Clinical scenario prediction method and related equipment for patients with acute myocardial infarction during hospitalization

Abstract

This application relates to a method and related equipment for predicting clinical scenarios during hospitalization of patients with acute myocardial infarction. The related equipment includes a system comprising multiple agents driven by a large model to perform different tasks. These agents communicate and iteratively optimize each other through prompts. The multiple agents include a first agent for performing medical text structuring, a second agent for performing medical knowledge graph construction, a third agent for performing clinical scenario prediction, and a fourth agent for performing evaluation result feedback optimization. This application can quickly and accurately predict the clinical scenarios of patients' subsequent disease progression based on their admission and medical records, helping to quickly and accurately identify acute myocardial infarction, avoiding misdiagnosis, enabling timely and effective treatment when patients experience acute myocardial infarction, reducing adverse cardiovascular events, and improving prognosis.

Description

Technical Field

[0001] The present application relates to the intersection of artificial intelligence and medicine, and in particular to a clinical scenario prediction method and related equipment for patients with acute myocardial infarction during hospitalization. Background Art

[0002] Acute myocardial infarction is a potentially life-threatening cardiovascular disease characterized by a sharp reduction or interruption of coronary blood flow, leading to myocardial ischemic necrosis. During the hospitalization of patients with acute myocardial infarction, reasoning about the clinical scenarios in their course of illness is crucial because it involves rapid and accurate diagnosis and timely and effective treatment to reduce the occurrence of adverse events such as myocardial necrosis and improve prognosis. At present, the clinical scenarios of patients with acute myocardial infarction during hospitalization mainly rely on doctors to give relevant opinions through subjective judgment, however, the accuracy of this method is not high enough. Summary of the invention

[0003] In response to the above-mentioned deficiencies or shortcomings, the present application provides a clinical scenario prediction method and related equipment for patients with acute myocardial infarction during hospitalization.

[0004] According to a first aspect, the present application provides a clinical scenario prediction system for acute myocardial infarction patients during hospitalization. In some embodiments, the system includes multiple intelligent agents driven by a large model for performing different tasks, and the multiple intelligent agents communicate information and iteratively optimize through prompt words; the multiple intelligent agents include:

[0005] The first agent is used to extract clinical features related to clinical scenario prediction from the medical data of multiple patients with acute myocardial infarction and analyze the relationship between the clinical features to generate structured medical knowledge text;

[0006] The second agent is used to construct a first medical knowledge graph based on the medical knowledge text; the entities in the first medical knowledge graph include various clinical features extracted by the first agent, and the relationship between the entities in the first medical knowledge graph is used to reflect the properties between two entities;

[0007] The third agent is used to obtain a clinical scenario prediction task, take the task problem of the clinical scenario prediction task as the target problem, retrieve data matching the target problem as reference knowledge, the reference knowledge includes data matching the target problem retrieved from the medical knowledge graph constructed by the second agent, perform clinical scenario prediction based on the reference knowledge, and obtain the first prediction result of the target problem; upon receiving the guidance experience of the target problem returned by the fourth agent, perform clinical scenario prediction again in combination with the reference knowledge and the guidance experience, and obtain the final prediction result of the target problem;

[0008] The fourth intelligent agent is used to evaluate the prediction results of the task problem sent by the third intelligent agent and return the evaluation results to the third intelligent agent; when the evaluation result of the first prediction result of any task problem sent by the third intelligent agent indicates that the evaluation fails, the guidance experience generated based on the task problem, the first prediction result of the task problem and the correct answer is returned to the third intelligent agent.

[0009] In some embodiments, the reference knowledge also includes prediction records matching the target problem retrieved from a record database and guidance experience matching the target problem retrieved from an experience database; the record database is used to store prediction records corresponding to task problems processed by the third agent, and the experience database is used to store effective guidance experience corresponding to task problems processed by the third agent;

[0010] The third intelligent agent is also used to store the problem information, planning results, prediction results and evaluation results corresponding to the target problem as a prediction record in the record database when receiving the first prediction result or the evaluation result of the last prediction result of the target problem; and to store the guiding experience of the target problem in the experience database when the evaluation result of the last prediction result of the target problem indicates that the evaluation is passed.

[0011] In some embodiments, the first agent is further configured to periodically use a search tool to search for literature related to acute myocardial infarction;

[0012] The second intelligent agent is also used to construct a second medical knowledge graph based on the literature searched by the first intelligent agent; the entities in the second medical knowledge graph include medical entity words extracted by the second intelligent agent from the literature, and the relationship between the entities in the second medical knowledge graph is used to reflect whether the two entities are positively correlated, negatively correlated or unrelated.

[0013] In some embodiments, the fourth agent is also used to evaluate the clinical feature extraction results of the first agent using pre-specified evaluation indicators, and provide optimization suggestions for the first agent based on the evaluation results, and / or to evaluate the entity relationship extraction results of the second agent, and provide optimization suggestions for the second agent based on the evaluation results.

[0014] In some embodiments, the first intelligent agent is specifically used to obtain a list of medical names provided by a doctor from a pre-specified file or database, and to extract clinical features from the medical data of each acute myocardial infarction patient with reference to the list of medical names; it is also used to send a medical name confirmation request containing the uncertain medical name and context information to the doctor when an uncertain medical name is obtained from the medical data, and to determine the corresponding medical name based on the doctor's feedback information.

[0015] In some embodiments, clinical features related to clinical scenario prediction include symptom features, disease features, drug features, examination features, and clinical intervention features; the feature information of each clinical feature includes feature name, category, feature value, parsing and discrimination rules, and clinical scenario association identifier; the value of the clinical scenario association identifier is used to indicate whether the relevant clinical feature is associated with a specific clinical scenario.

