Medical history identification method and system based on large model and expert strategy

Abstract

The invention discloses a medical history recognition method and system based on a large model and an expert strategy, and the method comprises the steps: collecting data in a medical network, and carrying out the cleaning and standardization processing of the data; extracting medical entities and relationships thereof from the text, and optimizing the rule set; the extracted entities and relations are fused into a knowledge graph, a multi-relation graph is constructed, and attributes are defined; storing the knowledge graph in a graph database, and receiving new data through a real-time data stream processing system; basic information and symptom description of a patient are collected, and key information is extracted by using a large language model; an inquiry path is dynamically adjusted, and specific symptoms are clearly and fuzzily answered or deeply discussed by questioning; and generating a structured electronic medical record by using the large language model. Through an intelligent pre-inquiry process, the system can collect detailed medical history information and generate a standardized electronic medical record before a patient officially sees a doctor, so that the information verification and supplement time when a doctor receives the doctor is greatly shortened.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a medical history recognition method and system based on a large model and expert strategy. Background Art

[0002] With the increasing demand for efficient services in the medical and health field, technological innovation plays an increasingly important role in improving diagnosis and treatment efficiency and improving patient satisfaction. Against this background, traditional pre-consultation methods have begun to face many challenges. Traditional pre-consultation methods mainly rely on paper questionnaires or preliminary interviews with nursing staff. Patients need to fill in personal symptoms and past medical history information or describe their condition verbally before formal consultation. However, this method has significant problems such as low information collection efficiency, insufficient data integrity, and possible human errors. In addition, due to the lack of professional medical guidance, patients often find it difficult to accurately describe their symptoms, and even miss important information, causing doctors to spend more time verifying and supplementing medical history records during the first consultation. This not only affects the efficiency of diagnosis and treatment, but also increases the waiting time of patients, thereby reducing the patient's medical experience.

[0003] With the rapid development of smart medical technology, intelligent pre-consultation systems based on medical dialogue interaction have gradually become an important tool for improving the efficiency and quality of medical services. The intelligent pre-consultation system simulates the doctor's consultation process and interacts efficiently with patients. It can collect detailed information about the patient's condition before the patient's formal consultation, and integrate this data into a standardized electronic medical record for the doctor's reference. This method significantly shortens the time for doctors to verify information and makes the patient's medical process more efficient. However, the existing intelligent pre-consultation technology still has some limitations.

[0004] Traditional intelligent pre-diagnosis methods based on large models mainly rely on natural language processing technology to analyze patient complaints and generate consultation paths and related questions. Although these methods have good performance in semantic understanding and text generation, they often have difficulty coping with consultation scenarios for complex or rare diseases due to the lack of deep integration and structured processing of medical knowledge. The pure text-driven method of large models has advantages in flexibility, but it is insufficient in accuracy and professionalism, especially for the recognition of key information hidden in medical history and medical entity relationships. This deficiency may lead to omission or misunderstanding of medical history information, further increasing the diagnostic burden of doctors. To address these problems, there is an urgent need for a system that can combine the natural language understanding capabilities of large language models with the structured representation capabilities of medical knowledge graphs. Such a method can not only give full play to the advantages of large models in language understanding, but also use knowledge graphs to deeply analyze and associate medical knowledge, thereby improving the comprehensiveness and accuracy of the pre-diagnosis system. Summary of the invention

[0005] In order to overcome the shortcomings of the existing technology, the present invention proposes a medical history recognition method and system based on a large model and expert strategy. Through multi-channel support (such as voice, text and video interaction) and efficient pre-consultation services, the system shortens the patient's waiting time, simplifies the medical process, and improves the patient's satisfaction with the medical treatment.

[0006] To achieve the above object, the medical history recognition method based on a large model and expert strategy of the present invention comprises the following steps:

[0007] Step 1: Collect structured data, semi-structured data, and unstructured data from the medical network, and clean and standardize the data to provide a basis for building a knowledge graph;

[0008] Step 2: Extract medical entities and their relationships from text through pattern matching and rule-based methods, and optimize the rule set based on the knowledge of domain experts to improve extraction accuracy;

[0009] Step 3: Integrate the extracted entities and relationships into the knowledge graph, build a multi-relationship graph containing entity nodes (such as symptoms, diseases, and drugs) and relationship sets (such as "disease-symptoms"), and define attributes such as name, type, and metadata for each entity node;

[0010] Step 4: Store the constructed knowledge graph in the graph database, receive new data through the real-time data stream processing system (Kafka), dynamically update the knowledge graph and regularly retrain it to maintain the timeliness and accuracy of the knowledge;

[0011] Step 5: In the pre-consultation process, the system collects basic information and symptom descriptions of the patient through open-ended and closed-ended questions, and uses a large language model to parse the patient's answers to extract key information;

[0012] Step 6: The system dynamically adjusts the consultation path based on the extracted information combined with the knowledge graph, and clarifies vague answers or explores specific symptoms through further questions to ensure that comprehensive patient data is obtained;

[0013] Step 7: Use a large language model to generate structured electronic medical records, convert key information from patients’ answers into a standardized format, store them in the medical record database after scoring and evaluation to support future consultations, and regularly optimize the medical record generation process and knowledge base content.

[0014] Furthermore, step 1 is as follows:

[0015] Step 1.1: Obtain structured data from public medical databases (such as medical dictionaries, classification tables), including disease lists, drug tables, and standard terminology lists.

