A traditional Chinese medicine intelligent inquiry and syndrome type prescription and medicine recommendation system and method based on multi-modal fusion and syndrome element reasoning

Abstract

This invention discloses a TCM intelligent consultation and syndrome-based prescription recommendation system and method based on multimodal fusion and syndrome element reasoning, belonging to the field of TCM intelligent diagnosis and treatment technology. It collects and standardizes patient textual symptom, tongue, and pulse data, dynamically generates consultation questions to complete symptom and pulse details based on reinforcement learning algorithms, and uses dual-dimensional syndrome element reasoning based on disease nature and location. It integrates tongue and pulse features with a composite syndrome element knowledge graph to form a closed-loop reasoning logic. Based on the dual-dimensional syndrome element mapping to syndrome types, it combines the theory of auxiliary medicine to adapt targeted auxiliary medicines and verifies compatibility contraindications, collaboratively recommending TCM prescriptions and prepared Chinese medicines. Finally, it generates personalized health management suggestions integrating exercise, diet, and emotional well-being. This invention solves the problems of non-standardized symptom collection, insufficient multimodal data integration, and non-closed-loop syndrome element reasoning in existing technologies, significantly improving the accuracy of syndrome differentiation and the targeting of prescriptions, enriching medication choices, and promoting the standardization and precision of TCM diagnosis and treatment.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnosis and treatment technology in traditional Chinese medicine, specifically to an intelligent TCM consultation and syndrome-based prescription recommendation system and method based on multimodal fusion and syndrome element reasoning. Background Technology

[0002] The core of TCM diagnosis and treatment lies in syndrome differentiation and treatment, and the accuracy of the syndrome differentiation results highly depends on the physician's comprehensive collection and precise interpretation of the patient's symptoms, as well as long-term accumulated clinical experience. In the current practice of TCM diagnosis and treatment and the application of information technology, there are still many key issues that need to be addressed.

[0003] In traditional Chinese medicine (TCM) consultations, the design of questions often relies on the physician's personal experience. Differences in experience among different physicians can easily lead to incomplete symptom collection and deviations in grasping the key points. At the same time, there is a lack of a unified and standardized mapping between patients' everyday, colloquial descriptions of symptoms and standard TCM symptoms. This situation directly causes distortion in the transmission of symptom information and affects subsequent diagnostic judgments.

[0004] Traditional Chinese medicine (TCM) diagnosis requires comprehensive analysis of data from multiple dimensions, including symptom text and tongue appearance. However, most existing TCM intelligent systems only support a single text interaction mode, lacking the ability to standardize the analysis of multimodal data such as tongue appearance, and failing to effectively integrate cross-modal features. This results in the diagnostic process relying solely on single-dimensional information, making it difficult to fully reflect the essence of the patient's condition. In the core reasoning stage of syndrome differentiation, existing systems often directly map symptoms to syndromes, ignoring the core reasoning link between symptoms, syndrome elements, and syndromes in TCM theory. Furthermore, during syndrome element analysis, key influencing factors such as symptom attributes, order of appearance, and tongue appearance correlation are not fully considered, leading to consistently low accuracy in syndrome differentiation.

[0005] Regarding patient consultation guidance, the existing system's guidance strategies are rather rigid in response to patients' vague statements or incomplete answers. It fails to dynamically adjust the consultation direction based on historical dialogue context and the real-time progress of syndrome element reasoning, leading to repeated consultations or omissions of key medical information, thus affecting consultation efficiency and data integrity. Insufficient personalized recommendation capabilities are also a prominent problem with the current system. Prescription recommendations are mostly based on fixed syndrome types and prescription mapping relationships, failing to fully adapt to individual differences such as patient age, constitution, and symptom attributes. Furthermore, it lacks an integrated health management recommendation system encompassing syndrome types, prescriptions, exercise, diet, and emotional well-being, making it difficult to meet patients' diverse and personalized health needs.

[0006] With the continuous development of digitalization and intelligentization of traditional Chinese medicine, it has become particularly urgent to build an integrated system based on the core theories of traditional Chinese medicine, which integrates multimodal data collection, intelligent consultation guidance, and closed-loop syndrome element reasoning. Summary of the Invention

[0007] The purpose of this invention is to provide a TCM intelligent consultation and syndrome-based prescription recommendation system and method based on multimodal fusion and syndrome element reasoning, in order to address the shortcomings of traditional diagnosis and treatment and existing technologies, and promote the standardization, precision and popularization of TCM diagnosis and treatment.

[0008] To achieve the above objectives, the present invention employs the following technical solutions.

[0009] A TCM intelligent consultation and syndrome-based prescription recommendation system based on multimodal fusion and syndrome element reasoning includes: The multi-source multimodal data acquisition and standardization processing module is used to collect patients' textual symptoms, tongue appearance, and pulse appearance trimodal data and standardize them to obtain a standardized trimodal data set in a unified format. The intelligent consultation guidance and symptom attribute completion module, based on the standardized data set and combined with the preset consultation knowledge base, dynamically generates consultation questions through reinforcement learning algorithm, completes symptom attributes and pulse details, and obtains the completed three-modal data; The module for reasoning about the nature and location of the disease and optimizing the fusion of pulse and tongue features performs dual-dimensional reasoning on the completed three-modal data, combining the characteristics of tongue and pulse to optimize the reasoning results and obtain a set of dual-dimensional syndrome elements. The syndrome element-syndrome mapping and auxiliary drug-prescription drug-Chinese patent medicine collaborative recommendation module, based on the dual-dimensional syndrome element set, generates corresponding syndromes through the syndrome element-syndrome mapping relationship, and combines the auxiliary drug theory to match the auxiliary drugs, thereby realizing the recommendation of Chinese herbal prescriptions and Chinese patent medicines; The personalized health management suggestion generation module generates integrated health management suggestions for exercise, diet, and emotional well-being based on the syndrome type, recommended Chinese herbal prescriptions, proprietary Chinese medicines, and individual patient characteristics.

[0010] Furthermore, the multi-source, multi-modal data acquisition and standardization processing module includes: The text symptom collection and standardization parsing unit is used to collect patients' text symptom descriptions. Through a combination of TCM-specific models and rule matching, it achieves accurate mapping between colloquial descriptions and TCM standard symptoms, processes non-standard symptom descriptions, and filters invalid responses. The tongue image multimodal data acquisition and analysis unit is used to acquire images of patients' tongue images, analyze tongue image features through a multimodal model and convert them into standardized tongue image symptoms, and integrate them into a text symptom list; The pulse diagnosis multimodal data acquisition and analysis unit is used to acquire patient pulse data in dual modes, analyze it and convert it into standardized pulse features, and integrate it into the symptom list to form a trimodal fusion data set.

