Emergency auxiliary decision-making system based on large model agent and multiple knowledge bases

Abstract

The invention discloses an emergency auxiliary decision-making system based on a large model agent and multiple knowledge bases, and the system comprises an input end which is used for obtaining patient information from multiple dimensions; the central processing module is used for intelligent agents of different scenes to construct based on different large models and cue word projects, and calling at an output server of the system through an embedded structure; the knowledge base comprises a knowledge graph base and a retrieval enhancement generation document base, a knowledge graph stock structured rule ensures reliable decision, and a retrieval enhancement generation document stock unstructured text realizes dynamic knowledge fusion; and the service node stores knowledge according to a fault-tolerant demand menu library or double libraries and outputs the knowledge to the server side. The server side comprises an intelligent triage auxiliary module and an intelligent diagnosis auxiliary module, and provides assistance in a mode of calling an intelligent agent service. The method has the advantages that the triage efficiency is improved, the triage decision-making time of nurses is shortened, the medical history combing and examination scheme making time of doctors is shortened, the reliability of medical decision-making is ensured, and the maintenance cost is lower.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, and in particular to an emergency decision-making support system based on a large model intelligent agent and multiple knowledge bases. Background Art

[0002] With the rapid development of medical information technology, emergency decision-making support systems are crucial for improving the efficiency and quality of emergency treatment. Differences in experience among different personnel in triage decisions lead to large variances, and doctors' medical history review and examination plan development are time-consuming, which in turn affects the overall efficiency and reliability of emergency treatment. However, most existing decision-making support systems primarily use natural language processing and text feature extraction technologies to characterize patient conditions, and provide decision support through rule-based reasoning under knowledge graph constraints and probabilistic prediction based on historical data. On the one hand, these systems have relatively limited functional coverage and often focus on a single specific link in the emergency service, such as pre-examination and triage. They do not provide customized support for multiple links in the emergency service chain, which easily leads to information silos within the service chain. In the actual emergency process, from triage to diagnosis and treatment, each link is closely connected, and information needs to flow smoothly, but the existing system cannot meet this demand; on the other hand, it is difficult to update the system based on the knowledge graph library: because the maintenance of the knowledge graph relies on the annotation of specialists, it requires a high level of professionalism from the maintenance personnel, increases maintenance costs, reduces the flexibility of the system, and makes it difficult to quickly update the knowledge base; in addition, the existing natural language processing and knowledge graph path mapping modules are insufficient in processing non-standard complaints. For colloquial descriptions or local descriptions, the failure rate of standardized mapping is high. In emergency scenarios, the expressions of patients or emergency personnel are not standardized. If the system cannot accurately understand this information, it may lead to triage errors and reduce the quality of medical care.

[0003] Therefore, technicians in this field are committed to providing a new emergency decision-making support system that can overcome the above-mentioned defects, comprehensively cover the emergency process, achieve effective coordination of multiple links, and at the same time have efficient knowledge base updating capabilities and flexible deployment for different business links to improve the level of emergency treatment and protect the life and health of patients. Summary of the Invention

[0004] In view of the above-mentioned shortcomings of the existing technology, the technical problem to be solved by the present invention is to provide an emergency decision-making support system that can effectively cover the entire emergency process, realize information collaboration in multiple links, and have a flexible and easy-to-use knowledge base mechanism. To this end, the present invention provides an emergency decision-making support system based on a large model intelligent agent cluster and multiple knowledge bases, including:

[0005] Input end, used to obtain patient information from multiple dimensions;

[0006] The central processing module is used to build intelligent agents for different scenarios based on different large models and prompt word projects. The system's output server calls the intelligent agents to complete tasks.

[0007] The knowledge base includes a knowledge graph library and a search-enhanced document library. The structured rules of the knowledge graph library ensure reliable decision-making, and the unstructured text of the search-enhanced document library realizes dynamic knowledge fusion. Business nodes select single library or dual library knowledge as the basis for decision-making based on fault tolerance requirements.

[0008] The output service end includes an intelligent triage assistance module and an intelligent diagnosis assistance module, which are embedded in the triage nurse station and the emergency doctor station in the form of buttons and call-out windows respectively, providing intelligent assistance by calling intelligent body services.

