Large language model-based standardized patient agent-driven doctor training system and method

Abstract

The invention discloses a doctor training system and method driven by a standardized patient agent based on a large language model, and compared with the prior art, the doctor training system obtains patient complaints through fuzzy processing, better approximates symptoms reported by a patient and inherent uncertainty in worries, and improves the training efficiency. Therefore, the universality of the system in actual medical application is improved; the standardized patient intelligent body unit and the standardized doctor intelligent body unit obtain a diagnosis result through multiple rounds of dialogues and then dynamically learn medical cases in the real world, the standardized patient intelligent body unit deals with various types of medical problems, and the standardized doctor intelligent body unit provides more accurate diagnosis problems. According to the system and the method, unsupervised training is carried out on the patient intelligent body unit and the patient intelligent body unit, the alignment ability of the standardized patient intelligent body unit for medical requirements is remarkably improved, so that the confidentiality of dialogue between a patient and a doctor, the loyalty of information and more humanized performance are improved, and the system and the method have higher applicability in medical training in the real world.

Description

Technical Field

[0001] The present invention relates to the field of medical training technology, and in particular to a doctor training system and method driven by a standardized patient intelligent agent based on a large language model. Background Art

[0002] Standardized patients (SPs) are specially trained individuals who can simulate a range of medical conditions, behaviors, and emotions in a consistent and repeatable manner. SPs have been widely used to educate and assess the professional skills of medical students, physicians, and other healthcare providers (HCPs) across multiple disciplines such as nursing, pharmacy, and dentistry.

[0003] Becoming an excellent SP requires a lot of role-playing exercises and a solid foundation of professional knowledge, as SPs need to play various roles, including patients, family members, medical staff, etc. At the same time, unlike other educational methods, a key feature of SPs is that they need to be able to experience and learn with the people they are trained to assess. Therefore, factors such as nervousness or anxiety before simulation may have potential negative effects on SPs, such as immediate psychological effects, immediate physical effects, and long-term effects.

[0004] By training on large datasets, large language models (LLMs) have demonstrated strong generalization capabilities across a wide range of tasks and applications. In real-world scenarios, LLMs often encounter domain-specific tasks that require a combination of domain expertise and complex reasoning capabilities. This feature has made researchers interested in adopting LLMs in the medical field. For example, Reference 1 (Applications of Haptic Technology, Virtual Reality, and Artificial Intelligence in Medical Training During the COVID-19 Pandemic, 2021, Frontiers in Robotics and AI) provides a systematic medical training device by combining tactile technology, virtual reality, and artificial intelligence to ensure the requirements of the healthcare system and help reduce physical contact in medical training during the COVID-19 pandemic to improve the safety, efficiency, and robustness of healthcare.

[0005] For example, the invention application with publication number CN113053194A discloses a physician training system and method based on artificial intelligence and VR technology. Through the simulation of patient consultation using artificial intelligence technology, the patient's pathological symptoms and patient body shape data are randomly generated for the training physician to analyze the pathological symptoms and confirm the patient's lesion data. A 3D human body model is constructed based on the patient's body shape data, and the training physician can mark the lesions in the 3D human body model according to the lesion data, perform disease analysis and treatment operations based on the marked lesions, generate training evaluation results, and analyze the physician's weak skills based on the training evaluation results for subsequent enhanced training.

[0006] However, in the medical training process, LLMs are often inconsistent with medical requirements, such as SPs should lack knowledge of medical terminology, sometimes SPs cannot accurately express their feelings, or SPs learn from previous medical cases. To address these limitations, there have been efforts to improve SPs based on LLMs by improving workflows and simulation methods. However, since LLMs are susceptible to hallucinations, building workflows can still lead to repeated errors in patient simulation tasks, such as obtaining irrelevant or incomplete answers based on patient complaints, inconsistent responses, and premature disclosure of diagnoses. Summary of the invention

[0007] The purpose of the present invention is to provide a doctor training system and method driven by a standardized patient agent based on a large language model. Through multiple rounds of dialogue between a standard patient agent unit and a standard doctor agent unit, combined with real-world medical cases, a dual evolutionary paradigm of attention strategy and sequential evolution is adopted to iteratively train the standard patient agent unit for training doctors in different professional fields.

