Multi-agent medical examination recommendation system and method

Abstract

The invention provides a multi-agent medical examination recommendation system and method, and relates to the technical field of medical intellectualization. Comprising a medical examination recommendation data set building module used for extracting key medical entities and personal information of patients from electronic medical records of an intensive care medical information database, and generating enhanced patient sequence data by using the key medical entities and a preprocessing mode; the medical knowledge graph building module is used for building a medical knowledge graph according to the disease and the triple generation module based on the large language model; and the medical examination recommendation module is used for processing the enhanced patient sequence data and personal information by utilizing a plurality of intelligent agents and a medical knowledge graph to obtain a target medical examination item. In this way, structured medical examination data can be provided, and the quality of a data set is high; multiple agents are introduced, so that the flexibility and adaptability of the recommended target medical examination items are improved; and the reasonability of the recommendation result is improved by utilizing the established medical knowledge graph.

Description

Technical Field

[0001] The present invention relates to the technical field of medical intelligence, and in particular to a multi-agent medical examination recommendation system and method. Background Art

[0002] Medical examination recommendations are a key means for health risk assessment based on patients' electronic medical records, aiming to provide personalized examination recommendations according to the unique situation of each patient, so as to ensure that the most appropriate medical examinations can be carried out in the next stage. This process not only requires medical examinations to have accurate risk prediction capabilities, but also needs to consider multi-dimensional information such as the patient's age, gender, medical history, genetic factors, etc. to achieve precise personalized recommendations. High-performance medical examination recommendation algorithms can help doctors better plan patients' examination plans by comprehensively analyzing this information, ensuring that necessary medical examinations are carried out at the most appropriate time. Such precise recommendations not only help avoid over-examination or missed diagnoses, but also improve the treatment effect and health prognosis of patients.

[0003] With the aging of China's population, the problem of shortage of medical resources has become increasingly serious, and the research on medical examination recommendation algorithms has become particularly important. First of all, medical examination recommendation algorithms can help medical institutions reasonably allocate limited medical resources, optimize the examination process, and reduce unnecessary waiting time and medical costs. Secondly, it can reduce the potential risks brought by misdiagnosis or delayed diagnosis through accurate risk prediction, and improve the efficiency and accuracy of medical services. For doctors, medical examination recommendation algorithms are not only an auxiliary decision-making tool, but also an important means to improve the diagnosis and treatment level. It can provide decision-making support based on a large amount of data analysis to help doctors make more scientific and comprehensive judgments. The research on medical examination recommendation algorithms not only helps improve the quality of medical services, but also provides more efficient health management for patients, promoting the medical industry to develop in a more precise and intelligent direction, with great application prospects and social value.

[0004] Currently, existing medical examination recommendations are based on artificial intelligence-based medical examination recommendation methods. However, most detailed medical examination information in the existing methods' databases is recorded in doctor's order text files, lacking structured medical examination data; there is a large amount of noise in the data in the database and there are incorrect entities or entity ambiguities, resulting in a low quality of the dataset; existing recommendation methods only focus on mining users' personalized needs, relying on high-quality static data-driven, and cannot fully cope with the complexity of the examination recommendation scenario, especially in a dynamically changing medical environment, lacking sufficient flexibility and adaptability; existing methods do not fully consider the problem of generation hallucinations of large language models, resulting in impaired performance of the recommendation results and lower rationality of the recommendation results. Summary of the Invention

[0005] The objective of the embodiments of the present invention is to provide a multi-agent medical examination recommendation system and method, which solve the problems in the prior art, such as the lack of structured medical examination data, the low quality of the data set, the lack of sufficient flexibility and adaptability, and the low rationality of the recommendation results.

[0006] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0007] The present invention provides a multi-agent medical examination recommendation system, including:

[0008] A medical examination recommendation data set construction module, which is used to extract the key medical entities and personal information of patients from the electronic medical records in the intensive care medicine information database, and generate enhanced patient sequence data by using the key medical entities and preprocessing methods. The enhanced patient sequence data includes diseases, symptoms, and examinations;

[0009] A medical knowledge graph construction module, which is used to construct a medical knowledge graph according to diseases and a triple generation module based on a large language model;

[0010] A medical examination recommendation module, which is used to process the enhanced patient sequence data and personal information by using multiple agents and the medical knowledge graph to obtain the target medical examination items.

[0011] The second aspect of the present invention provides a multi-agent medical examination recommendation method, including:

[0012] Extract the key medical entities and personal information of patients from the electronic medical records in the intensive care medicine information database, and generate enhanced patient sequence data by using the key medical entities and preprocessing methods. The enhanced patient sequence data includes diseases, symptoms, and examinations;

[0013] Construct a medical knowledge graph according to diseases and a triple generation module based on a large language model;

[0014] Process the enhanced patient sequence data and personal information by using multiple agents and the medical knowledge graph to obtain the target medical examination items.