[0016] In some embodiments, the first intelligent agent is also used to obtain medical terms and their corresponding discrimination rules from medical literature and medical standards, obtain medical terms and their corresponding discrimination rules specified by doctors, and organize the obtained medical terms and discrimination rules into a first standard file in a specific format; when extracting clinical features, the first standard file is used as a reference to determine the parsing and discrimination rules of the extracted clinical features.

[0017] In some embodiments, the first intelligent agent is also used to obtain medical terms and their corresponding clinical scenario association identifiers from medical literature and medical standards, obtain medical terms specified by doctors and their corresponding clinical scenario association identifiers, and organize the obtained medical terms and clinical scenario association identifiers into a second standard file in a specific format; when extracting clinical features, the clinical scenario association identifiers of the extracted clinical features are determined with reference to the second standard file.

[0018] In some embodiments, the clinical scenario prediction task is used to instruct a third agent to predict the clinical scenario of a designated acute myocardial infarction patient during hospitalization, and the clinical scenario includes one or more of the patient's symptoms, signs, auxiliary examinations, disease diagnosis, drug treatment plan, surgical treatment decisions, nursing care, and complications of acute myocardial infarction.

[0019] According to a second aspect, the present application provides a method for predicting clinical scenarios during hospitalization of patients with acute myocardial infarction, the method using a system as provided in any of the above embodiments, the method comprising:

[0020] receiving prompt words for indicating a clinical scenario of a target acute myocardial infarction patient during hospitalization, and extracting a task question from the prompt words;

[0021] Retrieve data matching the task problem from the constructed knowledge graph, record database, and experience database as target reference knowledge;

[0022] Clinical scenario prediction is performed based on target reference knowledge to obtain target results.

[0023] The embodiment of the present application is based on the idea of ​​Multi Agent, and designs a unified Multi Agent framework for the reasoning of the course of disease of patients with acute myocardial infarction during hospitalization. Based on the framework, a clinical scenario prediction system for patients with acute myocardial infarction during hospitalization is implemented. The system includes multiple intelligent agents driven by a large model, namely, a first intelligent agent, a second intelligent agent, a third intelligent agent and a fourth intelligent agent. The intelligent agents communicate information and iteratively optimize through prompt words. Different intelligent agents are used to perform different tasks, among which the first intelligent agent is used to perform the task of structuring medical texts, the second intelligent agent is used to perform the task of building a medical knowledge graph, the third intelligent agent is used to perform the task of clinical scenario prediction, and the fourth intelligent agent is used to perform the task of optimizing the feedback of evaluation results. Let each intelligent agent focus on a specific task, so that each intelligent agent can work more efficiently and accurately, and through multiple intelligent agents to jointly complete a complex task through collaboration and coordination, there is a great improvement in ability compared to a single intelligent agent, and it can also achieve goals that a single intelligent agent cannot complete. Based on the collaboration between multiple intelligent agents, the system can quickly and accurately predict the clinical scenario of the patient's subsequent course of illness based on the admission and medical record text of the patient with acute myocardial infarction. This helps to quickly and accurately identify acute myocardial infarction and avoid misdiagnosis. When the patient suffers from acute myocardial infarction again, it can provide timely and effective treatment to the patient, so as to avoid or reduce the occurrence of adverse events such as myocardial necrosis and improve prognosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of the architecture of an intelligent agent provided by the present application according to one or more embodiments;

[0025] Figure 2 A schematic diagram of a multi-agent-based clinical scenario prediction architecture provided by the present application according to one or more embodiments;

[0026] Figure 3 A flowchart of a method for predicting clinical scenarios during hospitalization of patients with acute myocardial infarction provided in accordance with one or more embodiments of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solutions and advantages of the present application clearer, the embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings. It should be clear that the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.

[0028] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Instead, they are only examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims.

[0029] In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances. In addition, in the description of the present application, unless otherwise specified, "multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.

[0030] In view of the shortcomings of related technologies, this application is based on the idea of ​​Multi Agent, and designs a unified Multi Agent framework for the reasoning of the course of acute myocardial infarction patients during hospitalization. Based on this framework, a clinical scenario prediction system for acute myocardial infarction patients during hospitalization is implemented. The system includes multiple agents driven by a large model, and the agents communicate information and iteratively optimize through prompt words. The multiple agents in the system are used to perform different tasks, allowing each agent to focus on a specific task, so that each agent can work more efficiently and accurately, and multiple agents can complete a complex task together through collaboration and coordination, which has a great improvement in capabilities compared to a single agent, and can also achieve goals that a single agent cannot complete. Based on the collaboration between multiple agents, the system can quickly and accurately predict the clinical scenario of the patient's subsequent course of disease based on the admission and medical record text of the patient with acute myocardial infarction, which helps to quickly and accurately identify acute myocardial infarction, avoid misdiagnosis, and provide timely and effective treatment to patients when acute myocardial infarction occurs again, so as to avoid or reduce adverse events such as myocardial necrosis in patients, and improve prognosis. The system is described in detail below.

[0031] The system can be implemented using computer devices such as servers. In some embodiments, the internal structure of the computer device may include a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection.

[0032] Regarding the above-mentioned intelligent agent, the intelligent agent is also called an intelligent agent or agent. It is an artificial intelligence system with a certain degree of autonomy that can perceive the environment, make decisions and take actions. Figure 1 The figure shows the architecture of the agent. In this architecture, the big model is used as the logical engine of the agent, which includes four major elements: planning, memory, tools, and action. The agent is located in the center, and it handles complex tasks by coordinating various components. The descriptions of the major elements are as follows: Planning: that is, the ability to plan and make decisions, adjust strategies through reflection and self-feedback, use thinking chains to clarify ideas, and decompose sub-goals to effectively perform complex tasks and ultimately achieve the final goal efficiently. Memory: Short-term memory involves learning context and is part of prompt engineering. Long-term memory involves long-term retention and retrieval of information, which requires the use of external vector storage and fast retrieval. Tools: Various tools that the agent may call, including calendars, computers, code interpreters, and search functions. These tools expand the capabilities of the agent, enabling it to perform more complex tasks. Execution: The agent performs specific actions based on planning and memory.

