Intelligent triage diagnosis method, system, device, storage medium and product

Abstract

This application discloses an intelligent triage method, system, device, storage medium, and product. The method includes: obtaining the patient's chief complaint; responding to the patient's chief complaint, retrieving the patient's consultation process and patient information to be collected from a medical consultation knowledge base based on a first retrieval enhancer; the medical consultation knowledge base is constructed based on a large language model and the latest version of medical data; conducting multiple rounds of consultation with the patient based on a trained consultation model, the patient's consultation process, and the patient information to be collected, generating consultation information; and reasoning about the patient's condition based on the consultation information and thought chain technology, generating a reasoning process and a reasoning result; wherein the reasoning process includes the patient's condition reasoning process and the patient's department reasoning process; the reasoning result includes the patient's condition analysis result and the first target department recommended to the patient. This improves the accuracy, real-time performance, and interpretability of the intelligent triage process.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to an intelligent triage method, system, device, storage medium, and product. Background Technology

[0002] Currently, intelligent triage refers to a technology that, based on a patient's chief complaint / condition, collects information through multiple rounds of interaction with the patient and ultimately recommends a specific department for treatment. This technology can be applied to online triage scenarios in hospitals to help patients seek medical care efficiently.

[0003] However, the current intelligent triage methods have certain limitations. (1) Deep learning-based intelligent triage technology uses deep learning to extract patients' medical information, but it is limited by the capabilities of deep learning models and cannot effectively handle ambiguous or polysemous descriptions of medical conditions. At the same time, as it is a black box model, the decision-making process of its recommendation results may be questioned by doctors and patients. (2) Retrieval-based intelligent triage technology relies on medical knowledge graphs to match patients' medical conditions and uses interactive consultations to guide patients in order to improve matching accuracy. However, this approach lacks consideration of comprehensive medical information such as patients' lifestyle habits and family medical history, which may lead to a decrease in the accuracy of recommendation results; secondly, because the knowledge graph updates slowly, it is difficult to reflect the latest medical information and changes in doctors' areas of expertise in real time, resulting in a lag in recommendation results. Summary of the Invention

[0004] In view of this, embodiments of this application provide an intelligent triage method, system, device, storage medium, and product, aiming to solve the problems of insufficient accuracy and real-time performance of intelligent triage results and weak interpretability of the intelligent triage process.

[0005] The technical solution of this application embodiment is implemented as follows:

[0006] In a first aspect, embodiments of this application provide an intelligent triage method, the method comprising:

[0007] Obtain the patient's chief complaint;

[0008] In response to the patient's chief complaint, the patient's consultation process and patient information to be collected are retrieved from the medical consultation knowledge base based on the first retrieval enhancer. The medical consultation knowledge base is constructed based on a large language model and the latest version of medical data.

[0009] Based on the trained consultation model, the patient consultation process, and the patient information to be collected, the patient is consulted in multiple rounds to generate consultation information.

[0010] Based on the consultation information and the thought chain technology, the patient's condition is inferred, generating a reasoning process and a reasoning result; wherein, the reasoning process includes the reasoning process of the patient's condition and the reasoning process of the patient's department; the reasoning result includes the analysis result of the patient's condition and the first target department recommended for the patient.

[0011] In some embodiments, the method further includes:

[0012] Receive the list of departments at the target hospital;

[0013] Based on the patient's condition analysis results, the first target department, and the second search enhancer, a second target department matching the first target department is retrieved from the department list of the target hospital.

[0014] In some embodiments, the method further includes:

[0015] Obtain custom input information from the hospital;

[0016] Based on the large language model, the hospital's custom input information, and the latest version of medical data, the medical consultation knowledge base is constructed.

[0017] In some embodiments, the latest version of the medical data includes the latest version of medical textbooks, and the method further includes:

[0018] In response to the version change notification information of the medical textbook, update the medical consultation knowledge base.

[0019] In some embodiments, the method further includes:

[0020] Based on the disease database, the large language model, and the pre-built patient profiling framework, a patient profile is constructed. The large language model is a general-domain open-source large language model.

[0021] Based on the patient profile, the medical consultation knowledge base, and the large language model, the consultation corpus is constructed.

[0022] The base model is trained based on the consultation corpus until the trained consultation model is obtained.

[0023] In some embodiments, constructing the consultation corpus based on the patient profile, the medical consultation knowledge base, and the large language model includes:

[0024] Extract the chief complaint from the patient profile;

[0025] Based on the patient's chief complaint in the patient profile, the consultation process in the medical consultation knowledge base, and the large language model, multi-round question-and-answer interactions between the patient and the doctor are simulated to generate question-and-answer interaction data.

[0026] Determine whether the number of question-and-answer rounds in the multi-round question-and-answer interaction has reached the target number. If so, construct the diagnostic corpus based on the question-and-answer interaction data.

[0027] In some embodiments, the step of generating question-and-answer interaction data based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and the large language model simulating multi-round question-and-answer interactions between the patient and the doctor includes:

[0028] Based on the patient profile's chief complaint and the large language model, multi-round question-and-answer interactions of the patient profile are simulated to generate the first question-and-answer data.

[0029] Based on the consultation process of the medical consultation knowledge base and the large language model, the doctor's multi-round question-and-answer interaction is simulated to generate second question-and-answer data;

[0030] The question-and-answer interaction data is generated based on the first question-and-answer data and the second question-and-answer data.

[0031] In some embodiments, training the base model based on the consultation corpus until the trained consultation model is obtained includes:

[0032] The base model is pre-trained based on the consultation corpus to generate a pre-trained consultation model.

[0033] The pre-trained consultation model is fine-tuned to generate the trained consultation model.

[0034] In some embodiments, the step of reasoning about the patient's condition based on the consultation information and the thought chain technology, and generating a reasoning process and a reasoning result, includes:

[0035] Based on the thinking chain technology, the reasoning of the patient's condition is decomposed into multiple sub-tasks, which include at least: symptom matching task, medical history association task, and department priority ranking task.

[0036] Based on the consultation information and the multiple sub-tasks, step-by-step reasoning is performed to generate the reasoning process and reasoning results.

[0037] Secondly, embodiments of this application provide an intelligent triage system applied to the intelligent triage method described in the first aspect above, the system comprising:

[0038] The acquisition module is used to acquire the patient's chief complaint;

[0039] The first retrieval enhancer, connected to the acquisition module, is used to retrieve the patient's consultation process and patient information to be collected from the medical consultation knowledge base based on the patient's chief complaint; and to construct the medical consultation knowledge base based on the large language model and the latest version of medical data.

[0040] The consultation module is connected to the first retrieval enhancer. The consultation module is used to conduct multiple rounds of consultation with the patient based on the trained consultation model, the patient consultation process and the patient information to be collected, and generate consultation information that represents the patient's condition information.

[0041] The first triage and recommendation module is connected to the consultation module and is used to reason about the patient's condition based on the consultation information and the thinking chain technology, and generate a reasoning process and a reasoning result; wherein, the reasoning process includes the reasoning process of the patient's condition and the reasoning process of the patient's department; the reasoning result includes the analysis result of the patient's condition and the first target department recommended for the patient.

[0042] In some embodiments, the system further includes:

[0043] The second triage recommendation module is connected to both the hospital database and the first triage recommendation module. It receives the department list of the target hospital and, based on the patient's condition analysis results, the first target department, and the second search enhancer, retrieves a second target department from the department list of the target hospital that matches the first target department.

[0044] In some embodiments, the first retrieval enhancer includes a custom input information module for obtaining custom input information from the hospital;

[0045] The first retrieval enhancer is used to construct the medical consultation knowledge base based on the large language model, the hospital's custom input information, and the latest version of medical data.

[0046] In some embodiments, the first search enhancer further includes an update module for updating the medical consultation knowledge base in response to version change indication information of the medical textbook.

[0047] In some embodiments, the acquisition module includes:

[0048] A touch screen, connected to the first retrieval enhancer, is used to obtain the patient's chief complaint.

[0049] In some embodiments, the second triage recommendation module further includes a memory for storing a list of departments of the target hospital.