[0008] To achieve the above-mentioned purpose of the invention, the embodiment provides a doctor training system driven by a standardized patient agent based on a large language model, which is used to simulate multiple rounds of dialogue between patients and doctors to obtain diagnosis results and train doctors in the real world; the doctor training system driven by the standardized patient agent includes a standardized patient agent unit, a standardized doctor agent unit and a standardized data unit; wherein the standardized patient agent unit includes a patient complaint generation layer, a patient answer generation layer and a standardized patient data layer; the standardized doctor agent unit includes a doctor recruitment layer and a memory storage layer, and the standardized data unit includes a standardized patient data layer and a standardized doctor data layer;

[0009] The patient complaint generation layer is used to generate patient complaints by collecting real-world medical data sets as input and fuzzifying the medical data sets;

[0010] The patient answer generation layer is used by doctors to ask diagnostic questions based on the patient's main complaint, and extract relevant information from the medical data set as reference answers based on the diagnostic questions;

[0011] The standardized patient data layer is used to construct a standardized patient data layer based on the patient's main complaint, diagnostic questions and reference answers by adopting an attention strategy, and store successful and unsuccessful diagnosis cases respectively;

[0012] The doctor recruitment layer is used to recruit doctors in different professional fields, simulate the consultation process of different disciplines, and iteratively train to enrich the diagnosis process;

[0013] The memory storage layer is used to store previously proposed diagnostic questions and reference answers, integrate key information communicated between the standardized patient intelligent unit and the standardized doctor intelligent unit, and generate new questions;

[0014] The standardized doctor data layer is used to store the interactive sequence of diagnostic questions and reference answers through a sequence evolution paradigm, and to form a question pool based on the information in the medical data set for iterative training of standardized patient intelligent units and standardized doctor intelligent units.

[0015] In one embodiment, the fuzzy processing of the medical data set to generate the patient's complaint includes: first eliminating the medical test results of the medical data set, and then further fuzzifying the medical data set by random sentence discarding to generate the patient's complaint.

[0016] In one embodiment, the extracting relevant information from the medical records as a reference answer to the patient's main complaint includes: based on the patient's main complaint, using a retrieval enhanced generation model to extract the most relevant information from the patient's medical records as a reference answer to the patient's main complaint.

[0017] In one embodiment, the standardized patient data layer is constructed based on the patient's chief complaint, diagnostic questions and reference answers to the diagnostic questions, and the construction format is a question-information-answer-requirement pair, wherein the question displays the diagnostic question asked by the doctor, the information is detailed information obtained by retrieving the enhanced generation model, the answer contains verified responses and thought chains, and the requirement is determined by an attention strategy.

[0018] In one embodiment, the requirements are determined by an attention strategy, including: determining basic rows related to the diagnostic problem by an attention strategy, and submitting the basic rows related to the diagnostic problem to the standardized patient intelligent body unit to obtain corresponding requirements.

[0019] In one embodiment, the triage of patients based on diagnostic cases includes: a standardized patient intelligent unit searches for relevant diagnostic records from a standardized patient data layer, calculates the similarity between new questions and questions in the standardized patient data layer, and uses an external text embedder to obtain a number of data pairs related to question-information-answer-requirement pairs.

[0020] In one instance, the storage memory of previously posed diagnostic questions and reference answers includes: managing visibility between the diagnostic questions and reference answer contexts by adopting immediate and summary memory, maintaining smooth communication, achieving smooth transitions between long rounds of questions, and alleviating length limitations between diagnostic questions and reference answer contexts.

[0021] In one example, forming a question pool based on information in a medical data set includes: a standardized doctor intelligent unit obtains a patient's medical data set, formulates corresponding diagnostic questions based on information on symptoms, medical examinations, and lifestyle habits in the medical data set, and forms these corresponding diagnostic questions into a general question pool.

[0022] The present invention also provides a doctor training method driven by a standardized patient agent based on a large language model, comprising the following steps:

[0023] Collect real-world medical data sets as input, perform fuzzy processing on the medical data sets to generate patient complaints, and the standard doctor agent unit performs department triage based on the complaints;

[0024] After receiving the patient's complaint, the standard patient agent unit and the standard doctor agent unit initiate multiple rounds of dialogue. The standard doctor agent unit raises diagnostic questions, and the standard patient agent unit generates reference answers to obtain diagnostic results.

