Triage classification training system and method based on large language model and knowledge graph

Abstract

The invention specifically discloses a triage training system and method based on a large language model and a knowledge graph, and relates to the technical field of medical training. The system comprises a multi-modal data sensing and processing module which is used for accessing and processing diversified training data sources; the dynamic time sequence knowledge graph construction and updating module comprises an entity relationship extraction unit, a multi-modal alignment unit and a knowledge fusion and conflict resolution unit; the multi-agent collaborative service module is composed of large language model agents which are special in function and work cooperatively, and comprises an extraction agent, a construction agent, a reasoning agent and a teaching agent; and the immersive interactive training and feedback module comprises a virtual case generation unit, a multi-modal interactive interface and a deep feedback system. According to the invention, the authenticity, accuracy and effectiveness of training are significantly improved, it is ensured that the learned knowledge and skills can efficiently adapt to a real emergency treatment scene, and the accuracy and efficiency of triage classification in emergency treatment are improved.

Description

Technical Field

[0001] This invention relates to the field of medical training technology, and in particular to a triage training system and method based on a large language model and knowledge graph. Background Technology

[0002] Triage, a core component of emergency care, directly determines patient priority allocation and the efficiency of medical resource utilization, playing a crucial role in saving lives and reducing mortality. Currently, triage training for medical personnel primarily relies on traditional classroom lectures, standardized case (SP) drills, and computer-based static case learning systems. While these methods play a role in conveying basic theory, they are no longer adequate for the evolving needs of modern emergency medicine, exhibiting several significant shortcomings that hinder the connection between training effectiveness and clinical application.

[0003] The core problems of existing training models lie in three dimensions: the timeliness of knowledge, the authenticity of scenarios, and the effectiveness of feedback. On the one hand, the rapid pace of updates to medical guidelines and clinical evidence, coupled with the rigidity of traditional training content, often lags behind cutting-edge medical practice, leading to a disconnect between the knowledge learned by medical staff and the actual needs of emergency scenarios. On the other hand, traditional cases are mostly static texts or simple simulations, lacking the support of dynamic disease evolution logic and multimodal information (such as real-time physiological signals and medical images), failing to recreate the complex and rapidly changing real-world scenarios in the emergency environment, making it difficult for trainees to develop clinical thinking to deal with complex situations. At the same time, the feedback mechanisms of existing systems are mostly limited to binary judgments of "correct / incorrect," failing to deeply analyze the trainees' decision-making process, making it difficult to identify and correct cognitive biases such as anchoring effect and confirmation bias, and failing to fundamentally improve trainees' clinical decision-making abilities.

[0004] In recent years, the application of knowledge graphs and Large Language Models (LLM) technologies in the medical field has provided new directions for optimizing training systems. However, existing technical solutions are mostly simple superpositions of the two, failing to achieve deep synergy and complementarity. While LLM possesses powerful natural language interaction and text understanding capabilities, its inherent "illusion" problem (generating false or inaccurate information) poses serious security risks in high-risk medical training scenarios. On the other hand, while traditional static knowledge graphs can achieve structured knowledge storage, they lack dynamic updating capabilities and flexibility, making them unable to adapt to the dynamic knowledge needs of emergency medicine. This fragmented application at the technical level makes it difficult for existing solutions to break through the limitations of traditional training and meet the demands of triage training for high realism, high accuracy, and personalization. Innovative technical solutions are urgently needed to address these pain points. Summary of the Invention

[0005] The purpose of this invention is to propose a triage training system and method based on a large language model and knowledge graph. This aims to overcome the shortcomings of existing triage training technologies, addressing issues such as lagging knowledge updates, limited and monotonous contexts, superficial and ineffective feedback in traditional training, and the inability to achieve deep synergy through simple combinations of existing large language models and knowledge graph technologies. The invention significantly improves the authenticity, accuracy, and effectiveness of training, ensuring that the learned knowledge and skills can be efficiently adapted to real emergency scenarios, thereby enhancing the accuracy and efficiency of triage in emergency care.

[0006] To achieve the above objectives, this invention proposes a triage training system based on a large language model and knowledge graph, comprising: a multimodal data perception and processing module, a dynamic temporal knowledge graph construction and updating module, a multi-agent collaborative service module, and an immersive interactive training and feedback module; The multimodal data perception and processing module is used to access and process diverse training data sources, including text-based medical records, medical images, time-series physiological signals, operational behavior logs, user interaction data, and sensor data. The dynamic temporal knowledge graph construction and update module is the core knowledge engine of the system, including entity relationship extraction unit, multimodal alignment unit and knowledge fusion and conflict resolution unit; The multi-agent collaborative service module consists of large language model (LLM) agents with specialized functions that work collaboratively, including: extraction agent, construction agent, reasoning agent, and teaching agent; The immersive interactive training and feedback module includes a virtual case generation unit, a multimodal interactive interface, and a deep feedback system.