[0033] For further information, see Figure 2 , Figure 2The following is a multi-agent-based clinical scenario prediction architecture provided in some embodiments. As shown in the figure, these embodiments divide the goal of clinical scenario prediction into four tasks, namely, medical text structuring, construction of medical knowledge graph, clinical scenario prediction, and evaluation result feedback optimization. These tasks are respectively implemented by the first agent, the second agent, the third agent, and the fourth agent in the system. The specific task content is sent to each agent by the user (such as the system administrator, the system user, etc.) through a prompt word. After receiving the corresponding prompt word, each agent can perform the corresponding task according to the prompt word. Each agent can simulate the doctor's thinking process of diagnosing and treating the disease by performing the task, that is, on the one hand, it is necessary to refer to the patient's medical history, medical examination, and the diagnosis and treatment results after admission. On the other hand, it is necessary to combine the diagnosis and treatment guidelines and medical knowledge related to acute myocardial infarction disease to continuously improve the understanding of the patient's condition so as to more accurately perform diagnosis and treatment intervention. The following is a description of each agent.

[0034] The first agent (referred to as Agent 1) is used to perform the task of structuring medical texts. The task content includes: extracting clinical features related to clinical scenario prediction from the medical data of multiple patients with acute myocardial infarction and analyzing the relationship between the clinical features to generate structured medical knowledge texts.

[0035] The first agent will automatically construct prompts based on the clinical features specified by the doctor, and execute the prompts to extract clinical features from the medical data of each acute myocardial infarction patient and analyze the relationship between the clinical features, and generate structured medical knowledge text based on the extracted clinical features and the relationship between the clinical features. The first agent will also send the generated medical knowledge text to the second agent.

[0036] These patients with acute myocardial infarction can be selected from one or more hospitals. The medical data of patients with acute myocardial infarction include data generated during the diagnosis and treatment process, such as electronic medical records, examination reports, and test reports. The clinical features specified by the doctor refer to clinical features related to clinical scenario prediction. There are many clinical features related to clinical scenario prediction, and these clinical features can be divided into 5 categories, including symptom features, disease features, drug features, examination features, and clinical intervention features. The clinical features of each category usually include many clinical features. For example, symptom features may include clinical features such as no headache, chest pain, fever, nausea, and vomiting. Disease features may include clinical features such as cerebral infarction, hypertension, and esophageal cancer. Drug features may include clinical features such as esomeprazole, cefoperazone, and metformin sustained-release tablets. Examination features may include clinical features such as head CT scan, blood routine, white blood cell count, and oxygen partial pressure. Clinical intervention features may include clinical features such as intravenous infusion of citicoline sodium, oral metformin sustained-release tablets, and bilateral subtotal thyroidectomy.

[0037] When the first agent performs the clinical feature extraction operation, it needs to identify the clinical features from the medical data and extract the feature information of the clinical features. The feature information of each clinical feature includes the feature name (feature_name), category (feature_category), feature value (feature_value), analysis and discrimination rule (feature_analysis) and clinical scenario association identifier (clinical_context). Taking the clinical feature of whether the patient has chest pain as an example, the feature name of the clinical feature is "chest pain", the feature category is "symptom feature", the feature value is "0 or 1", the analysis and discrimination rule is "if the patient has had chest pain symptoms before, then output 1, if there is no chest pain, then output 0", and the clinical scenario association identifier is "yes". It should be noted that the clinical scenario association identifier indicates whether the relevant clinical feature is associated with a specific clinical scenario, which means that if the subsequent modeling needs to predict whether the patient will have a specific clinical scenario in the future based on the clinical feature, the clinical scenario association identifier of the clinical feature is marked as "yes", and if not, the clinical scenario association identifier of the clinical feature is marked as "no", indicating that the clinical feature only participates in model training and is not used as a prediction target. The specific clinical scenario can be any clinical scenario. The category of the clinical scenario is related to the items included in the clinical scenario, and the number of categories of clinical scenarios is equal to 2 N -1, N represents the maximum number of items that a clinical scenario can contain. For example, assuming that each clinical scenario includes one or more of the patient's symptoms, signs, and auxiliary examinations, it can be determined that N=3, the number of clinical scenario categories is equal to 7, and the various clinical scenarios are (1) "symptoms", (2) "signs", (3) "auxiliary examinations", (4) "symptoms and signs", (5) "symptoms and auxiliary examinations", (6) "signs and auxiliary examinations", and (7) "symptoms, signs, and auxiliary examinations".

[0038] Exemplarily, the prompt words automatically constructed by the first agent may be as follows:

[0039] -Target-

[0040] Based on the patient's electronic medical record or clinical guideline text provided, accurately extract the five categories of medical-related feature information: patient symptoms, diseases, drugs, examinations, and clinical interventions. Ensure that a detailed description is provided for each extracted feature, including but not limited to the feature name, category, feature value, parsing and discrimination rules, and whether the feature is applicable to a specific clinical scenario.

[0041] -step-

[0042] 1. Identify and classify all entities.

[0043] For each identified entity, extract the following information:

[0044] feature_name: feature name, in uppercase;

[0045] feature_category: one of the five categories to which the feature belongs (patient symptoms, diseases, medications, tests, clinical interventions);

[0046] feature_value: the specific value or description of the feature;

[0047] feature_analysis: analysis of the feature, including its meaning and importance in the text;

[0048] clinical_context: indicates whether the feature is associated with a specific clinical scenario;

[0049] Format each feature as: ("feature"{tuple_delimiter}<feature_name> {tuple_delimiter}<feature_category> {tuple_delimiter}<feature_value> {tuple_delimiter}<feature_analysis> {tuple_delimiter}<clinical_context> ).