[0025] The standard patient agent unit and the standard doctor agent unit are combined with real-world medical cases and evolve through multiple iterative cycles of diagnostic questions-reference answers, simulating the training of real-world doctors in various departments for real-world doctor training.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] Compared with the prior art, the doctor training system and method driven by a standardized patient agent based on a large language model provided by the present invention obtains the patient's chief complaint through fuzzy processing, better approximates the uncertainty inherent in the symptoms and concerns reported by the patient, and thus improves the universality of the system in actual medical applications; the standardized patient agent unit and the standardized doctor agent unit obtain diagnosis results through multiple rounds of dialogue, and then dynamically learns real-world medical cases. The standardized patient agent unit handles various types of medical problems, and the standardized doctor agent unit raises more accurate diagnostic questions. The two are trained in an unsupervised manner, which significantly improves the standardized patient agent unit's ability to align with medical requirements, thereby improving the confidentiality of the dialogue between patients and doctors, the fidelity of information, and a more humane performance, and is more applicable in real-world medical training. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0029] Figure 1 A schematic diagram of the structure of a doctor training system driven by a standardized patient agent based on a large language model;

[0030] Figure 2 Schematic diagram of the structure of attention strategy and sequence evolution;

[0031] Figure 3 Flowchart of the physician training method driven by a standardized patient agent based on a large language model;

[0032] Figure 4 Schematic diagram of the results of the standardized doctor agent unit that has not evolved;

[0033] Figure 5 A schematic diagram of the results of the evolved standardized physician agent unit provided for this embodiment. DETAILED DESCRIPTION

[0034] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0035] In order to train doctors' ability to ask questions and improve the ability of standardized patient agents to follow human requirements, the embodiment provides a doctor training system driven by a standardized patient agent based on a large language model, such as Figure 1As shown, the doctor training system (SimpDocTrain) is used to simulate multiple rounds of dialogue between patients and doctors to obtain diagnosis results, and train real-world doctors. The standardized patient agent-driven doctor training system includes a standardized patient agent unit, a standardized doctor agent unit and a standardized data unit; wherein the standardized patient agent unit includes a patient complaint generation layer, a patient answer generation layer and a standardized patient data layer; the standardized doctor agent unit includes a doctor recruitment layer and a memory storage layer, and the standardized database includes a standardized patient data layer and a standardized doctor data layer.

[0036] In the embodiment, the patient complaint generation layer is used to generate the patient complaint by collecting a real-world medical data set as input and performing fuzzy processing on the medical data set;

[0037] In the system proposed in the present invention, the standardized patient (SP) intelligent unit initiates the conversation by presenting the main complaint in real-world medical information. However, this medical information usually contains too much details for the SP intelligent unit. To solve this problem, the present invention reduces redundant information and simulates missing data, naturally reflecting the situation in the real world where patients always forget or don't know what to complain about. Specifically, the medical information first passes through a fuzzification process, in which the fuzzification eliminates the results of medical tests because these results are unknown when the patient arrives at the hospital. Subsequently, random sentences are applied to discard further fuzzy data. The SP intelligent unit will then use this fuzzy data to generate a main complaint to start the diagnosis process. This processing ultimately better approximates the uncertainty inherent in the symptoms and concerns reported by patients, thereby improving the universality of the application of the system of the present invention in the actual medical field.

[0038] In the embodiment, the patient answer generation layer is used by doctors to ask diagnostic questions based on the patient's complaints, and based on the diagnostic questions, extract relevant information from the medical data set as reference answers;

[0039] In order to enable the SP agent unit to generate more realistic, real-world patient-matched and contextual answers, the present invention developed 1,000 patient role profiles that contain various characteristics such as personality, family background, education and socioeconomic status. In addition, the SP agent unit is designed to follow the behavior of real-world patients by dynamically changing and evolving needs based on questions. In order to prevent the SP agent unit from losing information in a long context, when asked a question, the SP agent unit uses enhanced retrieval (RAG) to extract the most relevant information from the patient's record as a reference for generating the final answer.