[0007] Preferably, the multimodal data sensing and processing module is equipped with a heterogeneous data adapter to convert data from different sources into a unified format that conforms to the Fast Medical Interconnection Resource (FHIR) standard.

[0008] Preferably, the entity relation extraction unit uses a large language model fine-tuned from the medical corpus as the core extractor. Through structured prompt word templates and fuzzy JSON parsing technology, it accurately extracts medical entities and relations from unstructured text to form triples.

[0009] Preferably, the multimodal alignment unit is used to extract image features, process physiological signals using a temporal convolutional network, and perform cross-modal association and embedding with text entities.

[0010] Preferably, the knowledge graph has temporal characteristics, with each triple in the knowledge graph accompanied by a timestamp and confidence information, and knowledge fusion and conflict resolution are performed through an algorithm based on confidence weighting and time decay function.

[0011] Preferably, the extraction agent is responsible for parsing the raw data and calling the entity relationship extraction unit; the construction agent is responsible for updating the extracted knowledge to the knowledge graph and performing reasoning to enrich the graph relationships; the reasoning agent performs clinical reasoning based on the current knowledge graph and specific case context, generating triage decision suggestions and their explanations; the teaching agent is responsible for analyzing trainees' behavioral trajectories, assessing trainees' cognitive status, and generating personalized feedback and training content; and the agents communicate and collaborate through a message bus and a shared working memory area.

[0012] Preferably, the virtual case generation unit dynamically generates virtual cases whose symptoms, signs, and disease progression logic all conform to medical principles, based on conditional generative adversarial networks (CGAN) and knowledge graphs.

[0013] Preferably, the multimodal interaction interface allows trainees to interact with virtual patients, family members, and medical team members through natural language, and to conduct hands-on simulations through AR / VR devices.

[0014] Preferably, the deep feedback system records and analyzes the trainees' decision-making trajectory, compares it with a pre-set cognitive bias pattern library, and generates a multi-dimensional assessment and personalized correction training plan for the quality of the decision-making process.

[0015] This invention also provides a training method for triage based on a large language model and knowledge graph, the specific steps of which are as follows: Step S1: Access and standardization processing of multimodal training data; Step S2: Dynamic case context generation and evolution based on the collaboration of large language model LLM and knowledge graph KG; Step S3: Collect and analyze trainees' interactive behaviors and decision-making trajectories; Step S4: Real-time evaluation and feedback generation based on multi-agent collaborative reasoning; Step S5: Provide adaptive reinforcement training for cognitive weaknesses.

[0016] Therefore, this invention proposes a triage training system and method based on a large language model and knowledge graph, the beneficial effects of which are as follows: (1) This invention breaks through the limitations of the simple superposition of LLM and KG in the prior art and constructs an innovative architecture of “LLM multi-agent cluster responsible for perception, computation and interaction, and dynamic KG responsible for structured memory, verification and reasoning”, which effectively avoids the hidden danger of inaccurate information in high-risk medical training scenarios and improves the overall reliability and clinical adaptability of the system.

[0017] (2) This invention relies on multimodal data fusion technology and dynamic temporal knowledge graph to accurately simulate the core features of the real emergency environment. It can not only integrate multi-dimensional information such as text cases, medical images, and temporal physiological signals, but also realize the effect of dynamic evolution of the condition with the intervention measures of trainees. It breaks the limitations of traditional static cases, helps trainees build triage thinking that fits clinical practice, and shortens the gap between training and practice.

[0018] (3) This invention breaks through the single mode of traditional training that only evaluates the results of operations. It records the trainees’ decision-making trajectory in real time through a deep feedback system and accurately identifies thinking defects such as anchoring effect and confirmation bias by combining a cognitive bias pattern library. It not only points out whether the results are right or wrong, but also breaks down the problems in the decision-making process. It makes up for the trainees’ knowledge gaps from the root of cognition, helps trainees form scientific and efficient triage decision-making ability, and improves the accuracy and long-term effectiveness of training.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall structure of a triage training system based on a large language model and knowledge graph according to the present invention. Figure 2 This is a schematic diagram of the dynamic temporal knowledge graph construction and updating process in this invention; Figure 3 This is a flowchart of the multi-agent collaborative workflow in this invention; Figure 4 This is a flowchart of the immersive interactive training and feedback module in this invention. Detailed Implementation

[0021] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0022] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0023] Example 1 like Figure 1As shown, the present invention provides a triage training system based on a large language model and knowledge graph, including: a multimodal data perception and processing module, a dynamic temporal knowledge graph construction and updating module, a multi-agent collaborative service module, and an immersive interactive training and feedback module; The multimodal data perception and processing module is used to access and process diverse training data sources, including text-based medical records, medical images, time-series physiological signals, operational behavior logs, user interaction data, and sensor data. At the same time, the multimodal data perception and processing module has a built-in heterogeneous data adapter, which can convert data from different sources into a unified format that conforms to the Fast Medical Interconnection Resource (FHIR) standard.