[0009] Furthermore, the input terminal obtains patient information from three channels: 1) a real-time voice input module, which converts oral complaints into text, supporting dialect recognition and ambiguous expression correction; 2) electronic medical records, which connect to the electronic medical record system through patient information; and 3) a vital sign monitoring module, which connects to medical devices and continuously obtains vital sign information. This information exists in different forms and requires a dedicated agent in the central processing module to analyze, integrate, generate a unified report, and form a temporary archive, which is stored in the knowledge base for subsequent system calls and real-time updates.

[0010] Furthermore, the two types of knowledge bases are the knowledge graph base and the enhanced retrieval generation knowledge base: the knowledge graph base contains a structured graph of disease symptom associations and diagnosis and treatment standard knowledge; the enhanced retrieval generation knowledge base is a knowledge base of dynamically updated clinical guidelines, case literature and unstructured knowledge of the hospital's medical resource information.

[0011] Furthermore, the central processing module is composed of multiple intelligent agents optimized for tasks, and the intelligent agents include prompt word engineering and large models with different characteristics, which are respectively responsible for specific links such as triage and classification, and diagnosis suggestions. When the generative large model in the intelligent agent generates an answer, it will search from the knowledge graph library and the RAG knowledge base according to the scenario. For the knowledge graph library, the relationship pairs with high similarity to the input text are retrieved, and the output of the large model is constrained by probability mask restrictions; for the relevant text set retrieved from the retrieval-enhanced generative knowledge base, the additional attention mechanism is used to expand the attention range of the large model during generation from the input text to include relevant fragments retrieved from the RAG knowledge base. The two mechanisms can jointly affect the generation of large model answers in the intelligent agent.

[0012] Furthermore, the triage assistance module combines the triage standards and patient temporary files contained in the knowledge base, and generates disease classification suggestions and department allocation recommendations by the intelligent agent based on the distribution of doctors' expertise in the hospital and the current reception load; after the triage is completed, the diagnosis assistance module automatically receives the synchronized files of the front-end link in this system, which contains the integrated multi-source information of the patient, and displays the key structured information in the front-end process. The diagnosis assistance module is integrated in the doctor's station, so that the doctor can call various auxiliary functions provided by different intelligent agents in the central processing module of this system when browsing information in the workstation; these functions can overcome the illusion problem of large models to a large extent due to the existence of the knowledge base, provide knowledge sources, and improve the reliability and scientificity of decision-making recommendations.

[0013] Furthermore, for the natural language input to the agent, it is converted into a vector q through the embedding model:

[0014] q=Embed LLM (”Text”)∈R d

[0015] Among them, “text” refers to the input natural sentence, R d Indicates the dimension of the conversion vector. The Embeding model can be selected according to needs. The BAAI / bge-zh series models are commonly used for Chinese scenarios.

[0016] Furthermore, when generating answers, the generative model in the agent will select knowledge from the knowledge graph library and the enhanced retrieval generated knowledge base based on the actual situation of the business link. For links with clear restrictions, it is necessary to search for entity relationships from the knowledge graph library. After the natural language is converted into a vector q, the similarity S can be calculated with each entity in the knowledge graph. i =q T E KG [i], find e that satisfies a similarity greater than the specified threshold τ i Entity E KG A collection of:

[0017] ε query ={e i |q T E KG [i]>τ}

[0018] When private domain knowledge is needed, it is necessary to search for knowledge related to the input from the RAG knowledge base. With the help of the Retriever model, the vector q can be retrieved from the RAG knowledge base to obtain the relevant text set D:

[0019] D=Retriever(q,RAGChunk)={(text1,text2...text i)}

[0020] Furthermore, the central processing module is composed of multiple intelligent agents and knowledge base frameworks optimized for tasks. The intelligent agents include prompt word engineering and large models with different characteristics, which are responsible for specific links in the business chain such as triage and diagnosis recommendations. When the large model generates an answer, it will search from the knowledge graph library and RAG knowledge base according to the scenario. For the knowledge graph library: find the answer that meets the similarity S i =q T E KG [i] e greater than the specified threshold τ i Entity Set ε query , and then find the required relationship according to the current scenario. For example, if we need to find the conflicting effects between different drugs in the medication analysis, we need to find the set ε query Are there conflicting pairs of entities in:

[0021]

[0022] Here C is the set of conflicting pairs, and r c Represents a conflicting relationship; this method is also applicable to finding other relationships between entities in the knowledge graph. The generation probability mask of the conflicting relationship for the large language model is:

[0023]

[0024] For the relevant text set D retrieved from the retrieval-enhanced generated knowledge base, it is encoded through the embedded model:

[0025] H RAG =Embed LLM (D)∈R m×d

[0026] Get the vectorized representation H of the relevant text fragments in the relevant knowledge base RAG After that, the attention vector is generated through the weight matrix W:

[0027] Q RAG =H RAG W Q ,K RAG =H RAG W K ,V RAG =H RAG W V

[0028] For a large language model based on the Transformer architecture, its attention mechanism is:

[0029]

[0030] Among them, d k is a scaling factor. The Softmax function normalizes the vector to ensure that each element in the vector has a value between 0 and 1, thereby reflecting the probability distribution of each content when the large language model generates the result:

[0031]

[0032] After considering the RAG retrieval content, the attention mechanism needs to additionally introduce attention to the vector of the retrieved knowledge base fragment:

[0033] H LLM+RAG =CrossAttention(H LLM ,H RAG )

[0034] Right now:

[0035]

[0036] Through cross-attention, the agent's large model's attention span is expanded from the input text to include relevant text fragments retrieved from the RAG knowledge base. When generating answers, the knowledge in the RAG knowledge base is fully considered as the basis for the large model's answer generation. The above two mechanisms enable the content of the RAG knowledge base and the KG knowledge base to influence the large model's answer generation in different ways. When performing certain specific tasks, the agent needs to retrieve relevant knowledge from the knowledge base according to the situation and influence the generation of the large model through the corresponding mechanism, namely, a joint generation probability influence:

[0037] P final (w t )=Normalize(P LLM (w t )·P RAG (w t )·Mask(w t ))

[0038] Among them, P final is the final generation probability; P LLM is the generation probability of the large model itself; Mask(w t ) is defined as only 0 and 1 in the conflict relationship, that is, the probability of generating conflicting content is 0. The relationship to be used to retrieve entities in the KG database needs to be selected according to the actual business process; RAG (w t )=Softmax(W RAG H LLM+RAG ) is used to enhance the generation of impact probability for retrieval, so that the final content generation probability fully considers the retrieval fragment content of the RAG knowledge base, It is a normalization operation similar to Softmax, which limits the final generation probability of each element to between 0 and 1.

[0039] It also includes a remote collaboration module and a system monitoring module: the remote collaboration module is used for real-time data sharing, enabling medical staff at all levels to exchange information in real time, share various patient information, and obtain expert diagnostic advice and treatment plans; the monitoring module is used to monitor the operating status of the system in real time, including the operating efficiency of the intelligent body, the update status of the knowledge base, and the stability of data transmission.

[0040] The beneficial effects of the present invention are:

[0041] 1) The triage assistance module is embedded in the nurse workstation. By combining the knowledge base with the patient's temporary file, triage standards, hospital resources and other information, it intelligently generates disease classification suggestions and department allocation recommendations. This reduces deviations caused by triage staff's lack of experience and unfamiliarity with hospital resources, improves triage efficiency, and shortens decision-making time.

[0042] 2) The diagnostic assistance module is integrated into the doctor's workstation, automatically receiving synchronized files from previous links, displaying key structured information, reducing the doctor's text retrieval time, and providing various auxiliary functions such as examination item recommendations, intelligent Q&A, and medication reminders to improve diagnostic efficiency.

[0043] 3) The knowledge graph database stores structured rules, which strictly constrain the decision-making of intelligent agents in rigid links such as the implementation of diagnosis and treatment standards, ensuring the reliability of medical decision-making recommendations.

[0044] 4) The retrieval-enhanced generated document library stores unstructured texts, such as clinical guidelines, case literature, and medical resource information of the hospital. It can quickly expand the knowledge base through agile updates and improve system flexibility.

[0045] 5) A hybrid knowledge management strategy reduces maintenance costs compared to a single knowledge graph solution. Agents at different business nodes can choose single or dual knowledge bases based on fault tolerance requirements, improving the flexibility and scientific nature of knowledge management.

[0046] 6) At the input end, a dedicated intelligent agent pre-processes and cleans the data, performs denoising, error correction, consistency checking, missing value filling, and outlier processing on various types of patient information. A data smoothing algorithm is used to handle abnormal fluctuations in vital sign data, improve data quality, and ensure the accuracy of subsequent intelligent agent analysis.