[0016] Step 1.2: Collect semi-structured data (such as medical classification systems ICD and SNOMED CT) and extract association information such as symptom-disease and disease-drug.

[0017] Step 1.3: Obtain unstructured data, such as medical literature, electronic medical records, medical records, etc., and extract key information through natural language processing technology.

[0018] Step 1.4: Clean the data to remove duplicate, incomplete, or conflicting records, and map and unify non-standardized terminology.

[0019] Step 1.5: Standardize the processed data to ensure that each piece of data has a fixed field format (such as entity type, attribute value, timestamp, etc.).

[0020] Furthermore, step 2 is as follows:

[0021] Step 2.1: Define a set of pattern templates (e.g., “drug A treats disease B”) and use regular expressions to automatically identify relation instances that match these patterns from the text;

[0022] Step 2.2: For structured and semi-structured data, directly match the existing relational template and perform data verification;

[0023] Step 2.3: Analyze the syntactic structure in the unstructured data and identify the relationship between the subject, predicate, and object through the syntax tree analysis tool;

[0024] Step 2.4: Design a rule set to identify the semantic relationship between medical entities by combining specific verbs (such as "inhibit", "relieve") and contextual information;

[0025] Step 2.5: Use semantic similarity analysis techniques to distinguish different relationship types (such as "disease-symptom" and "drug-side effect");

[0026] Step 2.6: Submit the preliminary results of pattern matching and rule-based methods to domain experts for review and annotation;

[0027] Step 2.7: Optimize the rule set based on expert feedback to ensure that it can adapt to more complex text scenarios and diverse medical expressions.

[0028] Furthermore, step 3 is as follows:

[0029] Step 3.1: Use the graph model G = (V, E), where V represents the set of entity nodes and E represents the set of edges connecting the entity nodes;

[0030] Step 3.2: Define attributes for each entity node, including entity name (such as "hypertension"), entity type (such as disease, drug, symptom), and detailed description and source metadata;

[0031] Step 3.3: Based on the extracted entity relationships, generate the edges of the graph and label their relationship types, such as "symptom-disease", "disease-drug", etc.

[0032] Step 3.4: Perform consistency check on all nodes and edges to ensure that the entities of each relationship are defined in the knowledge graph and avoid isolated nodes or invalid edges;

[0033] Step 3.5: Use graph visualization tools to visualize the knowledge graph to verify the intuitiveness and logical rationality of the structure.

[0034] Furthermore, step 4 is as follows:

[0035] Step 4.1: Choose a suitable graph database (such as Neo4j, JanusGraph) to store the knowledge graph to support efficient query and data association operations;

[0036] Step 4.2: Convert the knowledge graph data into a storage format supported by the graph database and import the data through the graph database API;

[0037] Step 4.3: Deploy a real-time data stream processing system (Kafka) to monitor updates from external data sources and automatically push new data to the knowledge graph update module;

[0038] Step 4.4: Fine-tune the model in a small range through the online learning algorithm, and update the knowledge graph structure and content based on the incremental data;

[0039] Step 4.5: Set a regular retraining cycle (such as monthly or quarterly), integrate historical data with new data to retrain the model, and deploy the optimized results to the production environment;

[0040] Step 4.6: Perform performance evaluation (such as node association accuracy and query response time) before and after the update to ensure the stability and efficiency of the knowledge graph after the update.

[0041] Furthermore, step 5 is as follows:

[0042] Step 5.1: The system starts the consultation process and asks the patient preset open-ended questions (such as "Please describe your main symptoms") or closed-ended questions (such as "How long does your pain last?") to collect preliminary information;

[0043] Step 5.2: The large language model performs natural language parsing on the patient’s responses and extracts key information from the semantics, including specific symptoms (such as back pain, neck pain), duration, and medical history;

[0044] Step 5.3: The system structures the patient’s answer, identifies medical entities (such as disease names, drug names), and matches them with nodes in the knowledge graph to verify validity;

[0045] Step 5.4: For answers that are ambiguous or lack sufficient information, the system generates clarifying follow-up questions through the large language model, prompting the patient to further provide a description (e.g., "Is the pain specifically in the lumbar sacral region, bilaterally, or in the middle?");

[0046] Step 5.5: All the preliminary information collected is stored in the memory layer as short-term storage to support the subsequent decision-making of dynamically adjusting the consultation path.

[0047] Furthermore, step 6 is as follows:

[0048] Step 6.1: The system calls the rule engine to analyze the patient's initial answer and generates a set of possible next questions based on the knowledge graph and the preset consultation path;

[0049] Step 6.2: If the symptoms mentioned by the patient point to a specific disease (such as "persistent low back pain"), the system will combine the knowledge graph to ask in-depth questions (such as "Is the low back pain accompanied by numbness in the lower limbs?");

[0050] Step 6.3: If the patient's answer is unclear (e.g., "I feel pain all over my waist"), the system generates context-based clarifying questions (e.g., "Is this pain dull, stabbing, aching, or bloating?") and continuously adjusts the consultation path;

[0051] Step 6.4: The system scores the completeness of the answer to each question (e.g. 60% completion is considered a partial answer), and incomplete content will trigger the subsequent question generation mechanism;

[0052] Step 6.5: In the process of dynamically adjusting the consultation path, the system refers to similar cases in the medical record library (such as those with similar symptoms and medical history) as additional auxiliary information to optimize the consultation efficiency and accuracy;

[0053] Step 6.6: The collected information is converted into structured data in real time through a large language model and updated to the memory layer to maintain the continuity of the consultation and the integrity of the data.

