can adjust the feature weights to reduce future misjudgments and promote human-machine collaborative decision-making. Through the differential question mechanism, the system not only assists doctors in diagnosis but also learns tacit knowledge from doctors, forming a virtuous cycle of AI assistance - doctor decision-making - system optimization.

[0047] The present invention is further configured such that the difference report is sent to the medical record library corresponding to the case entry in the disease library. The disease library is configured with a data correction unit. When the number of difference reports for the same disease entry is higher than a preset number, the data correction unit sends the case entry to several traditional Chinese medicine experts with a preset professional title or above for a review of the diagnosis results of the same disease entry.

[0048] By adopting the above technical solutions, expert review ensures data authority. When the difference report of a certain disease entry exceeds the threshold, the expert review process is automatically triggered. For example, multiple chief physicians confirm the diagnostic criteria for spleen deficiency with dampness entrapment to ensure the authority and clinical applicability of the disease database. The association rules or weights of disease entries are corrected based on the expert review results. For example, if the correlation between a certain disease and thick and greasy tongue coating is overestimated, its weight can be lowered to improve the accuracy of subsequent diagnoses and dynamically optimize the disease database. A professional title threshold is preset, such as above deputy chief physician, to ensure the qualifications of review experts and avoid misoperations by junior doctors affecting the core data of the system, meeting the compliance requirements of the medical industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic structural diagram of a traditional Chinese medicine intelligent consultation method based on an AI large language model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0050] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0051] Embodiment:

[0052] A traditional Chinese medicine intelligent consultation method based on an AI large language model includes the following steps:

[0053] Step S1: Construct a disease database that stores a number of disease entries. Each disease entry is associated with causes, symptoms, medical history, living habits, and corresponding traditional Chinese medicine professional terms, and data standardization is achieved by deconstructing diseases into four elements: causes, symptoms, medical history, and living habits; when terms with the same semantics but different names are input in different schools of thought, they are normalized through a synonym table and merged into a unified semantic entry;

[0054] Step S2: Obtain the facial image, body image, and tongue coating image of the patient through an image collector, and perform denoising, standardization, and feature extraction on the images to generate image data;

[0055] Step S3: Match the symptom features of the disease database based on the image data, generate an initial question set through an AI large language model, and dynamically adjust the questioning strategy according to the patient's answers; the question generation adopts a BERT model combined with a graph convolutional network to optimize the consultation logic and ensure that the questions cover key differential points;

[0056] Step S4: Collect the patient's voice answers through a voice collector, extract the semantic content and language breath features, including speech rate, pitch, pause frequency, and breathing rhythm, and use Mel frequency cepstral coefficients and long short-term memory networks to analyze the emotional tendency;

[0057] Step S5: Convert the semantic content into traditional Chinese medicine professional terms and perform normalization processing by associating with the synonym table of the disease database;

[0058] Step S6: Integrate the image data, normalized terms, and linguistic breath features through a multimodal neural network model to generate a comprehensive feature vector;

[0059] Step S7: Retrieve the matching disease entries in the disease library based on the comprehensive feature vector, screen the top N candidate diseases with confidence, and the sorting basis is the weighted score of cosine similarity and symptom coverage rate;

[0060] Step S8: Conduct a differential analysis on the remaining symptoms of the candidate diseases, generate a minimized set of quadratic problems through a sequence generation model, and maximize the distinction of the remaining diseases with each question, and iterate and update until a unique disease is matched or the top N disease results are output;

[0061] Step S9: Output the diagnosis result and treatment plan, accompanied by the term explanations generated by the AI large language model for the patient to confirm;

[0062] Step S10: After the patient takes the medicine, upload the efficacy feedback. The efficacy feedback obtains the facial image, body image, and tongue coating image of the patient after taking the medicine through an image acquisition device, and analyzes the relevance between the prescription and the efficacy through the LIME algorithm to dynamically adjust the dosage or replace the medicinal materials.

[0063] In step S1, the construction method of the synonym table includes:

[0064] Step S101: Collect traditional Chinese medicine terms and dialect expressions in different regions, and invite experts from regional schools to annotate the interpretations of the terms;

[0065] Step S102: Calculate the semantic similarity of the terms through the Word2Vec model, and establish a synonym relationship graph in combination with the expert annotations;

[0066] Step S103: Use knowledge graph technology to store the term nodes and relationships, and support dynamic update and fuzzy query.

[0067] Among them, the disease library refers to the knowledge library that stores standardized disease entries. Each entry is associated with four elements: cause, symptom, medical history, and living habits. Specifically, it can be implemented by a relational database. By deconstructing the disease elements, the differences in school expressions are eliminated. The synonym table refers to the mapping tool for term normalization. Specifically, it can be constructed through semantic similarity calculation and expert annotation to solve the problem of inconsistent regional terms. The multimodal neural network model refers to the processing architecture that integrates image, text, and speech data. Specifically, it can be implemented by a cross-modal attention mechanism to enhance the correlation analysis ability of unstructured data. The sequence generation model refers to the algorithm that optimizes the questioning strategy. Specifically, it can be implemented by the Transformer architecture to improve the disease discrimination efficiency through differential analysis.