[0050] 2. Analyze the relationships between the features identified in step 1.

[0051] For each pair of features that are significantly associated, extract the following information:

[0052] source_feature: the name of the source feature, as identified in step 1;

[0053] target_feature: the name of the target feature, as identified in step 1;

[0054] relationship_explanation: explains the relationship between the source feature and the target feature and its importance;

[0055] relationship_strength: Use a score from 1 to 5 to indicate the closeness of the relationship between two features, where 5 means very close;

[0056] Format the relationship between each pair of features as: ("relationship"{tuple_delimiter}<source_feature> {tuple_delimiter}<target_feature> {tuple_delimiter}<relationship_explanation> {tuple_delimiter}<relationship_strength> ).

[0057] 3. Return the Chinese output, summarizing all the features identified in steps 1 and 2 and the relationships between them, using **{record_delimiter}** as the delimiter between list items.

[0058] 4. When finished, end the output with {completion_definer}.

[0059] The second agent (referred to as Agent 2) is used to perform the task of building a medical knowledge graph. The task content includes: building a first medical knowledge graph based on medical knowledge text.

[0060] The first medical knowledge graph refers to a medical knowledge graph constructed based on medical knowledge text. The entities in the first medical knowledge graph include various clinical features extracted by the first agent, and the relationship between the entities in the first medical knowledge graph is used to reflect the properties between two entities.

[0061] Specifically, the second agent analyzes the structured text obtained from the first agent, that is, the medical knowledge text, and automatically identifies and extracts entities, relationships, attributes, etc., and constructs the first medical knowledge graph based on the extracted entities, relationships, and attributes, and continuously optimizes and improves the accuracy and effect of entity relationship extraction. Exemplarily, the first medical knowledge graph can be loaded into a database of visual knowledge graphs (such as Neo4j), and the construction of entities, relationships, etc. in the knowledge graph can be manually checked and adjusted according to actual conditions, such as optimizing prompt words to improve the second agent's sensitive recognition of drugs, and for example, the categories of extracted information can be increased, such as Claim Date—declared date, etc.

[0062] The entities extracted by the second agent from the medical knowledge text may be feature names in feature information of each clinical feature, and the attributes extracted by the second agent from the medical knowledge text may be other feature information of the clinical feature.

[0063] In order to enrich the content of the first medical knowledge graph and improve its practicality and accuracy, in addition to entities, relationships and attributes, more information can be further extracted. In some embodiments, the information extracted by the second agent from the medical knowledge text for constructing the medical knowledge graph includes:

[0064] -entity_name: entity name;

[0065] -entity_type: entity label or category;

[0066] -source_entity: from entity name, the name of the source entity;

[0067] -target_entity: the name of the target entity from entity name;

[0068] -relationship: the relationship between the source entity and the target entity;

[0069] -claims: A statement of facts, relationships, or attributes about a specific entity, such as a company, person, or organization.

[0070] The third agent (which can be called Agent 3) is used to perform the task of clinical scenario prediction. The task content includes: obtaining the clinical scenario prediction task, taking the task problem of the clinical scenario prediction task as the target problem, retrieving the data matching the target problem as the reference knowledge, the reference knowledge includes the data matching the target problem retrieved from the medical knowledge graph constructed by the second agent, performing clinical scenario prediction based on the reference knowledge, and the obtained clinical scenario prediction result is the first prediction result of the target problem; when receiving the guidance experience of the target problem returned by the fourth agent, combining the reference knowledge and the guidance experience to perform clinical scenario prediction again, and the obtained clinical scenario prediction result is the final prediction result of the target problem.

[0071] The clinical scenario prediction task is used to instruct the third agent to predict the clinical scenario of the designated acute myocardial infarction patient during hospitalization, specifically the prompt word (prompt) received by the third agent, such as "Hello, what examination does the patient with symptom X need to undergo next?" The clinical scenario includes one or more of the patient's symptoms, signs, auxiliary examinations, disease diagnosis, drug treatment plan, surgical treatment decision, nurse care, and complications of acute myocardial infarction (such as heart failure, arrhythmia, mechanical complications, and death, etc.).

[0072] In this embodiment, the third agent will take the task problem of the clinical scenario prediction task (such as "what examinations do patients with certain symptoms need to undergo") as the target problem, parse the target problem, and construct a clinical scenario prediction model based on the parsing results. It can autonomously retrieve the medical knowledge graph constructed by the second agent to obtain clinical features that match the target problem. These clinical features are used to train the constructed clinical scenario prediction model. Among them, the large model of the third agent is a general large model, which does not recognize knowledge in specific fields and therefore cannot accurately answer questions involving knowledge in specific fields. To this end, this embodiment uses the medical knowledge graph constructed by the second agent as an external knowledge base for retrieval and enhanced generation, which can enhance the knowledge of the large model question-answering system in vertical fields to enhance the factuality of the generated answers and reduce hallucinations.

[0073] The third agent has a reasoning module, which will perform reasoning based on the clinical scenario prediction model to obtain the clinical scenario prediction result of the target problem. The specific content of the clinical scenario prediction result depends on the content of the target problem. For example, if the target problem is "what examinations do patients with certain symptoms need", the clinical scenario prediction result can be "certain examination items".