[0054] Furthermore, step 7 is as follows:

[0055] Step 7.1: The system calls the medical record generation module to summarize the answers and key information accumulated during the patient consultation process and convert them into standard medical terminology;

[0056] Step 7.2: The large language model generates structured electronic medical records based on the medical record database template (such as chief complaint, current medical history, and past medical history) to ensure the standardization of the format and fields;

[0057] Step 7.3: The system automatically scores the generated medical record content and conducts quantitative evaluation based on dimensions such as completeness, detail, relevance, and logical consistency (if the score > 90% is considered a high-quality medical record);

[0058] Step 7.4: Store high-quality structured medical records in the medical record database and associate them with the entity and relationship nodes in the knowledge graph to further enhance the professionalism and practicality of the knowledge base;

[0059] Step 7.5: Collect patient case data and expert feedback, and regularly adjust the medical record generation algorithm and scoring criteria to ensure that the generated medical records meet the latest medical standards and diagnosis and treatment needs;

[0060] Step 7.6: Use the newly added medical records to dynamically expand the knowledge graph, extract potential relationships and nodes, and continuously improve the completeness of the knowledge graph and the accuracy of intelligent medical consultation.

[0061] The medical history recognition system based on a large model and an expert strategy is applicable to the medical history recognition method based on a large model and an expert strategy, and includes:

[0062] Input module: The input module receives patient medical history information through text, voice, video, etc., and supports multi-channel access such as enterprise WeChat and mini programs.

[0063] Data preprocessing module,The data preprocessing module integrates structured, semi-structured and unstructured data, and uses NLP technology to clean, parse and standardize patient input.

[0064] Knowledge graph construction module,The knowledge graph construction module extracts entities and relationships from the data, fuses them to generate a multi-relationship graph and stores it in the graph database to support dynamic updates.

[0065] Expert strategy engine: The expert strategy engine combines diagnostic rules and consultation logic to dynamically adjust the consultation path and optimize the interactive content based on the patient's answers.

[0066] Large model parsing module,The large model parsing module uses a pre-trained language model to extract medical entities and semantic information from patient descriptions, and generates personalized questions to complete ambiguous information.

[0067] Medical history generation module,The medical history generation module extracts key information from the medical consultation records, generates structured electronic medical records, and scores and stores them for system optimization.

[0068] Output module, the output module provides doctors with consultation records and auxiliary diagnosis suggestions in real time, supports doctors to modify medical records and interact with patients.

[0069] Dynamic update and learning module: The dynamic update and learning module optimizes the knowledge graph and rule engine through online learning and real-time data stream processing to ensure continuous improvement of system performance.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] 1. The present invention provides a medical history recognition method and system based on a large model and expert strategy. Through an intelligent pre-consultation process, the system can collect detailed medical history information and generate standardized electronic medical records before the patient formally sees the doctor, greatly reducing the time for doctors to verify and supplement information during consultation.

[0072] 2. The present invention provides a medical history recognition method and system based on a large model and expert strategy, which uses a knowledge graph to deeply analyze patient descriptions, accurately extract key information and medical entities, and avoid incomplete medical histories due to unclear descriptions or missing information.

[0073] 3. The present invention provides a medical history recognition method and system based on a large model and expert strategy. By dynamically adjusting the rule-based consultation logic, the system can generate personalized follow-up questions, conduct detailed exploration of vague or complex symptoms, and improve the pertinence and professionalism of the consultation process.

[0074] 4. The present invention provides a medical history recognition method and system based on a large model and expert strategy. Combining medical knowledge and semantic reasoning capabilities, the system can associate the potential relationship between patient symptoms and diseases, accurately capture the characteristics of complex conditions or rare diseases, and reduce the probability of misdiagnosis due to information omissions or misunderstandings. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0076] Figure 1 Schematic diagram of the process steps of the present invention

[0077] Figure 2 This is a schematic diagram of the data flow of the system of the present invention.

[0078] Figure 3 This is a diagram of a structured electronic medical record.

[0079] Figure 4 Diagram of doctor-patient dialogue on WeChat for Enterprises

[0080] Figure 5 Diagram of the mini program doctor-patient dialogue

[0081] Figure 6 This is a real-time voice conversation diagram

[0082] Figure 7 It is a flow chart of a specific embodiment of the present invention DETAILED DESCRIPTION

[0083] The technical solution of the present invention will be more clearly and completely explained below through description of preferred embodiments of the present invention in combination with the accompanying drawings.