[0068] In the data standardization stage, the disease is deconstructed into a four-element framework and a unified knowledge representation is established. When receiving terminology input from different schools, synonymous terms are mapped to standardized entries through semantic similarity calculation and combination with expert knowledge. In the image processing stage, traditional visual diagnosis is converted into computable features, and key indicators such as facial color and tongue coating morphology are extracted. The dynamic questioning strategy adjusts the question dimension based on real-time feedback, and generates a logically coherent follow-up question sequence through a large language model. Speech analysis not only extracts semantic content, but also captures sub-health indicators such as breathing rhythm, which complement image features. In the multimodal fusion stage, cross-domain feature alignment is used to eliminate noise interference from a single data source. In the difference analysis stage, the residual symptom discrimination of candidate diseases is calculated to generate a question combination with maximum information gain.

[0069] Compared with existing technologies, traditional systems rely on fixed terminology libraries, resulting in poor cross-school compatibility and an inability to effectively integrate dialect expressions. Conventional medical consultation robots use a static questionnaire model and lack the ability to dynamically adjust individual signs. Existing multimodal systems often simply splice data and fail to achieve deep feature associations. Most AI diagnostic tools directly output a single result and lack differentiated decoupling and secondary verification mechanisms.

[0070] This application realizes the cross-school integration of TCM diagnosis and treatment data, eliminates the interference of terminology differences on AI model training, and uses a dynamic questioning mechanism to significantly improve the recognition accuracy of atypical symptoms. It narrows the diagnosis scope through iterative optimization, and effectively improves the utilization rate of physical sign information through multimodal feature fusion, making up for the limitations of a single diagnostic basis. Standardized terminology conversion ensures the accuracy of doctor-patient communication and avoids the risk of misdiagnosis due to differences in expression. The differential analysis algorithm optimizes the consultation path, reducing the number of invalid questions while ensuring diagnostic accuracy.

[0071] A method for constructing a thesaurus is proposed, including collecting traditional Chinese medicine terms and dialect expressions in different regions, and inviting experts from regional schools to annotate the meanings of the terms; calculating the semantic similarity of terms through the Word2Vec model, and establishing a synonymous relationship graph in combination with expert annotations; using knowledge graph technology to store term nodes and relationships, supporting dynamic updates and fuzzy queries. Regional school experts refer to professionals with the practicing qualifications of specific traditional Chinese medicine regional schools, and their identities can be confirmed specifically through the verification of practicing certificates, which is used to ensure the authority and accuracy of term interpretations. The Word2Vec model refers to a semantic calculation model based on the spatial distribution of word vectors, which can be specifically implemented by using the Skip-gram neural network structure. Through training with a large-scale corpus, terms are mapped into high-dimensional vectors, which are used to quantify the semantic correlation between terms. Knowledge graph technology refers to a semantic network construction method based on a graph database, which can be specifically implemented by using the Neo4j graph database. Through the node-relationship data structure, terms and their associated attributes are stored, which is used to support the visual retrieval and incremental update of term relationships. In the data collection stage, terms are extracted through traditional Chinese medicine literature and clinical records covering multiple regions, and combined with the annotations of school experts to form a set of terms with authoritative semantic annotations. In the semantic modeling stage, the Word2Vec model is used to vectorize the terms, calculate the cosine similarity to obtain preliminary synonymous relationships, and then correct the wrong associations through expert review. In the storage and application stage, a knowledge graph containing term nodes, synonymous edges, and attribute fields is constructed, and the dynamic iteration of the term library is realized through a version control mechanism, and non-standard term inputs are processed in combination with a fuzzy matching algorithm. These three stages form a complete technical chain from raw data collection to intelligent application, enabling synonymous terms in different regions to be accurately mapped to a unified semantic space.

[0072] The standardization of traditional Chinese medicine terms mainly relies on manual compilation of dictionaries, which has problems such as lagging updates and limited coverage. This solution automatically mines term associations through a machine learning model, combines expert review to improve accuracy, and uses a knowledge graph to achieve dynamic storage, enabling the standardization process of terms to adapt to regional diversity and meet the needs of real-time updates of clinical data.

[0073] Through the above technical solution, this application realizes the accurate mapping of cross-regional traditional Chinese medicine terms, solves the problem of inconsistent disease descriptions caused by dialect differences, provides a reliable data basis for the subsequent normalization process of the disease library, and effectively improves the term matching accuracy and clinical applicability of the traditional Chinese medicine intelligent consultation system.