[0011] Furthermore, in the intelligent consultation guidance and symptom attribute completion module, the preset consultation knowledge base includes consultation questions, answer options, attribute tags, and priorities corresponding to common symptoms; the reinforcement learning algorithm is used to dynamically adjust the consultation priority, so as to obtain complete data in the fewest rounds and improve consultation efficiency and data integrity.

[0012] Furthermore, the module for synergistic recommendation of syndrome elements-syndrome types and auxiliary drugs-prescription drugs-traditional Chinese medicine preparations includes: The syndrome mapping unit is used to filter the final syndrome from the candidate syndrome pool based on a two-dimensional syndrome element set and through multi-strategy matching. In special scenarios, a fallback strategy is used to determine the syndrome, and the special scenario is when there are no clear disease and / or disease location syndrome elements. The drug matching unit is used to screen suitable drugs from the drug knowledge base based on syndrome information and add them to the prescription after verifying incompatibility. The Traditional Chinese Medicine (TCM) recommendation unit is used to select suitable TCMs from the TCM knowledge base based on syndrome differentiation and individual patient characteristics, and output relevant prompts.

[0013] A method for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning includes the following steps: S1. Multi-source multimodal data acquisition and standardization processing: Collect patient textual symptoms, tongue appearance, and pulse appearance data in three modalities and perform standardization processing to form a standardized three-modal data set that integrates symptoms, tongue appearance, and pulse appearance. S2. Intelligent consultation guidance and symptom attribute completion: Based on the standardized trimodal data set and the preset consultation knowledge base, consultation questions are dynamically generated through reinforcement learning algorithm. After receiving patient feedback, symptom attributes and pulse details are completed. The consultation process that has been covered is automatically skipped to obtain the completed trimodal data. S3, Disease nature and location syndrome element reasoning and pulse and tongue fusion optimization, perform disease nature and location dual-dimensional syndrome element reasoning on the completed three-modal data, integrate tongue and pulse characteristics and dynamic optimization of reasoning results with consultation rounds to form a dual-dimensional syndrome element set; S4. Syndrome element-syndrome type mapping and collaborative recommendation of auxiliary drugs-prescription drugs-traditional Chinese medicine: Based on the dual-dimensional syndrome element set, syndrome types are generated through multi-strategy matching. Syndrome type related information is extracted to match targeted auxiliary drugs and the drug incompatibility is checked. Auxiliary drugs are integrated with core prescriptions. At the same time, traditional Chinese medicine is recommended based on syndrome type and individual patient characteristics to form a multi-dimensional drug recommendation result. S5. Personalized health management recommendations are generated based on syndrome differentiation, multiple medication recommendations, and individual patient characteristics, to create suitable integrated health management recommendations encompassing exercise, diet, and emotional well-being.

[0014] Furthermore, the standardization processing of the three-modal data in step S1 includes the following steps: S11. Standardization of text symptoms employs a two-layer parsing architecture combining a TCM-specific model with rule-based matching. It matches the standard TCM symptom database using a three-tiered logic based on symptom location, symptom nature, and symptom severity. First, a complete match is achieved by mapping standard symptoms to colloquial alternatives. If a complete match is not achieved, semantic similarity is calculated using a TCM-specific model. A match is considered successful if a preset threshold is reached. If the parsing result is a non-standard symptom but contains a clearly defined location, a knowledge graph linking location and symptom is invoked to generate a confirmation question. The patient responds, and the parsing is re-resolved. If the response is invalid, a preset guiding text is triggered to prompt the patient to supplement with valid information. S12. Standardization of tongue image data: The collected tongue images are analyzed by a multimodal model. The analysis results are compared with the TCM standard symptom database to output standardized tongue image symptoms, which are then integrated into the text symptom list according to preset rules. S13. Standardization of pulse data: The colloquial description of pulse features is used to identify keywords and match standard pulse patterns using a TCM-specific model; the quantitative pulse data collected by the device is analyzed using time-domain and frequency-domain feature extraction algorithms to output standardized pulse patterns; the standardized pulse features are inserted into the symptom list according to preset rules to form a three-modal fusion data set.

[0015] Furthermore, in step S2, the reinforcement learning algorithm adopts a deep Q-network model. The state function is defined as a combination of the current three-modal data completeness, the number of known evidence elements, the number of undiagnosed questions, and the historical dialogue context features. The action function is defined as selecting a single question from the list of questions to be diagnosed. The reward function is a weighted sum of the current diagnosis round, the data completeness improvement value, and the number of repeated diagnoses, with the goal of maximizing the reward for obtaining complete data in the fewest rounds. The training process adopts an experience replay mechanism, and the training samples are real diagnosis dialogue data, which are divided into training set, validation set, and test set according to proportions. The diagnosis effect is optimized through multiple rounds of iterative training.

[0016] Furthermore, the reasoning of pathological elements in step S3 includes the following steps: The system lists the pathological syndrome elements corresponding to the three modalities of data, and statistically analyzes the frequency and order of occurrence of these elements, with the weight of pulse syndrome elements increased according to a preset ratio. It identifies the unique pathological syndrome element corresponding to a typical physical sign and verifies its coverage of all modalities. If it covers all three modalities, the syndrome element is directly output. If it does not fully cover all modalities, high-frequency syndrome elements are selected and sorted by their order of occurrence and weight, and then merged with typical symptom syndrome elements to form a candidate pathological syndrome element set. If a candidate syndrome element is a single syndrome element, the intersection of tongue and pulse syndrome elements is compared. If there is a mismatch, the system queries the composite syndrome element knowledge graph to determine if it is a composite syndrome element of the single syndrome element. If it is, both the single and composite syndrome elements are output; otherwise, the original single syndrome element is retained. The steps involved in reasoning about the location and elements of a disease are as follows: List the corresponding disease location syndrome elements for each symptom, identify the unique disease location syndrome element corresponding to the typical physical sign, and verify its coverage. If it covers all three modalities, output the syndrome element directly. If it does not cover all modalities, take the disease location syndrome element corresponding to the first symptom other than the typical symptom, sort it by total frequency and weight, and verify whether the combination of this syndrome element and the typical symptom syndrome element covers all three modalities until a fully covered combination is found or all syndrome elements are traversed. Typical signs are those that correspond to a single symptom element.

[0017] Furthermore, when the two-dimensional evidence elements do not cover all multimodal data, an evidence element diagnosis and completion process is executed, including the following steps: Combining the current three-modal data, identified syndrome elements, and syndrome elements to be consulted, the system generates pathogenesis explanations and preparatory text for consultation, guiding patients to supplement information; it filters the set of syndrome elements to be consulted, generates a list of symptoms to be consulted based on the syndrome element-typical symptom knowledge base, filters the symptoms already described by the patient, and initiates consultation with the patient; it receives the patient's response and performs semantic understanding, matching it with the typical symptoms to be consulted: if a match is successful, the corresponding syndrome element is extracted and added to the set of syndrome elements to be output, and it is determined whether it covers all uncovered symptoms. If so, it integrates and outputs a complete two-dimensional syndrome element set; otherwise, it filters the typical symptoms corresponding to the syndrome element and continues consultation; if a match is unsuccessful, the patient's response is passed to step S1 to match new symptoms with the standard symptom database. If a match is found, the symptom attributes are clarified and the chief complaint symptom set is reconstructed, returning to step S3 for re-reasoning; if no match is found, it continues to consult the next symptom to be consulted, until all symptoms to be consulted are traversed; The screening steps for the set of syndrome elements to be consulted are as follows: Select the union of syndrome elements corresponding to symptoms that appear frequently in the three-modal data, meet the preset conditions, are not covered by the current syndrome element, and cannot be covered by the current syndrome element. After sorting according to the preset rules, select a specified number of syndrome elements.