[0084] like Figure 1 As shown, the present invention is:

[0085] Step 1: Collect structured data, semi-structured data, and unstructured data from the medical network, and clean and standardize the data to provide a basis for building a knowledge graph;

[0086] Step 2: Extract medical entities and their relationships from texts through pattern matching and rule-based methods, and optimize the rule set in combination with the knowledge of domain experts;

[0087] Step 3: Integrate the extracted entities and relationships into the knowledge graph, build a multi-relationship graph containing entity nodes and relationship sets, and define attributes for each entity node;

[0088] Step 4: Store the constructed knowledge graph in the graph database, receive new data through the real-time data stream processing system, and dynamically update and regularly retrain the knowledge graph;

[0089] Step 5: In the pre-consultation process, the system collects basic information and symptom descriptions of the patient through open-ended and closed-ended questions, and uses a large language model to parse the patient's answers to extract key information;

[0090] Step 6: Dynamically adjust the consultation path based on the extracted information and knowledge graph, and clarify vague answers or explore specific symptoms in depth by asking questions;

[0091] Step 7: Use a large language model to generate structured electronic medical records, convert key information from patients’ answers into a standardized format, store them in the medical record database after scoring and evaluation to support future consultations, and regularly optimize the medical record generation process and knowledge base content.

[0092] The details are as follows:

[0093] Given the complexity of the patient's chief complaint information during the pre-consultation process, traditional methods are difficult to fully and accurately understand this information, which can easily lead to omissions or misjudgments of key data. To this end, the present invention adopts a pre-consultation question-and-answer system that combines a large language model with a knowledge graph. Through the medical knowledge graph, the system can obtain a professional knowledge base to ensure in-depth analysis of the patient's chief complaint information and accurately identify key medical entities and their relationships. The system includes:

[0094] Input module: The input module receives patient medical history information through text, voice, video, etc., and supports multi-channel access such as enterprise WeChat and mini programs.

[0095] Data preprocessing module,The data preprocessing module integrates structured, semi-structured and unstructured data, and uses NLP technology to clean, parse and standardize patient input.

[0096] Knowledge graph construction module,The knowledge graph construction module extracts entities and relationships from the data, fuses them to generate a multi-relationship graph and stores it in the graph database to support dynamic updates.

[0097] Expert strategy engine: The expert strategy engine combines diagnostic rules and consultation logic to dynamically adjust the consultation path and optimize the interactive content based on the patient's answers.

[0098] Large model parsing module,The large model parsing module uses a pre-trained language model to extract medical entities and semantic information from patient descriptions, and generates personalized questions to complete ambiguous information.

[0099] Medical history generation module,The medical history generation module extracts key information from the medical consultation records, generates structured electronic medical records, and scores and stores them for system optimization.

[0100] Output module, the output module provides doctors with consultation records and auxiliary diagnosis suggestions in real time, supports doctors to modify medical records and interact with patients.

[0101] Dynamic update and learning module: The dynamic update and learning module optimizes the knowledge graph and rule engine through online learning and real-time data stream processing to ensure continuous improvement of system performance.

[0102] In addition, the system uses a set of scoring criteria to achieve self-learning of the knowledge graph, continuously improving its ability to understand patient information.

[0103] The data flow in the system is as follows Figure 2As shown in the figure, the construction of knowledge graph involves three main links: knowledge extraction, knowledge fusion and knowledge storage. The data sources of the construction process mainly include structured data, semi-structured data and unstructured data obtained from medical networks, among which structured and semi-structured data are the main ones. In this process, extracting the relationship between entities from unstructured data such as text and medical records is a key task. This process is mainly achieved through pattern matching and rule-based methods. Pattern matching is a method of identifying specific relationship expressions. It relies on a predefined series of patterns or templates that describe the expression form of common relationships in text. For example, when processing medical literature, a pattern can be defined to capture expressions such as "drug A is used to treat disease B". The regular expression system can automatically identify and extract relationship instances that conform to these patterns. This method is simple and direct and is suitable for relationship types with clear and fixed expressions.

[0104] The rule-based approach is more flexible and powerful. It not only uses simple string matching, but also combines advanced natural language processing techniques such as grammatical analysis and semantic understanding. This approach usually includes a complex set of rules that cover multiple levels such as vocabulary, syntax, and contextual information to more accurately identify relationships in the text. For example, a rule can be set to indicate that when a verb (such as "inhibit") appears between two medical entities, there is a certain relationship between the two entities; or based on sentence structure analysis, determine the specific connection between subject-predicate-object. The rule-based approach allows for more detailed control of the relationship extraction process and can handle more complex and diverse text situations.

[0105] By combining pattern matching with rule-based methods, we can effectively extract the relationship between entities from unstructured medical texts, laying a solid foundation for building a rich and accurate knowledge graph. At the same time, in order to further improve the quality of extraction, we integrate the expertise of domain experts to design and optimize rules.

[0106] The knowledge graph is defined as a multi-relation graph G = (V, E), where V represents the set of entity nodes. Represents the set of edges connecting the nodes in V, and R represents the set of relationship types. Common relationship types include "symptom-disease", "disease-drug", "drug-side effect", etc.

[0107] For a given question (q)-answer (a) pair, define V q,a =V q ∪V a , represents all entities contained in the question (q) or answer (a). We further define a subgraph in G, which This subgraph contains All paths between nodes in G. Use the pre-trained language model to score the relevance of each node V in G, using the formula ρ v =f rel (f enc ([q; a])) calculates the relevance score, where f enc Vector representation for question (q)-answer (a) pairs, f rel Reasoning and subgraph pruning tasks are then performed based on this vector representation.