[0074] In step S8, the sequence generation model is the Transformer model architecture, the input is the remaining symptom feature vectors of the candidate diseases, and the output is the minimized question set; the objective function is designed as:

[0075]

[0076] Among them, is the objective function for model optimization; T is the termination index, representing the total number of rounds of questions; t = 1 is the starting index, representing the start of the first round of questions; Conditional probability, representing the probability of generating question qt at the t-th round of questioning, given all previous questions q <t and the comprehensive feature vector F; λ is the expert knowledge distillation weight, is used to measure the difference between two probability distributions; is the question priority distribution generated by the model; is the question priority distribution annotated by traditional Chinese medicine experts; the base of the logarithm symbol log is the natural constant e, and the same applies to the logarithmic operation in the KL divergence.

[0077] After inputting the remaining symptom feature vectors of the candidate diseases into the Transformer model, by calculating the correlation weights between symptoms, key questions that can effectively distinguish the remaining diseases are generated. In each round of questioning, the model updates the feature vectors according to the collected answer information and generates subsequent questions with the maximum information gain through the decoder. The conditional probability term in the objective function controls the minimization of the number of questioning rounds, while the KL divergence term corrects the clinical rationality of the questioning strategy by comparing the question priority predicted by the model with the standard priority annotated by experts.

[0078] When traditional methods use recurrent neural networks for sequence generation, there are problems such as weak long-range dependence capture ability and redundant questioning rounds, while the questioning strategy based on the rule engine cannot dynamically adapt to individual patient differences. This solution improves the symptom correlation analysis efficiency by two orders of magnitude through the parallel computing characteristics of the Transformer architecture. At the same time, combined with the expert knowledge distillation mechanism, the model's questioning logic is highly consistent with the traditional Chinese medicine syndrome differentiation thinking while maintaining the ability of autonomous decision-making.

[0079] This application can reduce the number of questions required to distinguish similar diseases to 30% - 50% of traditional methods while ensuring the integrity of traditional Chinese medicine theory. At the same time, through the collaborative optimization of expert experience and model decision-making, the generated question set covers key differential points while avoiding deviating from the common clinical interrogation path, improving the accuracy of patient responses and the diagnostic efficiency of the system.

[0080] In step S9, the generation method of the term explanation is: inputting traditional Chinese medicine professional terms and classical annotations into the large language model to generate popular explanations, and optimizing the explanation content through patient feedback.

[0081] Traditional Chinese medicine (TCM) professional terms refer to academic vocabulary with specific meanings in the field of TCM. Specifically, standardized entries from classics such as "Huangdi Neijing" and "Shanghan Lun" can be adopted to ensure the academic accuracy of the generated content. Among them, classical annotation refers to the interpretation and textual research of TCM classic literature. Specifically, it can be achieved by using the annotated versions of medical experts in past dynasties or the content of modern authoritative publications to establish the semantic connection between terms and explanations. Among them, the large language model refers to a deep neural network with natural language generation ability. Specifically, it can be implemented by using pre-trained models with GPT-3 or BERT architectures to transform professional expressions through semantic parsing and generation techniques. Among them, popularized explanation refers to text that conforms to modern spoken language habits. Specifically, it can be achieved by replacing professional terms with daily life expressions, adding metaphorical explanations, etc., to reduce the difficulty for patients to understand. Among them, patient feedback optimization refers to a dynamic adjustment mechanism based on patient interaction data. Specifically, it can be implemented by means of questionnaires, comprehensibility scores, or dialogue correction records, etc., to continuously improve the adaptability of the explanatory content.

[0082] Traditional Chinese medicine professional terms and classical annotations are input into the large language model. Through the model's in-depth parsing of the semantics of ancient books, preliminary explanatory texts are generated. During the process of explanation generation, classical annotations are used as constraints to ensure that the generated explanations conform to the TCM theoretical system. The generated popularized explanations are presented to patients through the interaction interface. At the same time, data on the confusion points marked or semantic misunderstandings of patients regarding the explanatory content are collected. These feedback data are cleaned to form training samples, which are used to fine-tune the generation strategy of the large language model, enabling it to actively avoid expression methods that are prone to misunderstanding in subsequent explanations and gradually improving the clarity and acceptance of the explanations.

[0083] Compared with the prior art, traditional TCM term explanation systems rely on a fixed entry library compiled manually and cannot dynamically adapt to the cognitive differences of different patients. However, this solution combines the generation ability of the large language model with a real-time feedback mechanism to achieve personalized adaptation of the explanatory content while maintaining the accuracy of the terms. At the same time, through the data closed-loop, the update bottleneck of the static knowledge base is broken through.

[0084] Through the above technical solution, this application effectively reduces the cognitive threshold of TCM terms, enables patients to accurately understand the professional concepts involved in the diagnosis conclusion, reduces the problems of repeated consultations or decreased treatment compliance caused by term misunderstandings. At the same time, the optimized explanatory content improves the semantic consistency of the consultation interaction data and provides high-quality labeled data for the continuous training of the AI model.