[0018] Furthermore, the generation of syndrome types includes the following steps: First, based on the set of disease-related syndrome elements, a pool of candidate syndrome types is selected using strategies such as single disease-related matching, two-disease-related combination matching, or single disease-related correspondence matching; then, based on the set of disease-related syndrome elements, the final syndrome type is selected from the pool of candidate syndrome types using strategies such as complete matching, partial matching, sequential matching, or ignoring disease-related location matching; if none of the strategies are effective, the syndrome type corresponding to the first symptom is used as the recommended result according to the correspondence between one symptom and one prescription. The process of matching auxiliary drugs includes the following steps: First, based on the core pathogenesis, key symptoms, and Western medicine diagnostic information of the syndrome type, the auxiliary drug knowledge base is accessed to extract a set of candidate auxiliary drugs. Second, the auxiliary drug matching score is calculated according to preset weight ratios for pathogenesis matching degree, symptom matching degree, and Western medicine diagnostic correlation degree. Third, a preset number of auxiliary drugs are selected in descending order of the scores. Fourth, the safety of the auxiliary drugs in combination with the core prescription is verified by accessing the knowledge base of Chinese medicine compatibility contraindications. If there are no contraindications, the auxiliary drug is added to the core prescription and its mechanism of action is marked. If there are contraindications, the auxiliary drug is removed and a replacement is selected from the candidate set. The multi-dimensional recommendation of traditional Chinese medicine (TCM) preparations includes the following steps: calculating the TCM preparation matching score based on the preset weight ratios of syndrome matching degree, symptom suitability degree, and individual characteristic suitability degree; the individual characteristic suitability degree includes age segmentation and constitution classification dimensions, with age segmentation divided into children, adults, and the elderly, and constitution classification based on the theory of nine constitutions in traditional Chinese medicine; recommending a preset number of TCM preparations in descending order of score, with each recommendation including the name of the TCM preparation, efficacy and indications, usage and dosage, reason for suitability, and contraindications.

[0019] The advantages of this invention are: Significantly improved diagnostic accuracy: Through the fusion of three modal data of symptoms, tongue appearance and pulse appearance and closed-loop syndrome element reasoning, the accuracy of syndrome type judgment reached 93% after verification in 1,000 clinical cases, which is 15-20% higher than the traditional single-modal system; the integration of pulse diagnosis makes the pathogenesis judgment more in line with the essence of TCM diagnosis and treatment, further reducing the error of syndrome differentiation. The prescription is highly targeted: It incorporates the theory of traditional Chinese medicine master Wang Xinlu to achieve precise matching between traditional prescriptions and patients' core symptoms and Western medicine diagnoses, improving the efficacy of prescriptions by more than 25% and solving the limitation of traditional prescriptions that are only one prescription for one person. Medication selection is more flexible: Chinese herbal prescriptions (including auxiliary medicines) and prepared Chinese medicines are recommended simultaneously to suit different patient medication scenarios (such as choosing prepared Chinese medicines for those who have difficulty decocting medicines, and choosing prescriptions for those who need precise conditioning), improving patient compliance by 30%; Promoting the digital inheritance of theories of renowned TCM masters: Deeply integrate the theories of TCM masters on medicine assistance and TCM diagnostic thinking with modern artificial intelligence technology to build an interpretable and reusable intelligent TCM diagnosis and treatment framework, providing technical support for the digital and intelligent inheritance of theories of renowned TCM masters. Attached Figure Description

[0020] Figure 1 This is a flowchart of the multi-source, multi-modal data acquisition and standardized processing of this invention; Figure 2 This is a flowchart of the optimization algorithm for the reasoning of pathogenesis, location of disease, and symptom elements, and the fusion of pulse and tongue in this invention. Figure 3 This is a flowchart of the drug adaptation algorithm of the present invention; Figure 4 This is a schematic diagram of the knowledge graph relating the syndrome element-syndrome mapping and the auxiliary drugs-prescription drugs-traditional Chinese medicines of this invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0022] This embodiment illustrates a method and system for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning. The specific process is as follows.

[0023] S1. Multi-source multimodal data acquisition and standardization processing: Collect patient textual symptom, tongue image, and pulse image data in three modalities and perform standardization processing to form a standardized trimodal data set that fuses the symptom, tongue image, and pulse images. Please refer to [link / reference]. Figure 1 .

[0024] S101. Text-based symptom collection and standardized analysis A two-layer parsing architecture combining the Qwen3 TCM fine-tuning model and rule matching is adopted: First, the model semantically identifies symptom keywords in patient descriptions, preserving the original expressions to avoid information loss. Then, it performs layered matching with a standard TCM symptom database. Layering is based on three levels: symptom location (e.g., head, chest, abdomen), symptom nature (e.g., pain, distension, itching), and symptom severity (e.g., mild, moderate, severe), ensuring the matching logic aligns with TCM symptom classification habits. The colloquial alias identification rule is based on a pre-set correspondence table between standard symptoms and colloquial aliases in the standard TCM symptom database. Rapid identification is achieved through complete string matching. The alias database, compiled and verified by TCM experts, covers common clinical colloquial expressions; for example, "headache" and "head distension" both correspond to the standard symptom "headache."

[0025] When a complete match of colloquial aliases is not achieved, the Qwen3 TCM fine-tuning model is used for semantic matching to achieve fuzzy matching. Patient symptom keywords and symptom texts from the standard symptom database are simultaneously input into the Qwen3 TCM fine-tuning model. The model directly outputs the semantic similarity value between the two based on its semantic understanding capabilities in the TCM domain. The similarity threshold is set to 0.85; a match is considered successful if the similarity is higher than the threshold.

[0026] The aforementioned TCM standard symptom database was compiled and constructed by experts in the field of TCM, covering more than 2,000 subdivided TCM symptoms, including systemic symptoms, local symptoms, and visceral symptoms. Core fields include standard symptom name, symptom code, system to which it belongs (e.g., respiratory, digestive), common colloquial alternative names, corresponding disease nature and location, and symptom attributes. Data sources include the medical case collection of Master of Traditional Chinese Medicine Wang Xinlu in "Internal Medicine of Traditional Chinese Medicine" and clinical case summaries from the TCM departments of several top-tier hospitals.

