Patient information collection and medical record construction system and method based on multiple rounds of dialogues

Abstract

The invention provides a patient information collection and medical record construction system and method based on multiple rounds of dialogues, and the system comprises an intelligent guide interaction module which receives the natural language input of a patient; the context-aware question and answer engine adopts a dialogue state representation method based on a graph structure to construct entities, relationships and attributes of each round of dialogue into knowledge sub-graphs; the medical record information dynamic builder monitors an updating event of the dialogue state diagram in real time; the abnormal information detection module adopts a mixed conflict detection method combining rules and learning; and the medical knowledge graph support system maps the oral expression of the patient to a standard medical term system in real time through a multi-level semantic matching strategy. According to the invention, the doctor does not need to distract the record in the inquiry process, and can pay more attention to patient observation and clinical thinking. The quality of the medical record first draft automatically generated by the system is high, a doctor only needs to perform a small amount of auditing and supplementing, and the medical record writing time is greatly shortened.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a system and method for patient information collection and medical record construction based on multi-turn dialogue. Background Technology

[0002] In modern healthcare systems, accurate and complete medical records are fundamental to clinical diagnosis and treatment decisions. Traditional medical information collection primarily relies on doctors obtaining patient histories through face-to-face consultations and manually recording them into the medical record system. This approach faces numerous challenges in practical application.

[0003] First, manual record-keeping places a heavy workload on medical staff. Statistics show that doctors spend approximately 35% of their daily work time writing medical records and entering information. This not only consumes valuable consultation time but also easily leads to doctor fatigue and affects the quality of care. Especially during peak outpatient periods, the consultation time for each patient is strictly limited, and doctors often need to quickly record information while asking about their symptoms, making it difficult to ensure both the comprehensiveness and accuracy of information collection.

[0004] Secondly, traditional consultation methods are prone to incomplete medical history information. Due to time constraints and communication barriers, doctors may overlook important details of the medical history. Patients often use non-professional, everyday language when describing their symptoms, such as "chest tightness" or "dizziness," requiring doctors to ask follow-up questions to obtain accurate medical information. However, in busy clinical environments, thorough follow-up questioning is often difficult to achieve. Furthermore, patients may not proactively mention certain information due to anxiety, forgetfulness, or deeming it unimportant, leading to the loss of crucial information. Studies show that approximately 20% of misdiagnosis cases are related to incomplete medical history collection.

[0005] Existing medical record systems have a low degree of structure. Although electronic medical record systems are widely used, doctors still primarily record medical history information in free text format. This unstructured recording method poses difficulties for subsequent data analysis, quality control, and clinical decision support. Medical institutions struggle to quickly extract valuable information from massive amounts of text-based medical records, failing to fully leverage the value of big data in healthcare. Furthermore, unstructured medical records also hinder information sharing and interoperability between different medical institutions.

[0006] Therefore, there is an urgent need for a new technological solution that can automatically understand patients' natural language expressions, intelligently guide multi-turn dialogues to collect complete medical history information, and transform it into high-quality structured medical records. This would reduce the burden on medical staff, improve consultation efficiency and medical record quality, and provide strong support for the development of intelligent healthcare. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention proposes a patient information collection and medical record construction system and method based on multi-turn dialogue. All collected information undergoes standardized processing and is mapped to a standardized medical terminology system, eliminating differences caused by doctors' individual writing habits. Each part of the medical record is logically clear, complete, and no key information is omitted, providing a high-quality data foundation for subsequent clinical diagnosis and treatment. This standardized medical record also greatly promotes the sharing and secondary use of medical data.

[0008] To achieve the above objectives, this invention proposes a patient information collection and medical record construction system based on multi-turn dialogue, characterized by comprising:

[0009] The intelligent guidance and interaction module receives the patient's natural language input, extracts medical entities and semantic features from the current input, and, based on a pre-trained medical domain large language model, calculates and generates follow-up questions most relevant to the current condition through an attention mechanism, according to the completeness score of the currently collected information and the priority weight of the fields to be filled.

