A role-driven method for synthesizing medical text data

Abstract

The present invention discloses a method for synthesizing medical text data based on role-driven, which relates to the fields of text data synthesis and artificial intelligence technology, and includes the following steps: S100, generating original questions; S200, screening original questions; S300, synthesizing medical text data. The role description and interest content generation of the present invention enhance the adaptability to the needs of different medical roles, utilize text encoding models and vector databases to achieve semantic deduplication and relevance evaluation, remove semantically repeated content, utilize large language models to accurately evaluate relevance, improve data quality and diversity, and generate variant questions and answers to enrich the diversity and practicality of medical text data. The present invention improves the personalization and pertinence of medical text data, enhances the quality and deduplication efficiency of medical text data, enriches the types of questions, optimizes the generated content, and comprehensively covers various information needs of medical text data.

Description

Technical Field

[0001] The present invention relates to the field of text data synthesis and artificial intelligence technology, and in particular to a role-driven medical text data synthesis method. Background Art

[0002] With the acceleration of the digitalization process in the medical field, the demand for high-quality medical text data is growing. This data is not only used for diagnosis, treatment and nursing in clinical practice, but also widely used in medical research, drug development and health management. However, there are many challenges in obtaining high-quality medical text data: manual annotation is costly and time-consuming, and it is difficult to meet the needs of large-scale applications; traditional text generation methods are insufficient in personalization and professionalism, and it is difficult to adapt to the specific needs of different medical roles.

[0003] Existing solutions mainly focus on automated or semi-automated text generation systems, using natural language processing (NLP) technology and large language models (such as BERT, GPT, etc.) to generate medical texts. Although these systems have made some progress, they still have some limitations. For example, existing methods usually rely on preset templates or fixed knowledge bases, which makes it difficult for them to generate personalized questions and answers for different medical roles (such as doctors, nurses, patients, etc.). In addition, the generated questions and answers often lack a high degree of relevance to the specific content of interest, resulting in low data quality. At the same time, due to the lack of an effective semantic deduplication mechanism, repeated or similar content often appears in the generated medical texts, reducing the diversity and practicality of the data. Moreover, the medical texts generated by existing systems are relatively simple and fail to fully consider the diverse needs in different scenarios. In complex medical environments, the patient's condition needs to be evaluated from multiple perspectives. Finally, although the automated system reduces some of the workload of manual annotation, a lot of manual review and correction is still required when generating high-quality medical texts, especially when sensitive information or complex cases are involved, which not only increases time and cost, but also limits the large-scale application of the system.

[0004] In recent years, with the development of large language models (LLMs), some advanced medical text generation systems have begun to combine the capabilities of large models to try to solve the above problems. Med-PaLM launched by Google can perform well in a variety of medical tasks, such as medical record summaries, diagnostic recommendations, etc., but it still relies on a large amount of pre-training data and does not perform well when generating personalized content. Meta's Galactica is a multimodal large model. Although it performs well in cross-modal tasks, it still faces challenges in generalization and relevance in medical text generation. Google's Flan-T5 can perform well in a variety of tasks after extensive fine-tuning, but when generating medical text, it still relies on fixed templates and preset rules, and it is difficult to cope with complex medical scenarios.

[0005] In summary, the data synthesis methods of medical texts in the prior art mainly have the following disadvantages:

[0006] 1. Lack of role-driven personalized processing. Existing data synthesis methods usually do not take into account the different concerns and needs of different roles (such as doctors, nurses, patients, etc.) in medical texts. This results in the lack of targetedness of the generated data and the inability to effectively simulate the diverse needs in actual medical scenarios;

[0007] 2. Data redundancy and duplication: Existing technologies often do not have an effective deduplication mechanism when generating questions and answers, resulting in a large amount of duplicate content in the generated data. This not only wastes storage resources, but also reduces the quality and practicality of the data;

[0008] 3. Imperfect screening mechanism. After generating questions and answers, the existing technology lacks an effective screening mechanism to ensure the accuracy and relevance of the content, which easily generates irrelevant or inaccurate information, affecting the practicality and reliability of the data;

[0009] 4. The question types are single and lack optimization. When generating questions, the existing technology does not classify and optimize the question types, resulting in a lack of diversity in the generated questions, which cannot fully cover the various information needs in medical texts. At the same time, the generated content also lacks rewriting and optimization for specific roles, which affects the practicality and effectiveness of the data.