[0074] Furthermore, the reasoning module will be adjusted by the fourth agent to continuously learn and optimize the results. Among them, the core part of the third agent and the fourth agent to achieve reasoning prediction is "memory-planning-reflection". The large model-driven agent handles tasks through interrelated modules. The third agent will plan based on the target problem (the planning method can be subtask decomposition, thinking chain, etc.) and perform subtasks, use tools to enhance related functions, and learn and improve through reflective behavior. For example, if the evaluation result of the clinical scenario prediction result (i.e., the first prediction result) initially generated by the third agent for the target problem is not passed, the fourth agent will return the guidance experience, and the third agent will use the reasoning module, combined with reference knowledge and guidance experience to predict the clinical scenario again, obtain a new clinical scenario prediction result (i.e., the last prediction result), and send the new clinical scenario prediction result to the fourth agent for re-evaluation. Under the continuous feedback optimization of the fourth agent, the prediction result generated by the third agent will eventually meet the preset output requirements.

[0075] In some embodiments, the prompt words provided by the user to the third intelligent agent to instruct the execution of the task of clinical scenario prediction can include instruction information for implementing thought planning, so as to instruct the third intelligent agent to use planning methods such as sub-goal decomposition and thought chain to gradually think about and implement predictions of various clinical scenarios when performing clinical scenario predictions.

[0076] For example, the prompt words provided by the user may be as follows:

[0077] You are a professional medical assistant responsible for building and optimizing clinical scenario prediction models. You need to follow the following steps and principles:

[0078] 1. **Loading knowledge graph data**: - When performing reasoning, first load the medical knowledge graph constructed by the second agent, including entities, attributes, relationships, and text units. - Use the data in the knowledge graph to parse the keywords, entity names, etc. involved in the task question, and retrieve and match them with the entity relationships in the knowledge graph to obtain a comprehensive and accurate answer.

[0079] 2. **Analyze the task questions**: - Preprocess the task questions, including word segmentation, removal of stop words, etc. - Identify the keywords in the questions (i.e. important words or phrases that can reflect the core content of the questions, which are usually used to describe the topic, goal or condition of the question) to ensure that the core content of the question is understood.

[0080] 3. **Build a prediction model**: - Based on the analysis results, build a prediction model and use the relevant features in the knowledge graph to train the model. - Use planning methods such as sub-goal decomposition and thinking chain to gradually think about and achieve predictions for various clinical scenarios (symptoms, signs, auxiliary examinations, disease diagnosis, drugs, surgery, complications, etc.).

[0081] 4. Reasoning Module: - Use the reasoning module to infer the results based on the information in the model and knowledge graph. - The reasoning module can be adjusted by the fourth agent to continuously learn and optimize the results.

[0082] 5. **Evaluation and optimization**: - The results of the current round iteration are passed to the fourth agent through prompt for evaluation (the evaluation involves accuracy, completeness, consistency, etc.). - Accept the experience guidance of the fourth agent (for example: the necessity of a certain disease for a certain drug). - Save the entire process information of this round (including task problems, thinking chain planning, predicted results, and the fourth agent's result evaluation) in the "record database". - Save the effective guidance experience returned by the fourth agent in the "experience database" as the experience for the next round of planning.

[0083] 6. **Learning through iteration**: - In the next round, relevant records and experiences are retrieved from the record database and experience database and applied to the planning stage. - This allows for continuous learning through iteration and gradually improves the accuracy and effectiveness of the prediction model.

[0084] 7. **Specific operation principles**: - **When the information is insufficient**, do not make a diagnosis too early. - **Ask the patient questions many times and actively** to obtain sufficient information. - **Only ask one question each time**, and keep it as short as possible. - **Ask the patient to undergo an examination when necessary**, and wait for the patient's feedback. - **Finally, give the diagnosis result, corresponding diagnostic basis and treatment plan based on the patient's physical condition and examination results**. - **The diagnosis result needs to be accurate to the specific disease**.

[0085] The fourth agent (which may be referred to as Agent 4) is used to perform the task of optimizing the feedback of the evaluation results. The task content includes evaluating the prediction results of the task problem sent by the third agent and returning the evaluation results to the third agent; when the evaluation result of the first prediction result of any task problem sent by the third agent indicates that the evaluation fails, the guidance experience generated based on the task problem, the first prediction result of the task problem and the correct answer is returned to the third agent.

[0086] The fourth agent guides the third agent to conduct model training by evaluating the clinical scenario prediction results generated by the third agent, thereby achieving the goal of continuous optimization of autonomous learning.

[0087] Among them, the fourth agent receives the reasoning result of the third agent, that is, the clinical scenario prediction result, through the prompt word, and then compares the result with the correct answer of the relevant case, and conducts multiple evaluations, such as the accuracy of the predicted time of death, the consistency of symptom prediction, etc. If the evaluation of each aspect meets the preset conditions, for example, the accuracy of the predicted time of death reaches the preset threshold and the consistency of the symptom prediction reaches the preset threshold, then the evaluation result is passed, otherwise the evaluation result is failed. If the evaluation result is passed, the fourth agent will directly return the evaluation result to the third agent, and the third agent will start the next round of iteration, that is, obtain a new clinical scenario prediction task and process the task. The relevant processing flow is as mentioned above and will not be repeated here; if the evaluation result is failed, the fourth agent needs to summarize the initial question, result, and correct answer to obtain guiding experience, and then return it and the evaluation result to the third agent, so that the third agent can answer again. Subsequently, if the evaluation result of the clinical scenario prediction result regenerated by the third agent (i.e., the last prediction result of the task problem) is passed, indicating that the indicator experience has been verified to be effective, then the guidance experience will be added to the "experience data" of the third agent; if the evaluation result of the clinical scenario prediction result regenerated by the third agent is still failed, it means that the guidance experience is invalid, and the evaluation result will be returned to the third agent to allow it to start the next round of iteration.

[0088] Furthermore, when the fourth agent evaluates the clinical scenario prediction results generated by the third agent, it needs to obtain the task information of the specific clinical scenario prediction task, formulate corresponding evaluation criteria for the task, and then design prompt words for evaluating the clinical scenario prediction results based on the evaluation criteria.

[0089] For example, assuming that the diagnosis and treatment results are divided into four levels of consistency, the prompt words can be as follows:

[0090] You are a senior medical expert.