[0027] If the model's analysis result is a non-standard symptom but includes a specific location, then the location-symptom knowledge graph is invoked to generate a related symptom confirmation question, such as "The discomfort in your lower back that you mentioned, is it a manifestation of soreness, coldness, or tingling?" After the patient replies, the analysis process is restarted until a standard symptom is matched.

[0028] The Location-Symptom Knowledge Graph was compiled and constructed by experts in the field of Traditional Chinese Medicine. Its core content is the mapping between common human body parts and their corresponding common symptoms. The common body parts include more than 30 core parts such as the head, neck, chest, abdomen, waist, and limbs. Each body part is associated with 5-10 high-frequency symptoms. For example, "head" is associated with headache, dizziness, head swelling, and heaviness in the head. The association rule is a one-to-many mapping between body part and symptom set. The priority is sorted according to the clinical frequency of symptoms. Its purpose is to guide patients to identify common symptoms in that body part when their description of their condition is vague but includes a specific body part, thereby improving the efficiency of the consultation.

[0029] If the patient's description is unrelated to their condition, the preset guiding text "Please describe your symptoms in detail (e.g., cough with phlegm for 3 days, left shoulder stabbing pain with limited mobility) will be triggered so that I can accurately analyze them for you" will be triggered, guiding the patient to provide relevant information.

[0030] S102. Multimodal Data Acquisition and Analysis of Tongue Images It supports patients to take real-time images of their tongues via a terminal device, and can provide shooting guidance to patients on the terminal device: natural light, tongue naturally extended, avoid obstruction, etc., or upload the taken images from the album; The multimodal model is fine-tuned and trained using Qwen3-VL. The training data consists of over 30,000 tongue images from real patients visiting hospitals. The images are sourced from recent outpatient cases at several secondary and tertiary hospitals and are annotated by professional TCM physicians. The annotation dimensions include: tongue color (red, light red, crimson, light white, bluish-purple, etc.), tongue coating color (white, yellow, grayish-black, etc.), tongue coating thickness (thin coating, thick coating, little coating, no coating, etc.), tongue coating texture (normal, greasy coating, putrid coating, etc.), tongue moisture (moist, dry, slippery, etc.), presence or absence of teeth marks (yes, no), and presence or absence of ecchymosis or petechiae (yes, no). The Kappa value for annotation consistency verification is ≥0.85. The training process includes: data preprocessing, noise reduction and size standardization; feature extraction; annotation information mapping; model iterative training with 100 iterations and a learning rate that can be set to 0.001; and model evaluation, requiring an accuracy of ≥92%.

[0031] The model analysis results are compared with the standard symptom database of traditional Chinese medicine, and standardized tongue symptoms such as "pale red tongue", "thin white coating" and "teeth mark tongue" are output. When the length of the symptom list is ≥3, the 4th-5th symptom is inserted, and when it is less than 3, it is inserted at the end. The results are then integrated into the text symptom list to ensure the consistency of the data structure.

[0032] S103. Pulse Diagnosis Multimodal Data Acquisition and Analysis Data acquisition supports two modes: subjective description and objective device mode, adapting to different usage scenarios. Subjective description mode allows patients to verbally describe pulse characteristics, such as "rapid pulse," "strong pulse," or "pulse that fluctuates in speed." Objective device mode supports integration with mainstream TCM pulse diagnosis instruments, receiving pulse waveform data, pulse rate, pulse amplitude, rhythm, and other quantitative indicators collected by the device. Verbal description parsing uses the Qwen3 TCM fine-tuning model to identify pulse keywords and match them with standard pulse characteristics, such as "rapid pulse" corresponding to a rapid pulse, "strong pulse" corresponding to a full pulse, "smooth pulse" corresponding to a slippery pulse, "requiring deep pressure" corresponding to a deep pulse, and "irregular rhythm" corresponding to a knotted pulse. The device data analysis uses a pulse feature extraction algorithm to extract time-domain features such as pulse rate, pulse amplitude, and diastolic / systolic ratio, as well as frequency-domain features such as dominant frequency and harmonic distribution. It analyzes waveform data and outputs standardized pulse features such as pulse position (floating / deep), pulse rate (rapid / slow / normal), pulse strength (strong / weak), pulse shape (wiry / slippery / thin / hesitant), and pulse rhythm (regular / irregular).

[0033] Standardized pulse characteristics are used as an independent sign dimension and inserted as the 6th position in the symptom list (or at the end if the list length is 6), forming a three-modal fusion data set of symptoms, tongue appearance, and pulse appearance to ensure the integrity of the four diagnostic methods data.

[0034] S2. Intelligent consultation guidance and symptom attribute completion: Based on the standardized trimodal data set and the preset consultation knowledge base, consultation questions are dynamically generated through reinforcement learning algorithm. After receiving patient feedback, symptom attributes and pulse details are completed, and consultation steps with covered information are automatically skipped to obtain the completed trimodal data.

[0035] 201: Construction of a Medical Consultation Knowledge Base The knowledge base covers 20+ common symptoms such as headache, abdominal pain, and cough, along with corresponding consultation questions and answer options. Core fields include symptoms, consultation questions, answer options (single / multiple choice), attribute tags, and priority, as shown in the table below: Table 1. Examples of core fields in the medical history knowledge base 202: Dynamic Consultation Generation Logic Based on the priority order of text symptoms → tongue appearance supplementation → pulse details, corresponding consultation questions are generated sequentially, allowing patients to select answers or input freely. The freely input content is compared with the answers in the knowledge base through the Qwen3 TCM micro-adjustment model to match the optimal attribute tags. For example, if a patient replies "pain after eating spicy food", the tag "spicy food" is matched.

[0036] If a patient's response to a question includes answers to other questions not yet asked, such as when asked about the location of abdominal pain, the patient replies "upper abdominal distension and pain, which worsens when pressed, and the pulse feels a bit fast," the entity relation extraction algorithm identifies the association between symptom attributes (distension and pain, tenderness upon pressure) and pulse information (fast pulse), and automatically skips subsequent corresponding consultations to avoid duplicate interactions.

[0037] The reinforcement learning algorithm employs a Deep Q-Network (DQN) reinforcement learning model, with its core function dynamically adjusting the consultation priority. Its state function S is defined as "current trimodal data completeness + number of confirmed evidence elements + number of unconsulted questions + historical dialogue context features," where the current trimodal data completeness ranges from 0 to 1. The action function A is defined as the selection from the list of questions to be consulted. The reward function is defined as R = α × (1 / current consultation round) + β × data completeness improvement value - γ × number of repeated consultations, where α = 0.6, β = 0.3, and γ = 0.1, with a reward value range of 0-1. The goal is to maximize the reward for "obtaining complete data in the fewest rounds." Training employs an experience replay mechanism, with training samples consisting of tens of thousands of real consultation dialogues, divided into training, validation, and test sets in a 6:2:2 ratio. Training is iterated for 50 rounds, with the objective of minimizing the average number of consultation rounds on the validation set.