[0047] 7) Utilize the hospital resource information and patient temporary files in the dual knowledge base to synchronously obtain the patient's diagnosis results and disease severity in the business process, and reasonably generate suggestions based on the hospital's resource situation. For example, in the triage stage, appropriate department and expert recommendations are proposed based on the patient's condition and the characteristics of the hospital to optimize resource utilization.

[0048] 8) The remote collaboration module supports real-time data sharing. Medical staff can exchange information with experts in real time through the knowledge base, synchronize patient information and treatment plans, improve the efficiency of information exchange in the emergency chain, improve emergency efficiency, and avoid information congestion.

[0049] 9) The system monitoring module monitors the system operation status in real time, including the intelligent agent operation load, knowledge base update and data transmission stability. It can detect faults or anomalies in time by setting thresholds, and has automatic fault diagnosis and repair functions to regularly maintain and optimize the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic diagram of the system structure of the present invention;

[0051] Figure 2 It is a schematic diagram of the input module and the central processing module agent connected to it;

[0052] Figure 3 It is a structural diagram of the triage auxiliary module;

[0053] Figure 4 It is a structural diagram of the diagnosis auxiliary module. DETAILED DESCRIPTION

[0054] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0055] In the description of the present invention, it should be noted that the terms "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings and are intended solely to facilitate and simplify the description of the present invention. They are not intended to indicate or imply that the devices or components referred to must have, be constructed, or operate in a specific orientation, and therefore should not be construed as limitations on the present invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0056] In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "installed," "installed," and "connected" should be understood in a broad sense. For example, they may refer to fixed, detachable, or integral connections; mechanical or electrical connections; direct or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on the specific circumstances.

[0057] like Figures 1 to 4As shown in the figure, an emergency decision-making support system based on a large-scale intelligent agent and multiple knowledge bases is proposed. This system is a comprehensive support platform covering the entire emergency process. It uses the customized functions of multiple knowledge bases and large-scale intelligent agent groups to improve the efficiency of each emergency node, including:

[0058] The input end includes a voice conversion module, an electronic medical record docking module, and a vital sign monitoring data docking module, which are used to obtain patient information from multiple dimensions;

[0059] The central processing module (CPM) stores multiple pre-built large-scale intelligent agents for different emergency links. The intelligent agents for different scenarios are built based on different large-scale models and prompt word projects, and are called through embedded structures on the output server of the system;

[0060] The knowledge base comprises a dedicated Retrieval-Augmented Generation (RAG) document repository, storing unstructured text for dynamic knowledge fusion, and a Knowledge Graph (KG) repository, storing structured rules to ensure decision reliability. Business nodes can choose between a single or dual knowledge base based on fault tolerance requirements. The knowledge base is continuously updated, allowing flexible selection of knowledge import methods depending on the situation.

[0061] The output server includes an intelligent triage assistance module and an intelligent diagnosis assistance module, embedded as buttons and call-out windows in the triage nurse station and emergency doctor station, respectively, to provide assistance by invoking intelligent agent services. The modules are integrated into the existing HIS emergency doctor and nurse workstations, using buttons and call-out menus to operate without changing the doctor's existing workflow.

[0062] A performance evaluation index system can also be established. Triage accuracy is a key indicator, calculated by comparing the consistency between the system's triage results and the actual diagnosed conditions. The formula is: Triage accuracy = (number of accurately triaged cases / total number of triaged cases) × 100%. The effectiveness of diagnostic assistance is measured by assessing the proportion of doctors adopting the system's diagnostic recommendations and the improvement in patient treatment outcomes. System response time refers to the time from user input to system feedback, and is evaluated by taking the average value of multiple tests. Resource utilization is determined by comparing the actual allocation of medical resources in the hospital with the theoretical optimal allocation plan.

[0063] In terms of security assessment, the system analyzes potential data leakage risks, such as patient information theft through cyberattacks and data leaks due to illegal internal operations. To mitigate these risks, data encryption technology is used to encrypt the storage and transmission of sensitive patient information, such as medical records and test results. Strict access control is implemented, with system access permissions assigned based on the roles and responsibilities of medical staff to ensure only authorized personnel have access to the data. Network security protection equipment, such as firewalls and intrusion detection systems, is deployed to monitor and prevent cyberattacks in real time. Regular security vulnerability scanning and repair are performed to ensure the safe and stable operation of the system.