[0010] Therefore, technicians in this field are committed to developing a role-driven medical text data synthesis method. Summary of the invention

[0011] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to improve the personalization and pertinence of data, enhance the quality and deduplication efficiency of data, enrich the types of questions, and optimize the generated content.

[0012] The applicant analyzed that the existing technology failed to fully consider the specific needs and concerns of different medical roles, resulting in the lack of pertinence and practicality of the generated data, and was unable to effectively simulate real medical scenarios. The existing technology lacks an effective deduplication mechanism when generating questions and answers, resulting in a large amount of redundant and repeated content in the generated data, wasting storage resources and reducing the quality of the data. When generating questions, the existing technology did not classify and optimize the question types, resulting in a lack of diversity in the generated questions, and unable to fully cover the various information needs in medical texts. At the same time, the generated content also lacks rewriting and optimization for specific roles, affecting the practicality and effectiveness of the data.

[0013] In response to the above-mentioned defects of the prior art, the applicant introduces role-driven personalized processing, slices the documents according to the occupations and concerns of different medical roles and generates interest content, ensuring that the generated questions and answers can accurately reflect the needs of different roles. At the same time, by analyzing the document content from multiple angles, richer and more targeted interest content is generated, thereby improving the practicality and simulation of the data; the applicant adopts a multi-level deduplication mechanism, combined with semantic deduplication and relevance deduplication methods, to ensure that the generated questions and answers are not repeated at the semantic level and are highly relevant to the interest content; the applicant classifies the interest content according to preset question types to ensure that the generated questions cover multiple types (such as diagnosis, treatment, nursing, etc.), thereby comprehensively covering various information needs in medical texts. In one embodiment of the present invention, a role-driven medical text data synthesis method is provided, comprising the following steps:

[0014] S100, generating original questions, parsing collected medical information documents, extracting key information fragments, generating role descriptions based on the occupations and concerns of preset medical roles, generating original interest content for key information fragments based on the role descriptions and performing semantic deduplication, generating and classifying interest content, and generating original questions using a large language model;

[0015] S200, screening the original question, performing semantic deduplication on the original question to obtain the original generated question, evaluating and screening the relevance between the original generated question and the content of interest to obtain the generated question, rewriting the generated question using the large language model and screening generalized questions to obtain the targeted generated question;

[0016] S300, medical text data synthesis, through role description, use the large language model to rewrite the targeted generated questions and interest content, generate variant questions and answers, use the large language model and re-ranking model to score and screen the variant questions and answers, and synthesize medical text data.

[0017] Optionally, in the role-driven medical text data synthesis method in the above embodiment, the large language model uses Qwen2.5-72B-Instruct.

[0018] Optionally, in the role-driven medical text data synthesis method in any of the above embodiments, the reranking model uses BGE-Rerank-Large.

[0019] Optionally, in the role-driven medical text data synthesis method in the above embodiment, step S100 includes:

[0020] S110, parsing medical information documents, using a document parser to split the collected medical information documents and extract key information fragments;

[0021] S120, generating interest content, generating a role description according to the occupation and focus of the preset medical role, and generating original interest content for key information fragments based on the role description;

[0022] S130, semantically de-duplicating the interest content, using a text encoding model and a vector database to encode the original interest content and remove semantically repeated content, obtain the de-duplicated original interest content, obtain the interest content and save it in the vector database;

[0023] S140, classifying the interest content, using a large language model to classify the interest content according to the preset question type and the question type classification prompt words;

[0024] S150, generating the original question, generating prompt words based on the interest content classification and role description combined with the original question, and using the large language model to generate the original question.

[0025] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the text encoding model uses BGE M3-Embedding.

[0026] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the vector database uses Milvus.

[0027] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, step S110 includes:

[0028] S111, document slicing, slicing the collected medical information documents according to line breaks and serial number codes to improve the efficiency and accuracy of subsequent processing;

[0029] S112, extracting key information fragments. For the sliced ​​text, calculate the string length, take 512 characters as the upper limit, process slice by slice, and store the excess part as a new slice. All slice texts constitute the key information fragment.