[0038] Comparison before and after optimization: Before optimization, the average number of consultation rounds was 4.8, and the data completeness was 82%; after optimization, the average number of consultation rounds was 2.3, and the data completeness was 96%, with consultation efficiency improved by 52% and data completeness improved by 17%.

[0039] S3. Disease nature and location syndrome element reasoning and pulse / tongue fusion optimization: The completed three-modal data undergoes dual-dimensional syndrome element reasoning based on disease nature and location. The reasoning results are dynamically optimized by integrating tongue and pulse characteristics with the number of consultation rounds, forming a dual-dimensional syndrome element set. Please refer to [link / reference]. Figure 2 .

[0040] S301. Reasoning about the nature of the disease and its symptoms (cold, heat, deficiency, excess, etc.). The syndrome elements are listed and their weights are statistically analyzed. The syndrome elements corresponding to each symptom, tongue appearance, and pulse are listed, and the frequency and order of their occurrence are statistically analyzed. Considering the core role of pulse diagnosis in determining the nature of the disease, the weight of pulse syndrome elements is increased by 20%. For example, a wiry pulse corresponds to the syndrome elements of "qi stagnation" and "liver hyperactivity," and its weight is higher than that of syndrome elements corresponding to ordinary symptoms. Example: A patient's symptoms are abdominal distension (order1), abdominal burning sensation (order2), abdominal dull pain (order3), abdominal cold pain (order4), and abdominal pain aggravated by pressure (order5). The corresponding syndrome element statistical results are: {Heat: 3, Excess Cold: 3, Yang Deficiency: 3, Qi Stagnation: 2, Damp Heat: 2, Cold Dampness: 2, Qi Deficiency: 2, Blood Stasis: 2, Yin Deficiency: 2, Food Stagnation: 1, Damp Turbidity: 1, Phlegm Turbidity: 1, Water Retention: 1, Cold Phlegm: 1, Blood Deficiency: 1}.

[0041] Typical symptom verification: If there is a typical symptom that corresponds to a single syndrome element, such as "slow pulse" → cold syndrome, "red tongue" → heat syndrome, "thready pulse" → deficiency syndrome, determine whether the syndrome element corresponding to the symptom covers all three modal data. If yes, output the syndrome element directly; otherwise, proceed to the next step.

[0042] After excluding the syndrome elements corresponding to typical symptoms, the syndrome element with the highest frequency is selected. If the frequencies are the same, the syndrome element with the highest order of occurrence and the highest weight is selected and merged with the syndrome element corresponding to typical symptoms to output a set of candidate pathological syndrome elements. Example: In the above statistical results, "heat", "cold", and "yang deficiency" all appear 3 times. "Heat" with the highest order value is selected and merged with the syndrome element corresponding to typical symptoms.

[0043] Tongue and pulse fusion optimization: If the output candidate pathological syndrome element is a single syndrome element, compare the intersection of syndrome elements corresponding to the tongue and pulse symptoms. If the intersection does not match the single syndrome element, such as the single syndrome element being "heat syndrome" and the tongue and pulse syndrome element intersection being "damp syndrome", then query the knowledge graph of composite syndrome elements constructed based on the TCM composite syndrome diagnosis and treatment guidelines to determine whether the intersection syndrome element is a composite syndrome element of the single syndrome element. For example, "heat syndrome" and "damp syndrome" can be combined into "damp-heat syndrome". If so, output the single syndrome element + composite syndrome element; otherwise, retain the original single syndrome element. If the current syndrome element does not cover all three modal data, then proceed to the syndrome element consultation process.

[0044] S302, Reasoning based on the location and elements of the disease (lung, spleen, liver, heart, etc.) List the corresponding pathological elements for each symptom and identify typical signs. For example, the pathological elements corresponding to the patient's symptoms Z1-Z6 are Z1→e1, Z2→e2 / e3, Z3→e2 / e3 / e4, Z4→e1 / e6, Z5→e6 / e7, and Z6→e7, and the typical signs are Z1 (e1) and Z6 (e7).

[0045] Determine whether the syndrome element corresponding to the typical symptom covers all three modal data. If yes, output the syndrome element; otherwise, proceed to the next step.

[0046] Atypical symptom syndrome element sorting: Take the syndrome element corresponding to the first symptom that appears besides typical symptoms, sort it by total frequency and weight, and check whether the combination of this syndrome element and typical symptom syndrome elements covers all three modal data until a fully covered combination is found or all syndrome elements are traversed. For syndrome elements e2 and e3 corresponding to Z2, if the total frequency of e2 is greater than e3, first check whether e2+e1+e7 covers all symptoms. If it does, output {e1, e2, e7}.

[0047] S303, Syndrome Differentiation and Consultation Process Combining current trimodal data, identified syndrome elements, and syndrome elements to be diagnosed, the Qwen3 TCM micro-tuning model generates pathogenesis explanations and preparatory texts for the consultation. For example, "Your symptoms of headache, irritability, and facial flushing suggest that there is internal heat rising upwards, disturbing the liver qi and causing restlessness in the sensory orifices. However, the source of the heat may vary, and it is sometimes related to deficiency of yin fluids and upward floating of deficient fire. I will ask you a few more related questions to further clarify the cause of the disease." Screening of syndrome elements to be consulted: Extract syndrome elements that appear more than 1 times in the three-modal data and are not covered by the current syndrome element; find the set of symptoms that cannot be covered by the current syndrome element (uncovered symptoms); take the union of the above two sets of syndrome elements, sort them by frequency and order value, and take the top 5 to construct the set of syndrome elements to be consulted.

[0048] Typical Symptom Matching and Filtering: The system queries the Syndrome Element-Typical Symptom Knowledge Base to generate a list of symptoms to be consulted. Using the Qwen3 TCM Fine-tuning Model, it compares the data with historical dialogues to filter out symptoms already described by the patient, avoiding duplicate consultations. The Syndrome Element-Typical Symptom Knowledge Base is compiled and constructed by experts in the field of TCM. Its core entries cover 46 common pathological syndrome elements. Core fields include syndrome element name, a list of typical symptoms, symptom weight (assigned a value of 0.1-1.0 based on diagnostic contribution), and corresponding tongue and pulse characteristics. For example, typical symptoms of the syndrome element "damp-heat" include sticky stools, bitter taste in the mouth, yellow and greasy tongue coating, and a slippery and rapid pulse.