[0091] Please judge the quality of the diagnosis and treatment results based on the auxiliary examinations and diagnostic results in the clinical scenario prediction results.

[0092] Please refer to the following rules for evaluation.

[0093] 1. Completeness of medical examination items\n(A) Very complete\n(B) Quite a part complete\n(C) A small part complete\n(D) Mostly incomplete\;

[0094] 2. Consistency of diagnostic results\n(A) Completely consistent, the diagnosis is correct\n(B) Quite partially consistent, the diagnosis is basically correct\n(C) A small part of the diagnosis is consistent, and there is an error in the diagnosis\n(D) Completely inconsistent, the diagnosis is completely wrong\n.

[0095] In some embodiments, the first agent is also used to regularly use the Search() search tool to search for literature related to acute myocardial infarction, which can provide a continuously expanding external knowledge base for the data source of the entire framework.

[0096] The documents searched by the first agent belong to external text data, mainly including clinical guidelines and academic papers, such as "2020 ESC Guidelines for the management of acute coronary syndromes in patients presenting without persistent ST-segment elevation", "2023 ESC Guidelines for the management of acute coronary syndromes", "Guidelines for the Prevention and Treatment of Hypertension in China (2018 Revised Edition)", "Guidelines for the Diagnosis and Treatment of Heart Failure in China 2018", etc. Specifically, the first agent can use the keywords related to the prediction of clinical scenarios of acute myocardial infarction specified by the doctor to search for relevant texts matching the keywords (meaning the same as the keywords or with a high similarity to the word vector of the keywords) from academic search engines or databases, and crawl qualified clinical guidelines and academic papers based on the text.

[0097] Correspondingly, the second agent is also used to construct a second medical knowledge graph based on the literature searched by the first agent. The entities in the second medical knowledge graph include medical entity words extracted by the second agent from the literature, and the relationship between the entities in the second medical knowledge graph is used to reflect whether the two entities are positively correlated, negatively correlated, or unrelated.

[0098] The second medical knowledge graph refers to a medical knowledge graph constructed based on the literature searched by the first agent. Specifically, the second agent will automatically extract medical entity words from the clinical guidelines and academic papers searched by the first agent, including diseases, symptoms, treatments, examinations and other medical knowledge. It should be noted that the scope of entity words here is not pre-specified by doctors or tasks, but the medical information contained in the text is extracted as completely as possible for knowledge graph construction and modeling. After the second agent extracts medical entity words from external text data, it needs to analyze the relationship between medical entity words. In this embodiment, the second agent only analyzes three types of relationships between medical entity words, namely: positive correlation, negative correlation, and irrelevant. The extracted medical entity words and relationship information can be stored in the Neo4j graph relational database.

[0099] In some embodiments, the third agent is also used to construct a record database and an experience database. The record database is used to store prediction records corresponding to task problems handled by the third agent, and the experience database is used to store effective guidance experience corresponding to task problems handled by the third agent.

[0100] Correspondingly, the reference knowledge retrieved by the third agent based on the target problem also includes prediction records matching the target problem retrieved from the record database and guidance experience matching the target problem retrieved from the experience database; the third agent is also used to store the problem information, planning results, prediction results and evaluation results corresponding to the target problem as a prediction record in the record database when receiving the first prediction result or the evaluation result of the last prediction result of the target problem; and, when the evaluation result of the last prediction result of the target problem indicates that the evaluation is passed, store the guidance experience of the target problem in the experience database.

[0101] In this embodiment, the data in the record database and the experience database are both long-term memories in the memory of the third agent. After the record database and the experience database are constructed, the third agent can save the information of the entire reasoning process of each round, including task problems, planning results, prediction results and evaluation results, into the record database in multiple rounds of model training, and when the guidance experience returned by the fourth agent is verified to be valid, it will be stored in the experience database; therefore, when the third agent processes a new clinical scenario prediction task in the next round, in addition to retrieving relevant data from the medical knowledge graph constructed by the second agent, it can also retrieve relevant records and experiences from the record database and experience data, and add the retrieved data to the planning stage for combined application, realizing continuous learning in a cycle of iteration.

[0102] In some embodiments, the fourth agent is also used to evaluate the clinical feature extraction results of the first agent using pre-specified evaluation indicators, and provide optimization suggestions for the first agent based on the evaluation results, and / or to evaluate the entity relationship extraction results of the second agent, and provide optimization suggestions for the second agent based on the evaluation results.

[0103] In this embodiment, the fourth agent is also used to communicate with the first agent and the second agent 2 to decide whether to optimize the extraction process of clinical features.

[0104] After the construction of the first and second agents is completed, the fourth agent receives the clinical feature extraction results and related data, and uses pre-specified evaluation indicators (such as accuracy, recall, F1 score, etc.) to evaluate the clinical feature extraction results (including the feature information of the clinical features extracted by the first agent), and then generates optimization suggestions based on the corresponding evaluation results, such as adding new clinical features, adjusting parameters, etc. The optimization suggestions are passed to the first agent through prompt words to guide the first agent to optimize the clinical feature extraction process. The fourth agent can also evaluate the entity relationship extraction results of the second agent (including the relationship between entities), evaluate the entity relationship extraction results with evaluation indicators such as accuracy, recall, F1 score, etc., and provide optimization suggestions to the second agent based on the evaluation results.

[0105] For example, the entity relationship extracted by the second agent from the external text data is:

[0106] { "entities": [ {"text": "Hypertension", "type": "Disease"}, {"text": "Heart disease", "type": "Disease"}, {"text": "Heart failure", "type": "Complication"} ], "relations": [ {"subject": "Hypertension", "object": "Heart disease", "relation": "Positive correlation"}, {"subject": "Heart disease", "object": "Heart failure", "relation": "Positive correlation"} ]}.