[0050] In some embodiments, the consultation module includes:

[0051] The consultation corpus construction module, connected to the consultation model training module, includes a disease database, a consultation knowledge base, and a patient profile construction module. The patient profile construction module is connected to the disease database and is used to construct patient profiles based on the disease database, the large language model, and a pre-built patient profile framework. The large language model is a general-domain open-source large language model. The consultation module is also used to construct consultation corpora based on the large language model, the patient profiles, and the consultation knowledge base.

[0052] The consultation model training module is used to train the base model based on the consultation corpus until the trained consultation model is generated.

[0053] In some embodiments, the consultation model training module further includes:

[0054] The pre-training module, connected to the consultation corpus construction module, is used to pre-train the base model based on the consultation corpus to generate a pre-trained consultation model.

[0055] The fine-tuning module, connected to the pre-training module, is used to fine-tune the pre-trained consultation model to generate the trained consultation model.

[0056] In some embodiments, the medical history corpus construction module further includes:

[0057] The extraction module, connected to the patient profile construction module, is used to extract the chief complaint from the patient profile;

[0058] A multi-role simulation question-and-answer module, connected to the extraction module, is used to simulate multi-round question-and-answer interactions between the patient and the doctor based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and the large language model, and generate question-and-answer interaction data.

[0059] The round control module, connected to the multi-role simulated question-and-answer module, is used to determine whether the number of question-and-answer rounds of the multi-round question-and-answer interaction has reached the target number. If so, the diagnostic corpus is constructed based on the question-and-answer interaction data.

[0060] In some embodiments, the multi-role simulated question-and-answer module includes:

[0061] The patient simulation question-and-answer module, connected to the extraction module, is used to simulate multi-round question-and-answer interactions of the patient profile based on the chief complaint of the patient profile and the large language model, and generate the first question-and-answer data.

[0062] The doctor simulation question-and-answer module is connected to the medical consultation knowledge base and is used to simulate the doctor's multi-round question-and-answer interaction based on the consultation process of the medical consultation knowledge base and the large language model to generate second question-and-answer data.

[0063] The generation module is used to generate the question-and-answer interaction data based on the first question-and-answer data and the second question-and-answer data.

[0064] In some embodiments, the round control module further includes a counting module for calculating the number of question-and-answer rounds.

[0065] In some embodiments, the first triage recommendation module includes:

[0066] The reasoning module, connected to the consultation module, is used to decompose the patient's condition reasoning into multiple sub-tasks based on the thought chain technology. These sub-tasks include at least: a symptom matching task, a medical history association task, and a department priority ranking task. Based on the consultation information and the multiple sub-tasks, step-by-step reasoning is performed to generate a reasoning process and a reasoning result. The reasoning process includes the patient's condition reasoning process and the patient's department reasoning process. The reasoning result includes the patient's condition analysis result and a recommended first target department for the patient.

[0067] The reasoning process visualization module is used to display the reasoning process of the patient's condition and the reasoning process of the patient's department.

[0068] In some embodiments, the inference module further includes:

[0069] The decomposition module, connected to the consultation module, is used to decompose the patient's condition into multiple sub-tasks based on the thinking chain technology.

[0070] Thirdly, embodiments of this application provide an intelligent triage device, including: a processor and a memory for storing a computer program capable of running on the processor, wherein the processor, when running the computer program, performs the steps of the method described in the first aspect above.

[0071] Fourthly, embodiments of this application provide a computer storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect above.

[0072] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by an intelligent triage device, implements the steps of the method described in the first aspect above.

[0073] This application provides an intelligent triage method, which includes: obtaining the patient's chief complaint; responding to the patient's chief complaint, retrieving the patient's consultation process and patient information to be collected from a medical consultation knowledge base based on a first retrieval enhancer, wherein the medical consultation knowledge base is constructed based on a large language model and the latest version of medical data; conducting multiple rounds of consultation with the patient based on a trained consultation model, the patient's consultation process, and the patient information to be collected, generating consultation information; and reasoning about the patient's condition based on the consultation information and thought chain technology, generating a reasoning process and a reasoning result; wherein the reasoning process includes the patient's condition reasoning process and the patient's department reasoning process; and the reasoning result includes the patient's condition analysis result and the first target department recommended to the patient.

[0074] Thus, this application embodiment improves the accuracy, interpretability, and real-time performance of intelligent triage results by integrating RAG technology, large language models, and thought chain technology. (1) In terms of accuracy, by combining Retrieval Enhanced Generation (RAG) technology and large language models, it is possible to collect and understand patients' medical information more comprehensively, thereby improving the accuracy of triage. Compared with traditional methods based on deep learning or retrieval, this application can handle more complex medical descriptions and reduce the possibility of misdiagnosis. (2) In terms of interpretability, by introducing thought chain technology, the reasoning process of triage is made more transparent. In this way, not only can recommended departments be given, but the reasoning process can also be shown in detail, which helps to enhance the trust of doctors and patients in the system's recommended results. (3) In terms of real-time performance, this application can retrieve the latest medical consultation standards in real time by utilizing RAG technology, ensuring the timeliness of the recommended results. Compared with methods based on static knowledge graphs, this application embodiment can adapt to the updates of medical knowledge more quickly and provide more timely triage suggestions. Attached Figure Description

[0075] Figure 1 A flowchart illustrating the intelligent triage method provided in this application embodiment;

[0076] Figure 2 This is a schematic diagram of the intelligent triage system provided in the embodiments of this application;

[0077] Figure 3 A schematic diagram of the workflow of the intelligent triage system provided as an application example of this application;

[0078] Figure 4 This is a schematic diagram of the structure of the intelligent triage and guidance device provided in the embodiment of this application. Detailed Implementation

[0079] The present application will now be described in further detail with reference to the accompanying drawings and embodiments.

[0080] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0081] This application provides an intelligent triage method, such as... Figure 1 As shown, the method includes:

[0082] Step 110: Obtain the patient's chief complaint.

[0083] In this embodiment, intelligent triage refers to collecting information through multiple rounds of interaction based on the patient's chief complaint or condition, and ultimately recommending a suitable department for treatment. This intelligent triage method can be applied to online triage scenarios in hospitals.

[0084] In this embodiment, multiple information entry methods can be supported, such as patient personal information entry and QR code scanning, to meet the needs of different medical treatment scenarios and thereby obtain the patient's chief complaint. For example, users can enter information through a touch screen display.

[0085] In this embodiment, the patient's chief complaint refers to the description of symptoms or problems provided by the patient. For example, in a real-world scenario, a patient would give a chief complaint based on their condition, such as "I have recently had a sore throat" or "My arm is numb, which department should I go to?"

[0086] Step 120: In response to the patient's complaint, retrieve the patient's consultation process and the patient information to be collected from the medical consultation knowledge base based on the first retrieval enhancer. The medical consultation knowledge base is built based on a large language model and the latest version of medical data.

[0087] In this embodiment, the first retrieval enhancer is a RAG (Retrieval Augmented Generation) retrieval tool, used to retrieve the patient's consultation process and the patient information to be collected from the medical consultation knowledge base. For example, RAG technology can be used to search and organize the consultation process, information to be collected, etc., from the medical consultation knowledge base.

[0088] In this embodiment, the latest version of medical data includes the latest versions of medical textbooks and works such as Internal Medicine, Surgery, Obstetrics and Gynecology, Pathology, and Pediatrics.

[0089] In this embodiment, the large language model refers to a large language model (LLM), also known simply as a large language model. It is a language model with a large parameter scale, designed to understand and generate human language. Trained on a large amount of text data, it can perform a wide range of tasks, including text summarization, translation, and sentiment analysis. The medical consultation knowledge base is built based on the large language model and the latest version of medical data.

[0090] Thus, this embodiment utilizes RAG technology, combined with patient complaint information, to retrieve the latest medical consultation guidelines from the medical consultation knowledge base, and summarizes the patient information to be collected, thereby enhancing the completeness and relevance of information collection.

[0091] Step 130: Based on the trained consultation model, patient consultation process and patient information to be collected, conduct multiple rounds of consultations with the patient to generate consultation information.