[0015] Compared with the prior art, a multi-agent medical examination recommendation system and method provided by the present invention include: a medical examination recommendation dataset construction module, which is used to extract key medical entities and personal information of patients from the electronic medical records in the intensive care medicine information database, and use the key medical entities and preprocessing methods to generate enhanced patient sequence data; a medical knowledge graph construction module, which is used to construct a medical knowledge graph according to diseases and a triple generation module based on a large language model; a medical examination recommendation module, which is used to process the enhanced patient sequence data and personal information by using multiple agents and the medical knowledge graph to obtain target medical examination items. In this way, the enhanced patient sequence data and personal information can be generated through the medical examination recommendation dataset construction module, and structured medical examination data can be provided; the enhanced patient sequence data is generated through key medical entities and preprocessing methods, so that the quality of the dataset is relatively high; multiple agents are introduced, and the powerful decision-making ability of the multiple agents and the advantage of processing complex tasks other than simple tasks are utilized to significantly improve the flexibility and adaptability of the recommended target medical examination items; the constructed medical knowledge graph can make full use of the structured medical knowledge in the graph to enhance the input representation of the multiple agents, effectively guide the multiple agents to generate more accurate medical content, effectively solve the hallucination problem that may occur in the agents based on the large language model, and thus further improve the rationality of the recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown by way of illustration and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein:

[0017] Figure 1 Schematically shows the structural diagram of the multi-agent medical examination recommendation system;

[0018] Figure 2 Schematically shows the framework diagram of the medical examination recommendation dataset construction module;

[0019] Figure 3 Schematically shows the three-stage few-shot entity recognition framework diagram in the named entity recognition sub-module;

[0020] Figure 4 Schematically shows the structural diagram of the auxiliary information collection sub-module;

[0021] Figure 5 Schematically shows the structural diagram of the data correction sub-module based on the large language model;

[0022] Figure 6Schematically shows an example prompt diagram of the data correction sub-module based on the large language model;

[0023] Figure 7 Schematically shows the framework diagram of the medical knowledge graph construction module;

[0024] Figure 8 Schematically shows the flowchart of the multi-agent medical examination recommendation method;

[0025] Figure 9 Schematically shows the operation flowchart of outputting the target medical examination items. Detailed implementation manners

[0026] The following further describes in detail the implementation manners of the present invention in conjunction with the drawings and embodiments. The detailed descriptions and drawings of the following embodiments are used to exemplarily illustrate the principles of the present invention, but cannot be used to limit the scope of the present invention. The present invention can be implemented in many different forms, not limited to the specific embodiments disclosed in the text, but including all technical solutions falling within the scope of the claims.

[0027] The following details a multi-agent medical examination recommendation system and method in an embodiment of the present invention.

[0028] See Figure 1 as shown Figure 1 Schematically shows the structure diagram of the multi-agent medical examination recommendation system. An embodiment of the present invention proposes a multi-agent medical examination recommendation system, including:

[0029] A medical examination recommendation data set construction module 101, configured to extract key medical entities and personal information of patients from the electronic medical records in the intensive care medicine information database, and generate enhanced patient sequence data by using the key medical entities and preprocessing methods. The enhanced patient sequence data includes diseases, symptoms, and examinations;

[0030] A medical knowledge graph construction module 102, configured to construct a medical knowledge graph according to diseases and a triple generation module based on the large language model;

[0031] A medical examination recommendation module 103, configured to process the enhanced patient sequence data and personal information by using multiple agents and the medical knowledge graph to obtain target medical examination items.

[0032] Specifically, the multi-agent medical examination recommendation system proposed by the present invention mainly involves three parts: a medical examination recommendation data set construction module, a medical knowledge graph construction module, and a medical examination recommendation module.

[0033] Referring to the real-world diagnosis process, doctors usually monitor the changes in patients' conditions and formulate medical examination plans based on heterogeneous data such as patients' electronic medical records and personal information. At the same time, in order to better evaluate the progress of patients' conditions, more attention should be paid to the time information of patients and medical heterogeneous data in the electronic health record. Therefore, the constructed dataset has the characteristics of heterogeneity and temporality. Specifically, the medical examination recommendation dataset proposed by the present invention mainly includes enhanced patient sequence data and personal information, and has the characteristics of heterogeneity and temporality. The enhanced patient sequence data consists of three heterogeneous key medical entities: diseases, symptoms, and examinations. The personal information of patients contains two heterogeneous entities: age information and gender information. The enhanced patient sequence is arranged in chronological order according to the heterogeneous entities extracted from each patient's medical record.

[0034] In this embodiment, Figure 2 A schematic framework diagram of the medical examination recommendation dataset construction module is shown schematically. Refer to Figure 2 As shown, the medical examination recommendation dataset construction module 101 includes a named entity recognition sub-module, an auxiliary information collection sub-module, and a data correction sub-module based on a large language model;

[0035] The named entity recognition sub-module is used to extract the key medical entities of patients from the electronic medical records in the Medical Information Mark for Intensive Care III (MIMIC-Ⅲ) and sort the key medical entities in chronological order to generate patient sequence data;

[0036] The auxiliary information collection sub-module is used to extract the personal information of patients from the electronic medical records in the intensive care medicine information database;

[0037] The data correction sub-module based on the large language model is used to perform denoising, disambiguation, and standardization processing on the patient sequence data in sequence to obtain enhanced patient sequence data.