[0108] To build an efficient and reliable medical knowledge graph system, we first prepared a dataset containing question-answer pairs and annotated the relevant entities and their relevance scores for each entry. The dataset was divided into training, validation, and test sets. A joint model was constructed by combining a pre-trained language model (BERT) with a graph neural network (GCN). The pre-trained language model was used to extract the semantic features of the question-answer pairs, while the graph neural network was responsible for generating node embeddings and performing reasoning tasks. The model was trained on the training set, and the validation set was used to adjust the hyperparameters to optimize the performance. Finally, the accuracy, recall, and F1 score of the model were evaluated on the test set. K-fold cross validation was used to ensure the consistency of model performance.

[0109] In order to achieve dynamic update of the graph, the present invention sets up a real-time data stream processing system (Kafka) to receive new data. Use online learning algorithms to fine-tune the existing model and regularly save the updated model version. Set a regular retraining cycle based on the frequency of data updates. Integrate all new data into the current data set, retrain the model, and evaluate its performance on the validation set. If the performance improves, deploy the new model.

[0110] During the inference process, the pre-trained language model first generates a vector representation of the question-answer pair, which is then passed to the graph neural network to generate node embeddings. Based on this embedded information, the graph neural network performs the inference task and outputs the relevance score. The cross entropy loss function is defined to optimize the entire model. In the inference stage, the relevance threshold is set to 0.9, and nodes and their connecting edges with scores higher than this threshold are retained, thereby improving query efficiency and accuracy.

[0111] In addition, the present invention integrates the diagnostic rules and clinical guidelines formulated by medical experts into the knowledge graph. These rules can guide the system to select the appropriate consultation path during the pre-consultation process. Collect a large amount of clinical medical records to form a medical record library. Generate structured electronic medical records expressed in professional terms with the help of a large language model, such as Figure 3As shown in the figure, Drools and other rule engines are used to store formalized pre-consultation logic and processes. The system calls the rule engine through the API interface and dynamically adjusts the consultation path based on the patient's input. Combining the patient's answer and the output of the rule engine, a large language model is used to generate high-quality structured medical records to ensure the professionalism and accuracy of the content.

[0112] In the intelligent pre-consultation system, building an expert strategy to guide the pre-consultation path is a comprehensive and dynamic process. First, a series of initial questions are designed as the starting point of the pre-consultation process, which are designed to collect basic information and preliminary symptom descriptions of patients. These questions are open-ended or closed-ended to accommodate different types of user input and ensure that the patient's health status can be widely covered.

[0113] The system uses advanced large language models to parse patients' responses. Such models have strong natural language understanding capabilities and can extract key information from complex semantic structures, such as specific symptoms, medical history, etc., and identify implicit meanings. In this way, the system can not only accurately grasp the patient's direct feedback, but also capture potential important details, thereby gaining a more comprehensive understanding of the patient's health status.

[0114] Based on the understanding of the user's input, the system will dynamically decide the next most appropriate inquiry direction. This decision-making process involves evaluating the patient's specific situation based on the information currently collected. For example, if the patient mentions persistent low back pain, the system may further explore the nature of the low back pain, such as whether it is accompanied by symptoms such as numbness or radiating pain in the lower limbs; if the symptoms are more severe or point to a specific disease, it may directly ask for detailed information about the relevant disease. In addition, for unclear answers, the system will take further clarification measures to ensure that the information obtained is as accurate as possible.

[0115] In order to support this complex decision-making process, the system needs a large and comprehensive knowledge base. This knowledge base contains detailed information, typical symptoms, diagnosis and treatment guidelines of various diseases, etc., to help the algorithm make more scientific and reasonable judgments and ensure that the design of the pre-diagnosis path is both in line with medical logic and efficient and practical.

[0116] As each pre-consultation is completed, the system can collect the user's final diagnosis results and other relevant data to continuously improve the performance of the model. By analyzing which paths are more efficient and which problems can be quickly located, the design of the entire pre-consultation process is gradually optimized. At the same time, throughout the process, the system strictly abides by medical privacy regulations, properly handles personal health information, protects user data security, and ensures that all operations are carried out under the premise of legality and compliance.

[0117] In this way, the intelligent pre-consultation system can not only provide a personalized consultation experience and improve diagnostic efficiency, but also enhance the user experience. This approach helps to reduce the workload of medical staff, allowing them to focus on dealing with more complex situations, thereby improving the quality and efficiency of medical services as a whole.

[0118] The big model is called through dynamic prompt words to generate personalized responses to patients, achieving efficient pre-consultation. First, the interaction goal must be clearly defined, whether it is to collect preliminary information or provide health advice. Based on this, a series of question templates are designed, combining open and closed questions to better guide the conversation. Natural language processing technology is used to understand user input and automatically generate corresponding responses based on it, including keyword recognition and sentiment analysis. The rule engine will confirm the next consultation path based on the user's answer, and dynamically adjust the subsequent questioning method and content to ensure the coherence and relevance of the conversation. For example, if the patient mentions a headache, the subsequent questioning will focus on the specific nature of the headache. Throughout the process, strictly abide by data protection laws and regulations, ensure the security of patient information, and clearly inform them of how their information is used and obtain necessary consent. Although AI can assist in preliminary screening, patients should still be encouraged to seek professional medical advice and provide reliable resource links. In this way, customized pre-consultation services can be provided to each patient, improving efficiency and accuracy while ensuring safety.