[0092] In this embodiment, to obtain complete consultation information, multiple rounds of interactive consultation are required to collect in-depth details of the patient's condition. Based on a trained consultation model, patient consultation process, and patient information to be collected, this embodiment conducts multiple rounds of consultation with the patient to generate consultation information that represents the patient's condition.

[0093] In this way, based on the trained consultation model, the retrieved patient consultation process, and the patient information to be collected, multiple rounds of consultation are conducted with the patient. In each round of consultation, the questions are dynamically adjusted based on the patient's answers, gradually guiding the patient to provide more comprehensive and accurate information about their condition, ultimately generating consultation information that fully characterizes the patient's condition.

[0094] It should be noted that the mind chain technique will not be used during the consultation process based on the trained consultation model, because it is too time-consuming and the consultation task is relatively simple. The mind chain technique will only be used after the consultation is completed, when providing the disease analysis and preliminary department recommendations to generate the final analysis. This is because the departmental task is relatively complex, and the mind chain technique can improve the efficiency and interpretability.

[0095] Step 140: Based on the consultation information and thinking chain technology, reason about the patient's condition and generate the reasoning process and reasoning result; the reasoning process includes the reasoning process of the patient's condition and the reasoning process of the patient's department; the reasoning result includes the analysis result of the patient's condition and the first target department recommended to the patient.

[0096] In this embodiment, after collecting complete consultation information, it is necessary to further analyze the patient's condition and determine the appropriate department to visit. Furthermore, to present the decision-making basis of intelligent triage in an intuitive way, the transparency and explainability of the diagnostic process are enhanced.

[0097] The Mind Chain technique is a technique used to improve the performance of large language models in complex reasoning tasks. It requires the model to output a series of progressive reasoning steps as a reasoning path before giving the final answer, and the final answer is the output sequence.

[0098] This embodiment utilizes thought chain technology to visualize the reasoning process of the aforementioned disease analysis and department recommendation, generating a detailed disease reasoning process that demonstrates how to derive the disease analysis results from the consultation information. Simultaneously, it generates the patient's department reasoning process, explaining the rationale and basis for recommending the target department. These reasoning processes are then presented to the user through an appropriate display interface, allowing the user to understand and verify the results of the intelligent triage system.

[0099] Thus, this application embodiment improves the accuracy, interpretability, and real-time performance of intelligent triage results by integrating RAG technology, large language models, and thought chain technology. (1) In terms of accuracy, by combining Retrieval Enhanced Generation (RAG) technology and large language models, it is possible to collect and understand patients' medical information more comprehensively, thereby improving the accuracy of triage. Compared with traditional methods based on deep learning or retrieval, this application can handle more complex medical descriptions and reduce the possibility of misdiagnosis. (2) In terms of interpretability, by introducing thought chain technology, the reasoning process of triage is made more transparent. In this way, not only can recommended departments be given, but the reasoning process can also be shown in detail, which helps to enhance the trust of doctors and patients in the system's recommended results. (3) In terms of real-time performance, this application can retrieve the latest medical consultation standards in real time by utilizing RAG technology, ensuring the timeliness of the recommended results. Compared with methods based on static knowledge graphs, this application embodiment can adapt to the updates of medical knowledge more quickly and provide more timely triage suggestions.

[0100] In some embodiments, the method further includes:

[0101] Receive the list of departments at the target hospital;

[0102] Based on the patient's condition analysis results, the first target department, and the second search enhancer, the second target department matching the first target department is retrieved from the department list of the target hospital.

[0103] In this embodiment, the first target department recommended to the patient is highly flexible (i.e., not limited by a fixed list of departments, but mainly relying on the model's generalization ability). Therefore, to further optimize the user experience and provide targeted department recommendation results, this embodiment receives a list of departments from the target hospital. Furthermore, combining the patient's condition analysis results, the first target department, and the second search enhancer, the list of departments from the target hospital is searched to retrieve a second target department that matches the first target department.

[0104] In this embodiment, the hospital database stores department list information for each hospital, including but not limited to department names, department specialties, and doctor information, thereby supporting hospitals to provide personalized department lists and detailed descriptions.

[0105] For example, after receiving the list of departments of the target hospital, a second target department matching the first target department is retrieved from the list of departments of the target hospital based on the second searcher, the patient's condition analysis results, and the first target department. For instance, if the initial recommendation is to recommend the patient to see an ENT specialist (first target department), but the search results show that the ENT department in the target hospital's department list has too many patients or is not a dominant department in the target hospital, this embodiment will recommend other more suitable departments, such as general internal medicine (second target department), based on the patient's specific condition and the hospital's departmental setup.

[0106] Thus, this embodiment, when providing department recommendations, not only considers the model's generalization ability but also combines the actual department situation of the target hospital, further optimizing the user experience and ensuring the relevance and effectiveness of the recommendation results.

[0107] In some embodiments, the method further includes:

[0108] Obtain custom input information from the hospital;

[0109] A medical consultation knowledge base is constructed based on a large language model, custom input information from hospitals, and the latest version of medical data.

[0110] In this embodiment, in addition to the latest version of medical data, custom inputs from hospitals can also be obtained. These custom inputs include: patient information to be collected, consultation process, and other requirements. Based on the large language model, the hospital's custom input information, and the latest version of medical data, a medical consultation knowledge base is constructed, thereby generating a comprehensive medical consultation knowledge base 12.

[0111] For example, hospitals can use large-scale language models to sort and integrate custom inputs, further enriching and improving the medical consultation knowledge base built on the latest version of medical data, thereby enhancing the personalized service capabilities of triage and guidance, and better meeting the actual needs of hospitals.

[0112] In some embodiments, the latest version of medical data includes the latest version of medical textbooks, and the method further includes:

[0113] In response to the version change notices for medical textbooks, update the medical consultation knowledge base.

[0114] In this embodiment, in order to replace outdated information or add new content to ensure the accuracy of triage results, the medical consultation knowledge base needs to be updated regularly or as needed.

[0115] For example, the latest version of medical data includes the latest version of medical textbooks. In response to version change instructions for medical textbooks, the medical consultation knowledge base is updated to integrate the latest medical information into the medical consultation knowledge base to replace outdated information or supplement new content.

[0116] In some embodiments, the method further includes:

[0117] Based on a disease database, a large language model, and a pre-built patient profiling framework, patient profiles are constructed. The large language model is an open-source large language model for general domains.

[0118] Based on patient profiles, medical consultation knowledge bases, and large language models, a consultation corpus is constructed.

[0119] The base model is trained based on the consultation corpus until a well-trained consultation model is obtained. The base model is the Ernie-Speed ​​model.

[0120] In this embodiment, in order to obtain a high-quality consultation model, the consultation corpus must first be constructed. After the consultation corpus is constructed, the base model is trained based on the consultation corpus until a well-trained consultation model is obtained.

[0121] In this embodiment, the large language model includes a general-domain large language model, which has scalability and multi-task processing capabilities and is suitable for a wide range of application scenarios. It includes DeepSeek V3, DeepSeekR1, GPT-4, Qwen1.5 and Ernie-4.0 models, etc.

[0122] In this embodiment, for the medical consultation corpus, a patient profile is first constructed, and then the medical consultation corpus is constructed based on the patient profile, the medical consultation knowledge base, and the large language model. The patient profile can be constructed based on a disease database, a large language model, and a pre-built patient profile framework.

[0123] Here, the pre-built patient profile framework includes, but is not limited to, age, gender, basic symptoms, present medical history, allergy history, family history, and examination and test reports. The large language model optimizes the patient profile framework. For example, specific prompts can be designed to guide the large language model in generating patient profiles that meet the needs of actual applications.

[0124] In this embodiment, after the consultation corpus is constructed, the base model can be trained using the consultation corpus until a trained consultation model is obtained.

[0125] In this embodiment, the base model includes a lightweight large language model, which includes ChatGLM, LLaMa, RoBERTa-Lite, and Ernie-Speed ​​models, etc.

[0126] In some embodiments, a consultation corpus is constructed based on patient profiles, a medical consultation knowledge base, and a large language model, including:

[0127] Extract the chief complaint from the patient profile;

[0128] Based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and the large language model, multi-round question-and-answer interactions between patients and doctors are simulated to generate question-and-answer interaction data.