[0085] In step S4, the language breath features further include sentiment analysis and voiceprint recognition. Sentiment analysis uses the pre-trained model RoBERTa to identify positive or negative emotional states, and voiceprint recognition uses the i-vector algorithm to distinguish patient identities to prevent data confusion.

[0086] Emotion analysis refers to the identification of the implicit emotional state in the patient's speech. Specifically, it can be achieved by using natural language processing techniques based on the RoBERTa pre-trained model. The semantic understanding ability formed by training this model on a large-scale corpus enables it to capture the emotional factor characteristics required for traditional Chinese medicine syndrome differentiation. Speaker recognition refers to distinguishing different speakers by analyzing the biometric features of speech. Specifically, it can be achieved by using the i-vector algorithm combined with the Gaussian mixture model to model the acoustic feature space. This algorithm extracts the speaker identity vector to establish a unique identifier for the patient's speech data.

[0087] In the speech data processing stage, the patient's speech signal is decomposed into two dimensions: semantic content and acoustic features. For the semantic content, the RoBERTa model analyzes the emotional tendency in the sentence through the self-attention mechanism and identifies emotion indicators such as anxiety index and fatigue level. These indicators are quantified into feature vectors for assisting syndrome differentiation analysis. At the same time, the i-vector algorithm extracts physiological feature parameters such as vocal tract length and fundamental frequency from the speech signal, and generates a discriminative identity vector through factor analysis to ensure that the speech data of different patients will not be confused in the subsequent processing process. The two technologies achieve data isolation through a parallel processing architecture. The emotional feature vector and the identity identification information are respectively transmitted to the multi-modal fusion module to realize the auxiliary decision-making of the emotional state for traditional Chinese medicine syndrome differentiation on the premise of maintaining the independence of patient data.

[0088] Traditional methods mostly use keyword matching to identify the emotional state, which cannot capture the deep semantic associations in the sentences and do not establish a patient identity verification mechanism. This solution can effectively identify the implicit emotions in traditional Chinese medicine emotion-related expressions such as "irritability" and "shortness of breath" through the deep semantic understanding ability of RoBERTa. Combined with the speaker recognition technology of i-vector, it can meet the identity recognition needs of patients with different dialect accents, solve the problem of insufficient generalization ability in traditional fixed speaker template matching, accurately identify the emotional state and identity characteristics in the patient's speech, avoid syndrome differentiation deviation caused by misjudgment of emotions, and prevent cross-contamination of speech data of multiple patients, ensuring the accurate correspondence between the pathological characteristics and identity information of each patient, and effectively improving the diagnostic reliability of the traditional Chinese medicine intelligent consultation system.

[0089] This application further proposes steps for patients to upload efficacy feedback after taking medicine. The efficacy feedback obtains the facial images, body images, and tongue coating images of patients after taking medicine through an image acquisition device, analyzes the correlation between the prescription and the efficacy through the LIME algorithm, dynamically adjusts the dosage or replaces the medicinal materials. The efficacy feedback refers to the physical sign change data of patients after taking medicine, and specifically, standardized data acquisition can be achieved through a multi-modal image acquisition device for tracking the evolution of physiological characteristics under the action of the prescription. The image acquisition device includes facial, tongue coating, and body shooting modules, and uses fixed light sources and angle control to achieve image comparability across time points. The LIME algorithm refers to the local interpretable model-agnostic explanation technique, and specifically, it can identify the influence intensity of the key medicinal material dosage on the efficacy based on the correlation weight analysis between the prescription ingredients and the improvement of physical signs, so as to provide an interpretable basis for dosage adjustment.

[0090] During the medication cycle, patients regularly upload facial, tongue coating, and body images. Sign feature vectors are extracted through standardized image processing. The feature vectors are compared with the baseline data in the initial diagnosis stage to generate a physical sign change trajectory. The LIME algorithm conducts attribution analysis on the dosage parameters of each medicinal material in the prescription and the physical sign changes, calculates the contribution degree of each medicinal material to the improvement of specific physical signs. When a certain physical sign fails to achieve the expected improvement, the algorithm preferentially adjusts the dosage of the medicinal material with a low contribution degree or replaces the medicinal material with the same efficacy, forming a closed-loop feedback mechanism for prescription optimization. For example, if the improvement of thick and greasy tongue coating is insufficient, and the algorithm detects that the dosage contribution degree of the dampness-resolving medicinal materials in the prescription is lower than the threshold, it will automatically increase the dosage or replace them with the same type of medicinal materials with stronger medicinal properties.