[0038] Specifically, the ChatGLM3-6B model is adopted in the data correction sub-module based on the large language model, and it can also be a more powerful general large language model or a professional medical large language model, such as ChatGPT4.0, HuatuoGPT, etc.

[0039] In this embodiment, the named entity recognition sub-module is specifically configured to train the clinical BERT model using the first labeled data to obtain a trained clinical BERT model; construct three dictionaries based on web crawling technology, where the three dictionaries contain diseases, examinations, and symptoms respectively; generate second labeled data using a matching method based on the three dictionaries; use the trained clinical BERT model, confidence evaluation method, and noise-aware loss function to perform completion and enhancement processing on the second labeled data to obtain processed second labeled data; use the first labeled data and the processed second labeled data to train the trained clinical BERT model to obtain a trained first clinical BERT model; use the first labeled data to train the trained first clinical BERT model to obtain a trained target clinical BERT model; input the electronic medical records of the intensive care medicine information database into the trained target clinical BERT model so that the trained target clinical BERT model outputs key medical entities; sort the key medical entities according to the time sequence to generate patient sequence data. The first labeled data is data obtained by adding strong labels to the original data using a labeling tool, and the second labeled data is data with weak labels.

[0040] Performing three trainings on the clinical BERT model can be called three-stage progressive training.

[0041] The named entity recognition sub-module receives the unstructured text medical records of the patient and extracts the key medical entities therein to construct patient sequence data.

[0042] Figure 3 Schematically shows the three-stage few-shot entity recognition framework diagram in the named entity recognition sub-module. See Figure 3As shown, in order to extract key entities in electronic medical records with high quality on the basis of consuming a small amount of labor costs, the present invention uses a three-stage few-shot entity recognition framework. This framework generates a large amount of weakly labeled data to enrich the existing labeled data, and adopts techniques such as a noise-aware loss function to suppress the noise in the weakly labeled data, so as to enhance the quality of the weakly labeled data, and finally achieve the goal of alleviating the few-shot problem. For this purpose, the present invention creates a few-shot entity recognition dataset for electronic medical records, namely the first labeled data. The dataset includes approximately 5,000 manually annotated entity labels, involving 170,000 words, including three types of entities: diseases, symptoms, and examinations, and creates three corresponding dictionaries. The dictionaries will be used to match and generate a large amount of weakly labeled data, namely the second labeled data, in the electronic medical record corpus. In the selection of the backbone network, the present invention considers using the Clinical BERT model (Clinical Bidirectional Encoder Representations from Transformers, ClinicalBERT). The reason is that this network uses the MIMIC-Ⅲ database for pre-training, which enables it to learn and master a large number of context dependencies in the MIMIC-Ⅲ database during the pre-training stage. In view of this, when applied to downstream tasks also built based on the MIMIC-Ⅲ database, the ClinicalBERT model can often show more excellent performance and generalization ability.

[0043] In the named entity recognition module, the present invention takes the manually annotated strongly labeled data, namely the first labeled data and the weakly labeled data, as input, uses the CilnicalBERT model as the backbone network, and through a three-stage progressive training strategy, efficiently and accurately extracts the structured key medical entities in the electronic medical records of MIMIC-Ⅲ, and makes them into patient sequence data in the form of sequence data according to the time order.

[0044] In this embodiment, the personal information includes gender information and age information. The auxiliary information collection sub-module is specifically used to obtain the patient information registration form in the electronic medical records of the intensive care medicine information database, and query the patient information registration form according to the patient's identity identification number to obtain the gender information; set multiple templates according to the context form, format and keywords corresponding to the patient; in the medical order text data corresponding to the electronic medical records of the intensive care medicine information database, use the multiple templates to extract the age information through regular expressions.

[0045] Figure 4 Schematically shows the structural schematic diagram of the auxiliary information collection sub-module, see Figure 4As shown, the auxiliary information mainly consists of the patient's personal information, especially the patient's age information and gender information. The MIMIC-III electronic medical record provides a patient information registry that includes data such as gender information, date of birth, and date of death, enabling direct query of gender information based on the patient's identity number (Identity, ID). Although the MIMIC-III electronic medical record does not directly provide structured age information, this age information is usually embedded in the medical order text data. Given the consistent pattern of age representation in the medical order text data, a template matching strategy is proposed to accurately identify and extract the specific numerical values representing the patient's age information. Specifically, by making full use of specific context forms, formats, and keywords, more than a dozen sets of templates are developed, and the patient's age information is extracted through regular expressions using these more than a dozen sets of templates.