[0129] Determine whether the number of question-and-answer rounds in the multi-round question-and-answer interaction has reached the target number. If so, construct a diagnostic corpus based on the question-and-answer interaction data.

[0130] In this embodiment, in order to construct a consultation corpus for training and optimizing the consultation model, so as to improve the accuracy and efficiency of intelligent triage and guidance.

[0131] This embodiment extracts key chief complaint information from the patient profile, utilizes the patient's chief complaint, the consultation process in the medical consultation knowledge base, and a large language model to simulate a multi-round question-and-answer process between the patient and the doctor, generating question-and-answer interaction data. This process can produce rich dialogue samples, covering consultation situations under different conditions.

[0132] Simultaneously, a target number of question-and-answer rounds is determined. When the simulated multi-round question-and-answer interactions reach this number, the collected question-and-answer data is considered sufficiently comprehensive and representative. Based on the question-and-answer interaction data that meets the target number of rounds, a complete consultation corpus is constructed. This corpus can be used to train the consultation model, enabling it to better understand patient conditions and doctor consultation logic, thereby improving the performance of intelligent triage.

[0133] In some embodiments, based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and a large language model, multi-round question-and-answer interactions between the patient and the doctor are simulated to generate question-and-answer interaction data, including:

[0134] Based on the chief complaint and large language model of the patient profile, multi-round question-and-answer interaction of the patient profile is simulated to generate the first question-and-answer data;

[0135] Based on the medical consultation knowledge base, the consultation process and large language model are used to simulate the doctor's multi-round question and answer interaction and generate second question and answer data.

[0136] Based on the first and second question-and-answer data, question-and-answer interaction data is generated.

[0137] In this embodiment, the system can generate the patient's first question-and-answer data during the consultation process based on the acquired patient's chief complaint information and a large language model. For example, if the patient's chief complaint is "I've been feeling chest tightness recently," the simulated first question-and-answer data might include questions like "How long has the chest tightness lasted?" and "Are there any other symptoms?" By simulating the patient's responses, the system can generate more realistic dialogue scenarios, thereby improving the quality and diversity of the consultation corpus.

[0138] In this embodiment, based on the consultation process and standard questions in the medical consultation knowledge base and the large language model, the system guides the patient step-by-step to provide a detailed description of their condition. Each interaction generates corresponding question-and-answer data. For example, when handling a consultation task related to cardiovascular diseases, second question-and-answer data is generated based on the medical consultation knowledge base and the large language model, such as "Do you have a family history of cardiovascular disease?" and "Have you had an electrocardiogram (ECG)?". In this way, by simulating a doctor's professional consultation process, the system ensures that it adheres to medical norms and standards, generating high-quality consultation data.

[0139] In some embodiments, the base model is trained based on the consultation corpus until a trained consultation model is obtained, including:

[0140] The base model is pre-trained based on the consultation corpus to generate a pre-trained consultation model.

[0141] Fine-tune the pre-trained consultation model to generate a trained consultation model.

[0142] In this embodiment, the training process of the consultation model includes two processes: the first process is pre-training, and the second process is supervised fine-tuning.

[0143] First, the base model is pre-trained based on the patient consultation corpus to generate a pre-trained consultation model. Then, further supervised fine-tuning is performed using labeled data for specific application scenarios to fine-tune the pre-trained consultation model, generating a trained consultation model. In this way, the accuracy and adaptability of the consultation model can be improved through the two steps of pre-training and fine-tuning.

[0144] In some embodiments, based on consultation information and thought chain technology, reasoning is performed on the patient's condition to generate a reasoning process and reasoning results, including:

[0145] Based on the thinking chain technology, the reasoning of a patient's condition is decomposed into multiple sub-tasks, which include at least the following: symptom matching task, medical history association task, and department priority ranking task.

[0146] Based on the consultation information and multiple sub-tasks, step-by-step reasoning is performed to generate the reasoning process and reasoning results.

[0147] In this embodiment, the reasoning of the patient's condition is decomposed into multiple sub-tasks based on the thinking chain technology. The multiple sub-tasks include at least: symptom matching task, medical history association task, and department priority ranking task. Based on the consultation information and multiple sub-tasks, step-by-step reasoning is performed to generate and display the reasoning process of the patient's condition and the reasoning process of the patient's department.

[0148] In this embodiment, after receiving the patient's condition information, the reasoning of the patient's condition can be decomposed into multiple sub-tasks based on the thinking chain technology, such as symptom association sub-task, disease matching sub-task, and department recommendation sub-task. Based on the consultation information, these sub-tasks are processed in a certain logical order, thereby performing step-by-step reasoning on multiple sub-tasks, generating and displaying the reasoning process of the patient's condition and the reasoning process of the patient's department.

[0149] For example, the system first performs correlation analysis on the patient's symptoms based on the consultation information to determine the correlation between the various symptoms; then, it matches the symptoms with possible related diseases; finally, it combines disease information with the hospital's departmental setup to recommend the most suitable department for consultation. Throughout the reasoning process, the system records the disease reasoning process and the departmental reasoning process in detail, and can present these reasoning processes to doctors and patients in an intuitive way, improving the interpretability and transparency of intelligent triage.

[0150] Furthermore, this application also provides an intelligent triage system applied to the aforementioned intelligent triage method, such as... Figure 2 As shown, the intelligent triage system includes: an acquisition module, a first search enhancer, a consultation module, and a first triage recommendation module.

[0151] In this embodiment, the acquisition module is used to acquire the patient's chief complaint.

[0152] In this embodiment, it is connected to the acquisition module and is used to retrieve the patient's consultation process and patient information to be collected from the medical consultation knowledge base based on the patient's chief complaint; and to construct the medical consultation knowledge base based on the large language model and the latest version of medical data.

[0153] Here, the first search enhancer is the RAG (Retrieval Augmented Generation) searcher, which is used to retrieve the patient's consultation process and patient information to be collected from the medical consultation knowledge base based on the patient's chief complaint.

[0154] In this embodiment, the first retrieval enhancer is used to process the latest version of medical data based on a large language model to construct a medical consultation knowledge base. For example, RAG technology can be used to search and organize consultation processes, information to be collected, etc., from the medical consultation knowledge base.

[0155] In this way, the intelligent triage system utilizes RAG technology, combined with patient complaint information, to retrieve the latest medical consultation guidelines from the medical consultation knowledge base, and summarizes the patient information to be collected, thereby enhancing the completeness and relevance of information collection.

[0156] In this embodiment, the consultation module is connected to the first retrieval enhancer; it is used to conduct multiple rounds of consultations on the patient based on the trained consultation model, the patient consultation process and the patient information to be collected, and generate consultation information that represents the patient's condition.

[0157] In this embodiment, the first triage recommendation module is connected to the consultation module and is used to reason about the patient's condition based on consultation information and thinking chain technology, and generate a reasoning process and a reasoning result; wherein, the reasoning process includes the reasoning process of the patient's condition and the reasoning process of the patient's department; the reasoning result includes the patient's condition analysis result and the first target department recommended to the patient.

[0158] Here, the application of thought chain technology can significantly improve the model's analysis and recommendation performance. At the same time, it can clearly demonstrate the decision-making process to doctors and patients, further enhancing the system's interpretability.

[0159] Therefore, this application embodiment innovatively designs an intelligent triage system by integrating RAG technology, large language model, and thought chain technology. This system achieves improvements in accuracy, interpretability, and real-time performance. (1) In terms of accuracy, by combining Retrieval Enhanced Generation (RAG) technology and large language model, the system can more comprehensively collect and understand patients' medical information, thereby improving the accuracy of triage. Compared with traditional methods based on deep learning or retrieval, this system can handle more complex medical descriptions and reduce the possibility of misdiagnosis. (2) In terms of interpretability, by introducing thought chain technology, the reasoning process of triage is made more transparent. This system can not only provide recommended departments, but also display the reasoning process in detail, including medical analysis and reasons for recommendation. This helps to enhance the trust of doctors and patients in the system's recommendation results. (3) In terms of real-time performance, by utilizing RAG technology, the system can retrieve the latest medical consultation guidelines and medical works in real time, ensuring the timeliness of the recommendation results. Compared with methods based on static knowledge graphs, this system can adapt to the updates of medical knowledge more quickly and provide more timely triage suggestions.