[0091] The system involved in the traditional Chinese medicine intelligent consultation method based on the AI large language model includes:

[0092] Data acquisition device: including an image collector and a sound collector, used to obtain the facial images, body images, tongue coating images, and voice responses of patients;

[0093] Image processing module: denoises, standardizes the images, and extracts structured features;

[0094] Semantic analysis module: extracts semantic content and language breath features from the voice response, converts the response into traditional Chinese medicine professional terms through natural language processing technology, and associates the synonym table. The language breath features include sentiment analysis and voiceprint recognition;

[0095] Disease library: stores disease entries, synonym mappings, and dynamically updated clinical weights;

[0096] Multi-modal fusion module: used to fuse image, text, and voice features to generate a comprehensive feature vector;

[0097] Neural network retrieval module: screens candidate diseases through a deep matching model and sorts them according to the confidence level;

[0098] Interaction Optimization Module: Generate targeted questions and optimize term explanations;

[0099] Output Module: Provide diagnostic results, treatment plans, and a visual explanation interface;

[0100] Collaborative Inquiry Module: The collaborative inquiry module includes a data collaborative collection unit and a collaborative inquiry unit;

[0101] The collaborative collection unit is connected to the data collection module, obtains the facial image, body image, and tongue coating image of the patient through an image collector, and synchronously records the conversation content between the doctor and the patient through a voice collector.

[0102] The collaborative inquiry unit is used to obtain an initial question set and a secondary question set, obtains the doctor's question set through voiceprint recognition, compares the initial question set and the secondary question set with the doctor's question set respectively, and generates a set of questions not asked by the doctor in real time.

[0103] Each disease entry in the disease library is associated with a case library. The collaborative inquiry module stores the doctor's question set, the patient's answer set, the doctor's diagnosis result, the doctor's treatment plan, and the disease information obtained by the data collection module in the case library corresponding to the disease entry in the disease library, and establishes a traceable electronic medical record card.

[0104] The collaborative inquiry module also includes a system follow-up unit. The system follow-up unit semantically compares the doctor's diagnosis result and the system's diagnosis result. When the semantic comparison result is different semantics, the system follow-up unit asks the doctor for differences and records the reasons for the doctor's description differences, generating a difference report.

[0105] The difference report is sent to the medical record library corresponding to the case entry in the disease library. The disease library is configured with a data correction unit. When the number of difference reports for the same disease entry is higher than the preset number, the data correction unit sends the case entry to several traditional Chinese medicine experts with a preset professional title or above to review the diagnosis results of the same disease entry.

[0106] Among them, the data acquisition device refers to an image acquisition device for synchronously obtaining facial images, body images, and tongue coating images, and an audio input device for collecting voice responses. Specifically, it can be implemented using a camera device with multi-spectral imaging capabilities and a microphone array with noise reduction functions. By collecting multi-modal data, it overcomes the limitation of traditional medical inquiries relying on single information. The image processing module refers to an algorithm module for denoising, filtering, unifying resolution, and extracting features from the collected images. Specifically, it can be implemented using an image enhancement model based on a convolutional neural network and edge computing devices to eliminate the influence of shooting environment differences on the recognition of physical signs. The semantic analysis module refers to a text processing unit that converts natural language expressions into standardized traditional Chinese medicine terms. Specifically, it can be implemented using a BERT model combined with a synonym mapping table to solve the problem of differences in school terms through automatic conversion of dialect terms. The disease database refers to a structured database that stores disease entries and their associated clinical weights. Specifically, it can be implemented using a graph database combined with a dynamic weight update algorithm to support the continuous absorption of clinical data and the optimization of diagnostic logic. The multi-modal fusion module refers to a deep learning model that integrates visual features, text features, and voice features. Specifically, it can be implemented using a cross-modal attention mechanism combined with a feature concatenation method to simulate the comprehensive diagnostic thinking of traditional Chinese medicine's "inspection, auscultation and olfaction, interrogation, and palpation". The neural network retrieval module refers to a disease screening system based on deep matching. Specifically, it can be implemented using a siamese network combined with cosine similarity calculation to achieve multi-dimensional matching of symptoms rather than simple keyword comparison. The interaction optimization module refers to a dialogue management system that dynamically generates questioning strategies. Specifically, it can be implemented using a reinforcement learning framework combined with a sequence generation model to make up for the one-way collection defect of traditional AI medical inquiries through two-way information interaction. The output module refers to a human-computer interaction interface that displays the diagnostic results and treatment plans. Specifically, it can be implemented using visual charts combined with natural language generation technology to convert professional terms into intuitive expression forms.

[0107] Specifically, after the image acquisition device obtains the patient's physical sign images, the image processing module eliminates device differences through noise suppression and size normalization, and extracts structured features such as the color of the tongue body and the distribution of cracks. The semantic analysis module performs entity recognition and relationship extraction on the voice response, and converts dialect expressions such as "hyperactivity of liver fire" into standard terms in combination with the synonym table. While storing the basic disease data, the disease database updates the clinical weight coefficients of each symptom through an online learning mechanism. The multi-modal fusion module performs cross-modal association on the tongue coating texture features and voice emotion features to generate a comprehensive vector representing the overall pathogenesis. The neural network retrieval module performs similarity ranking in the disease database based on the comprehensive vector, and screens out the top N candidate diseases. The interaction optimization module generates a minimized set of questions for secondary confirmation based on the remaining symptom differences of the candidate diseases. The output module converts professional terms such as "Shaoyang syndrome" in the final diagnostic results into visual heat maps and natural language explanations.