[0119] The intelligent medical pre-consultation system integrates the real-time audio and video call functions of enterprise WeChat and mini programs for dialogue, such as Figure 4-6 As shown in the figure, it provides users with a multi-channel communication platform with digital doctors. The system focuses on efficiently collecting patients' preliminary symptom information, supports text communication and high-definition and smooth audio and video interaction, and ensures that users can choose the most appropriate communication method according to their needs. Whether it is through corporate WeChat for quick consultation or using mini programs to enter symptom data, the system can accurately record and organize patient information and provide reliable data support for subsequent diagnosis and treatment. This multi-dimensional support not only improves the convenience and accuracy of information collection, but also enhances user experience and optimizes the use of medical resources.

[0120] like Figure 7 As shown in the figure, the doctor-patient dialogue is processed in a branch node of the decision tree, including dialogue parsing, dialogue memory, process control, etc.

[0121] 1. Dialogue Parser:

[0122] The conversation parser will dynamically generate prompt words, and combined with the professional data in the conversation knowledge base, call the large language model to parse the entire conversation between the doctor and the patient at the current node. The system analyzes the completion of the patient's answer to each question in the question queue of the current node. Only questions with a completion rate of more than 60% are considered to have been answered, and the answer to the question is recorded in the memory layer. For questions with a completion rate of more than 90%, the system will extract the question and reply content of the question and store it in the knowledge base to achieve self-learning of the knowledge base. When the memory layer records all the questions in the question queue, the process controller will be triggered and automatically jump to the next node through rule matching in the decision tree. Specifically, the conversation parser dynamically generates prompt words containing scoring criteria and example question-answer pairs based on the question type of the current node and the examples in the doctor-patient conversation knowledge base to help the large language model understand the scoring criteria and evaluation methods. The system calls the large language model to parse the conversation and scores the patient's answer based on the prompt words and scoring criteria, and gives a detailed scoring explanation. The system analyzes the completion of the answer based on the scoring results of the large language model. Only questions with a completion rate of more than 60% are considered to be answered and recorded in the memory layer. For questions with a completion rate of more than 90%, the system stores the questions and responses in the knowledge base for future scoring examples, continuously enriching and improving the knowledge base. When the memory layer records all the questions in the question queue, the process controller will trigger and automatically jump to the next node through rule matching in the decision tree to continue the consultation process. This not only improves the flexibility and accuracy of the system, but also ensures that sufficient information is collected and realizes self-learning of the knowledge base.

[0123] 2. Dialogue Builder:

[0124] For questions in the question queue with a completion rate of less than 60%, since the memory layer does not have the memory of this question, the dialogue generator will dynamically generate prompt words for this question again, and then the large model will ask the patient again about this question from another angle.

[0125] The same applies to unasked questions.

[0126] 3. Memory layer:

[0127] In order to ensure the updating and maintenance of the medical record database and generate structured electronic medical records in combination with a large language model, the system adopts a comprehensive approach. Short-term memory is used as a temporary storage mechanism to save all relevant information in the current session, including the patient's answers and the doctor's questions. This information is only valid during the session and will be cleared or archived after the session ends. The data structure of short-term memory is in the form of a dictionary, where the key is the question ID or question description and the value is the patient's answer. Whenever the patient provides an answer or the doctor asks a new question, the system will update the short-term memory in real time.

[0128] To avoid redundancy, before recording a new question-answer pair, the system checks whether the same or similar content already exists; if so, it will not record it again, but will choose to update the original entry. The information in short-term memory is arranged in chronological order, with the latest information at the front, while allowing important information to be marked for quick access. This design ensures the timeliness and accuracy of information, while avoiding the existence of redundant and outdated data, and improving the flexibility and practicality of the system.

[0129] In terms of implementation, the system initializes an empty short-term memory dictionary. Whenever a new question-and-answer interaction occurs, the information is dynamically maintained by parsing the conversation and updating the short-term memory. During the update process, the system automatically detects and processes duplicates to ensure that the short-term memory is always up-to-date and non-redundant. At the end of the session, the information in the short-term memory will be cleared or archived to free up memory resources. In this way, the system can provide immediate and accurate information support in each session, thereby improving the quality and efficiency of medical services.

[0130] 4. Medical record analysis:

[0131] In order to ensure the updating and maintenance of the medical record database and generate structured electronic medical records in combination with a large language model, the system adopts a comprehensive approach. The initial medical record database is manually created by medical experts and contains detailed information such as the patient's chief complaint, current medical history, past history, personal history, menstrual marital history, and allergy history. The expert team will regularly review the content of the medical record database to ensure its accuracy and timeliness, and can also add new medical record entries or correct existing entries. The system automatically obtains the latest medical record information of patients through integration with the hospital's electronic medical record system (EMR), and obtains the latest medical knowledge and case data from academic literature and public databases to enrich the medical record database. The acquired data is cleaned and standardized using natural language processing technology to extract key information and convert it into a structured format.