[0010] The context-aware question-answering engine adopts a graph-based dialogue state representation method, which constructs the entities, relationships and attributes of each round of dialogue into a knowledge subgraph. It dynamically maintains the dialogue history information through an incremental update mechanism of dialogue state. Newly identified entities are added to the graph as nodes, and medical relationships between entities are used as edges to connect nodes. Global information propagation is achieved through graph neural networks to ensure that relevant information in the historical dialogue can be referenced when generating subsequent questions and answers.

[0011] The medical record information dynamic builder monitors the update events of the dialogue state graph in real time. When a new medical entity node is detected, it calculates the final confidence score through a multi-source confidence fusion mechanism. This mechanism uses log-odds transformation to fuse the confidence scores of multiple information sources and introduces a context consistency bias term based on cosine similarity to ensure the semantic consistency between new information and existing information. The information is only filled into the medical record template when the fused confidence score exceeds a preset threshold.

[0012] The abnormal information detection module adopts a hybrid conflict detection method that combines rules and learning. First, it detects hard contradictions through a medical logic constraint rule base. For entities of the same type, it judges semantic-level conflicts through a deep learning model, thereby achieving multi-level information consistency assurance.

[0013] The medical knowledge graph support system uses a multi-level semantic matching strategy to map patients' colloquial expressions to a standard medical terminology system in real time.

[0014] Furthermore, the questioning strategy of the intelligent guided interaction module is generated based on the principle of maximizing information gain. By calculating the information entropy of each unfilled field and the conditional entropy between the field and the filled field, the field that can minimize the overall uncertainty is selected for questioning. The influence between related fields is weighted by the inverse of the graph distance.

[0015] Furthermore, the context-aware question-answering engine's dialogue state graph update includes: initializing feature vectors based on entity type and attributes when nodes are created; automatically reasoning and establishing implicit relationships between entities based on a medical knowledge base; updating node features and propagating information through a multi-layer graph convolutional network; and intelligently resolving conflicts when new and old information conflict, taking into account timestamps, information source reliability, and confidence.

[0016] Furthermore, the medical record information dynamic builder performs incremental information filling, including: mapping medical information extracted from the dialogue to standard fields of the medical record template through a machine learning model; setting dependencies and triggering rules between fields to achieve cascading updates of related fields; and recording detailed version information for each information update operation to support complete tracing of the medical record modification history.

[0017] Furthermore, it also includes an active learning module, which designs a reward mechanism for medical semantic perception. This module provides higher rewards for actions that collect key clinical information, with the reward value decreasing exponentially with the duration of the dialogue, thereby guiding the system to prioritize and quickly collect the information most important for diagnosis.

[0018] The method for patient information collection and medical record construction based on multi-turn dialogue, applicable to the aforementioned patient information collection and medical record construction system based on multi-turn dialogue, includes the following steps:

[0019] Step S1: Load the pre-trained language model for the medical field, standard medical record template, medical knowledge graph and logical constraint rule base, create an empty dialogue state graph, and initialize the medical record completeness threshold and the clinical priority of each field;

[0020] Step S2: Generate a personalized greeting based on the current time, department, and other information to guide the patient in describing their main symptoms;

[0021] Step S3: After receiving patient input, preprocessing is performed first, including speech recognition, word segmentation, and time standardization; then, a medical entity recognition model is used to extract entities such as symptoms, diseases, and drugs and their attributes; finally, the medical relationships between entities are identified to form the structured information of this round of dialogue.

[0022] Step S4: Update the dialogue state graph with the newly identified entities and relationships incrementally, update the attributes of existing entities, and create corresponding nodes for new entities; propagate information throughout the graph using a graph neural network so that each node can perceive the relevant context.

[0023] Step S5: Traverse the updated dialogue state graph, extract medical information and map it to medical record template fields; when a field has multiple candidate values, calculate the final confidence score through a multi-source confidence fusion mechanism.

[0024]

[0025] in:

[0026] Conf fused The final confidence level after fusion, with a value ranging from [0,1].

[0027] σ is the sigmoid activation function, used to normalize the output to the [0,1] interval.

[0028] n represents the number of information sources participating in the integration.

[0029] w i Let the weight of the i-th information source satisfy the following condition: Dynamically adjusted based on the historical accuracy of the information source.

[0030] conf i The initial confidence level provided for the i-th information source, with a value ranging from [0,1].

[0031] For log-odds transformation, the confidence level is mapped to the real number domain.