[0108] This application solves the standardization problem caused by term differences in the process of traditional Chinese medicine diagnosis. For example, different expressions such as "liver depression transforming into fire" and "wood stasis transforming into fire" are unified into standardized disease entries, achieving the effective integration of multi-source data. For example, the characteristics of yellow and greasy tongue coating and the feature of rapid speech are jointly used as the judgment basis for damp-heat syndrome, optimizing the dynamic interaction process. For example, differentiating questions are automatically generated according to the preliminary diagnosis results, reducing the repeated inquiry of ineffective questions, and finally improving the recognition accuracy of atypical symptoms. For example, in the syndrome of intermingled cold and heat, misjudgment is avoided through cross-validation of multi-modal features.

[0109] The collaborative acquisition unit refers to the full-element recording of the doctor's interrogation process through a multi-modal data synchronous acquisition device. Specifically, the timestamp alignment technology can be used to achieve the spatio-temporal consistency of image and voice data. This unit integrates visual and auditory data streams to ensure that term expressions from different sources form associated mappings on the unified time axis. The collaborative interrogation unit refers to the question difference detection module based on voiceprint feature recognition and semantic comparison. Specifically, the speaker separation algorithm can be used to extract the doctor's voice segment, and the text similarity calculation model is combined to identify the uncovered question dimensions. The case database refers to the dynamic database that structurally stores the data of the diagnosis and treatment process. Specifically, the graph database technology can be used to establish multi-dimensional associations between disease entries and clinical data, and fast retrieval is achieved through hash indexing. The traceable electronic medical record card refers to the recording unit with data source identification and operation logs. Specifically, the blockchain technology can be used to achieve the immutable storage of diagnosis and treatment data, and the data change process is recorded through the timestamp chain structure.

[0110] During the traditional interrogation process of doctors, the image collector and the sound collector synchronously collect the patient's physiological characteristics and the content of the doctor-patient conversation. The voiceprint recognition algorithm separates the voice segments of doctors and patients in real time, converts the doctor's question content into structured text. The initial question set is generated by the AI system according to the patient's image characteristics, and the secondary question set is dynamically optimized based on the difference points of candidate diseases. By comparing the question set generated by the AI with the actual question content of the doctor, the system identifies the symptom inquiry points not covered by the doctor, such as the discrimination of the tongue coating color or the pursuit of specific body posture characteristics. The identified differential question items are associated and stored with the corresponding disease entries through the case database, forming complete medical record data including the interrogation path, diagnosis basis, and treatment plan. When doctors in different regions use dialect terms to describe the same disease, the electronic medical record card realizes the context backtracking of term expressions through timestamps and data source identifications, providing a correction basis for subsequent synonym normalization.

[0111] Traditional Chinese medicine electronic medical record systems only record diagnostic results and prescription information, and cannot completely preserve the doctor's inquiry path and decision-making logic. In the existing technology, the independently stored imaging reports and text medical records have data fragmentation problems, resulting in a lack of context correlation during term normalization. This solution enables cross-verification of doctors' experience data and machine analysis results in the spatio-temporal dimension through multi-modal data synchronous acquisition and structured storage. For example, when the AI system detects that a doctor has not asked about key symptoms, it can automatically complete the inquiry dimension and record the differences. In contrast, traditional systems can only unidirectionally receive doctors' input information. This application solves the problem of symptom omission caused by differences in doctors' experience during Chinese medicine inquiries, such as avoiding the lack of syndrome differentiation elements caused by differences in school terms. By comparing the AI suggestions with the doctors' actual question content in real time, it ensures the complete coverage of key disease symptom characteristics. The structured electronic medical record card provides a data basis for subsequent case backtracking and knowledge correction. For example, when different doctors have diagnostic disagreements on the same disease item, the historical inquiry path can be retrieved to analyze the root cause of the differences. In addition, the spatio-temporal alignment mechanism of multi-modal data effectively eliminates feature contradictions caused by asynchronous acquisition times, such as the consistency verification of the tongue coating image and the breathing rhythm in the voice description on the time axis.

[0112] When the system detects a semantic disagreement between the doctor's diagnosis and the system's recommendation, it triggers a follow-up mechanism. The two diagnostic texts are vectorized and encoded by a comparison model, and the semantic similarity threshold is calculated. For example, when the cosine similarity is set to be lower than 0.7, it is determined as a substantial difference. The system automatically generates an interactive interface containing the classification of the difference type, guiding the doctor to select or fill in the specific reasons, such as differences in term dialects, school syndrome differentiation, or unrecognized patient individual characteristics. All difference data is stored in the medical record database in a structured form and a two-way index is established with the corresponding disease item. When the number of difference reports accumulated for the same disease item reaches the preset quantity threshold, it triggers a data correction process, and the relevant cases are sent to the Chinese medicine expert committee for review. The term mapping relationship or the syndrome differentiation logic weight of the disease database is updated according to the opinions of the majority of experts.