[0024] like Figure 2 As shown, the dynamic temporal knowledge graph construction and update module is the core knowledge engine of the system, including an entity relationship extraction unit, a multimodal alignment unit, and a knowledge fusion and conflict resolution unit; The entity relation extraction unit uses a large language model fine-tuned from the medical corpus as the core extractor. Through structured prompt word templates and fuzzy JSON parsing technology, it accurately extracts medical entities and relations from unstructured text and forms triples.

[0025] The multimodal alignment unit uses models such as Vision Transformer to extract image features, uses temporal convolutional networks to process physiological signals, and performs cross-modal association and embedding with text entities.

[0026] The knowledge fusion and conflict resolution unit employs an algorithm based on confidence weighting and a time decay function to fuse knowledge extracted from different sources, automatically resolving knowledge conflicts and ensuring the authority and timeliness of the knowledge graph. Simultaneously, the knowledge graph possesses temporal characteristics; each triple is accompanied by a timestamp and confidence level, enabling it to record and reflect the evolution of medical knowledge over time.

[0027] like Figure 3 As shown, the multi-agent collaborative service module consists of large language model (LLM) agents with specialized functions that work collaboratively, including: Extraction Agent: Responsible for parsing the raw data and calling the entity relationship extraction unit; Constructing intelligent agents: responsible for updating the extracted knowledge to the knowledge graph and performing reasoning to enrich the graph relationships; Reasoning agent: Based on the current knowledge graph and specific case context, it performs clinical reasoning to generate triage decision suggestions and their explanations; The teaching agent is responsible for analyzing learners' behavioral patterns, assessing their cognitive state, and generating personalized feedback and training content. Each agent communicates and collaborates through a message bus and a shared working memory area, forming an efficient decision-making loop.

[0028] like Figure 4 As shown, the immersive interactive training and feedback module utilizes the support provided by other modules to offer training services to learners, including: Virtual case generation unit: Based on conditional generative adversarial network (CGAN) and knowledge graph, it dynamically generates virtual cases whose symptoms, signs, and disease progression logic all conform to medical laws.

[0029] Multimodal interaction interface: It supports trainees to interact with virtual patients, family members and medical team members through natural language, and to conduct practical simulations through AR / VR devices.

[0030] Deep Feedback System: Records trainees' decision-making trajectory in real time (such as operation sequence, response time, and information retrieval order), compares and analyzes it through a cognitive bias pattern library, and generates multi-dimensional evaluation reports and targeted corrective training programs that not only focus on the correctness of the results but also on the quality of the decision-making process.

[0031] Example 2 This invention also provides a training method for triage based on a large language model and knowledge graph, the specific steps of which are as follows: Step S1: Access and standardization processing of multimodal training data; Step S2: Dynamic case context generation and evolution based on the collaboration of large language model LLM and knowledge graph KG; Step S3: Collect and analyze trainees' interactive behaviors and decision-making trajectories; Step S4: Real-time evaluation and feedback generation based on multi-agent collaborative reasoning; Step S5: Provide adaptive reinforcement training for cognitive weaknesses.

[0032] The following specific implementation examples will further illustrate this application.

[0033] The system is deployed on a cloud server. Students log in to the system through AR devices. The system's teaching agent first retrieves the student's historical performance data, analyzes their knowledge mastery in depth, and accurately identifies knowledge gaps (such as "identification of atypical myocardial infarction in the elderly"). Then, it instructs the case generation unit to create an elderly acute heart failure case that meets the training requirements of that gap.

[0034] After the trainee receives the patient, the LLM-driven virtual patient complains of "fatigue and loss of appetite," while also exhibiting slightly shortness of breath. Vital sign data is generated in real-time through a data perception module, ensuring the data's authenticity and dynamism. The trainee can interact with the virtual patient using natural language, for example, by asking, "Do you have chest pain?" The virtual patient will answer based on the case settings and medical logic, "No, it doesn't hurt, but I can't breathe properly and I feel weak." At the same time, the trainee can view the system-simulated electrocardiogram waveform (showing sinus tachycardia) and virtual chest X-ray (showing signs of pulmonary edema) in the AR view, obtaining multimodal case information.

[0035] All the patient's consultations, examinations, and procedures during the training process are recorded in real time by the system, forming a complete decision-making trajectory. If the patient rules out heart problems because the virtual patient has "no chest pain" and considers other diagnostic directions, the reasoning agent will immediately determine, based on medical knowledge and clinical guidelines in the knowledge graph, that the decision-making path has an "anchoring effect" (over-reliance on typical symptoms).