[0132] Before the process controller is ready to jump to the next node, the system uses the cosine similarity algorithm to obtain the three structured medical records that are most similar to the current conversation from the medical record library. Then, the large language model is called to parse the conversation between the doctor and the patient at the current node into a structured medical record with professional medical terms. The final structured medical record is filled and generated according to the template in the medical record library and the content generated by the large model. The generated structured medical record will be scored, and the scoring criteria include completeness, detail, relevance and logical consistency. If the score of the generated electronic medical record exceeds 90%, it is considered to be of high quality and automatically stored in the medical record library, thereby continuously enriching the content of the medical record library and ensuring that the information in the medical record library is always kept up to date and accurate. Through this combination of manual updates by experts, automatic data acquisition, and self-learning mechanism based on scoring, the medical record library can maintain a high degree of professionalism and accuracy, while improving the flexibility and practicality of the system, ensuring that sufficient information is collected, and realizing continuous optimization of the knowledge base, so as to better serve the clinical diagnosis and treatment process.

[0133] As a specific implementation method, the system receives the patient's medical history information input by text, voice or video, and uniformly accesses it through enterprise WeChat or mini-programs. The received patient information is cleaned and parsed by the data preprocessing module, and medical-related entities are extracted through natural language processing technology. The unstructured or semi-structured data is standardized into a structured data format and passed to the core module. The core module uses a large language model to semantically understand the patient's input, extract key medical entities in the medical history, such as disease names, symptoms, and drugs, and match related nodes in combination with the knowledge graph to derive the potential relationship between entities. Based on the patient's answer, the expert strategy engine adjusts the consultation path according to preset rules, dynamically generates follow-up questions, and especially further clarifies vague or incomplete answers to ensure the integrity of the consultation information.

[0134] During the consultation process, the system compares the patient's answers with the knowledge graph and historical medical records in real time, selects the most relevant consultation path and generates a consultation record, and uses a large language model to generate supplementary questions to guide the patient to complete the medical history description. After the consultation, the system automatically integrates the patient's answers, extracts key information and generates a structured electronic medical record, including fields such as chief complaint, current medical history, past history and allergy history. The generated medical records are evaluated for quality through a scoring mechanism. The scoring criteria include logic, completeness and relevance. Medical records with qualified scores will be stored in the medical record library for subsequent use. The content in the medical record library will be regularly used to optimize the knowledge graph structure and expand medical entities and relationships.

[0135] On the doctor's side, the system visualizes the generated medical records and analysis results, and provides diagnostic suggestions for suspected diseases and recommendations for further consultation paths. Doctors can modify the content of medical records and further communicate with patients through corporate WeChat or mini programs to improve medical record information. The system receives new data through real-time data stream processing tools (such as Kafka), and updates the big model and rule engine through online learning. At the same time, it optimizes the knowledge graph structure based on data accumulation to ensure the accuracy and real-time nature of the consultation. The entire system combines the powerful semantic understanding capabilities of the big model with the expertise of expert strategies to provide patients with efficient and accurate medical history identification services, while reducing the workload of doctors and improving the efficiency of medical resource utilization.

[0136] The above specific implementations are only descriptions of the preferred implementations of the present invention, and do not limit the protection scope of the present invention. Without departing from the design concept and spirit of the present invention, various modifications, substitutions and improvements made by ordinary technicians in this field to the technical solution of the present invention based on the text description and drawings provided by the present invention should all fall within the protection scope of the present invention. The protection scope of the present invention is determined by the claims.

Claims

1. A medical history recognition method based on a large model and expert strategy, characterized in that: The following steps are involved: Step 1: Collect structured data, semi-structured data, and unstructured data from the medical network, and clean and standardize the data to provide a basis for building a knowledge graph; Step 2: Extract medical entities and their relationships from texts through pattern matching and rule-based methods, and optimize the rule set in combination with the knowledge of domain experts; Step 3: Integrate the extracted entities and relationships into the knowledge graph, build a multi-relationship graph containing entity nodes and relationship sets, and define attributes for each entity node; Step 4: Store the constructed knowledge graph in the graph database, receive new data through the real-time data stream processing system, and dynamically update and regularly retrain the knowledge graph; Step 5: In the pre-consultation process, the system collects basic information and symptom descriptions of the patient through open-ended and closed-ended questions, and uses a large language model to parse the patient's answers to extract key information; Step 6: Dynamically adjust the consultation path based on the extracted information and knowledge graph, and clarify vague answers or explore specific symptoms in depth by asking questions; Step 7: Use a large language model to generate structured electronic medical records, convert key information from patients’ answers into a standardized format, store them in the medical record database after scoring and evaluation to support future consultations, and regularly optimize the medical record generation process and knowledge base content.

2. The medical history recognition method based on a large model and expert strategy according to claim 1 is characterized in that: Step 1 is as follows: Step 1.1: Obtain structured data from public medical databases; Step 1.2: Collect semi-structured data and extract relevant information; Step 1.3: Obtain unstructured data and extract information from it through natural language processing technology; Step 1.4: Clean the data to remove duplicate, incomplete or conflicting records, and map and unify non-standardized terms; Step 1.5: Standardize the processed data to ensure that each piece of data has a fixed field format.

3. The medical history recognition method based on large model and expert strategy according to claim 1 is characterized in that: Step 2 is as follows: Step 2.1: Define a set of pattern templates and use regular expressions to automatically identify relation instances that match the pattern from the text; Step 2.2: For structured and semi-structured data, match the existing relational template and perform data verification; Step 2.3: Analyze the syntactic structure in the unstructured data and identify the relationship between the subject, predicate, and object through syntax tree analysis tools; Step 2.4: Design a rule set to identify the semantic relationship between medical entities by combining specific verbs and context information; Step 2.5: Use semantic similarity analysis techniques to distinguish different relationship types; Step 2.6: Submit the preliminary results of pattern matching and rule-based methods to domain experts for review and annotation; Step 2.7: Optimize the rule set based on expert feedback.