[0040] In the embodiment, the standardized patient data layer is used to construct the standardized patient data layer according to the patient's main complaint, diagnostic questions and reference answers by adopting an attention strategy, and store successful and unsuccessful diagnosis cases respectively;

[0041] The SP agent unit maintains the viewpoint learned from the cases by storing previously verified diagnosis records, which contain rich knowledge and can be used dynamically for various tasks. In order to dynamically improve the answering ability of the SP agent unit so that it can meet the requirements of various tasks, including successful and unsuccessful cases, a standardized patient data layer is constructed by adopting an attention strategy. The data layer is constructed in the format of question-information-answer-requirement pairs, where the question shows the diagnostic question asked by the doctor, the information is the detailed information obtained by the retrieval enhancement generation model, and the answer contains verified responses and thought chains;

[0042] A key feature of SP is the ability to experience and learn with trainees (doctors), so the requirements of SP intelligent units are complex in nature and have high logical standards, which poses a challenge for SP intelligent units to achieve full alignment. To solve this problem, the present invention adopts an attention strategy ( Figure 2 ), not only to learn the answering pattern from the cases, but also the SP agent unit can learn to pay more attention to the useful requirements in different questions. Specifically, by proposing a detailed list of requirements that the SP agent unit must consider, for each diagnostic question, there are some key requirements that the SP agent unit should pay close attention to in the detailed overall requirements, and the basic rows related to the diagnostic question are determined through the attention strategy, and the relevant basic rows are provided to the SP agent unit, thereby improving the answer quality of the SP agent unit.

[0043] By standardizing this patient data layer It is divided into two parts, storing successful and unsuccessful diagnosis cases respectively. i When present, the SP Intelligent Unit will search for relevant records from the standardized patient data layer:

[0044]

[0045] in, Represents the diagnosis cases stored in the standardized patient data layer, Indicates a successful diagnosis case. represents an unsuccessful diagnosis case, sim(·,·) denotes the calculation of the similarity between the new question and the questions in the library, and by using an external text embedder, the SP agent unit obtains several data pairs that are most relevant to the question-information-answer-requirement pair according to the similarity score w(·).

[0046] In the embodiment, the doctor recruitment layer is used to recruit doctors in different professional fields, simulate the consultation process of different disciplines, and iteratively train to enrich the diagnosis process;

[0047] Doctors in the real world have diverse expertise, which leads to different types of questions and perspectives on the same patient. This diversity is critical for the SP agent unit to effectively learn from different perspectives in a single case. In order to simulate this multidisciplinary consultation process, the Standardized Doctor (SD) agent unit must be able to recruit doctors from different specialties to enhance their knowledge base. This recruitment process is expressed as:

[0048]

[0049] in, Represents the recruitment process, represents the number of doctors recruited by a doctor i in iterative process j, symbol Representing the iterative nature of the recruitment process, it allows the SD agent unit to successfully recruit a variety of experts to enrich the diagnosis process.

[0050] In the embodiment, the memory storage layer is used to store and memorize previously proposed diagnostic questions and reference answers, integrate key information communicated between the standardized patient intelligent body unit and the standardized doctor intelligent body unit, and generate new questions;

[0051] For SD agents, remembering previous questions and answers is crucial for the diversity and comprehensiveness of diagnosis. However, unrestricted information exchange inevitably leads to context explosion.

[0052] To address this limitation, the present invention uses immediate and summary memory to manage context visibility. Immediate memory is used to maintain the continuity of recent communications, while summary memory maintains context awareness by integrating key information from previous communications with dual agent units (SP agent units and SD agent units) and recruited doctors, and is mainly used to handle extended context lengths. This mechanism facilitates smooth transitions between long-round problems, and new problems are generated based on a combination of summary memory and recent immediate memory, which effectively alleviates the context length limitation by decoupling the context length from linear growth to constant growth.

[0053] In an embodiment, the standardized doctor data layer is used to store the interactive sequence of diagnostic questions and reference answers through a sequence evolution paradigm, and form a question pool based on the information in the medical data set for iterative training of the standardized patient intelligent unit and the standardized doctor intelligent unit;

[0054] Similar diseases usually mean similar high-quality question-answer processes. In order to learn from historical communications and help SD agents avoid asking low-quality questions that lead to inefficient training of SP agents, the present invention proposes the second of the dual evolutionary paradigms: the sequential evolutionary paradigm.