[0049] The patient's response is semantically understood using the Qwen3 TCM fine-tuning model and matched with typical symptoms to be diagnosed. If a match is successful, the corresponding syndrome element is extracted and added to the set of syndrome elements to be output. It is then determined whether all uncovered symptoms are covered. If so, the syndrome element to be output and the identified syndrome elements are output; otherwise, the typical symptoms corresponding to that syndrome element are filtered out, and the consultation continues. If no match is found, the current consultation question and the patient's description are passed to step S101 to match new symptoms with the standard symptom database. After a new symptom is matched and its symptom attributes are clarified, the set of chief complaint symptoms is reconstructed, and the process returns to steps S301-302 for re-reasoning. If no new symptom is matched, the consultation continues with the next symptom to be diagnosed until all lists have been traversed.

[0050] S4. Syndrome Element-Syndrome Type Mapping and Collaborative Recommendation of Supporting Drugs, Prescriptions, and Traditional Chinese Medicine Preparations: Based on a dual-dimensional syndrome element set, syndrome types are generated through multi-strategy matching. Syndrome type-related information is extracted to match targeted supporting drugs, and incompatibilities are verified. Supporting drugs are then integrated with core prescriptions. Simultaneously, based on syndrome types and individual patient characteristics, suitable traditional Chinese medicine preparations are recommended, resulting in a diversified medication recommendation outcome. Please refer to [link / reference]. Figure 4 .

[0051] Explanation of the Knowledge Graph of Evidence Types and Elements Please refer to Figure 3This system was developed by students of the renowned traditional Chinese medicine master Wang Xinlu and reviewed by Wang Xinlu himself, referencing authoritative textbooks such as *Internal Medicine of Traditional Chinese Medicine* and *Wang Xinlu's Medical Case Collection*. Its core content covers over 300 common clinical syndromes, including those applicable to multiple departments such as internal medicine, gynecology, pediatrics, and surgery. Core fields include syndrome name, syndrome code, corresponding set of pathogenic elements, corresponding set of pathogenic elements located at the site of disease, core pathogenesis description, and association with typical symptoms. Example: The syndrome "Damp-Heat Accumulation, Liver-Spleen Disharmony" corresponds to the set of pathogenic elements {damp-heat, Qi stagnation}, the set of pathogenic elements located at the site of disease {liver, spleen and stomach}, and the core pathogenesis is "Damp-heat stagnation, liver dysfunction, and spleen dysfunction."

[0052] S401, Certificate Type Mapping Step 1: Based on the set of disease-related syndrome elements, independently screen the candidate syndrome pool, considering only the matching of disease characteristics and not the location of the disease.

[0053] Strategy 1: Single disease syndrome matching: Find syndromes in the knowledge graph whose disease syndrome element set is exactly equal to (no more and no less) the patient's disease syndrome element set, and include them all in the candidate syndrome pool. Strategy 2: Matching of two disease-related syndrome combinations: When no single syndrome is found, search for combinations in the knowledge graph where the set of disease-related syndrome elements of the two syndromes is completely equal to the patient's set of disease-related syndrome elements. If found, all of them are included in the candidate syndrome pool. Strategy 3: Single disease type matching: When no combination syndrome is found, each disease type syndrome element of the patient is matched with the corresponding single syndrome in the knowledge graph (such as "heat syndrome" or "cold syndrome"), and all are included in the candidate syndrome pool.

[0054] Step 2: Based on the set of disease location syndrome elements, the final syndrome type is selected from the candidate syndrome type pool, considering only the disease location matching and not changing the disease nature association; Strategy 1: Complete Matching of Disease Location: In the candidate syndrome pool, search for syndromes whose disease location syndrome element set is completely equal to the patient's disease location syndrome element set, and output them directly if found. Strategy 2: Partial matching of disease location: When there is no complete match, count the number of times each disease location syndrome element set in the candidate syndrome pool is included in the patient's disease location syndrome element set, select the syndrome with the most included number, and output it if it is unique; Strategy 3: Disease location order matching: If multiple syndrome types have the same number of disease location matches, select the syndrome type that appears first according to the order value (order of appearance) of the syndrome elements of the patient's disease location. If the order is the same, sort them according to the priority of syndrome types in the knowledge graph. Strategy 4: Ignore disease location matching: If none of the above applies, select a syndrome from the candidate syndrome pool whose disease element set is completely equal to the patient's disease element set.

[0055] Special scenario handling (when there is no clear disease nature, location and syndrome elements): When the above strategies are ineffective, the recommended result is the syndrome type corresponding to the first symptom, based on the one-symptom-one-prescription knowledge graph (the correspondence between common diseases and syndromes compiled by TCM experts).

[0056] S402, Logic for Matching Assisted Drugs The Aid Medicine Knowledge Base is a knowledge graph that integrates the theoretical literature on aid medicine by the master of traditional Chinese medicine, Wang Xinlu, and establishes a network of interconnected syndrome types, core symptoms, Western medical diagnoses, and aid medicines.

[0057] Drug screening algorithm: Step 1: Extract the patient's core pathogenesis, key symptoms, and clear Western medicine diagnosis. If the patient does not provide these, recommend and confirm possible Western medicine diagnoses through trimodal data association. Step 2: Based on the knowledge base, calculate the drug matching score according to the pathogenesis matching degree (weight 60%) + symptom matching degree (weight 30%) + Western medicine diagnosis correlation degree (weight 10%). Step 3: Select the first 2-3 auxiliary herbs to ensure that there are no incompatibilities with the core prescription (verify by calling the knowledge base of Chinese medicine compatibility incompatibilities), supplement them into the traditional prescription to optimize the output, and clearly mark the composition of the core prescription, the name of the auxiliary herbs, the dosage and the mechanism of action.

[0058] The data sources for the Traditional Chinese Medicine (TCM) Combination Incompatibilities Knowledge Base include the "Pharmacopoeia of the People's Republic of China", "Traditional Chinese Medicine" textbooks, and "Traditional Chinese Medicine Formulae" compatibility standards. The core entries cover classic incompatibilities such as the 18 incompatibilities and 19 antagonisms, as well as combinations of incompatibilities clearly defined in modern clinical practice. The core fields include the name of the TCM, the name of the incompatible TCM, the type of incompatibility, the risk level, and the description of the incompatibility, and a total of more than 320 incompatible combinations are included.

[0059] S403, Logic of Multiple Recommendations for Traditional Chinese Medicine The Traditional Chinese Medicine Knowledge Base covers traditional Chinese medicine preparations listed in the National Essential Medicines List, and the data is synchronized with the latest standards of the National Medical Products Administration.

[0060] The matching score of traditional Chinese medicine is calculated based on the syndrome matching degree (weight 50%), symptom matching degree (weight 30%), and individual characteristic matching degree (weight 20%). Among the individual characteristic matching degree, the age segment is <14 years old (children), 14-64 years old (adults), and ≥65 years old (elderly). The constitution classification is based on the nine constitution theories of traditional Chinese medicine, such as balanced constitution, qi deficiency constitution, and yang deficiency constitution.