[0107] The fourth agent evaluated it, and the evaluation results showed that the accuracy rate was 85% and the recall rate was 75%.

[0108] The generated optimization suggestions can be: "suggestions": [ "Add new entity types, such as 'symptom' and 'check'.", ,"Adjust model parameters to improve recall.", ].

[0109] In some embodiments, the first intelligent agent is specifically used to obtain a list of medical names provided by a doctor from a pre-specified file or database, and to extract clinical features from the medical data of each acute myocardial infarction patient with reference to the list of medical names; it is also used to send a medical name confirmation request containing the uncertain medical name and context information to the doctor when an uncertain medical name is obtained from the medical data, and to determine the corresponding medical name based on the doctor's feedback information.

[0110] For the structuring of medical knowledge, the first agent needs to be able to correctly identify the names of medical entities such as diseases and drugs. In this embodiment, the first agent can obtain the list of medical names provided by the doctor from a specified file or database. The format of the file can be CSV, JSON, etc., and the database can be MySQL, etc. When the first agent extracts clinical features from the patient's medical data, it will use the list of medical names provided by the doctor as a reference, which can improve the accuracy and comprehensiveness of feature recognition. In addition, for uncertain medical names, you can mark them and request the doctor to confirm. For example, when encountering an uncertain medical name, send a request through a dialogue, and the request should contain the specific medical name and context information so that the doctor can quickly confirm.

[0111] In some embodiments, the first intelligent agent is also used to obtain medical terms and their corresponding discrimination rules from medical literature and medical standards, obtain medical terms and their corresponding discrimination rules specified by doctors, and organize the obtained medical terms and discrimination rules into a first standard file in a specific format; when extracting clinical features, the first standard file is used as a reference to determine the parsing and discrimination rules of the extracted clinical features.

[0112] When the first agent extracts the characteristic information of clinical features, it needs to accurately determine the parsing and discrimination rules of clinical features. In this embodiment, the first agent can obtain authoritative medical terms and corresponding rules from the medical literature and medical standards (such as ICD-10, SNOMED CT, etc.) involved in the prompt words, and can also cooperate with doctors to obtain medical terms and specific discrimination rules specified by doctors by sending requests to doctors. The first agent can organize the acquired medical terms and discrimination rules into a standard file (i.e., a first standard file), such as a file in CSV, JSON or XML format. Such an operation can ensure that the information in the standard file is structured and easy to read. When performing the clinical feature extraction operation, the first agent will automatically obtain the first standard file as a reference to determine the parsing and discrimination rules of clinical features.

[0113] In some embodiments, the first intelligent agent is also used to obtain medical terms and their corresponding clinical scenario association identifiers from medical literature and medical standards, obtain medical terms specified by doctors and their corresponding clinical scenario association identifiers, and organize the obtained medical terms and clinical scenario association identifiers into a second standard file in a specific format; when extracting clinical features, the clinical scenario association identifiers of the extracted clinical features are determined with reference to the second standard file.

[0114] When the first agent extracts the characteristic information of the clinical feature, it needs to accurately determine the clinical scenario association identifier of the clinical feature. In this embodiment, the first agent can obtain authoritative medical terms and corresponding clinical scenario association identifiers from the medical literature and medical standards (such as ICD-10, SNOMED CT, etc.) involved in the prompt word, and can also cooperate with the doctor to obtain the medical terms and specific clinical scenario association identifiers specified by the doctor by sending a request to the doctor. The first agent can organize the acquired medical terms and clinical scenario association identifiers into a standard file (i.e., a second standard file), such as a file in CSV, JSON or XML format. Such an operation can ensure that the information in the standard file is structured and easy to read. When performing the clinical feature extraction operation, the first agent will automatically obtain the second standard file as a reference to determine the clinical scenario association identifier of the clinical feature.

[0115] Based on the same inventive concept, the present application also provides a clinical scenario prediction method for acute myocardial infarction patients during hospitalization, which can be applied to the clinical scenario prediction system for acute myocardial infarction patients during hospitalization provided in any of the above embodiments. In some embodiments, the method includes: Figure 3 The steps shown are described below.

[0116] S110: receiving a prompt word for indicating a clinical scenario for predicting a target acute myocardial infarction patient during hospitalization, and extracting a task question from the prompt word;

[0117] S120: Retrieve data matching the task problem from the constructed knowledge graph, record database, and experience database as target reference knowledge;

[0118] S130: Predict clinical scenarios based on target reference knowledge to obtain target results.

[0119] The target outcome refers to the predicted outcome of the clinical scenario corresponding to the target acute myocardial infarction patient.

[0120] This embodiment can use the clinical scenario prediction system provided in the aforementioned embodiment to predict clinical scenarios that may occur for target acute myocardial infarction patients during hospitalization. The target acute myocardial infarction patients are real patients who need to predict clinical scenarios, rather than patients in cases used for model training. The system will send prompt words of the clinical scenarios of the target acute myocardial infarction patients during hospitalization to the third intelligent agent, which will respond to the prompt words and feedback the clinical scenario prediction results to the user. It should be noted that before predicting the clinical scenarios that may occur in the future for the target acute myocardial infarction patients, the third intelligent agent has met the training end conditions of the prediction models of various clinical scenarios under the continuous feedback optimization of the fourth intelligent agent and reached the pre-set output requirements.

[0121] In some embodiments, the prompt word also includes or corresponds to the admission and medical record text of the target acute myocardial infarction patient. The third intelligent agent will perform clinical scenario prediction based on the admission and medical record text, which can further improve the accuracy of the prediction results.

[0122] With respect to the various steps included in the clinical scenario prediction method for acute myocardial infarction patients during hospitalization provided in any of the above embodiments, unless otherwise clearly stated in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least a part of these steps may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0123] It should be noted that the specific limitations on the clinical scenario prediction method for patients with acute myocardial infarction during hospitalization can also be found in the above limitations on the clinical scenario prediction system for patients with acute myocardial infarction during hospitalization, which will not be repeated here.