[0064] In summary, the purpose of this system is to significantly reduce the time it takes nurses to make triage decisions and reduce the variance caused by differences in personnel experience; to help doctors shorten the time it takes to sort out patient information, provide intelligent advice on diagnosis and treatment, and overall improve the efficiency and reliability of emergency treatment.

[0065] The core of the present invention lies in the customization and calling of intelligent bodies for different business links, and the method model of intelligent body text generation through dual knowledge bases in different ways.

[0066] The input terminal obtains patient information from three channels: 1) The real-time voice input module converts the patient's or first responder's verbal description into text, supporting dialect recognition and ambiguous expression correction; 2) It automatically connects to electronic medical records based on identity information, extracting the patient's historical medical records and key information such as allergy history; 3) The vital sign monitoring module directly connects to medical equipment to continuously obtain structured vital sign data such as blood pressure and blood oxygen. This information exists in different forms and requires the dedicated intelligent agent of the central processing module to analyze, integrate, generate a unified report, and form a temporary archive, which is stored in the knowledge base for subsequent system calls and real-time updates.

[0067] The central processing module is composed of multiple task-optimized intelligent agents, including specific prompt word engineering and large models with different characteristics, responsible for tasks such as triage and diagnosis recommendations. Regarding prompt word engineering, prompt word design for different business processes precisely guides model generation based on the current business process. For example, in the triage process, prompt words such as "Based on the patient's symptoms, vital signs, and medical history, and in accordance with our hospital's emergency triage standards, determine the patient's emergency level and recommend the triage department." In diagnostic scenarios, prompt words prompt the model to use the knowledge base framework for knowledge retrieval and combine private domain knowledge and rule constraints for diagnosis. For example, "Based on the patient's current symptoms, vital sign monitoring data, historical electronic medical records, and the latest clinical guidelines and case literature, a possible disease diagnosis is given, and the corresponding diagnostic evidence is listed." These precisely designed prompt words guide the large model to generate answers that better meet actual needs.

[0068] The system has two built-in knowledge bases: the knowledge graph library contains structured graphs of knowledge such as disease symptom associations and diagnosis and treatment standards; the RAG library is a knowledge base of unstructured knowledge such as dynamically updated clinical guidelines, case literature and medical resource information of the hospital; after the patient enters the emergency process, the system will generate a temporary exclusive file through the corresponding intelligent agent in the input module and the central processing module, dynamically record the patient's symptoms, signs and medical history and temporarily store them in the knowledge base to provide support for subsequent decision-making.

[0069] The input end converts the natural language of the input agent into a vector q through the embedding model:

[0070] q=Embed LLM (”Text”)∈R d

[0071] Among them, R d Indicates the dimension of the conversion vector. "text" refers to the input natural sentence. The Embed model can be selected according to needs. The BAAI / bge-zh series models are commonly used for Chinese scenarios.

[0072] Furthermore, the two types of knowledge bases are the knowledge graph base and the enhanced retrieval generation knowledge base: the knowledge graph base contains a structured graph of disease symptom associations and diagnosis and treatment standard knowledge; after the natural language is converted into a vector q, the similarity S can be calculated with each entity in the knowledge graph i =q T E KG [i], find e that satisfies a similarity greater than the specified threshold τ i Entity E KG A collection of:

[0073] ε query ={e i |q T E KG [i]>τ}

[0074] The enhanced retrieval generation knowledge base is a dynamically updated knowledge base of clinical guidelines, case literature, and unstructured knowledge of the hospital's medical resources. For the RAG knowledge base, the Retriever model is used to perform relevance retrieval on the vector q in the knowledge base to obtain the relevant text set D:

[0075] D=Retriever(q,RAGChunk)={(text1,text2...text i )}

[0076] Furthermore, the central processing module is composed of multiple intelligent agents optimized for tasks. The intelligent agents include prompt word engineering and large models with different characteristics, which are responsible for specific links such as triage and diagnosis recommendations. When the generative large model in the intelligent agent generates an answer, it will search from the knowledge graph library and RAG knowledge base according to the scenario. For the knowledge graph library: find the answer that meets the similarity S i =q T E KG [i] e greater than the specified threshold τ i Entity set; then find the required relationship according to the current scenario. For example, in the medication analysis link, if you need to find the conflicting effects between different drugs, you need to find the entity set with higher correlation ε query Are there conflicting pairs of entities in:

[0077]

[0078] Here C is the set of conflicting pairs, and r c Represents a conflicting relationship; this method is also applicable to finding other relationships between entities in the knowledge graph. The generation probability mask of the conflicting relationship for the large language model is:

[0079]

[0080] For the relevant text set D retrieved from the retrieval-enhanced generated knowledge base, it is encoded through the embedded model:

[0081] H RAG =Embed LLM (D)∈R m×d

[0082] Get the vectorized representation H of the relevant text fragments in the relevant knowledge base RAG After that, the attention vector is generated through the weight matrix W:

[0083] Q RAG =H RAG W Q ,K RAG =H RAG W K ,V RAG =H RAG W V

[0084] For a large language model based on the Transformer architecture, its attention mechanism is:

[0085]

[0086] Among them, d kis a scaling factor. The Softmax function normalizes the vector to ensure that each element in the vector has a value between 0 and 1, thereby reflecting the probability distribution of each content when the large language model generates the result:

[0087]

[0088] After considering the RAG retrieval content, the attention mechanism needs to additionally introduce attention to the vector of the retrieved knowledge base fragment:

[0089] H LLM+RAG =CrossAttention(H LLM ,H RAG )

[0090] Right now:

[0091]

[0092] Through cross-attention, the agent's large model's attention span is expanded from the input text to include relevant text fragments retrieved from the RAG knowledge base. When generating answers, the knowledge in the RAG knowledge base is fully considered as the basis for the large model's answer generation. The above two mechanisms enable the content of the RAG knowledge base and the KG knowledge base to influence the large model's answer generation in different ways. When performing certain specific tasks, the agent needs to retrieve relevant knowledge from the knowledge base according to the situation and influence the generation of the large model through the corresponding mechanism, namely, a joint generation probability influence:

[0093] P final (w t )=Normalize(P LLM (w t )·P RAG (w t )·Mask(w t ))

[0094] Among them, P final is the final generation probability; P LLM is the generation probability of the large model itself; Mask(w t ) is defined as only 0 and 1 in the conflict relationship, that is, the probability of generating conflicting content is 0. The relationship to be used to retrieve entities in the KG database needs to be selected according to the actual business process; RAG (w t )=Softmax(W RAG H LLM+RAG ) is used to enhance the generation of impact probability for retrieval, so that the final content generation probability fully considers the retrieval fragment content of the RAG knowledge base, It is a normalization operation similar to Softmax, which limits the final generation probability of each element to between 0 and 1.

[0095] Furthermore, the triage auxiliary module of the system is embedded in the nurse workstation during the triage process. Its structure is as follows: Figure 3 As shown. By invoking a dedicated agent, the triage assistance module combines triage criteria and patient profiles contained in the knowledge base, along with information such as the hospital's physician expertise and current patient load, to generate condition classification and department assignment recommendations. This reduces scheduling errors caused by triage staff's lack of experience and unfamiliarity with hospital resources, and improves triage efficiency. The triage results and condition summary are simultaneously transmitted to the corresponding doctor's station in the office.

[0096] Furthermore, the system's diagnostic auxiliary function module is deeply integrated into the doctor's workstation, and its structure is as follows: Figure 4 As shown. After the triage is completed, the system automatically receives the synchronized files of the front-end link, which contains the integrated multi-source information of the patient, and displays the key structured information in the front-end process to help doctors quickly understand the condition. The diagnostic assistance module integrates a button in the doctor's station, so that doctors can call various auxiliary functions provided by different intelligent bodies in the central processing module of this system when browsing information in the workstation; for example: the inspection item recommendation function can provide inspection suggestions and provide a one-click adoption button to quickly prescribe inspections; after the test results are returned, the doctor can use the button to call the intelligent question and answer body in the system to summarize and generate suggestions; the medication reminder function can provide real-time reminders to doctors based on information such as drug conflict data and medication specification documents. All conclusions can be marked with the source in the knowledge base to facilitate doctor review.

[0097] In addition, the system can play a key role in public health emergencies such as large-scale infectious disease outbreaks. By connecting to public health databases, it can obtain real-time epidemic-related information, such as the characteristics of virus transmission and symptoms. During the triage process, the system prioritizes identifying patients with infectious disease symptoms and directs them to dedicated isolation areas for further examination and diagnosis. At the same time, based on the special allocation of medical resources during the epidemic, patients are rationally allocated to corresponding departments or temporary medical points to ensure the efficient use of medical resources. For example, when an influenza outbreak occurs in a certain area, the system can quickly screen suspected influenza patients, arrange rapid testing for them, and allocate them to different levels of treatment areas based on the severity of their condition.