[0032] b context This is the context consistency bias term, obtained by calculating the cosine similarity between the new information vector and the average vector of its graph neighbor nodes, with a value ranging from [-1, 1].

[0033] Step S6: Perform multi-level consistency checks on the newly populated information, including rule-based medical logic checks, temporal rationality verification, and deep learning-based semantic conflict identification, and generate a conflict list sorted by severity.

[0034] Step S7: Determine the next action based on the current completeness and conflict status of the medical record: If the information is complete and there are no conflicts, end the data collection; if there are conflicts, prioritize generating clarification questions; otherwise, continue with routine data collection.

[0035] Step S8: For the most serious conflicts, combine the conflict type and contextual information to generate natural and tactful clarifying questions using a language model, avoiding making the patient feel questioned;

[0036] Step S9: Select the next query target based on the principle of maximizing information gain, and generate a natural language question containing appropriate context by comprehensively considering the uncertainty of the field, its dependence on other fields, and its clinical importance.

[0037] Step S10: After the data collection is completed, a global consistency check and medical reasoning supplement are performed on the entire dialogue information. The structured medical record is output in a standard format, and a quality assessment report containing multiple dimensions such as completeness and accuracy is generated.

[0038] Furthermore, multiple rounds of dialogue are achieved by cyclically executing steps S3 to S9. Each round of dialogue is incrementally updated based on the preceding information until the medical record information reaches a preset level of completeness and is free of logical conflicts.

[0039] Furthermore, it supports multimodal inputs such as voice, text, and medical images, and converts multimodal information into a structured representation through corresponding recognition and understanding modules.

[0040] Furthermore, the feature is that it includes an online learning mechanism, which continuously optimizes the model performance by collecting doctors' feedback on the system output and using a medical semantic perception reward function, in which higher rewards are given for the behavior of quickly and accurately collecting key clinical information.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] 1. This invention provides a system and method for patient information collection and medical record construction based on multi-turn dialogue. Through a global information integration mechanism using a dialogue state graph, it effectively avoids the information omission problem in traditional consultations. The system can continuously track all information provided by the patient throughout the dialogue process and establish relationships between information through a graph structure. Even when the patient's description is scattered across multiple turns of dialogue, the system can still accurately capture and integrate this fragmented information to construct a complete picture of the patient's condition. A follow-up questioning strategy based on maximizing information gain ensures that the system always prioritizes asking the most valuable questions, avoiding aimless questioning and enabling the collection of the most critical clinical information within a limited consultation time.

[0043] 2. This invention provides a system and method for patient information collection and medical record construction based on multi-turn dialogue, achieving a deep understanding of patients' natural language expressions. By integrating pre-trained language models and medical domain knowledge, the system can accurately understand patients' colloquial descriptions, local slang, and even vague expressions.

[0044] 3. This invention provides a patient information collection and medical record construction system and method based on multi-turn dialogue. Through the global information propagation mechanism of graph neural networks, the system considers all acquired information when generating each question, achieving true context-aware dialogue. The system does not mechanically follow a fixed questioning process but dynamically adjusts its consultation strategy according to the patient's specific situation. When contradictory information is detected, the system can intelligently generate clarifying questions to subtly guide the patient to provide accurate information. Attached Figure Description

[0045] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the system architecture of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be more clearly and completely explained below with reference to the accompanying drawings and through the description of preferred embodiments of the present invention.

[0048] like Figure 1 As shown, this invention provides a patient information collection and medical record construction system based on multi-turn dialogue. This system realizes intelligent collection and structured output of patient information by simulating the doctor's consultation thinking.

[0049] The system workflow begins with the patient's input, where the patient can describe their symptoms using natural language, such as "I've been feeling dizzy and a little nauseous lately." This unstructured natural language information is first processed by the intelligent guidance and interaction module.

[0050] The intelligent guided interaction module is the core of the system's front-end processing, undertaking two key tasks. First, it performs entity recognition, accurately extracting medical entities from the patient's natural expressions, such as recognizing "dizziness" as a symptom entity and "recently" as a time entity. Second, it generates follow-up question strategies, intelligently determining the next question to ask based on the currently collected information. The module's output includes the recognized entity information and the generated follow-up questions, which are returned to the patient via the dotted feedback path shown in the diagram, forming a multi-turn dialogue interaction.