[0160] In some embodiments, such as Figure 2 As shown, the system also includes:

[0161] The second triage recommendation module is connected to both the hospital database and the first triage recommendation module. It receives the list of departments of the target hospital and, based on the patient's condition analysis results, the first target department, and the second search enhancer, retrieves the second target department that matches the first target department from the list of departments of the target hospital.

[0162] In this embodiment, the first target department recommended by the first triage recommendation module is highly flexible (i.e., not limited by a fixed department list, but mainly relying on the model's generalization ability). Therefore, to further optimize the user experience, the second triage recommendation module is connected to the hospital database, receiving the department list of the target hospital, thereby supporting the hospital to provide personalized department lists and detailed descriptions. Furthermore, the second triage recommendation module is connected to the first triage recommendation module, and its main function is to, based on the second retrieval enhancer, combine the patient's condition analysis results with the first target department recommended by the first triage recommendation module, retrieve the second target department matching the first target department from the target hospital's department list.

[0163] In this embodiment, the hospital database stores a list of departments in each hospital, including but not limited to department names, department specialties, and doctor information.

[0164] In some embodiments, the first retrieval enhancer includes a custom input information module for obtaining custom input information from the hospital;

[0165] The first search enhancer is used to build a medical consultation knowledge base based on a large language model, custom input information from hospitals, and the latest version of medical data.

[0166] In this embodiment, in addition to the latest version of medical data, the intelligent triage system also supports custom inputs from hospitals, including: patient information to be collected, consultation process, and other requirements.

[0167] In this embodiment, the custom input information module in the first retrieval enhancer 10 is responsible for collecting this information and combining it with the latest medical data to generate a comprehensive medical consultation knowledge base.

[0168] In some embodiments, the first search enhancer further includes an update module for updating the medical consultation knowledge base in response to version change instructions for medical textbooks.

[0169] In this embodiment, the update module is used to update the medical consultation knowledge base in response to the version change instruction information of the medical textbook, so as to ensure the real-time update of the knowledge base information and thereby improve the accuracy and timeliness of the triage.

[0170] In some embodiments, the acquisition module includes:

[0171] A touch screen, connected to a first retrieval enhancer, is used to obtain the patient's chief complaint.

[0172] In this embodiment, the acquisition module also includes a touch screen. The main function of the touch screen is to receive the patient's chief complaint information and send this information to the first search enhancer.

[0173] In some embodiments, the second triage recommendation module further includes a memory for storing a list of departments of the target hospital.

[0174] In this embodiment, the second triage recommendation module of the intelligent triage system is equipped with a memory specifically responsible for storing the department list information of the target hospital. In this way, the system can save the department information of a specific hospital and accurately recommend suitable departments to patients based on their specific conditions and the actual situation of the hospital.

[0175] In some embodiments, such as Figure 2 As shown, the consultation module also includes:

[0176] The consultation corpus construction module, connected to the consultation model training module, includes a disease database, a consultation knowledge base, and a patient profile construction module. The patient profile construction module is connected to the disease database and is used to construct patient profiles based on the disease database, a large language model, and a pre-built patient profile framework. The large language model is a general-domain open-source large language model. The consultation module is also used to construct consultation corpora based on the large language model, patient profiles, and consultation knowledge base.

[0177] The consultation model training module is used to train the base model based on the consultation corpus until a trained consultation model is generated.

[0178] In this embodiment, the consultation module includes a consultation corpus construction module and a consultation model training module. In order to obtain a high-quality consultation model, the consultation corpus is first constructed based on the consultation corpus construction module, and then the generated consultation corpus is provided to the consultation model training module to train an efficient consultation model.

[0179] In this embodiment, the consultation corpus construction module includes a disease database, a consultation knowledge base, and a patient profile construction module. The patient profile construction module is connected to the disease database and is used to construct patient profiles based on the disease database, a large language model, and a pre-built patient profile framework. The large language model is a general-domain open-source large language model. The consultation module is also used to construct consultation corpora based on the large language model, patient profiles, and consultation knowledge base.

[0180] In this embodiment, the consultation model training module is used to train the base model based on the consultation corpus until a trained consultation model is generated.

[0181] In this embodiment, the large language model includes a general-domain large language model, which has scalability and multi-task processing capabilities and is suitable for a wide range of application scenarios. It includes DeepSeek V3, DeepSeekR1, GPT-4, Qwen1.5 and Ernie-4.0 models, etc.

[0182] In this embodiment, the base model includes a lightweight large language model, which includes ChatGLM, LLaMa, RoBERTa-Lite, and Ernie-Speed ​​models, etc.

[0183] In some embodiments, the medical history model training module further includes:

[0184] The pre-training module, connected to the consultation corpus construction module, is used to pre-train the base model based on the consultation corpus and generate a pre-trained consultation model.

[0185] The fine-tuning module, connected to the pre-training module, is used to fine-tune the pre-trained consultation model to generate a trained consultation model.

[0186] In this embodiment, the training process of the consultation model training module includes two processes: the first is pre-training, and the second is supervised fine-tuning. Correspondingly, this intelligent triage system includes a pre-training module and a fine-tuning module. The pre-training module is connected to the consultation corpus construction module and is used to pre-train the base model based on the consultation corpus to generate a pre-trained consultation model.

[0187] The fine-tuning module connects to the pre-training module, receives the pre-trained diagnostic model, and performs further supervised fine-tuning based on labeled data for specific application scenarios. Through this fine-tuning, the model can better adapt to the specific needs of real-world applications, generating the final trained diagnostic model.

[0188] In this way, the accuracy and adaptability of the consultation model can be improved through the two steps of pre-training and fine-tuning.

[0189] In some embodiments, the medical history corpus construction module further includes:

[0190] The extraction module, connected to the patient profile construction module, is used to extract the chief complaint from the patient profile.

[0191] The multi-role simulation question-and-answer module, connected to the extraction module, is used to simulate multi-round question-and-answer interactions between patients and doctors based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and the large language model, and generate question-and-answer interaction data.

[0192] The round control module, connected to the multi-role simulated question-and-answer module, is used to determine whether the number of question-and-answer rounds in the multi-round question-and-answer interaction has reached the target number. If so, a diagnostic corpus is constructed based on the question-and-answer interaction data.

[0193] In this embodiment, the extraction module extracts the patient's chief complaint from the patient profile generated by the patient profile construction module. This chief complaint information includes, but is not limited to, the patient's symptoms, present medical history, and allergy history.

[0194] The multi-role simulation question-and-answer module simulates multiple rounds of question-and-answer interactions between patients and doctors based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and the large language model, generating question-and-answer interaction data.

[0195] In this embodiment, the round control module monitors the number of question-and-answer rounds in real time. Once the preset target is reached, the interaction stops, and the final diagnostic corpus is constructed based on the generated data and corpus.

[0196] In some embodiments, the multi-role simulated question-and-answer module includes:

[0197] The patient simulation question-and-answer module, connected to the extraction module, is used to simulate multi-round question-and-answer interactions based on the patient profile's chief complaint and large language model, generating the first question-and-answer data.

[0198] The doctor simulation question-and-answer module is connected to the medical consultation knowledge base. It is used to simulate the doctor's multi-round question-and-answer interaction based on the consultation process and large language model of the medical consultation knowledge base, and generate second question-and-answer data.

[0199] The generation module is used to generate question-and-answer interaction data based on the first question-and-answer data and the second question-and-answer data.

[0200] In this embodiment, the patient simulation question-and-answer module is connected to the extraction module to simulate multiple rounds of question-and-answer interaction by the patient and generate the first question-and-answer data. For example, if the patient's chief complaint is "I have been feeling chest tightness recently," the patient simulation question-and-answer module will generate a series of possible answers from the patient, such as "How long has the chest tightness lasted?" and "Are there any other symptoms?"

[0201] In this embodiment, the doctor-simulated question-and-answer module is connected to the medical consultation knowledge base. It simulates multi-round question-and-answer interactions by a doctor to generate second question-and-answer data. Based on the consultation process and standard questions in the medical consultation knowledge base, this module guides the patient step-by-step to provide a detailed description of their condition. For example, when handling a consultation task related to cardiovascular disease, the doctor-simulated question-and-answer module will generate second question-and-answer data based on the guidance in the medical consultation knowledge base 12, such as "Do you have a family history of cardiovascular disease?" or "Have you had an electrocardiogram (ECG)?"