[0046] Figure 5 Schematically shows the structural diagram of the data correction sub-module based on the large language model. See Figure 5 As shown, there is still a large amount of noise in the extracted key medical entities, including problems such as incorrect entities or entity ambiguities. To significantly enhance the quality of the dataset at a relatively low cost, the data correction sub-module based on the large language model considers introducing advanced large language models. Through the designed specific guiding prompts, the data correction sub-module based on the large language model can make full use of the medical domain expertise embedded in the large language model. Therefore, by adopting the data correction sub-module based on the large language model, that is, the organic integration of the large language model and manual intervention, efficient denoising, disambiguation, and standardization of the entities in the dataset can be achieved.

[0047] Figure 6 Schematically shows an example prompt diagram of the data correction sub-module based on the large language model. See Figure 6As shown, the present invention uses a data correction sub-module based on a large language model to correct example prompts for entity design, including entity disambiguation, entity denoising, and entity standardization. The guiding prompt in entity disambiguation is: Do CXR and Chest X-ray refer to the same medical examination? Data correction sub-module based on a large language model: Yes, CXR and Chest X-ray refer to the same medical examination. 'CXR' is an abbreviation of 'Chest X-ray', which is a common imaging examination used to generate images of the internal structures of the chest, including the heart, lungs, and blood vessels. The guiding prompt in entity denoising: Is a chest CT scan a medical examination? Data correction sub-module based on a large language model: Yes, a chest CT scan is a medical examination used to generate detailed images of the chest area, including the heart, lungs, blood vessels, and other structures. It is commonly used to diagnose diseases such as lung cancer, pneumonia, and heart disease. The guiding prompt in entity standardization: What is the medical standard term for a head MRI? Data correction sub-module based on a large language model: The medical standard terms for magnetic resonance imaging of the head are: cranial magnetic resonance imaging, magnetic resonance imaging of the brain. In different medical scenarios, these two terms can be selected and used as needed. Figure 6 It clearly shows that the data correction sub-module based on a large language model can accurately understand the intention and effectively improve the quality of the dataset.

[0048] The above-mentioned medical examination recommendation dataset construction module of the present invention can generate a high-quality medical examination recommendation dataset, which is composed of heterogeneous sequence information of 37,774 patients, namely enhanced patient sequence data and personal information. Among them, the heterogeneous sequence information fully refers to the form of the sequence recommendation dataset and is composed of three types of medical key entity information: the patient's past medical history, currently presented symptoms, and all medical examination items that have been carried out. Moreover, the types of entity information included are rich, involving a total of 1,698 disease entities, 701 examination entities, and 409 symptom entities. At the same time, the average number of interactions per patient is as many as 22.33, and the overall density of the dataset is 0.07%, effectively overcoming the data sparsity and cold start challenges commonly existing in the recommendation system, and providing rich and insightful information features for the data correction sub-module based on the large language model to accurately depict the patient's condition. In addition, the personal information part details the patient's age information and gender information. Among them, the gender information is evenly distributed without significant deviation. Particularly prominent is that the patient age span is extremely large, covering from the young age of 4 to the old age of 98. Such a broad age range fully verifies the comprehensive representativeness and universal value of the dataset. An example of the medical examination recommendation dataset is shown in Table 1 below. Table 1 shows the generated personal information and enhanced patient sequence data. In clinical practice, physicians frequently perform repetitive examinations on patients to evaluate the patient's disease development by comparing the changes in their examination results before and after. In view of this, the present invention believes that the extraction of multiple homogeneous examination entities in the same electronic medical record has substantial utility for model learning, so they are retained. For example, the sequence data interacted by patient #4 contains multiple MRI examination instances.

[0049] Table 1 Generated personal information and enhanced patient sequence data

[0050]

[0051]

[0052] In this embodiment, the medical knowledge graph construction module 102 is specifically configured to generate various types of structured triple data for diseases by using the triple generation module based on the large language model; and use the various types of structured triple data to construct a medical knowledge graph.

[0053] In this embodiment, the various types of structured triple data include first structured triple data, second structured triple data, and third structured triple data. The first structured triple data is triple data composed of diseases, complications, and diseases. The second structured triple data is triple data composed of diseases, required examinations, and preset examinations. The third structured triple data is triple data composed of diseases, possessed symptoms, and preset symptoms.

[0054] Triple data can comprehensively describe the complex relationships between diseases and related medical entities, providing rich knowledge support for medical examination recommendation systems. The medical knowledge graph contains approximately 62,000 triples and 64,000 entity nodes, involving three types of entities: diseases, symptoms, and examinations, as well as three types of relationship edges: "has complication", "has symptom", and "requires examination".

[0055] The ChatGLM3-6B model is adopted in the medical knowledge graph construction module, which can also be replaced by other more powerful general large language models or professional medical large language models, such as ChatGPT4.0, HuatuoGPT, etc.