[0051] The identified entity information is then passed to the context-aware question-answering engine, a key module for the system to understand and maintain dialogue coherence. This engine does not process each round of dialogue in isolation, but rather integrates all dialogue information into a unified dialogue state graph G(t). The dialogue state graph is the core innovation of this system; it organizes all the information provided by the patient in multiple rounds of dialogue into a graph structure, where medical entities are nodes and the relationships between entities are edges. For example, there might be a relationship edge of "accompanying symptoms" between the two symptom nodes "dizziness" and "nausea".

[0052] The dialogue state graph is dynamically updated, evolving with each round of dialogue. When the patient provides new information in subsequent conversations, the system intelligently determines whether to create a new node or update the attributes of an existing node. This graph structure design allows the system to have a global grasp of the patient's health status, avoiding information omissions or duplicate inquiries.

[0053] Based on the continuously updated dialogue state graph, the system performs two important tasks in parallel. On one hand, the medical record information dynamic builder extracts relevant information from the graph and maps it to a standardized medical record template through a confidence fusion mechanism. Confidence fusion is a key technology for handling information uncertainty. When the same information comes from multiple sources or is mentioned multiple times, the system comprehensively considers various factors to calculate the final confidence level. Only information exceeding a threshold is entered into the medical record.

[0054] On the other hand, the anomaly detection module monitors the consistency between newly added and existing information in real time. This module identifies logical contradictions through conflict detection, such as a patient claiming "no high blood pressure" but then mentioning "taking antihypertensive medication daily"—a clear conflict. When an anomaly is detected, the system generates corresponding clarification questions, which are then fed back to the patient through the intelligent guidance and interaction module.

[0055] The entire system operates with full support from a medical knowledge graph support system. As shown by the dotted lines in the figure, the medical knowledge graph provides domain knowledge to multiple modules, including the correlation between diseases and symptoms, drug interactions, and the standardization of medical terminology. This knowledge support enables the system not only to understand patient expressions but also to perform medical reasoning and terminology standardization.

[0056] After multiple rounds of dialogue and information processing, the results from the dynamic medical record builder and the abnormal information detection module converge to generate the final structured medical record. This medical record includes complete standard sections such as chief complaint, present illness, and past medical history. All information has undergone semantic understanding, logical verification, and standardization processing, and can be directly used for subsequent clinical diagnosis and treatment.

[0057] This system ensures information integrity through a multi-turn dialogue mechanism and guarantees medical accuracy through knowledge graph support. Ultimately, it achieves intelligent conversion from patient natural language input to standardized medical record output, significantly improving the efficiency and quality of medical information collection.

[0058] As one specific implementation method, the specific implementation method

[0059] This implementation was deployed in a top-tier hospital, employing a front-end and back-end separated system architecture. The front-end includes a mobile app for patients and a web management system for doctors, while the back-end uses a microservices architecture deployed on the hospital's private cloud platform. The system runs in a cluster environment consisting of 8 application servers and 4 GPU servers, using MongoDB to store conversation data and medical records, and Redis to provide high-speed caching support.

[0060] The core of the intelligent guided interaction module is a medical entity recognizer. This recognizer is based on the Chinese pre-trained language model RoBERTa-wwm-ext and has been fine-tuned on a dictionary of 200,000 medical terms and 50,000 annotated datasets. The recognizer can accurately identify 15 categories of medical entities, including symptoms, diseases, medications, examination items, body parts, time expressions, and frequency descriptions. When a patient inputs "I've been coughing a lot lately, and sometimes my chest hurts," the system will recognize "cough" as a symptom entity (confidence 0.98), "chest" as a body part (confidence 0.95), "pain" as a symptom entity (confidence 0.96), and "recently" as a time entity (confidence 0.85).

[0061] After entity recognition is complete, the system updates this information in the dialogue state graph. The dialogue state graph is one of the core innovations of this invention; it models the entire dialogue process as a dynamically evolving knowledge graph structure. In this example, the system creates four new nodes, each corresponding to one of the four identified entities, and establishes a "chest-pain" relationship edge, indicating that the pain occurs in the chest area. Each node stores information such as the entity's type, text content, confidence level, and timestamp, while maintaining a 768-dimensional feature vector obtained through a BERT encoder, which fully expresses the semantic information of the entity.