[0202] Thus, the patient simulation question-and-answer module helps the system generate more realistic dialogue scenarios by simulating patient responses, thereby improving the quality and diversity of the consultation data. The doctor simulation question-and-answer module ensures that the system adheres to medical norms and standards and generates high-quality consultation data by simulating the professional consultation process of doctors.

[0203] In some embodiments, the round control module further includes a counting module for calculating the number of question-and-answer rounds.

[0204] In this embodiment, the round control module includes a counting module, which is used to calculate the number of question-and-answer rounds to provide a basis for determining whether to stop subsequent question-and-answer sessions.

[0205] In some embodiments, such as Figure 2 As shown, the first triage recommendation module includes:

[0206] The reasoning module, connected to the consultation module, is used to decompose the patient's condition reasoning into multiple sub-tasks based on the thinking chain technology. These sub-tasks include at least: symptom matching, medical history association, and department priority ranking. Based on the consultation information and these sub-tasks, step-by-step reasoning is performed to generate the reasoning process and results. The reasoning process includes the patient's condition reasoning process and the patient's department reasoning process; the reasoning results include the patient's condition analysis results and the recommended first target department for the patient.

[0207] The reasoning process visualization module is used to display the reasoning process of a patient's condition and the reasoning process of the patient's department.

[0208] In this embodiment, the reasoning process visualization module is used to present the patient's condition reasoning and department recommendation process to doctors and patients in an intuitive way, so that they can clearly understand the origin of the system's recommendation results, thereby improving the transparency and credibility of the system.

[0209] In some embodiments, the inference module further includes:

[0210] The decomposition module, connected to the consultation module, is used to decompose the patient's condition into multiple sub-tasks based on the thinking chain technology.

[0211] The technical solution of this application will be described in detail below with reference to an application example.

[0212] Among related technologies, intelligent triage and patient guidance technology has the following drawbacks:

[0213] (1) Intelligent triage method based on deep learning

[0214] While deep learning-based triage systems are technologically advanced, these models are, on the one hand, black-box models with poor interpretability. Due to their nature, doctors and patients find it difficult to understand the model's decision-making process, which may reduce their trust in the results.

[0215] In addition, with the rapid development of large-scale language models, traditional deep learning models have become insufficient in terms of semantic understanding and context processing capabilities. Large models are better able to understand complex patient descriptions and generate more accurate triage results, while traditional deep learning models have relatively limited performance in this regard and struggle to handle ambiguous or polysemous descriptions of symptoms.

[0216] (2) Search-based intelligent consultation method

[0217] While search-based triage systems possess certain practicality and interpretability, they also have some significant shortcomings. First, patient profiles are often not comprehensive enough. The system relies primarily on basic patient information and the patient's chief complaint, lacking more comprehensive information such as medical history, lifestyle habits, and family medical history. This may lead to decreased accuracy in recommendations. Second, the system's effectiveness heavily depends on the construction and maintenance of a medical knowledge graph and a physician database. Building such a knowledge base requires extensive medical knowledge and expert annotation, resulting in high costs and difficulty in covering all diseases and symptoms.

[0218] Furthermore, because knowledge graphs are updated slowly, they struggle to reflect the latest medical information and changes in doctors' areas of expertise in real time, which can lead to delays in recommendation results. Finally, retrieval-based systems may lack matching accuracy when handling complex or multi-symptom cases, especially when patient descriptions are inaccurate or incomplete, in which case the system's recommendation effectiveness will significantly decrease.

[0219] To address the aforementioned issues, this application example constructs a comprehensive intelligent triage system based on large-scale language models, thought chain technology, and Retrieval Augmentation (RAG) technology. This solution utilizes RAG technology, combined with patient complaints, to search for the latest medical consultation guidelines and summarize the patient information to be collected. It collects and summarizes patient information based on a large-scale language model and provides a final department recommendation and justification based on thought chain technology. Simultaneously, this solution can also receive consultation process requirements and real department lists provided by hospitals, collecting patient information and making final department recommendations as needed. Compared to traditional deep learning-based methods, this solution utilizes a large-scale language model to improve the understanding and summarization of patient information and uses thought chain technology to enhance interpretability. Compared to retrieval-based intelligent triage methods, this solution uses RAG technology to improve the completeness of patient information collection and enhances the real-time nature of consultation and department recommendations based on hospital-defined consultation processes and current medical literature. Addressing the problems of existing technologies, this application example proposes an intelligent triage system based on large-scale language models, thought chain technology, and retrieval augmentation generation (RAG) technology to improve the accuracy, interpretability, and real-time performance of triage.

[0220] The following is combined with Figure 3 This application example provides a detailed description of the intelligent triage system. For example... Figure 3 As shown, the intelligent triage system is mainly divided into stages such as RAG retrieval, corpus construction, model training, and intelligent triage.

[0221] In practical applications, Figure 3 In the RAG retrieval, corpus construction, model training, and intelligent triage process, each can be considered an independent module or workflow step. The first three steps primarily involve model training and preparation. The final intelligent triage stage involves model inference, i.e., the application stage.

[0222] (1) RAG search stage.

[0223] In practical applications, the default workflow of the RAG retrieval module is to organize and construct a medical consultation knowledge base based on the latest versions of medical textbooks and works such as *Internal Medicine*, *Surgery*, *Obstetrics and Gynecology*, *Pathology*, and *Pediatrics*, using a large-scale language model (Ernie-4.0) module. This knowledge base includes, but is not limited to, disease names, common symptoms, patient information to be collected, and basic consultation procedures. This medical consultation knowledge base can be used in subsequent corpus construction and intelligent triage.

[0224] (2) Hospital-defined input requirements were analyzed using a large-scale language model and comprehensively organized in conjunction with a medical consultation knowledge base built from default workflows. Hospital-defined inputs include patient information to be collected, consultation information, and other requirements.

[0225] In other words, the input to the RAG retrieval module is the patient's chief complaint, and the output is the consultation process obtained from the medical consultation knowledge base (which is more logically ordered, such as asking about basic symptoms first and then medical history) and the information to be collected (which is more like a checklist, specifying what information needs to be collected to be considered sufficient).

[0226] For example, such as Figure 3 As shown, when a patient provides their chief complaint (such as "I've recently had a sore throat" or "Which department should I see for numbness in my arm?"), the system searches the medical consultation knowledge base using RAG technology based on these complaints. Simultaneously, this knowledge base also incorporates specific requirements from the hospital's custom input criteria. Ultimately, RAG retrieval generates a detailed consultation process and further relevant patient information for the patient.

[0227] (2) Corpus construction stage.

[0228] like Figure 3 As shown, a prompt can be designed based on a disease database, and a patient profile can be generated randomly based on a predefined patient profile framework (including but not limited to age, gender, basic symptoms, current medical history, allergy history, family history, examination and test reports, etc.) and the second major language model module.

[0229] Specifically, a large-scale language model (Ernie-4.0) was used to simulate (play the role of) patients and doctors, respectively, and to construct a consultation corpus through interactive dialogue. In this process, the patient role extracted the chief complaint based on a pre-generated patient profile, while the doctor role constructed the consultation process based on a medical consultation knowledge base and conducted the consultation step by step.

[0230] Once all the information to be collected has been gathered, the conversation will end.

[0231] (3) Model training phase.

[0232] Here, the base model chosen is the Ernie-Speed ​​model. Ernie-Speed ​​is a high-efficiency pre-trained language model specifically designed to improve processing speed while maintaining performance.

[0233] The diagnostic corpus constructed in the above corpus construction stage is used to perform supervised fine-tuning (SFT) on the base model to construct the fine-tuned diagnostic model (the fine-tuned Ernie-Speed ​​model).

[0234] (4) Intelligent triage stage.

[0235] 1. Initial intelligent triage phase.