4. The medical history recognition method based on large model and expert strategy according to claim 1 is characterized in that: Step 3 is as follows: Step 3.1: Use the graph model G = (V, E), where V represents the set of entity nodes and E represents the set of edges connecting the entity nodes; Step 3.2: Define properties for each entity node, including entity name, entity type, detailed description and source metadata; Step 3.3: Based on the extracted entity relationships, generate the edges of the graph and label their relationship types; Step 3.4: Perform consistency check on all nodes and edges to ensure that the entities of each relationship are defined in the knowledge graph and avoid isolated nodes or invalid edges; Step 3.5: Use graph visualization tools to visualize the knowledge graph and verify the intuitiveness and logical rationality of the structure.

5. The medical history recognition method based on large model and expert strategy according to claim 1 is characterized in that: Step 4 is as follows: Step 4.1: Select a graph database to store the knowledge graph; Step 4.2: Convert the knowledge graph data into a storage format supported by the graph database and import the data through the graph database API; Step 4.3: Deploy a real-time data stream processing system to monitor updates from external data sources and automatically push new data to the knowledge graph update module; Step 4.4: Fine-tune the model in a small range through the online learning algorithm, and update the knowledge graph structure and content based on the incremental data; Step 4.5: Set up retraining cycles regularly, integrate historical data and new data to retrain the model, and deploy the optimized results to the production environment; Step 4.6: Perform performance evaluation before and after the update.

6. The medical history recognition method based on large model and expert strategy according to claim 1 is characterized in that: Step 5 is as follows: Step 5.1: Start the interview process and ask the patient preset open-ended or closed questions to collect preliminary information; Step 5.2: The large language model performs natural language parsing on the patient’s responses and extracts key information from the semantics, including specific symptoms, duration, and medical history; Step 5.3: Structuring the patient’s answers, identifying medical entities, and matching them with nodes in the knowledge graph to verify validity; Step 5.4: For answers that are semantically ambiguous or lack sufficient information, the system generates clarifying follow-up questions through a large language model, prompting the patient to further provide additional descriptions; Step 5.5: All the collected preliminary information is stored in the memory layer as short-term storage to support the subsequent decision-making of dynamically adjusting the consultation path.

7. The medical history recognition method based on large model and expert strategy according to claim 1 is characterized in that: Step 6 is as follows: Step 6.1: Call the rule engine to analyze the patient's initial answer and generate the next set of questions based on the knowledge graph and the preset consultation path; Step 6.2: If the symptoms mentioned by the patient point to a specific disease, further questions are asked based on the knowledge graph; Step 6.3: If the patient’s answer is unclear, the system generates context-based clarifying questions and continuously adjusts the consultation path; Step 6.4: The system scores the completeness of the answer to each question, and the incomplete content will trigger the subsequent question generation mechanism; Step 6.5: In the process of dynamically adjusting the consultation path, the system refers to similar cases in the medical record library as additional auxiliary information; Step 6.6: The collected information is converted into structured data in real time through the large language model and updated to the memory layer.

8. The medical history recognition method based on large model and expert strategy according to claim 1 is characterized in that: Step 7 is as follows: Step 7.1: The system calls the medical record generation module to summarize the answers and key information accumulated during the patient consultation process and convert them into standard medical terminology; Step 7.2: The large language model generates structured electronic medical records based on the medical record library template to ensure the standardization of the format and fields; Step 7.3: The system automatically scores the generated medical record content and performs quantitative evaluation according to the dimensions; Step 7.4: Store high-quality structured medical records in the medical record database and associate them with the entity and relationship nodes in the knowledge graph; Step 7.5: Collect patient medical record data and expert feedback, and regularly adjust the medical record generation algorithm and scoring criteria; Step 7.6: Use the newly added medical records to dynamically expand the knowledge graph, extract potential relationships and nodes, and continuously improve the completeness of the knowledge graph and the accuracy of intelligent medical consultation.

9. A medical history recognition system based on a large model and an expert strategy, applicable to the medical history recognition method based on a large model and an expert strategy described in claims 1-8, characterized in that: include: An input module, the input module is used to receive patient medical history information; Data preprocessing module: The data preprocessing module is used to integrate structured, semi-structured and unstructured data, and use NLP technology to clean, parse and standardize patient input; The knowledge graph construction module is used to extract entities and relationships from data, fuse them to generate multi-relationship graphs and store them in the graph database; Expert strategy engine: The expert strategy engine combines diagnostic rules and inquiry logic to dynamically adjust the inquiry path and optimize the interactive content based on the patient's answers; Large model parsing module: The large model parsing module is used to extract medical entities and semantic information from patient descriptions using a pre-trained language model and generate personalized questions to complete ambiguous information; Medical history generation module: The medical history generation module is used to extract key information from the medical consultation records, generate structured electronic medical records, and score and store them for system optimization; Output module: The output module is used to provide doctors with consultation records and auxiliary diagnosis suggestions in real time, supporting doctors to modify medical records and interact with patients; Dynamic update and learning module, the dynamic update and learning module optimizes the knowledge graph and rule engine through online learning and real-time data stream processing.

Images