[0056] By integrating and structuring medical knowledge, the medical knowledge graph supports intelligent decision-making, personalized treatment, and interdisciplinary collaboration, promoting the development of precision medicine and medical intelligence. In addition, as an effective knowledge base, when combined with retrieval-augmented generation technology, a high-quality medical knowledge graph can effectively reduce the risk of hallucination problems in large language models. Therefore, the present invention proposes a medical knowledge graph construction module. Figure 7 Schematically shows a framework diagram of the medical knowledge graph construction module. See Figure 7 As shown, the medical knowledge graph construction module utilizes the medical knowledge embedded in the large language model. By designing appropriate prompt engineering and adopting the method of head entity plus relationship edge, the large language model directly generates the tail entity, thus forming triples. Finally, combined with manual correction, a high-quality large-scale medical knowledge graph is constructed.

[0057] See Figure 7 As shown, in the triple generation module based on the large language model, with the disease entity as the root node, by combining the root node with predefined relationship edges (such as "complication", "required examination", "has symptom"), structured prompts are designed. The large language model automatically infers and generates the corresponding tail nodes according to these prompts, thus directly generating complete triple data. For example, for the disease "coronary heart disease", the large language model can generate the tail node "angina pectoris" according to the prompt "What complications does coronary heart disease have?", forming the triple (coronary heart disease, complication, angina pectoris); or generate the tail node "electrocardiogram" according to the prompt "What examinations can be used to diagnose coronary heart disease?", forming the triple data (coronary heart disease, required examination, electrocardiogram). In this way, a large-scale and high-quality medical knowledge graph can be generated efficiently. In addition, the entity types involved in the constructed medical knowledge graph are consistent with the entity types of the enhanced patient sequence data in the medical examination recommendation data, including diseases, examinations, and symptoms. This consistency enables the medical knowledge graph to effectively provide context information, enhancing the reasoning ability and personalized recommendation effect of the recommendation system.

[0058] In this embodiment, the multiple agents include a manager agent, a patient analyst agent, a searcher agent, and a reflector agent. The manager agent interacts with the patient analyst agent, the searcher agent, and the reflector agent respectively.

[0059] In this embodiment, the medical examination recommendation module 103 is specifically configured to input the enhanced patient sequence data and personal information into the manager agent, so that the manager agent analyzes the enhanced patient sequence data and personal information and outputs initial medical recommendation items; use the manager agent to send the enhanced patient sequence data and personal information to the patient analyst agent; use the patient analyst agent to adopt a retrieval-enhanced generation strategy according to the enhanced patient sequence data, personal information, and medical knowledge graph, so that the patient analyst agent generates an analysis result of the patient's health status and feeds back the health status analysis result to the manager agent; use the manager agent to adjust the initial medical recommendation items according to the health status analysis result, generate the first medical recommendation items, and send the first medical recommendation items to the searcher agent; use the searcher agent to adopt a retrieval-enhanced generation strategy according to the medical knowledge graph and the first medical recommendation items, so that the searcher agent generates the introduction information of the first medical recommendation items and sends the introduction information of the first medical recommendation items to the manager agent; use the manager agent to output the preset medical examination items according to the introduction information of the first medical recommendation items; use the reflector agent to judge the rationality of the preset medical examination items, and when the preset medical examination items are reasonable, make the reflector agent output the target medical examination items.

[0060] The multiple agents in the medical examination recommendation module have strong decision-making and complex task processing capabilities. Each agent selects ChatGPT3.5-turbo, and can also be replaced with other more powerful general large language models, such as ChatGPT4.0.

[0061] Specifically, the manager agent: During the task it undertakes, the manager agent can decompose the main task and allocate it to other agents (i.e., the patient analyst agent, the searcher agent, and the reflector agent), and at the same time lead the execution process of the entire task. The manager agent can monitor the collaboration and coordination among agents and plays the role of the central brain. The manager agent always alternately performs the two operations of thinking and acting. Specifically, in the thinking stage, the manager agent evaluates the status of the medical examination recommendation task, including but not limited to analyzing whether the current data (i.e., patient input data or results feedback from other agents) is complete or whether other information outside the patient input data needs to be further obtained to evaluate the patient's health status; in addition, when performing the action operation, the manager agent decides whether to give a solution to terminate the task process or request the assistance of other agents through defined interaction prompts based on the evaluation results.

[0062] The patient analyst agent: The role of the patient analyst agent is to deeply explore and personally evaluate the patient's health risk status. To achieve this goal, the patient analyst uses its own medical reserve knowledge to deeply analyze the patient's health status and achieve a comprehensive assessment of the patient's disease risk.

[0063] The searcher agent: The task of the searcher agent is to perform information retrieval using corresponding search tools under the specific requirements proposed by the manager and feedback the results to the manager agent in the form of a text summary. For example, the manager agent can let the searcher agent collect relevant information on medical items such as magnetic resonance imaging.