[0236] First, streamline the consultation process. In real-world scenarios, patients will provide their chief complaint based on their symptoms (e.g., "I've recently had a sore throat," "My arm is numb, which department should I see?"). We first use the aforementioned RAG technology to search and streamline the consultation process and information to be collected from the medical consultation knowledge base.

[0237] Second, the consultation. After outlining the consultation process and gathering patient information, the consultation is conducted with the patient using a supervised, fine-tuned model (Ernie-Speed). The dialogue ends after all information has been collected.

[0238] Third, preliminary conclusions. For example... Figure 3 As shown, the collected patient information is organized and analyzed, and a preliminary analysis of the patient's condition and a department recommendation (first target department) are provided using the mind chain technique. In this process, the mind chain technique helps the model improve the effectiveness of its analysis and recommendations, while also demonstrating the decision-making process to doctors and patients, thus enhancing interpretability.

[0239] 2. Final intelligent triage stage.

[0240] Since the initial department recommendations generated by large language models are boundless (i.e., there is no fixed list of departments, and they are mainly based on the model's generalization effect), in order to improve the user experience, this application example allows hospitals to provide their own exclusive list of departments and descriptions, and uses RAG technology to retrieve the department most similar to the patient's condition as the final recommended department (second target department).

[0241] Below is a doctor-patient dialogue and an example of a thought process chain.

[0242] The patient complained of a persistent cough that worsened at night and sometimes accompanied by shortness of breath.

[0243] Doctor-patient dialogue

[0244] Doctor: How long have you had this cough? Is it a dry cough or a cough with phlegm?

[0245] Patient: It's been about five or six days. It's mainly a dry cough, and sometimes there's a little white phlegm.

[0246] Doctor: Do you have a fever, chills, or general weakness?

[0247] Patient: I had a slight fever a couple of days ago, around 37.8℃, but it seems to have subsided now. I just feel a bit tired.

[0248] Doctor: Have you been in contact with any patients with colds or flu recently?

[0249] Patient: My colleague caught a cold a few days ago, with a cough and runny nose, but I don't know if it's the flu.

[0250] Doctor: Does your shortness of breath worsen during activity, or do you also experience significant difficulty breathing when at rest?

[0251] Patient: I feel it more when I'm active, for example, when I'm climbing stairs I feel like I don't have enough air.

[0252] Doctor: Do you have any history of chronic lung disease, such as asthma or COPD?

[0253] Patient: No, I'm usually in good health and I don't smoke.

[0254] Thought process: The patient's main symptom is a persistent dry cough lasting five to six days, worsening at night, accompanied by mild shortness of breath and a brief low-grade fever. First, the cough could be caused by an infection (viral or bacterial), allergies, or airway disease, while the shortness of breath suggests potential involvement of the lungs or cardiovascular system. The patient has no history of chronic illness and is a non-smoker, reducing the likelihood of chronic lung disease or other serious lung conditions. Furthermore, his recent contact with someone who had a cold suggests a higher probability of viral infection. The shortness of breath worsens with activity, possibly related to airway inflammation or mild bronchoconstriction. Based on this comprehensive analysis, the most probable diagnosis is an upper respiratory tract infection or acute bronchitis, but pneumonia should still be ruled out. The patient is advised to visit a pulmonologist for a chest X-ray or complete blood count to further clarify the cause and receive targeted treatment.

[0255] Doctor: Okay, based on your symptoms, it could be an upper respiratory tract infection or bronchitis. Further examination is needed to determine if a lung infection is possible. I suggest you see a pulmonologist. The doctor may order a chest X-ray or blood tests to rule out pneumonia or other diseases.

[0256] Therefore, this application example, by combining large-scale language models, thought chain technology, and RAG technology, significantly improves the accuracy, real-time performance, and interpretability of the intelligent triage system. Compared to triage methods based on traditional deep learning or retrieval, this application example can collect patient information more comprehensively, combining the latest medical standards and hospital-customized processes to provide more personalized and accurate department recommendations. Simultaneously, the introduction of thought chain technology makes the model's reasoning process more transparent, enhancing doctors' and patients' trust in the system, thus creating a significant competitive advantage and improving user satisfaction and trust in the product.

[0257] Based on the hardware implementation of the above program modules, and in order to implement the method of the embodiments of this application, the embodiments of this application also provide an intelligent triage device. Figure 4 This is only an exemplary structure of the intelligent triage device, not the entire structure; implementation is possible as needed. Figure 4 The structure shown may be part of the structure or the entire structure. For example... Figure 4 As shown, the intelligent triage device 400 provided in this embodiment includes at least one processor 401, a memory 402, a user interface 403, and at least one network interface 404. The various components in the intelligent triage device 400 are coupled together via a bus system 405. It can be understood that the bus system 405 is used to realize the connection and communication between these components. In addition to a data bus, the bus system 405 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 4 The general designated all buses as Bus System 405.

[0258] The user interface 403 may include a monitor, keyboard, mouse, trackball, click wheel, buttons, touchpad, or touch screen.

[0259] The memory 402 in this embodiment is used to store various types of data to support the operation of the intelligent triage device. Examples of such data include any computer program used to operate on the intelligent triage device.

[0260] The intelligent triage method of the intelligent triage device disclosed in this application can be applied to or implemented by the processor 401. The processor 401 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the intelligent triage method of the intelligent triage device can be completed by the integrated logic circuit of the hardware in the processor 401 or by instructions in the form of software. The processor 401 mentioned above may be a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor 401 can implement or execute the methods, steps and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor, etc. The steps of the method disclosed in the embodiments of this application can be directly reflected as being executed by a hardware decoding processor, or being executed by a combination of hardware and software modules in the decoding processor. The software modules may be located in a storage medium, which is located in the memory 402. The processor 401 reads the information in the memory 402 and, in conjunction with its hardware, completes the steps of the intelligent triage method of the intelligent triage device provided in the embodiments of this application.

[0261] In an exemplary embodiment, the intelligent triage device may be implemented by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers (MCUs), microprocessors, or other electronic components to perform the aforementioned method.

[0262] It is understood that memory 402 can be volatile memory or non-volatile memory, or both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), ferromagnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM); magnetic surface memory can be disk storage or... Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as Static Random Access Memory (SRAM), Synchronous Static Random Access Memory (SSRAM), Dynamic Random Access Memory (DRAM), Synchronous Dynamic Random Access Memory (SDRAM), Double Data Rate Synchronous Dynamic Random Access Memory (DDRSDRAM), Enhanced Synchronous Dynamic Random Access Memory (ESDRAM), SyncLink Dynamic Random Access Memory (SLDRAM), and Direct Rambus Random Access Memory (DRRAM). The memories described in the embodiments of this application are intended to include, but are not limited to, these and any other suitable types of memory.

[0263] In an exemplary embodiment, this application also provides a computer storage medium, specifically a computer-readable storage medium storing a computer program thereon. This computer program can be executed by a processor to complete the steps of the method described in this application embodiment. The computer-readable storage medium can be a memory such as ROM, PROM, EPROM, EEPROM, Flash Memory, magnetic surface memory, optical disc, or CD-ROM.

[0264] In an exemplary embodiment, this application also provides a computer program product, including a computer program that can be executed by the processor 401 of the intelligent triage device to complete the steps of the method of this application embodiment.

[0265] It should be noted that terms such as "first" and "second" are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0266] Furthermore, the technical solutions described in the embodiments of this application can be combined arbitrarily without conflict.

[0267] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart triage and patient guidance method, characterized in that, The methods include: Obtain the patient's chief complaint; In response to the patient's chief complaint, the patient's consultation process and patient information to be collected are retrieved from the medical consultation knowledge base based on the first retrieval enhancer. The medical consultation knowledge base is constructed based on a large language model and the latest version of medical data. Based on the trained consultation model, the patient consultation process, and the patient information to be collected, the patient is consulted in multiple rounds to generate consultation information. Based on the consultation information and the thought chain technology, the patient's condition is inferred, generating a reasoning process and a reasoning result; wherein, the reasoning process includes the reasoning process of the patient's condition and the reasoning process of the patient's department; the reasoning result includes the analysis result of the patient's condition and the first target department recommended for the patient.