[0098] It also includes a remote collaboration module and a system monitoring module: the remote collaboration module is used for real-time data sharing, enabling medical staff at all levels to exchange information in real time, share various patient information, and obtain expert diagnostic advice and treatment plans; the monitoring module is used to monitor the operating status of the system in real time, including the operating efficiency of the intelligent body, the update status of the knowledge base, and the stability of data transmission.

[0099] This invention overcomes the traditional model of emergency assistance systems that rely on static knowledge graph rule reasoning and historical data probabilistic prediction. It divides the knowledge application forms of emergency scenarios - using knowledge graphs to achieve structured constraints for rigid links such as the implementation of diagnosis and treatment standards, and introducing retrieval enhancement generation technology to introduce knowledge references for dynamic needs such as complex disease assessment. At the same time, by constructing multiple large-model intelligent entities and corresponding prompt word projects, the capabilities of different business links are specialized. This hybrid knowledge management strategy not only ensures the safety of medical decision-making, but also has lower maintenance costs than a single knowledge graph solution.

[0100] The above describes in detail the preferred embodiments of the present invention. It should be understood that those skilled in the art can make numerous modifications and variations based on the concepts of the present invention without inventive effort. Therefore, any technical solutions that can be derived by those skilled in the art through logical analysis, reasoning, or limited experimentation based on the concepts of the present invention and the prior art should be within the scope of protection defined by the claims.

Claims

1. An emergency decision-making support system based on a large-scale intelligent agent and multiple knowledge bases, comprising: Input end, used to obtain patient information from multiple dimensions; The central processing module is used for intelligent agents in different scenarios, which are built based on different large models and prompt word projects and are called through embedded structures on the output server of the system; The knowledge base includes a knowledge graph library and a search-enhanced generated document library. The structured rules of the knowledge graph library ensure reliable decision-making, and the unstructured text of the search-enhanced generated document library realizes dynamic knowledge fusion. Business nodes select a single library or a dual library to store knowledge according to fault tolerance requirements. The output service end includes an intelligent triage assistance module and an intelligent diagnosis assistance module, which are embedded in the triage nurse station and the emergency doctor station in the form of buttons and call-out windows respectively, to provide assistance by calling intelligent body services.

2. The emergency decision support system based on a large model agent and multiple knowledge bases as claimed in claim 1 is characterized in that: The input terminal obtains patient information from three channels: 1) The real-time speech input module can convert the patient's or first responder's verbal description into text, supporting dialect recognition and ambiguous expression correction; 2) Automatically connect to electronic medical records based on identity information, extract patient history and key information such as allergy history; 3) The Vital Sign Monitoring Module directly interfaces with medical equipment to continuously acquire vital sign data such as blood pressure and blood oxygen levels. These three types of information are presented in different formats. A dedicated agent in the central processing module analyzes and integrates these information into a unified comprehensive report and temporary archive, which can be accessed and updated in real time by subsequent system steps.

3. The emergency decision support system based on a large model agent and multiple knowledge bases as claimed in claim 2 is characterized in that: The central processing module is composed of multiple intelligent agents built for tasks, which include prompt word engineering and large models with different expertise, which assist in various businesses such as information retrieval, triage recommendation, and diagnosis advice. The system has two built-in knowledge bases: the knowledge graph library contains structured graphs of knowledge such as disease symptom associations, diagnosis and treatment standard knowledge, hospital rules and regulations, and quality control rules; the RAG library is a knowledge base of unstructured knowledge such as dynamically updated clinical guidelines, case literature, and medical resources of the hospital. After the patient enters the emergency process, the system will generate a temporary exclusive file through the corresponding intelligent agents in the input module and the central processing module, dynamically record the patient's symptoms, signs, and medical history, and temporarily store them in the knowledge base.

4. The emergency decision support system based on a large model agent and multiple knowledge bases as claimed in claim 3 is characterized in that: The triage assistance module can combine the triage standards and patient temporary files contained in the knowledge base, and generate disease classification suggestions and department allocation recommendations based on the distribution of doctors' expertise in the hospital and the current reception load.