[0064] The reflector agent: The main responsibility of the reflector agent is to verify the correctness of the answers submitted by the manager agent. If the reflector agent determines that the answers given by the manager agent are reasonable, then the manager agent will stop further processing of the task. On the contrary, if there is room for improvement, the reflector agent will clearly point out the deficiencies of the manager in which aspects.

[0065] By designing reasonable prompt engineering and combining the collaborative cooperation of the four types of agents, it is possible to realize personalized and reasonable and effective medical examination item recommendations for patients.

[0066] Traditional techniques such as entity denoising and entity disambiguation have high computational costs, and some supervised training models also require the construction of relevant data sets, further increasing the labor cost. However, through the data correction sub-module based on the large language model, the present invention adopts an advanced large language model and designs precise guiding prompts, makes full use of the medical domain expertise embedded in the large language model, directly performs entity denoising, disambiguation, and regularization processing, and significantly improves the quality of the data set at the cost of a small amount.

[0067] The present invention fully takes into account the complexity and high reliability requirements of the medical examination recommendation task, and ingeniously introduces multi-agent technology. By leveraging its powerful decision-making ability and the advantage of handling complex tasks, the flexibility, intelligence level, and decision-making accuracy of the recommendation algorithm are significantly improved. At the same time, the present invention constructs a high-quality medical knowledge graph, combines retrieval-enhanced generation technology, and makes full use of the structured medical knowledge in the graph, effectively solving the hallucination problem that may occur in agents based on large language models, thereby further improving the rationality of the recommendation results.

[0068] Based on the above Figure 1 As can be seen from the above implementation manner, the multi-agent medical examination recommendation system in the embodiment of the present invention includes a medical examination recommendation dataset construction module, which is used to extract the key medical entities and personal information of patients from the electronic medical records in the intensive care medicine information database, and generate enhanced patient sequence data by using the key medical entities and preprocessing methods; a medical knowledge graph construction module, which is used to construct a medical knowledge graph according to diseases and a triple generation module based on a large language model; a medical examination recommendation module, which is used to process the enhanced patient sequence data and personal information by using multiple agents and the medical knowledge graph to obtain the target medical examination items. In this way, the enhanced patient sequence data and personal information can be generated through the medical examination recommendation dataset construction module, and structured medical examination data can be provided; the enhanced patient sequence data is generated by key medical entities and preprocessing methods, making the quality of the dataset relatively high; multiple agents are introduced, and by leveraging the powerful decision-making ability of multiple agents and the advantage of handling complex tasks other than simple tasks, the flexibility and adaptability of the recommended target medical examination items are significantly improved; the constructed medical knowledge graph can make full use of the structured medical knowledge in the graph to enhance the input representation of multiple agents, effectively guiding multiple agents to generate more accurate medical content, and effectively solving the hallucination problem that may occur in agents based on large language models, thereby further improving the rationality of the recommendation results.

[0069] Based on the same inventive concept, as an implementation of the above multi-agent medical examination recommendation system, the embodiment of the present invention also provides a multi-agent medical examination recommendation method. Figure 8 The flowchart of the multi-agent medical examination recommendation method in the embodiment of the present invention is shown in Figure 8 As shown, the multi-agent medical examination recommendation method may include:

[0070] S801. Extract the key medical entities and personal information of the patient from the electronic medical records in the intensive care medicine information database, and generate enhanced patient sequence data by using the key medical entities and preprocessing methods.

[0071] Among them, the enhanced patient sequence data includes diseases, symptoms, and examinations.

[0072] Specifically, step S801 includes:

[0073] Step A1: In the electronic medical records of the intensive care medicine information database, extract the key medical entities of the patient, and sort the key medical entities according to the time sequence to generate patient sequence data.

[0074] Specifically, in the electronic medical records of the intensive care medicine information database, extract the key medical entities of the patient, and sort the key medical entities according to the time sequence to generate patient sequence data, including:

[0075] Step A11: Use the first label data to train the clinical BERT model to obtain a trained clinical BERT model.

[0076] Among them, the first label data is the data obtained by adding strong labels to the original data through a labeling tool.

[0077] Step A12: Construct three dictionaries according to the web crawler technology.

[0078] Among them, the three dictionaries respectively contain diseases, examinations, and symptoms.

[0079] Step A13: Based on the three dictionaries, use the matching method to generate the second label data.

[0080] Among them, the second label data is the data with weak labels.

[0081] Step A14: Use the trained clinical BERT model, the confidence evaluation method, and the noise-aware loss function to perform complementation and enhancement processing on the second label data to obtain the processed second label data.

[0082] Step A15: Use the first label data and the processed second label data to train the trained clinical BERT model to obtain a trained first clinical BERT model.

[0083] Step A16: Use the first label data to train the trained first clinical BERT model to obtain a trained target clinical BERT model.

[0084] Step A17: Input the electronic medical records of the intensive care medicine information database into the trained target clinical BERT model so that the trained target clinical BERT model outputs the key medical entities.

[0085] Step A18: Sort the key medical entities according to the time sequence to generate patient sequence data.