2. The method according to claim 1, characterized in that, The method further includes: Receive the list of departments at the target hospital; Based on the patient's condition analysis results, the first target department, and the second search enhancer, a second target department matching the first target department is retrieved from the department list of the target hospital.

3. The method according to claim 1, characterized in that, The method further includes: Retrieve custom input information from the hospital; Based on the large language model, the hospital's custom input information, and the latest version of medical data, the medical consultation knowledge base is constructed.

4. The method according to claim 1, characterized in that, The latest version of medical data includes the latest version of medical textbooks, and the method further includes: In response to the version change notification information of the medical textbook, update the medical consultation knowledge base.

5. The method according to claim 1, characterized in that, The method further includes: Based on the disease database, the large language model, and the pre-built patient profiling framework, a patient profile is constructed. The large language model is a general-domain open-source large language model. Based on the patient profile, the medical consultation knowledge base, and the large language model, the consultation corpus is constructed. The base model is trained based on the consultation corpus until the trained consultation model is obtained.

6. The method according to claim 5, characterized in that, The process of constructing the consultation corpus based on the patient profile, the medical consultation knowledge base, and the large language model includes: Extract the chief complaint from the patient profile; Based on the patient's chief complaint in the patient profile, the consultation process in the medical consultation knowledge base, and the large language model, multi-round question-and-answer interactions between the patient and the doctor are simulated to generate question-and-answer interaction data. Determine whether the number of question-and-answer rounds in the multi-round question-and-answer interaction has reached the target number. If so, construct the diagnostic corpus based on the question-and-answer interaction data.

7. The method according to claim 6, characterized in that, The process, based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and the large language model simulating multi-round question-and-answer interactions between the patient and the doctor, generates question-and-answer interaction data, including: Based on the patient profile's chief complaint and the large language model, multi-round question-and-answer interactions of the patient profile are simulated to generate the first question-and-answer data. Based on the consultation process of the medical consultation knowledge base and the large language model, the doctor's multi-round question-and-answer interaction is simulated to generate second question-and-answer data; The question-and-answer interaction data is generated based on the first question-and-answer data and the second question-and-answer data.

8. The method according to claim 5, characterized in that, The step of training the base model based on the consultation corpus until the trained consultation model is obtained includes: The base model is pre-trained based on the consultation corpus to generate a pre-trained consultation model. The pre-trained consultation model is fine-tuned to generate the trained consultation model.

9. The method according to claim 1, characterized in that, The process of reasoning about the patient's condition based on the consultation information and the thought chain technology, generating a reasoning process and a reasoning result, includes: Based on the thinking chain technology, the reasoning of the patient's condition is decomposed into multiple sub-tasks, which include at least: symptom matching task, medical history association task, and department priority ranking task. Based on the consultation information and the multiple sub-tasks, step-by-step reasoning is performed to generate the reasoning process and reasoning results.

10. An intelligent triage and referral system, characterized in that, The system, applied to the intelligent triage method as described in any one of claims 1-9, comprises: The acquisition module is used to acquire the patient's chief complaint; The first retrieval enhancer, connected to the acquisition module, is used to retrieve the patient's consultation process and patient information to be collected from the medical consultation knowledge base based on the patient's chief complaint; and to construct the medical consultation knowledge base based on the large language model and the latest version of medical data. The consultation module is connected to the first retrieval enhancer. The consultation module is used to conduct multiple rounds of consultation with the patient based on the trained consultation model, the patient consultation process and the patient information to be collected, and generate consultation information that represents the patient's condition information. The first triage and recommendation module is connected to the consultation module and is used to reason about the patient's condition based on the consultation information and the thinking chain technology, and generate a reasoning process and a reasoning result; wherein, the reasoning process includes the reasoning process of the patient's condition and the reasoning process of the patient's department; the reasoning result includes the analysis result of the patient's condition and the first target department recommended for the patient.

11. The system according to claim 10, characterized in that, The system also includes: The second triage recommendation module is connected to both the hospital database and the first triage recommendation module. It receives the department list of the target hospital and, based on the patient's condition analysis results, the first target department, and the second search enhancer, retrieves a second target department from the department list of the target hospital that matches the first target department.

12. The system according to claim 10, characterized in that, The first retrieval enhancer includes a custom input information module for obtaining custom input information from the hospital; The first retrieval enhancer is used to construct the medical consultation knowledge base based on the large language model, the hospital's custom input information, and the latest version of medical data.

13. The system according to claim 10, characterized in that, The first search enhancer also includes an update module for updating the medical consultation knowledge base in response to version change indication information of the medical textbook.

14. The system according to claim 10, characterized in that, The acquisition module includes: A touch screen, connected to the first retrieval enhancer, is used to obtain the patient's chief complaint.

15. The system according to claim 11, characterized in that, The second triage recommendation module also includes a memory for storing a list of departments of the target hospital.

16. The system according to claim 10, characterized in that, The consultation module includes: The consultation corpus construction module, connected to the consultation model training module, includes a disease database, a consultation knowledge base, and a patient profile construction module. The patient profile construction module is connected to the disease database and is used to construct patient profiles based on the disease database, the large language model, and a pre-built patient profile framework. The large language model is a general-domain open-source large language model. The consultation module is also used to construct consultation corpora based on the large language model, the patient profiles, and the consultation knowledge base. The consultation model training module is used to train the base model based on the consultation corpus until the trained consultation model is generated.

17. The system according to claim 16, characterized in that, The diagnostic model training module also includes: The pre-training module, connected to the consultation corpus construction module, is used to pre-train the base model based on the consultation corpus to generate a pre-trained consultation model. The fine-tuning module, connected to the pre-training module, is used to fine-tune the pre-trained consultation model to generate the trained consultation model.

18. The system according to claim 16, characterized in that, The diagnostic corpus construction module also includes: The extraction module, connected to the patient profile construction module, is used to extract the chief complaint from the patient profile; A multi-role simulation question-and-answer module, connected to the extraction module, is used to simulate multi-round question-and-answer interactions between the patient and the doctor based on the patient's chief complaint, the consultation process in the medical consultation knowledge base, and the large language model, and generate question-and-answer interaction data. The round control module, connected to the multi-role simulated question-and-answer module, is used to determine whether the number of question-and-answer rounds of the multi-round question-and-answer interaction has reached the target number. If so, the diagnostic corpus is constructed based on the question-and-answer interaction data.

19. The system according to claim 18, characterized in that, The multi-role simulated question-and-answer module includes: The patient simulation question-and-answer module, connected to the extraction module, is used to simulate multi-round question-and-answer interactions of the patient profile based on the chief complaint of the patient profile and the large language model, and generate the first question-and-answer data. The doctor simulation question-and-answer module is connected to the medical consultation knowledge base and is used to simulate the doctor's multi-round question-and-answer interaction based on the consultation process of the medical consultation knowledge base and the large language model to generate second question-and-answer data. The generation module is used to generate the question-and-answer interaction data based on the first question-and-answer data and the second question-and-answer data.

20. The system according to claim 18, characterized in that, The round control module also includes a counting module for calculating the number of question-and-answer rounds.

21. The system according to claim 10, characterized in that, The first triage recommendation module includes: The reasoning module, connected to the consultation module, is used to decompose the patient's condition reasoning into multiple sub-tasks based on the thought chain technology. These sub-tasks include at least: a symptom matching task, a medical history association task, and a department priority ranking task. Based on the consultation information and the multiple sub-tasks, step-by-step reasoning is performed to generate a reasoning process and a reasoning result. The reasoning process includes the patient's condition reasoning process and the patient's department reasoning process. The reasoning result includes the patient's condition analysis result and a recommended first target department for the patient. The reasoning process visualization module is used to display the reasoning process of the patient's condition and the reasoning process of the patient's department.

22. The system according to claim 21, characterized in that, The reasoning module also includes: The decomposition module, connected to the consultation module, is used to decompose the patient's condition into multiple sub-tasks based on the thinking chain technology.

23. An intelligent triage and referral device, characterized in that, include: A processor and a memory for storing a computer program capable of running on the processor, wherein the processor, when running the computer program, performs the steps of the method according to any one of claims 1 to 9.

24. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 9.

25. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the intelligent triage device, it implements the steps of the method as described in any one of claims 1 to 9.

Images