[0062] The graph update process is not a simple node addition, but involves complex information fusion and reasoning mechanisms. When a patient mentions "I've had a cough for over a week" in a subsequent conversation, the system recognizes this as supplementing the attribute of an existing "cough" node, rather than creating a new node. The system uses entity disambiguation technology to determine that the two mentions refer to the same symptom, and then updates the duration attribute of that node. If the same information comes from multiple sources and there are discrepancies, the system will activate a multi-source confidence fusion mechanism.

[0063] The multi-source confidence fusion mechanism does not simply select the highest confidence value, but comprehensively considers multiple factors. First, the system performs a log-odds transformation on the original confidence of each information source, mapping it from the probability space to the real number space for easier subsequent calculations. Then, weights are dynamically assigned based on the historical accuracy of the information sources, with sources with better historical performance receiving higher weights. Most importantly, a context consistency bias term is introduced, which evaluates the semantic consistency between the new information and the existing context by calculating the cosine similarity between the vector representation of the new information and the average vector of its neighboring nodes in the graph. Finally, the fusion result is mapped back to the probability space using the sigmoid function to obtain the final fusion confidence.

[0064] The update of the dialogue state graph also includes reasoning about implicit relationships. The system has a rich built-in medical knowledge base. When it identifies the two symptom nodes "fever" and "cough," it automatically infers the possible disease node of "upper respiratory tract infection" and establishes corresponding association edges. This reasoning is not a hard rule, but a soft reasoning based on probability, and the reasoning results are labeled with corresponding confidence levels.

[0065] The system employs a three-layer graph convolutional network, with each layer aggregating neighbor information for each node, enabling every node to perceive information from multi-hop neighbors. This global information propagation mechanism ensures that the system fully considers all collected information when generating follow-up questions. For example, when the system knows that a patient has a history of hypertension, it will pay special attention to antihypertensive drugs when inquiring about medication. This correlation is naturally achieved through the information propagation of the graph neural network.

[0066] The consultation process is modeled as a Markov decision process. The state space includes the encoding of the current dialogue state diagram, the list of filled fields, the dialogue turn, and other information; the action space corresponds to all queryable medical record fields; the reward function comprehensively considers information completeness improvement, collection efficiency, and medical importance. In particular, the system provides higher rewards for the collection of key clinical information (such as chief complaint and present medical history), and introduces a time decay factor to encourage rapid information collection.

[0067] Calculate the information entropy of each unfilled field to measure its degree of uncertainty. Also consider the dependencies between fields; if querying a field helps infer the values ​​of several other fields, that field has a higher information gain. Dependencies are quantified using distances in a graph structure; fields that are closer together have a greater influence on each other.

[0068] The anomaly detection module employs a combination of rule-based and learning-based approaches. The system maintains a knowledge base containing over 1200 medical logic rules, covering common medical contradictions such as "male patient" versus "pregnancy," and "penicillin allergy" versus "use of amoxicillin." For cases not covered by the rules, the system uses a deep learning model for semantic conflict detection. This model takes as input the vector representations of two entities, their element-wise product, and the difference vector, and uses a multi-layer neural network to determine whether semantic conflicts exist.

[0069] When a conflict is detected, the system does not directly point out the patient's error, but instead generates tactful clarifying questions. For example, when a patient claims "I don't have diabetes" but mentions "I inject insulin every day," the system will ask, "You mentioned using insulin; what is the reason for using it?" This design fully considers the patient's feelings and avoids adversarial interactions.

[0070] The dynamic construction of medical records is an ongoing process. The system maintains a standard medical record template, which includes multiple parts such as chief complaint, present illness, past medical history, personal history, family history, physical examination, and auxiliary examinations. Whenever the dialogue state graph is updated, the system traverses all nodes in the graph, extracts relevant information, and maps it to the corresponding fields in the medical record template. This mapping is not a simple rule matching, but is achieved through a trained mapping model that has learned a large number of patterns in doctors' medical record entry.

[0071] The system has demonstrated significant effectiveness after three months of deployment in the respiratory department of a top-tier hospital. The average consultation time has decreased from 12 minutes to 4.5 minutes, medical record completeness has increased from 75% to 93%, and the rate of missing key information has decreased from 8.2% to 1.3%. The average response time for a single-turn conversation is 320 milliseconds, the system supports over 1000 concurrent sessions, and availability reaches 99.95%.