[0113] When there is a semantic difference between the system and the doctor in the diagnostic results for the same disease item, the difference report is automatically classified into the electronic medical record database corresponding to the disease item, forming traceable structured data. The data correction unit continuously monitors the number of difference reports under each disease item. When the cumulative amount of difference reports for a certain disease item exceeds the preset threshold, this unit will trigger a review process, and multiple Chinese medicine experts meeting the qualification requirements are selected from the expert database. Based on clinical experience and standardized evaluation criteria, the experts conduct a collective study of the controversial cases, focusing on checking for diagnostic deviations caused by regional term differences or atypical symptoms. The review results form a final conclusion through a weighted voting mechanism. If it is confirmed that there are data defects in the original disease item, the item revision process is initiated to update core parameters such as the cause association and symptom weight, realizing the dynamic optimization of the disease database.

[0114] When there is a semantic difference between the system's and the doctor's diagnosis results for the same disease entry, the difference report is automatically classified into the electronic medical record library corresponding to the disease entry, forming traceable structured data. The data correction unit continuously monitors the number of difference reports under each disease entry. When the cumulative amount of difference reports for a certain disease entry exceeds the preset threshold, this unit will trigger a review process, select multiple qualified traditional Chinese medicine experts from the expert database. Based on clinical experience and standardized evaluation criteria, the experts conduct a collective judgment on the controversial cases, focusing on verifying whether there are diagnostic deviations caused by regional term differences or atypical symptoms. The review result forms the final conclusion through a weighted voting mechanism. If it is confirmed that there are data defects in the original disease entry, the entry revision process is initiated to update core parameters such as etiology association and symptom weights, realizing the dynamic optimization of the disease library.

[0115] Traditional Chinese medicine AI systems usually adopt static disease libraries, lacking a difference feedback and active correction mechanism, resulting in the continuous accumulation of term differences and diagnostic contradictions, affecting data credibility. However, this solution can systematically eliminate data deviations caused by inconsistent regional terms or atypical cases while retaining the diversity of schools of thought by constructing a closed-loop correction system driven by difference reports, enabling the disease library to have the ability of adaptive evolution.

[0116] This application solves the problem of decreased data credibility caused by inconsistent names of traditional Chinese medicine diagnostic terms due to differences in regional schools of thought. It effectively eliminates the cumulative effect of diagnostic result differences through an expert review mechanism, while enhancing the adaptability of the AI consultation system to atypical cases. The directional filing and threshold triggering mechanism of difference reports ensure the timeliness of data correction, and the multi-expert double-blind review mode avoids the subjective influence of a single school of thought. The dynamically revised disease library can accurately reflect the true diagnosis and treatment rules in different clinical scenarios.

Claims

1. The intelligent TCM consultation method based on AI large language model is characterized by: The following steps are involved: Step S1: construct a symptom database, storing a number of symptom entries, each of which is associated with the cause of disease, symptoms, medical history, lifestyle habits and corresponding TCM professional terms, and realize data standardization by deconstructing the disease into four elements: cause of disease, symptoms, medical history and lifestyle habits; when different schools input terms with the same semantics but different names in the symptom entries, they are normalized through a synonym table and merged into symptom entries with unified semantics; Step S2: Obtain the patient's facial image, body image and tongue coating image through an image collector, and perform denoising, standardization and feature extraction on the image to generate image data; Step S3: Based on the image data, the disease characteristics of the disease database are matched, an initial set of questions is generated through the AI ​​large language model, and the questioning strategy is dynamically adjusted according to the patient's answer; Step S4: Collect the patient's voice response through a sound collector, extract the semantic content and language characteristics, including speech rate, pitch, pause frequency and breathing rhythm, and use Mel frequency cepstral coefficients and long short-term memory network to analyze the emotional tendency; Step S5: converting the semantic content into TCM professional terms, and normalizing the synonym table of the associated disease database to generate normalized terms; Step S6: fusing image data, normalized terms and language atmosphere features through a multimodal neural network model to generate a comprehensive feature vector; Step S7: searching for matching symptom entries in the symptom database based on the comprehensive feature vector, screening the top N candidate symptom with the highest confidence, and sorting them based on the weighted score of cosine similarity and symptom coverage; Step S8: Perform differential analysis on the remaining symptoms of the candidate diseases, generate a minimized set of secondary questions through a sequence generation model, and maximize the differentiation of the remaining diseases each time, iterating and updating until a unique disease is matched or the top N disease results are output; Step S9: Output the diagnosis results and treatment plan, with explanations of terms generated by the AI ​​large language model, for the patient to confirm.