[0124] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0125] Those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink), DRAM (SLDRAM), memory bus (Rambus), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0126] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0127] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A clinical scenario prediction system for patients with acute myocardial infarction during hospitalization, characterized in that: The system includes multiple intelligent agents driven by a large model for performing different tasks, wherein the multiple intelligent agents communicate information and iteratively optimize through prompt words; the multiple intelligent agents include: The first agent is used to extract clinical features related to clinical scenario prediction from the medical data of multiple patients with acute myocardial infarction and analyze the relationship between the clinical features to generate structured medical knowledge text; A second agent is used to construct a first medical knowledge graph based on the medical knowledge text; the entities in the first medical knowledge graph include various clinical features extracted by the first agent, and the relationship between the entities in the first medical knowledge graph is used to reflect the properties between two entities; The third agent is used to obtain a clinical scenario prediction task, take the task problem of the clinical scenario prediction task as the target problem, retrieve data matching the target problem as reference knowledge, the reference knowledge includes data matching the target problem retrieved from the medical knowledge graph constructed by the second agent, use the clinical scenario prediction model to perform clinical scenario prediction based on the reference knowledge, and obtain the first prediction result of the target problem, the clinical scenario prediction model is a large model; when receiving the guidance experience of the target problem returned by the fourth agent, the clinical scenario prediction is performed again in combination with the reference knowledge and the guidance experience to obtain the final prediction result of the target problem; a fourth agent, for evaluating the prediction results of the task problem sent by the third agent, and returning the evaluation results to the third agent; when the evaluation result of the first prediction result of any task problem sent by the third agent indicates that the evaluation fails, returning the guidance experience generated based on the task problem, the first prediction result of the task problem and the correct answer to the third agent; The reference knowledge also includes prediction records matching the target problem retrieved from a record database and guidance experience matching the target problem retrieved from an experience database; the record database is used to store prediction records corresponding to task problems processed by the third agent, and the experience database is used to store effective guidance experience corresponding to task problems processed by the third agent; The third intelligent agent is also used to store the problem information, planning results, prediction results and evaluation results corresponding to the target problem as a prediction record in the record database when receiving the first prediction result or the evaluation result of the last prediction result of the target problem; and to store the guiding experience of the target problem in the experience database when the evaluation result of the last prediction result of the target problem indicates that the evaluation is passed.

2. The system according to claim 1, characterized in that The first agent is further configured to periodically use a search tool to search for literature related to acute myocardial infarction; The second intelligent agent is also used to construct a second medical knowledge graph based on the literature searched by the first intelligent agent; the entities in the second medical knowledge graph include medical entity words extracted by the second intelligent agent from the literature, and the relationship between the entities in the second medical knowledge graph is used to reflect whether the two entities are positively correlated, negatively correlated or unrelated.

3. The system according to claim 1, characterized in that The fourth agent is also used to evaluate the clinical feature extraction results of the first agent using pre-specified evaluation indicators, and provide optimization suggestions for the first agent based on the evaluation results, and / or to evaluate the entity relationship extraction results of the second agent, and provide optimization suggestions for the second agent based on the evaluation results.

4. The system according to claim 1, characterized in that The first agent is specifically used to obtain a list of medical names provided by a doctor from a pre-specified file or database, and extract clinical features from the medical data of each of the acute myocardial infarction patients with reference to the list of medical names; It is also used to send a medical name confirmation request containing the uncertain medical name and context information to a doctor when an uncertain medical name is obtained from the medical data, and determine the corresponding medical name based on the doctor's feedback information.

5. The system according to claim 1, wherein: The clinical characteristics related to clinical scenario prediction include symptom characteristics, disease characteristics, drug characteristics, examination characteristics and clinical intervention characteristics; the characteristic information of each clinical characteristic includes characteristic name, category, characteristic value, parsing and discrimination rules and clinical scenario association identifier; the value of the clinical scenario association identifier is used to indicate whether the relevant clinical characteristic is associated with a specific clinical scenario.

6. The system according to claim 5, characterized in that The first agent is further used to obtain medical terms and corresponding discrimination rules from medical literature and medical standards, obtain medical terms and corresponding discrimination rules specified by doctors, and organize the obtained medical terms and discrimination rules into a first standard file in a specific format; When extracting clinical features, the first standard file is used as a reference to determine the analysis and discrimination rules of the extracted clinical features.

7. The system according to claim 5, characterized in that The first agent is further used to obtain medical terms and their corresponding clinical scenario association identifiers from medical literature and medical standards, obtain medical terms specified by doctors and their corresponding clinical scenario association identifiers, and organize the obtained medical terms and clinical scenario association identifiers into a second standard file in a specific format; When extracting clinical features, the clinical scenario association identifier of the extracted clinical features is determined with reference to the second standard file.

8. The system of claim 1, wherein: The clinical scenario prediction task is used to instruct the third agent to predict the clinical scenario of a designated acute myocardial infarction patient during hospitalization, wherein the clinical scenario includes one or more of the patient's symptoms, signs, auxiliary examinations, disease diagnosis, drug treatment plan, surgical treatment decision, nursing care, and complications of acute myocardial infarction.

9. A method for predicting clinical scenarios during hospitalization of patients with acute myocardial infarction, characterized in that: The method applies the system according to any one of claims 1 to 8, and the method comprises: receiving a prompt word for indicating a clinical scenario for predicting a target acute myocardial infarction patient during hospitalization, and extracting a task question from the prompt word; Retrieve data matching the task problem from the constructed knowledge graph, record database and experience database as target reference knowledge; A clinical scenario prediction is performed based on the target reference knowledge to obtain a target result.

Images