[0036] After the training, the teaching AI quickly generated a detailed feedback report. The report not only clearly indicated the correct triage level but also focused on analyzing the trainees' cognitive biases. For example, it stated, "You may have underestimated the risk of the condition because the patient lacked typical chest pain symptoms. Please note that elderly heart failure patients often present with atypical symptoms such as fatigue and loss of appetite." Furthermore, it immediately provided trainees with a series of case studies specifically targeting "atypical cardiovascular emergencies in the elderly" to facilitate intensive training and address any knowledge gaps.

[0037] Throughout the training process, the extracting agent, constructing agent, and reasoning agent maintain a highly efficient collaborative working state on the message bus, continuously performing knowledge retrieval, case simulation, and reasoning decision-making to ensure the real-time performance and accuracy of all operations, providing trainees with a high-quality training experience.

[0038] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0039] Therefore, this invention provides a triage training system and method based on large language models and knowledge graphs. By deeply collaborating with LLM and dynamic KG, it avoids the "illusion" problem of LLM and the rigidity of KG. At the same time, by integrating multimodal information and realizing the dynamic evolution of the condition, it improves the realism of the training scenario. It can also analyze the decision-making trajectory of trainees, identify cognitive biases, generate personalized feedback and reinforcement training, and ultimately significantly improve the reliability, adaptability and accuracy of triage training.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A triage training system based on a large language model and knowledge graph, characterized in that, include: The module includes a multimodal data perception and processing module, a dynamic temporal knowledge graph construction and update module, a multi-agent collaborative service module, and an immersive interactive training and feedback module. The multimodal data perception and processing module is used to access and process diverse training data sources, including text-based medical records, medical images, time-series physiological signals, operational behavior logs, user interaction data, and sensor data. The dynamic temporal knowledge graph construction and update module is the core knowledge engine of the system, including entity relationship extraction unit, multimodal alignment unit and knowledge fusion and conflict resolution unit; The multi-agent collaborative service module consists of large language model (LLM) agents with specialized functions that work collaboratively, including: extraction agent, construction agent, reasoning agent, and teaching agent; The immersive interactive training and feedback module includes a virtual case generation unit, a multimodal interactive interface, and a deep feedback system.

2. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, The multimodal data sensing and processing module is equipped with a heterogeneous data adapter to convert data from different sources into a unified format of the Fast Medical Interconnection Resource (FHIR) standard.

3. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, The entity relation extraction unit uses a large language model fine-tuned from the medical corpus as the core extractor. Through structured prompt word templates and fuzzy JSON parsing technology, it accurately extracts medical entities and relations from unstructured text and forms triples.

4. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, The multimodal alignment unit is used to extract image features, process physiological signals using a temporal convolutional network, and perform cross-modal association and embedding with text entities.

5. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, Knowledge graphs have temporal characteristics. Each triple in a knowledge graph is accompanied by a timestamp and confidence information, and knowledge fusion and conflict resolution are carried out through algorithms based on confidence weighting and time decay functions.

6. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, The extraction agent is responsible for parsing the raw data and calling the entity relationship extraction unit; the construction agent is responsible for updating the extracted knowledge to the knowledge graph and performing reasoning to enrich the graph relationships; the reasoning agent performs clinical reasoning based on the current knowledge graph and specific case context, generating triage decision suggestions and their explanations; the teaching agent is responsible for analyzing trainees' behavioral trajectories, assessing trainees' cognitive status, and generating personalized feedback and training content; the agents communicate and collaborate through a message bus and a shared working memory area.

7. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, The virtual case generation unit is based on conditional generative adversarial network CGAN and knowledge graph to dynamically generate virtual cases whose symptoms, signs and disease progression logic all conform to medical laws.

8. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, The multimodal interaction interface allows trainees to interact with virtual patients, family members, and medical team members through natural language, and to conduct hands-on simulations through AR / VR devices.

9. The triage training system based on a large language model and knowledge graph according to claim 1, characterized in that, The deep feedback system records and analyzes trainees' decision-making trajectories, compares them with a pre-set cognitive bias pattern library, and generates multi-dimensional assessments and personalized corrective training programs for the quality of the decision-making process.

10. A training method for triage based on a large language model and knowledge graph, characterized in that, The specific steps are as follows: Step S1: Access and standardization processing of multimodal training data; Step S2: Dynamic case context generation and evolution based on the collaboration of large language model LLM and knowledge graph KG; Step S3: Collect and analyze trainees' interactive behaviors and decision-making trajectories; Step S4: Real-time evaluation and feedback generation based on multi-agent collaborative reasoning; Step S5: Provide adaptive reinforcement training for cognitive weaknesses.

Images