[0086] Step A2: Extract the personal information of the patient from the electronic medical records in the intensive care medicine information database.

[0087] Among them, the personal information includes gender information and age information.

[0088] Specifically, Step A2 includes:

[0089] Step A21: In the electronic medical records of the intensive care medicine information database, obtain the patient information registration form, and query the patient information registration form according to the patient's identity identification number to obtain the gender information.

[0090] Step A22: Set multiple templates according to the context form, format, and keywords corresponding to the patient.

[0091] Step A23: In the medical order text data corresponding to the electronic medical records in the intensive care medicine information database, use multiple templates to extract the age information through regular expressions.

[0092] Step A3: Perform denoising, disambiguation, and standardization processing on the patient sequence data in turn to obtain enhanced patient sequence data.

[0093] S802. Build a medical knowledge graph according to the disease and the triple generation module based on the large language model.

[0094] Specifically, Step S802 includes:

[0095] Step B1: For the disease, use the triple generation module based on the large language model to generate various types of structured triple data.

[0096] Among them, various types of structured triple data include the first structured triple data, the second structured triple data, and the third structured triple data. The first structured triple data is the triple data composed of the disease, complications, and the disease. The second structured triple data is the triple data composed of the disease, required examinations, and preset examinations. The third structured triple data is the triple data composed of the disease, possessed symptoms, and preset symptoms.

[0097] Step B2: Use various types of structured triple data to construct a medical knowledge graph.

[0098] S803. Use multiple agents and the medical knowledge graph to process the enhanced patient sequence data and personal information to obtain the target medical examination items.

[0099] Among them, multiple agents include a manager agent, a patient analyst agent, a searcher agent, and a reflector agent. The manager agent interacts with the patient analyst agent, the searcher agent, and the reflector agent respectively.

[0100] Specifically, Figure 9 schematically shows an operation flowchart for outputting target medical examination items. Refer to Figure 9 as shown, step S803 includes:

[0101] S8031: Input the enhanced patient sequence data and personal information into the manager agent, so that the manager agent analyzes the enhanced patient sequence data and personal information and outputs initial medical recommendation items.

[0102] S8032: Use the manager agent to send the enhanced patient sequence data and personal information to the patient analyst agent.

[0103] S8033: Use the patient analyst agent, according to the enhanced patient sequence data, personal information, and medical knowledge graph, adopt a retrieval-enhanced generation strategy, so that the patient analyst agent generates an analysis result of the patient's health status and feeds back the analysis result of the health status to the manager agent.

[0104] S8034: Use the manager agent to adjust the initial medical recommendation items according to the analysis result of the health status, generate the first medical recommendation items, and send the first medical recommendation items to the searcher agent.

[0105] S8035: Use the searcher agent, according to the medical knowledge graph and the first medical recommendation items, adopt a retrieval-enhanced generation strategy, so that the searcher agent generates introduction information of the first medical recommendation items and sends the introduction information of the first medical recommendation items to the manager agent.

[0106] S8036: Use the manager agent to output preset medical examination items according to the introduction information of the first medical recommendation items.

[0107] S8037: Use the reflector agent to judge the rationality of the preset medical examination items. When the preset medical examination items are reasonable, make the reflector agent output the target medical examination items.

[0108] When the preset medical examination items are unreasonable, return to step S8031 to continue execution until the preset medical examination items are reasonable and stop the loop operation, so that the reflector agent outputs the target medical examination items.

[0109] It should be noted here that the description of the above embodiments of the multi-agent medical examination recommendation method is similar to the description of the above embodiments of the multi-agent medical examination recommendation system, and has beneficial effects similar to those of the embodiments of the multi-agent medical examination recommendation system. For the technical details not disclosed in the embodiments of the multi-agent medical examination recommendation method of the present invention, please refer to the description of the embodiments of the multi-agent medical examination recommendation system of the present invention for understanding.

[0110] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A multi-agent medical examination recommendation system, characterized in that: include: A medical examination recommendation data set building module is used to extract key medical entities and personal information of patients in the electronic medical records of the critical care medical information database, and generate enhanced patient sequence data using the key medical entities and preprocessing methods, wherein the enhanced patient sequence data includes diseases, symptoms and examinations; A medical knowledge graph building module, used to build a medical knowledge graph according to the disease and a triple generation module based on a large language model; The medical examination recommendation module is used to utilize multiple intelligent agents and the medical knowledge graph to process the enhanced patient sequence data and the personal information to obtain target medical examination items.

2. The multi-agent medical examination recommendation system according to claim 1, characterized in that: The medical examination recommendation data set building module includes a named entity recognition submodule, an auxiliary information collection submodule and a data correction submodule based on a large language model; The named entity recognition submodule is used to extract the key medical entities of the patient from the electronic medical records of the critical care medical information database, and sort the key medical entities according to the chronological order to generate patient sequence data; The auxiliary information collection submodule is used to extract the patient's personal information from the electronic medical records of the intensive care medical information database; The data correction submodule based on the large language model is used to perform denoising, disambiguation and standardization processing on the patient sequence data in sequence to obtain the enhanced patient sequence data.