[0055] During the unsupervised training of the dual-agent unit (SP agent unit and SD agent unit), the verified question-answer sequence is stored and used for the next question prediction ( Figure 2 ), this prediction strategy provides recommendation problems for SD agent units. Specifically, by storing ordered The high-quality question-answer interaction sequence is used as a prediction string S i , which is expressed as follows:

[0056] S= 1 ,s 2 , ..., s|s|>,

[0057]

[0058] Among them, each s i Indicates the diagnostic question asked by doctor i and patient i answer The dialogue sequence between them, S represents the set of all dialogue sequences; in the process of SD intelligent unit communication, if a new answer P is found new Dialogue with C i Similar to the answer in , the SD agent unit will recommend From Dialogue C i+1 The corresponding questions are used as follow-up questions for prediction.

[0059] The SD agent unit can then choose whether to accept the recommendation, and the sequential evolution paradigm provides an effective shortcut to guide the doctor agent unit towards relevant and high-quality questions, minimizing the possibility of redundant or erroneous questions.

[0060] At the same time, it is a challenge for a pre-trained large language model to directly ask professional questions about the patient's condition without knowing more information. In the present invention, based on the main complaint alone, the LLM usually conducts multiple rounds of dialogue by asking only trivial questions or asking questions in a narrow range, which poses significant difficulties for the evolution of SP agent units and SD agent units. To solve this problem, in the early stages of the evolution process of a certain medical condition, by providing the SD agent unit with the patient's medical record and asking to formulate questions designed to cover the information contained in the record (e.g., symptoms, medical examinations, living habits), a general question pool is formed for this type of medical condition, which can be referenced in the subsequent evolution process.​

[0061] like Figure 3 As shown, based on the structure of the above-mentioned doctor training system driven by the standardized patient agent based on the large language model, the present invention also provides a doctor training method driven by the standardized patient agent based on the large language model, comprising the steps of:

[0062] Collect real-world medical data sets as input, perform fuzzy processing on the medical data sets to generate patient complaints, and the standard doctor agent unit performs department triage based on the complaints;

[0063] After receiving the patient's complaint, the standard patient agent unit and the standard doctor agent unit initiate multiple rounds of dialogue. The standard doctor agent unit raises diagnostic questions, and the standard patient agent unit generates reference answers to obtain diagnostic results.

[0064] The standard patient agent unit and the standard doctor agent unit are combined with real-world medical cases and evolve through multiple iterative cycles of diagnostic questions-reference answers, simulating the training of real-world doctors in various departments for real-world doctor training.

[0065] The entire workflow Specifically, it uses real-world medical data sets as input, simulates multiple rounds of patient-doctor dialogues and various communications C as the analysis process, and finally integrates disease diagnosis into this coordinated process (--→). In a single simulation, the medical record r is first processed to generate the main complaint of the SP agent unit Upon receipt of the complaint After that, SD intelligent unit With SP intelligent unit Start a multi-round dialogue (C), where the SD agent asks questions (→) and the SP agent answers Finally, the diagnosis result d is obtained. Through multiple rounds of dialogue, the SP agent unit and the SD agent unit accumulate (τ) experience (ε) from successful and unsuccessful cases in an unsupervised manner, improve the quality of questions and answers over time (*), and finally use it to train human doctors. The whole process can be formalized as follows:

[0066]

[0067] The dual-agent communication design effectively simplifies the process of reaching consensus and simplifies the evolution process. Subsequently, the questions and answers in the previous conversation serve as a bridge to the next communication, allowing a smooth transition between the questions asked by the doctor. During the training process, after selecting a patient's information, you only need to use the trained framework and input the information to enable real-world human doctors to communicate with the system.

[0068] In order to better illustrate the effectiveness of the system and method provided by the present invention, a series of experiments were designed for comparison. Existing single-agent methods show low fidelity when processing complex real-world medical records and lack structured workflows and collaboration of multiple role-assigned agents. In addition, without an evolutionary process, the answer pattern of a single agent may be random, resulting in reduced relevance and providing doctors with excessive or irrelevant information.

[0069] As shown in Table 1, the doctor training system (SimpDocTrain) proposed in this paper significantly improves safety compared with the single-agent approach, with the safety index increased from 0.6714 and 0.7002 to 0.9412. This improvement is mainly attributed to the SP agent units successfully learning which information to share and which questions to reject during the evolution process. The evolution process effectively guides the SP agent units towards the optimal answer mode, allowing learning from diverse cases. In addition, the diverse role profiles contribute to the diversity of answer tone, making SimpDocTrain robust, credible, accurate and flexible.