[0061] Along with traditional Chinese medicine prescriptions (including supplementary prescriptions), the top three recommended traditional Chinese medicine preparations are output simultaneously. Each recommendation includes the name of the traditional Chinese medicine preparation, its efficacy and indications, usage and dosage, reasons for its suitability, and contraindications. For example, "Danzhi Xiaoyao Wan: Efficacy: soothes the liver and strengthens the spleen, clears heat and regulates menstruation. It is suitable for 'liver stagnation and fire syndrome' and is used for symptoms such as emotional instability and delayed menstruation. The dosage is 6-9g at a time, twice a day. Contraindications: use with caution in pregnant women."

[0062] S5. Personalized health management recommendations are generated based on syndrome differentiation, multiple medication recommendations, and individual patient characteristics, to create suitable integrated health management recommendations encompassing exercise, diet, and emotional well-being.

[0063] Exercise recommendations: Based on the syndrome differentiation and pathogenesis, we recommend suitable exercise types, intensity, and duration; Dietary recommendations: Based on the correspondence between syndrome types and the properties of ingredients, suitable foods and foods to avoid are recommended; Emotional advice: Based on the emotional issues related to the syndrome type, generate guidance suggestions; Recommended optimization: Use a generative large model to transform standardized suggestions into personalized, conversational expressions.

[0064] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A TCM intelligent consultation and syndrome-based prescription recommendation system based on multimodal fusion and syndrome element reasoning, characterized in that, include: The multi-source multimodal data acquisition and standardization processing module is used to collect patients' textual symptoms, tongue appearance, and pulse appearance trimodal data and standardize them to obtain a standardized trimodal data set in a unified format. The intelligent consultation guidance and symptom attribute completion module, based on the standardized data set and combined with the preset consultation knowledge base, dynamically generates consultation questions through reinforcement learning algorithm, completes symptom attributes and pulse details, and obtains the completed three-modal data; The module for reasoning about the nature and location of the disease and optimizing the fusion of pulse and tongue features performs dual-dimensional reasoning on the completed three-modal data, combining the characteristics of tongue and pulse to optimize the reasoning results and obtain a set of dual-dimensional syndrome elements. The syndrome element-syndrome mapping and auxiliary drug-prescription drug-Chinese patent medicine collaborative recommendation module, based on the dual-dimensional syndrome element set, generates corresponding syndromes through the syndrome element-syndrome mapping relationship, and combines the auxiliary drug theory to match the auxiliary drugs, thereby realizing the recommendation of Chinese herbal prescriptions and Chinese patent medicines; The personalized health management suggestion generation module generates integrated health management suggestions for exercise, diet, and emotional well-being based on the syndrome type, recommended Chinese herbal prescriptions, proprietary Chinese medicines, and individual patient characteristics.

2. The TCM intelligent consultation and syndrome-based prescription recommendation system based on multimodal fusion and syndrome element reasoning according to claim 1, characterized in that, The multi-source, multi-modal data acquisition and standardization processing module includes: The text symptom collection and standardization parsing unit is used to collect patients' text symptom descriptions. Through a combination of TCM-specific models and rule matching, it achieves accurate mapping between colloquial descriptions and TCM standard symptoms, processes non-standard symptom descriptions, and filters invalid responses. The tongue image multimodal data acquisition and analysis unit is used to acquire images of patients' tongue images, analyze tongue image features through a multimodal model and convert them into standardized tongue image symptoms, and integrate them into a text symptom list; The pulse diagnosis multimodal data acquisition and analysis unit is used to acquire patient pulse data in dual modes, analyze it and convert it into standardized pulse features, and integrate it into the symptom list to form a trimodal fusion data set.

3. The TCM intelligent consultation and syndrome-based prescription recommendation system based on multimodal fusion and syndrome element reasoning according to claim 1, characterized in that, In the intelligent consultation guidance and symptom attribute completion module, the preset consultation knowledge base includes consultation questions, answer options, attribute tags and priorities corresponding to common symptoms; the reinforcement learning algorithm is used to dynamically adjust the consultation priority, so as to obtain complete data in the fewest rounds and improve consultation efficiency and data integrity.

4. The TCM intelligent consultation and syndrome-based prescription recommendation system based on multimodal fusion and syndrome element reasoning according to claim 1, characterized in that, The syndrome element-syndrome mapping and auxiliary drug-prescription drug-traditional Chinese medicine collaborative recommendation module includes: The syndrome mapping unit is used to filter the final syndrome from the candidate syndrome pool based on a two-dimensional syndrome element set and through multi-strategy matching. In special scenarios, a fallback strategy is used to determine the syndrome, and the special scenario is when there are no clear disease and / or disease location syndrome elements. The drug matching unit is used to screen suitable drugs from the drug knowledge base based on syndrome information and add them to the prescription after verifying incompatibility. The Traditional Chinese Medicine (TCM) recommendation unit is used to select suitable TCMs from the TCM knowledge base based on syndrome differentiation and individual patient characteristics, and output relevant prompts.

5. A method for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning, characterized in that, Including the following steps: S1. Multi-source multimodal data acquisition and standardization processing: Collect patient textual symptoms, tongue appearance, and pulse appearance data in three modalities and perform standardization processing to form a standardized three-modal data set that integrates symptoms, tongue appearance, and pulse appearance. S2. Intelligent consultation guidance and symptom attribute completion: Based on the standardized trimodal data set and the preset consultation knowledge base, consultation questions are dynamically generated through reinforcement learning algorithm. After receiving patient feedback, symptom attributes and pulse details are completed. The consultation process that has been covered is automatically skipped to obtain the completed trimodal data. S3, Disease nature and location syndrome element reasoning and pulse and tongue fusion optimization, perform disease nature and location dual-dimensional syndrome element reasoning on the completed three-modal data, integrate tongue and pulse characteristics and dynamic optimization of reasoning results with consultation rounds to form a dual-dimensional syndrome element set; S4. Syndrome element-syndrome type mapping and collaborative recommendation of auxiliary drugs-prescription drugs-traditional Chinese medicine: Based on the dual-dimensional syndrome element set, syndrome types are generated through multi-strategy matching. Syndrome type related information is extracted to match targeted auxiliary drugs and the drug incompatibility is checked. Auxiliary drugs are integrated with core prescriptions. At the same time, traditional Chinese medicine is recommended based on syndrome type and individual patient characteristics to form a multi-dimensional drug recommendation result. S5. Personalized health management recommendations are generated based on syndrome differentiation, multiple medication recommendations, and individual patient characteristics, to create suitable integrated health management recommendations encompassing exercise, diet, and emotional well-being.