5. The emergency decision support system based on a large model agent and multiple knowledge bases as claimed in claim 4 is characterized in that: After the triage is completed, the diagnosis assistance module automatically receives the synchronized files of the front-end link in this system, which contains the integrated multi-source information of the patient and displays the key information in the front-end process. The diagnosis assistance module is integrated in the doctor's station, so that the doctor can call the customized assistance functions provided by different intelligent agents in the central processing module of this system when browsing information in the workstation.

6. The emergency decision support system based on a large model agent and multiple knowledge bases as claimed in claim 5 is characterized in that: The input end converts the natural language input into a vector through the embedding model, and the formula is: q=Embed LLM (”Text")∈R d Among them, R d Indicates the dimension of the conversion vector. "text" refers to the input natural sentence. The Embed model can be selected according to needs. The BAAI / bge-zh series models are commonly used in Chinese scenarios.

7. The emergency decision support system based on a large model agent and multiple knowledge bases as claimed in claim 6 is characterized in that: When the generative model in the agent generates answers, it draws knowledge from the knowledge graph and RAG knowledge base based on the scenario. The knowledge graph stores structured documents, and its constraints must be considered; the RAG knowledge base provides private domain knowledge that the model does not cover. After inputting the text translation, it can calculate similarity and match entities in the knowledge graph; the formula is: e query ={e i |q T E KG [i]>t} Where q is the vector of input text; E KG It is the embedding matrix of each entity in the knowledge graph library, and finds the entities that satisfy the similarity S i =q T E KG [i] e greater than the specified threshold τ i Entity set; then find the required relationship according to the current scenario. For example, in the medication analysis link, if you need to find the conflicting effects between different drugs, you need to find the entity set with higher correlation ε query Are there conflicting pairs of entities in: Here C is the set of conflicting pairs, r c Represents a conflicting relationship; this method is also applicable to finding other relationships between entities in the knowledge graph. The probability mask of the conflicting relationship for the intelligent agent's large language model is: For the RAG knowledge base, the Retriever model is needed to perform relevance retrieval on the input text vector q in the knowledge base to obtain a collection of text content: D=Retriever(q,RAGChunk)={(text1,text2...text i )} And encode the text content through the embedded model: H RAG =Embed LLM (D)∈R m×d Get the vector H of the relevant text fragment of the knowledge base RAG Finally, the attention vector is generated through the weight matrix: Q RAG =H RAG ·W Q ,K RAG =H RAG ·W K ,V RAG =H RAG ·W V For a large language model based on the Transformer architecture, its attention mechanism is: Among them, d k is a scaling factor. The Softmax function normalizes the vector to ensure that each number in the vector is between 0 and 1, thereby reflecting the probability distribution of each element: After considering the RAG retrieval content, the attention mechanism needs to additionally introduce attention to the knowledge base fragment vector: H LLM+RAG =CrossAttention(H LLM ,H RAG ) Right now: By using cross-attention, the attention scope of the large model in the agent is expanded from the input text to the relevant text fragments retrieved from the RAG knowledge base; When performing some specific tasks, the agent needs to retrieve relevant knowledge from the KG and RAG knowledge bases respectively, and influence the generation of the large model through different mechanisms, that is, the joint generation probability influence: P final (w t )=Normalize(P LLM (w t )·P RAG (w t )·Mask(w t )) Among them, Mask(w t ) is defined as only 0 and 1 in the conflict relationship, that is, the probability of generating conflicting content is 0; P final is the final generation probability; P LLM is the generation probability of the large model itself; P RAG (w t )=Softmax(W RAG H LLM+RAG ) is the RAG enhancement impact probability; so that the final content generation probability fully considers the RAG knowledge base retrieval fragment content; It is a normalization operation similar to Softmax, which limits the final generation probability of each element to between 0 and 1.

8. The emergency decision support system based on a large model agent and multiple knowledge bases as claimed in claim 7 is characterized in that: It also includes a remote collaboration module and a system monitoring module: the remote collaboration module is used for real-time data sharing, enabling medical staff at all levels to exchange information in real time, share various patient information, and obtain expert diagnostic advice and treatment plans; the monitoring module is used to monitor the operating status of the system in real time, including the operating efficiency of the intelligent body, the update status of the knowledge base, and the stability of data transmission.

Images