[0072] In practical applications, the system demonstrates excellent adaptability. By continuously collecting feedback from doctors, the system constantly optimizes its model parameters. Whenever a doctor corrects a medical record generated by the system, these corrections are used as training samples for incremental learning of the model. The system features a specially designed reward function for medical semantic awareness, which provides higher rewards for actions that quickly and accurately collect key clinical information, making the system's consultation strategy increasingly similar to that of experienced doctors.

[0073] To ensure system robustness, a comprehensive exception handling mechanism has been implemented. When entity recognition fails, the system will activate a template-based backup question-answering strategy; when an error occurs during graph update, the system will record the exception but will not interrupt the dialogue process; for any unexpected errors, the system has corresponding degradation strategies to ensure that a reasonable response is given and the continuity of the dialogue is maintained.

[0074] This embodiment implements an intelligent medical information acquisition system. The system can not only accurately understand patients' natural language expressions, but also conduct targeted follow-up questions like an experienced doctor, ultimately generating high-quality structured medical records to provide strong support for subsequent diagnostic and treatment decisions.

[0075] The above-described specific embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, and improvements made by those skilled in the art to the technical solutions of the present invention based on the provided textual description and drawings, without departing from the design concept and spirit of the present invention, should all fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. A patient information collection and medical record construction system based on multi-turn dialogue, characterized in that, include: The intelligent guidance and interaction module receives the patient's natural language input, extracts medical entities and semantic features from the current input, and, based on a pre-trained medical domain large language model, calculates and generates follow-up questions most relevant to the current condition through an attention mechanism, according to the completeness score of the currently collected information and the priority weight of the fields to be filled. The context-aware question-answering engine adopts a graph-based dialogue state representation method, which constructs the entities, relationships and attributes of each round of dialogue into a knowledge subgraph. It dynamically maintains the dialogue history information through an incremental update mechanism of dialogue state. Newly identified entities are added to the graph as nodes, and medical relationships between entities are used as edges to connect nodes. Global information propagation is achieved through graph neural networks to ensure that relevant information in the historical dialogue can be referenced when generating subsequent questions and answers. The medical record information dynamic builder monitors the update events of the dialogue state graph in real time. When a new medical entity node is detected, the final confidence score is calculated through a multi-source confidence fusion mechanism. The multi-source confidence fusion mechanism uses log-odds transformation to fuse the confidence scores of multiple information sources and introduces a context consistency bias term based on cosine similarity to ensure the semantic consistency between new information and existing information. The information is only filled into the medical record template when the fused confidence score exceeds a preset threshold. The abnormal information detection module adopts a hybrid conflict detection method that combines rules and learning. First, it detects hard contradictions through a medical logic constraint rule base. For entities of the same type, it judges semantic-level conflicts through a deep learning model, thereby achieving multi-level information consistency assurance. The medical knowledge graph support system uses a multi-level semantic matching strategy to map patients' colloquial expressions to a standard medical terminology system in real time.

2. The patient information collection and medical record construction system based on multi-turn dialogue according to claim 1, characterized in that, The intelligent guided interaction module generates follow-up questioning strategies based on the principle of maximizing information gain. By calculating the information entropy of each unfilled field and the conditional entropy between the unfilled field and the filled field, it selects fields that can reduce overall uncertainty for follow-up questioning. The degree of influence between related fields is weighted by the inverse of the graph distance.

3. The patient information collection and medical record construction system based on multi-turn dialogue according to claim 1, characterized in that, The context-aware question-answering engine's dialogue state graph update includes: When a node is created, its feature vector is initialized based on the entity type and attributes. Automatic reasoning and establishment of implicit relationships between entities are based on a medical knowledge base; The updating of node features and the propagation of information are achieved through multi-layer graph convolutional networks; When old and new information conflict, the system intelligently resolves the conflict by comprehensively considering timestamps, information source reliability, and confidence level.