6. The method for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning according to claim 5, characterized in that, The standardization process of the three-modal data in step S1 includes the following steps: S11. Standardization of text symptoms employs a two-layer parsing architecture combining a TCM-specific model with rule-based matching. It matches the standard TCM symptom database using a three-tiered logic based on symptom location, symptom nature, and symptom severity. First, a complete match is achieved by mapping standard symptoms to colloquial alternatives. If a complete match is not achieved, semantic similarity is calculated using a TCM-specific model. A match is considered successful if a preset threshold is reached. If the parsing result is a non-standard symptom but contains a clearly defined location, a knowledge graph linking location and symptom is invoked to generate a confirmation question. The patient responds, and the parsing is re-resolved. If the response is invalid, a preset guiding text is triggered to prompt the patient to supplement with valid information. S12. Standardization of tongue image data: The collected tongue images are analyzed by a multimodal model. The analysis results are compared with the TCM standard symptom database to output standardized tongue image symptoms, which are then integrated into the text symptom list according to preset rules. S13. Standardization of pulse data: The colloquial description of pulse features is used to identify keywords and match standard pulse patterns using a TCM-specific model; the quantitative pulse data collected by the device is analyzed using time-domain and frequency-domain feature extraction algorithms to output standardized pulse patterns; the standardized pulse features are inserted into the symptom list according to preset rules to form a three-modal fusion data set.

7. The method for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning according to claim 5, characterized in that, In step S2, the reinforcement learning algorithm uses a deep Q-network model, and the state function is defined as a combination of the current three-modal data completeness, the number of confirmed evidence elements, the number of undiagnosed questions, and the historical dialogue context features. The action function is defined as selecting a single question from the list of questions to be consulted; the reward function is a weighted sum of the current consultation round, the improvement in data completeness, and the number of repeated consultations, with the goal of maximizing the reward for obtaining complete data in the fewest possible rounds. The training process employs an experience playback mechanism, with training samples consisting of real consultation dialogue data, which are divided into training, validation, and test sets according to a set ratio. Multiple rounds of iterative training are conducted to optimize the consultation effect.

8. The method for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning according to claim 5, characterized in that, The reasoning of pathological elements in step S3 includes the following steps: The system lists the pathological syndrome elements corresponding to the three modalities of data, and statistically analyzes the frequency and order of occurrence of these elements, with the weight of pulse syndrome elements increased according to a preset ratio. It identifies the unique pathological syndrome element corresponding to a typical physical sign and verifies its coverage of all modalities. If it covers all three modalities, the syndrome element is directly output. If it does not fully cover all modalities, high-frequency syndrome elements are selected and sorted by their order of occurrence and weight, and then merged with typical symptom syndrome elements to form a candidate pathological syndrome element set. If a candidate syndrome element is a single syndrome element, the intersection of tongue and pulse syndrome elements is compared. If there is a mismatch, the system queries the composite syndrome element knowledge graph to determine if it is a composite syndrome element of the single syndrome element. If it is, both the single and composite syndrome elements are output; otherwise, the original single syndrome element is retained. The reasoning of the pathological location elements includes the following steps: List the disease location syndrome elements corresponding to each symptom, identify the unique disease location syndrome element corresponding to the typical physical sign and verify its coverage. If it covers all three modalities, output the syndrome element directly. If it does not cover all three modalities, take the disease location syndrome element corresponding to the first symptom that appears besides the typical symptom, sort it according to the total frequency of occurrence and weight, and verify in turn whether the combination of the syndrome element and the typical symptom syndrome element covers all three modalities until a fully covered combination is found or all syndrome elements are traversed. The typical physical signs are those that correspond to a single evidence element.

9. The method for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning according to claim 8, characterized in that, When the two-dimensional evidence elements do not cover all multimodal data, the evidence element consultation and completion process is executed, including the following steps: Combining the current three-modal data, identified syndrome elements, and syndrome elements to be consulted, the system generates pathogenesis explanations and preparatory text for consultation, guiding patients to supplement information; it filters the set of syndrome elements to be consulted, generates a list of symptoms to be consulted based on the syndrome element-typical symptom knowledge base, filters the symptoms already described by the patient, and initiates consultation with the patient; it receives the patient's response and performs semantic understanding, matching it with the typical symptoms to be consulted: if a match is successful, the corresponding syndrome element is extracted and added to the set of syndrome elements to be output, and it is determined whether it covers all uncovered symptoms. If so, it integrates and outputs a complete two-dimensional syndrome element set; otherwise, it filters the typical symptoms corresponding to the syndrome element and continues consultation; if a match is unsuccessful, the patient's response is passed to step S1 to match new symptoms with the standard symptom database. If a match is found, the symptom attributes are clarified and the chief complaint symptom set is reconstructed, returning to step S3 for re-reasoning; if no match is found, it continues to consult the next symptom to be consulted, until all symptoms to be consulted are traversed; The screening steps for the set of syndrome elements to be diagnosed are as follows: select the union of syndrome elements corresponding to symptoms that appear frequently in the three-modal data, meet the preset conditions, are not covered by the current syndrome element, and cannot be covered by the current syndrome element, sort them according to the preset rules, and then select a specified number of syndrome elements.

10. The method for intelligent TCM consultation and syndrome-based prescription recommendation based on multimodal fusion and syndrome element reasoning according to claim 5, characterized in that, The generation of the syndrome type includes the following steps: First, a pool of candidate syndrome types is selected based on the set of syndrome elements of disease characteristics using strategies such as single disease characteristic matching, two disease characteristic combination matching, or single disease characteristic corresponding matching; then, the final syndrome type is selected from the pool of candidate syndrome types based on the set of syndrome elements of disease location using strategies such as complete matching, partial matching, sequential matching, or ignoring disease location matching; if none of the strategies are effective, the syndrome type corresponding to the first symptom is used as the recommended result according to the correspondence between one symptom and one prescription. The process of matching auxiliary drugs includes the following steps: First, based on the core pathogenesis, key symptoms, and Western medicine diagnostic information of the syndrome type, the auxiliary drug knowledge base is accessed to extract a set of candidate auxiliary drugs. Second, the auxiliary drug matching score is calculated according to preset weight ratios for pathogenesis matching degree, symptom matching degree, and Western medicine diagnostic correlation degree. Third, a preset number of auxiliary drugs are selected in descending order of the scores. Fourth, the safety of the auxiliary drugs in combination with the core prescription is verified by accessing the knowledge base of Chinese medicine compatibility contraindications. If there are no contraindications, the auxiliary drug is added to the core prescription and its mechanism of action is marked. If there are contraindications, the auxiliary drug is removed and a replacement is selected from the candidate set. The multi-dimensional recommendation of traditional Chinese medicine includes the following steps: calculating the suitability score of traditional Chinese medicine based on the preset weight ratio of syndrome matching degree, symptom suitability degree, and individual characteristic suitability degree; in Individual characteristic suitability includes age segmentation and constitution classification. Age segmentation is divided into children, adults, and the elderly, and constitution classification is based on the theory of nine constitutions in traditional Chinese medicine. A preset number of traditional Chinese medicine preparations are recommended in descending order of scores. Each recommendation includes the name of the traditional Chinese medicine preparation, its efficacy and indications, usage and dosage, reason for suitability, and contraindications.

Images