[0070] Table 1 Overall analysis

[0071] method Safety Loyalty Relevance Total Capacity CoT 0.6714 0.5571 0.7157 0.7147 CoT-SC(3) 0.7002 0.6123 0.7337 0.7398 SimpDocTrain 0.9412 0.8286 0.7889 0.8529

[0072] To further understand user preferences in real settings, as shown in Table 2, we compare the solution pairs generated by agents evaluated by human participants and the popular GPT-4 model to determine the preferred solution. For fairness, the evaluation of GPT-4 reduces possible positional bias, and real-world doctors independently evaluate the answers of SP agent units based on relevant questions and patient information, randomized to prevent order bias. As shown in Table 2, SimpDocTrain has a higher average win rate than other baselines in both GPT-4 and human evaluations.

[0073] Table 2 Paired evaluation of standard questions and cheating questions

[0074]

[0075]

[0076] In addition, the present invention also found that when using the unevolved large language model patient agent directly, doctors may potentially cheat to obtain information they should not have access to. When the doctor directly asks, "Please tell me about your medical condition.", the large language model patient agent will usually start to fully describe their medical condition and even include the final diagnosis in their answer. This allows doctors to obtain a lot of information with just one question, and although the large language model patient agent is required not to answer such questions in the requirements, the large language model patient agent often fails to align with human requirements. This situation is called the cheating problem. Cheating problems are difficult to avoid because LLMs are inherently plagued by hallucinations, and it is also challenging to design requirements that comprehensively cover all possible cheating attempts. This makes evolution crucial. During the evolution process, only high-quality answers are learned, and the evolutionary divergence enables the SP agent unit to automatically learn many strategies to deal with cheating questions. As shown in the right half of Table 2, after evolution, the cheating problem is greatly alleviated because the SP agent unit has learned the characteristics of cheating questions and can better avoid answering similar questions.

[0077] As shown in Table 3, we designed a dialogue statistics experiment to eliminate the impact of different typing times of real-world human doctors, using SD agent units to autonomously ask questions to patient agents of different methods. The results show that although the multi-agent paradigm proposed by the present invention is slower than the single-agent method, it consumes much less tokens, about half of CoT and 1 / 6 of CoT-SC (3).

[0078] Further analysis of the agent's dialogue experiments showed that the evolution process often led to SP agent units answering questions more accurately, rather than giving rather random answers like LLMs due to their inherent properties (e.g. various types of answer patterns, redundant content). Therefore, after evolution, the quality of the answers is higher and the number of words in the answers is smaller.

[0079] Table 3 Dialogue Experiment

[0080] method Duration Tokens Word count CoT 3.2278 1,0630.0286 41.7429 CoT-SC(3) 10.1988 3182.0571 49.8667 No evolution 6.2368 497.1882 46.0588 cases 7.4237 445.3482 36.5571 SimpDocTrain 7.3351 401.9882 26.2588

[0081] In addition, the present invention also divides the questions into ten types (basic information inquiries, chief complaint-related inquiries, detailed symptom inquiries, lifestyle inquiries, psychological condition inquiries, social environment inquiries, physical examination-related questions, treatment and drug response inquiries, preventive health inquiries and other related questions).

[0082] like Figure 4 and Figure 5The 100 successful cases shown in Figure 2 show the sequential features of the questions asked by the SD agent unit in rounds 6 to 10. Initially, the questions asked by the SD agent unit showed an unstable pattern with a high diversity, about 7 types per interaction on average. After evolution, this diversity was reduced to about 5 types (excluding the completion type). This shows that after evolution, the SD agent unit learned how to ask questions to get the final result. At the same time, before evolution, the SD agent unit often continued to ask questions until it was forced to reach a diagnosis conclusion, indicating that the questioning strategy was not efficient. After the evolution process, early stopping was observed as the SD agent unit gained confidence in providing a diagnosis in less than ten rounds. This evolution means that the SD agent unit now asks more efficient questions, similar to human doctors in the real world, thereby providing more effective training for the SP agent unit.