4. The patient information collection and medical record construction system based on multi-turn dialogue according to claim 1, characterized in that, The medical record information dynamic builder performs incremental information population, including: The medical information extracted from the conversation is mapped to standard fields in the medical record template using a machine learning model. Set up dependencies and triggering rules between fields to enable cascading updates of related fields; It records detailed version information for each information update operation, supporting complete tracing of medical record modification history.

5. The patient information collection and medical record construction system based on multi-turn dialogue according to claim 1, characterized in that, It also includes an active learning module, which designs a reward mechanism for medical semantic perception. This module provides higher rewards for actions that collect key clinical information. The reward value decreases exponentially with the duration of the dialogue, thereby guiding the system to prioritize and quickly collect the information most important for diagnosis.

6. A method for patient information collection and medical record construction based on multi-turn dialogue, applicable to the patient information collection and medical record construction system based on multi-turn dialogue as described in any one of claims 1-5, characterized in that, Includes the following steps: Step S1: Load the pre-trained language model for the medical field, standard medical record template, medical knowledge graph and logical constraint rule base, create an empty dialogue state graph, and initialize the medical record completeness threshold and the clinical priority of each field; Step S2: Generate a personalized greeting based on the current time, department, and other information to guide the patient in describing their main symptoms; Step S3: After receiving patient input, preprocessing is performed first, including speech recognition, word segmentation and time normalization; Then, a medical entity recognition model is used to extract entities such as symptoms, diseases, and drugs, as well as their attributes; Finally, the medical relationships between entities are identified, forming the structured information for this round of dialogue; Step S4: Update the newly identified entities and relationships incrementally to the dialogue state graph, update the attributes of existing entities, and create corresponding nodes for new entities; By using graph neural networks to propagate information across the entire graph, each node can perceive the relevant context. Step S5: Dynamic medical record construction iterates through the updated dialogue state graph, extracts medical information, and maps it to medical record template fields; When a field has multiple candidate values, the final confidence score is calculated using a multi-source confidence fusion mechanism. in: Conf fused The final confidence level after fusion, with a value range of [0,1]; σ is the sigmoid activation function, used to normalize the output to the [0,1] interval; n represents the number of information sources participating in the integration; w i Let the weight of the i-th information source satisfy the following condition: Adjust dynamically based on the historical accuracy of the information source; conf i The original confidence level provided for the i-th information source, with a value range of [0,1]; For log-odds transformation, the confidence level is mapped to the real number domain; b context The context consistency bias term is obtained by calculating the cosine similarity between the new information vector and the average vector of its graph neighborhood nodes, and its value ranges from [-1, 1]. Step S6: Perform multi-level consistency checks on the newly populated information, including rule-based medical logic checks, temporal rationality verification, and deep learning-based semantic conflict identification, and generate a conflict list sorted by severity. Step S7: Determine the next action based on the current completeness and conflict status of the medical record: If the information is complete and there are no conflicts, end the data collection; if there are conflicts, prioritize generating clarification questions; otherwise, continue with routine data collection. Step S8: For the most serious conflicts, combine the conflict type and contextual information to generate natural and tactful clarifying questions using a language model, avoiding making the patient feel questioned; Step S9: Select the next query target based on the principle of maximizing information gain, and generate a natural language question containing appropriate context by comprehensively considering the uncertainty of the field, its dependence on other fields, and its clinical importance. Step S10: After the data collection is completed, a global consistency check and medical reasoning supplement are performed on the entire dialogue information. The structured medical record is output in a standard format, and a quality assessment report containing multiple dimensions such as completeness and accuracy is generated.

7. The method for patient information collection and medical record construction based on multi-turn dialogue according to claim 6, characterized in that, Multi-round dialogue is achieved by repeatedly executing steps S3 to S9. Each round of dialogue is incrementally updated based on the previous information until the medical record information reaches the preset completeness and has no logical conflicts.

8. The method for patient information collection and medical record construction based on multi-turn dialogue according to claim 6, characterized in that, It supports multimodal input including voice, text, and medical images, and converts multimodal information into a structured representation through corresponding recognition and understanding modules.

9. The method for patient information collection and medical record construction based on multi-turn dialogue according to claim 6, characterized in that, It includes an online learning mechanism that collects doctors' feedback on the system's output and continuously optimizes the model's performance using a medical semantic awareness reward function, with higher rewards given to behaviors that quickly and accurately collect key clinical information.

Images