3. The multi-agent medical examination recommendation system according to claim 2, characterized in that: The named entity recognition submodule is specifically used to train the clinical BERT model using the first label data to obtain a trained clinical BERT model; construct three dictionaries based on the crawler technology, and the three dictionaries respectively contain the disease, the examination and the symptom; based on the three dictionaries, generate the second label data using a matching method; use the trained clinical BERT model, the confidence assessment method and the noise-aware loss function to complete and enhance the second label data to obtain the processed second label data; use the first label data and the processed second label data to train the trained clinical BERT model to obtain the trained first clinical BERT model; use the first label data to train the trained first clinical BERT model to obtain a trained target clinical BERT model; input the electronic medical records of the intensive care medical information database into the trained target clinical BERT model, so that the trained target clinical BERT model outputs the key medical entities; sort the key medical entities according to the chronological order to generate the patient sequence data, the first label data is the data that adds strong labels to the original data through the annotation tool, and the second label data is the data with weak labels.

4. The multi-agent medical examination recommendation system according to claim 2, characterized in that: The personal information includes gender information and age information. The auxiliary information collection submodule is specifically used to obtain a patient information registration form in the electronic medical record of the intensive care medical information database, and query the patient information registration form according to the patient's identity identification number to obtain the gender information; set multiple templates according to the context form, format and keywords corresponding to the patient; and use the multiple templates to extract the age information through regular expressions in the medical sequence text data corresponding to the electronic medical record of the intensive care medical information database.

5. The multi-agent medical examination recommendation system according to claim 2, characterized in that: The medical knowledge graph building module is specifically used to generate multiple types of structured triple data for the disease using the triple generation module based on the large language model; and to construct the medical knowledge graph using the multiple types of structured triple data.

6. The multi-agent medical examination recommendation system according to claim 5, characterized in that: The multiple types of structured triple data include first structured triple data, second structured triple data and third structured triple data, the first structured triple data being triple data consisting of the disease, complications and the disease, the second structured triple data being triple data consisting of the disease, required examinations and preset examinations, and the third structured triple data being triple data consisting of the disease, existing symptoms and preset symptoms.

7. The multi-agent medical examination recommendation system according to claim 1, characterized in that: The multiple agents include a manager agent, a patient analyst agent, a searcher agent and a reflector agent, and the manager agent interacts with the patient analyst agent, the searcher agent and the reflector agent respectively.

8. The multi-agent medical examination recommendation system according to claim 7, characterized in that: The medical examination recommendation module is specifically used to input the enhanced patient sequence data and the personal information into the manager agent, so that the manager agent analyzes the enhanced patient sequence data and the personal information and outputs an initial medical recommendation item; The manager agent is used to send the enhanced patient sequence data and the personal information to the patient analyst agent; the patient analyst agent is used to adopt a retrieval enhancement generation strategy based on the enhanced patient sequence data, the personal information and the medical knowledge graph, so that the patient analyst agent generates a health status analysis result of the patient, and feeds the health status analysis result back to the manager agent; Utilizing the manager agent, the initial medical recommendation item is adjusted according to the health status analysis result to generate a first medical recommendation item, and the first medical recommendation item is sent to the searcher agent; utilizing the searcher agent, based on the medical knowledge graph and the first medical recommendation item, a retrieval enhancement generation strategy is adopted to enable the searcher agent to generate introduction information of the first medical recommendation item, and send the introduction information of the first medical recommendation item to the manager agent; utilizing the manager agent, based on the introduction information of the first medical recommendation item, the manager agent is enabled to output preset medical examination items; utilizing the reflector agent, the rationality of the preset medical examination items is judged, and when the preset medical examination items are reasonable, the reflector agent is enabled to output the target medical examination items.

9. A multi-agent medical examination recommendation method, characterized in that: include: Extracting key medical entities and personal information of a patient from an electronic medical record of an intensive care medical information database, and generating enhanced patient sequence data using the key medical entities and a preprocessing method, wherein the enhanced patient sequence data includes diseases, symptoms, and examinations; Build a medical knowledge graph based on the diseases and a triple generation module based on a large language model; The enhanced patient sequence data and the personal information are processed using multiple intelligent agents and the medical knowledge graph to obtain target medical examination items.

10. The multi-agent medical examination recommendation method according to claim 9, characterized in that: Extracting key medical entities and personal information of patients from the electronic medical records of the intensive care medical information database, and generating enhanced patient sequence data using the key medical entities and preprocessing methods, including: Extracting the key medical entities of the patient from the electronic medical records of the critical care medical information database, and sorting the key medical entities according to the chronological order to generate patient sequence data; Extracting the patient's personal information from the electronic medical record of the intensive care medicine information database; The patient sequence data is subjected to denoising, disambiguation and standardization processing in sequence to obtain the enhanced patient sequence data.

Images