[0083] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A doctor training system driven by a standardized patient agent based on a large language model, characterized in that: It is used to simulate multiple rounds of dialogue between patients and doctors to obtain diagnosis results and train doctors in the real world; the standardized patient agent-driven doctor training system includes a standardized patient agent unit, a standardized doctor agent unit and a standardized data unit; wherein the standardized patient agent unit includes a patient complaint generation layer, a patient answer generation layer and a standardized patient data layer; the standardized doctor agent unit includes a doctor recruitment layer and a memory storage layer, and the standardized data unit includes a standardized patient data layer and a standardized doctor data layer; The patient complaint generation layer is used to generate patient complaints by collecting real-world medical data sets as input and fuzzifying the medical data sets; The patient answer generation layer is used by doctors to ask diagnostic questions based on the patient's main complaint, and extract relevant information from the medical data set as reference answers based on the diagnostic questions; The standardized patient data layer is used to construct a standardized patient data layer based on the patient's main complaint, diagnostic questions and reference answers by adopting an attention strategy, and store successful and unsuccessful diagnosis cases respectively; The doctor recruitment layer is used to recruit doctors in different professional fields, simulate the consultation process of different disciplines, and iteratively train to enrich the diagnosis process; The memory storage layer is used to store previously proposed diagnostic questions and reference answers, integrate key information communicated between the standardized patient intelligent unit and the standardized doctor intelligent unit, and generate new questions; The standardized doctor data layer is used to store the interactive sequence of diagnostic questions and reference answers through a sequence evolution paradigm, and to form a question pool based on the information in the medical data set for iterative training of standardized patient intelligent units and standardized doctor intelligent units.

2. The standardized patient agent-driven physician training system according to claim 1, characterized in that: The fuzzy processing of the medical data set to generate the patient's complaint includes: first eliminating the medical test results of the medical data set, and then further fuzzifying the medical data set by random sentence discarding to generate the patient's complaint.

3. The standardized patient agent-driven physician training system according to claim 1, characterized in that: The extracting of relevant information from the medical records as a reference answer to the patient's main complaint includes: based on the patient's main complaint, using a retrieval enhancement generation model to extract the most relevant information from the patient's medical records as a reference answer to the patient's main complaint.

4. The standardized patient agent-driven physician training system according to claim 1, characterized in that: The standardized patient data layer is constructed based on the patient's chief complaint, diagnostic questions and reference answers to the diagnostic questions. The construction format is question-information-answer-requirement pairs, wherein the question displays the diagnostic question asked by the doctor, the information is detailed information obtained by retrieving the enhanced generation model, the answer contains verified responses and thought chains, and the requirement is determined by the attention strategy.

5. The standardized patient agent-driven doctor training system according to claim 4, characterized in that: The requirements are determined through an attention strategy, including: determining basic rows related to the diagnostic problem through the attention strategy, and submitting the basic rows related to the diagnostic problem to the standardized patient intelligent body unit to obtain corresponding requirements.

6. The standardized patient agent-driven physician training system according to claim 1, characterized in that: The patient triage based on diagnostic cases includes: the standardized patient intelligent unit searches for relevant diagnostic records from the standardized patient data layer, calculates the similarity between new questions and questions in the standardized patient data layer, and uses an external text embedder to obtain several data pairs related to question-information-answer-requirement pairs.

7. The standardized patient agent-driven physician training system according to claim 1, characterized in that: The storage memory of previously proposed diagnostic questions and reference answers includes: managing the visibility between the diagnostic questions and reference answer contexts by adopting immediate and summary memory, maintaining smooth communication, achieving smooth transitions between long rounds of questions, and alleviating the length restrictions between the diagnostic questions and reference answer contexts.

8. The standardized patient agent-driven physician training system according to claim 1, characterized in that: The forming of a question pool based on the information in the medical data set includes: the standardized doctor intelligent unit obtains the patient's medical data set, formulates corresponding diagnostic questions based on the symptoms, medical examinations and living habits information in the medical data set, and forms these corresponding diagnostic questions into a general question pool.

9. A doctor training method driven by a standardized patient agent based on a large language model, characterized in that: The following steps are involved: Collect real-world medical data sets as input, perform fuzzy processing on the medical data sets to generate patient complaints, and the standard doctor agent unit performs department triage based on the complaints; After receiving the patient's complaint, the standard patient agent unit and the standard doctor agent unit initiate multiple rounds of dialogue. The standard doctor agent unit raises diagnostic questions, and the standard patient agent unit generates reference answers to obtain diagnostic results. The standard patient agent unit and the standard doctor agent unit are combined with real-world medical cases and evolve through multiple iterative cycles of diagnostic questions-reference answers, simulating the training of real-world doctors in various departments for real-world doctor training.

Images