2. The intelligent TCM diagnosis method based on AI large language model according to claim 1 is characterized in that: In step S1, the method for constructing the synonym table includes: Step S101: Collect TCM terms and dialect expressions from different regions, and invite regional experts to annotate the terms; Step S102: Calculate the semantic similarity of terms through the Word2Vec model, and establish a synonym relationship map in combination with expert annotations; Step S103: Use knowledge graph technology to store term nodes and relationships, and support dynamic updates and fuzzy queries.

3. The intelligent TCM diagnosis method based on AI large language model according to claim 1 is characterized in that: In step S8, the sequence generation model is a Transformer model architecture, the input is the residual symptom feature vector of the candidate disease, and the output is the minimized question set; the objective function is designed as: in, is the objective function of model optimization; T is the end index, indicating the total number of rounds of questions; t=1 is the start index, indicating the start of the first round of questions; Conditional probability, indicating the generation of question q in round t t The probability of, given all the previous questions q <t and comprehensive feature vector F; λ is the expert knowledge distillation weight, It is used to measure the difference between two probability distributions; is the question priority distribution generated by the model; The priority distribution of questions marked by TCM experts; the base of the logarithmic symbol log is the natural constant e, and the logarithmic operation in KL divergence is the same.

4. The intelligent TCM diagnosis method based on AI large language model according to claim 1 is characterized in that: In step S9, the method for generating term explanations is: inputting TCM professional terms and classic annotations into the large language model, generating popular explanations, and optimizing the explanation content through patient feedback.

5. The intelligent TCM diagnosis method based on AI large language model according to claim 1 is characterized in that: In step S4, the language tone features further include sentiment analysis and voiceprint recognition. The sentiment analysis identifies positive emotional states or negative emotional states through the pre-trained model RoBERTa, and the voiceprint recognition uses the i-vector algorithm to distinguish the patient's identity.

6. The intelligent TCM diagnosis method based on AI large language model according to claim 1 is characterized in that: The process also includes step S10: the patient uploads feedback on the efficacy of the medicine after taking the medicine. The feedback on the efficacy is obtained by using an image acquisition device to obtain facial images, body images and tongue coating images of the patient after taking the medicine. The correlation between the prescription and the efficacy is analyzed by using the LIME algorithm to dynamically adjust the dosage or replace the medicinal materials.

7. The intelligent TCM consultation system based on AI large language model is applicable to the intelligent TCM consultation method based on AI large language model according to any one of claims 1 to 6, characterized in that: include: Data acquisition equipment: including image collectors and sound collectors, used to obtain the patient's facial image, body image, tongue coating image and voice answer; Image processing module: denoise, standardize, and extract structural features of images; Semantic analysis module: extracts semantic content and language characteristics based on voice answers, converts answers into TCM professional terms through natural language processing technology, and associates a synonym table. The language characteristics include sentiment analysis and voiceprint recognition; Symptom database: stores symptom entries, synonym mappings, and dynamically updated clinical weights; Multimodal fusion module: used to fuse image, text and speech features to generate a comprehensive feature vector; Neural network retrieval module: Screen candidate diseases through deep matching models and sort them by confidence; Interaction optimization module: generates targeted questions and optimizes term explanations; Output module: provides diagnostic results, treatment plans and visual explanation interface.

8. The intelligent TCM consultation system based on AI large language model according to claim 7 is characterized in that: It also includes a collaborative consultation module, which includes a collaborative data collection unit and a collaborative consultation unit; The collaborative acquisition unit is connected to the data acquisition module, and acquires the patient's facial image, body image and tongue coating image through the image collector, and synchronously records the conversation between the doctor and the patient through the sound collector; The collaborative consultation unit is used to obtain an initial question set and a secondary question set, obtain a collection of doctor's questions through the voiceprint recognition, compare the initial question set and the secondary question set with the doctor's question collection respectively, and generate a collection of questions not asked by the doctor in real time; Each symptom entry in the symptom library is associated with a case library. The collaborative consultation module associates and stores the collection of doctor's questions, the collection of patient's answers, the doctor's diagnosis results, the doctor's treatment plan and the symptom information obtained by the data acquisition module in the case library corresponding to the symptom entry in the symptom library to establish a traceable electronic medical record card.

9. The intelligent TCM consultation system based on AI large language model according to claim 8 is characterized in that: The collaborative consultation module also includes a system inquiry unit, which performs a semantic comparison between the doctor's diagnosis results and the system's diagnosis results. When the semantic comparison result is a different semantics, the system inquiry unit inquires about the difference with the doctor and records the doctor's description of the reason for the difference to generate a difference report.

10. The intelligent TCM consultation system based on AI large language model according to claim 9 is characterized in that: The difference report is sent to the medical record library of the corresponding case entry in the symptom library. The symptom library is equipped with a data correction unit. When the difference report of the same symptom entry exceeds a preset number, the data correction unit sends the case entry to several Chinese medicine experts with preset titles and above to review the diagnosis results of the same symptom entry.

Images