[0030] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, step S120 includes:

[0031] S121, role description generation, generating prompt words based on the occupation and focus of the preset medical role combined with the medical role description, and using the large language model to generate the role description;

[0032] S122, generating interest content, based on the role description, generating prompt words in combination with the interest content, and using a large language model to generate original interest content for key information fragments.

[0033] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the preset medical roles include doctors, nurses, and patients.

[0034] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the medical role description generates prompt words including:

[0035] Role name, including:

[0036] Doctor

[0037] Nurse

[0038] Patient

[0039] Pharmacist

[0040] Caregiver

[0041] Medical Researcher

[0042] Hospital Administrator

[0043] Professional background, including:

[0044] Educational background, including Doctor of Medicine, Bachelor of Nursing, Master of Pharmacy;

[0045] Professional qualifications, including practicing physician certificate, registered nurse certificate, and pharmacist qualification certificate;

[0046] Work experience, including years of clinical experience, specialist training experience, and participation in scientific research projects;

[0047] Job responsibilities include:

[0048] Physicians, including those responsible for diagnosing diseases, formulating treatment plans, performing surgeries, and supervising residents;

[0049] Nurses, including those responsible for patient care, executing doctor's orders, monitoring patient conditions, and assisting doctors in diagnosis and treatment;

[0050] Patients, including providing medical history, describing symptoms, following doctor's orders, and cooperating with treatment;

[0051] Pharmacists, including those responsible for drug dispensing, drug consultation, drug management, and drug efficacy monitoring;

[0052] Nursing staff, including those responsible for daily care, life care, rehabilitation training, and psychological support;

[0053] Medical researchers, including those responsible for conducting clinical trials, data analysis, writing research reports, and publishing academic papers;

[0054] Hospital management personnel, including those responsible for hospital operations, resource management, quality control, and policy making;

[0055] Common concerns include:

[0056] Doctors, including the accuracy of disease diagnosis, treatment effectiveness, patient prognosis, and the latest medical advances;

[0057] Nurses, including patient comfort, quality of care, patient safety, and optimization of nursing processes;

[0058] patients, including the causes of illness, treatment, recovery time, cost issues, and quality of life;

[0059] Pharmacists, including drug safety, effectiveness, drug interactions, and patient medication compliance;

[0060] caregivers, including patients’ daily living needs, recovery progress, mental health, and family support;

[0061] Medical researchers, including the selection of research topics, rationality of experimental design, reliability of data, and application value of research results;

[0062] Hospital managers, including hospital operational efficiency, quality of medical services, patient satisfaction, and medical cost control;

[0063] Typical behaviors include:

[0064] Doctors, including communicating with patients, reviewing medical records, prescribing, conducting consultations, and attending academic conferences;

[0065] Nurses, including recording patients’ vital signs, administering medications, changing dressings, accompanying patients, and conducting health education;

[0066] Patients, including taking medication on time, regular follow-up visits, recording symptom changes, and maintaining good living habits;

[0067] Pharmacists, including reviewing prescriptions, dispensing medications, answering patients’ medication questions, and conducting adverse drug reaction monitoring;

[0068] Nursing staff, including helping patients eat, bathe, turn over, conduct rehabilitation training, and provide psychological support;

[0069] Medical researchers, including designing experimental plans, collecting and analyzing data, writing research reports, and applying for scientific research funds;

[0070] Hospital management, including conducting management meetings, reviewing medical records, developing hospital rules and regulations, and handling emergencies;

[0071] Personalized needs, including:

[0072] Doctors, including those who want to obtain the latest medical knowledge, advanced diagnosis and treatment techniques, and efficient case management systems;

[0073] Nurses, including those who want to improve their nursing skills, improve their work environment, and gain more career development opportunities;

[0074] Patients, including those who want clearer explanations, more considerate care, and more convenient medical services;

[0075] Pharmacists, including those who want to learn the latest drug information, master more drug management skills, and improve patients' medication safety;

[0076] Nursing staff, including the desire for more professional training, better welfare benefits, and more social recognition;

[0077] Medical researchers, including those who hope to obtain more scientific research resources, broader cooperation opportunities, and higher academic influence;

[0078] Hospital managers, including those who hope to improve the hospital's management level, optimize resource allocation, and enhance the hospital's social reputation.

[0079] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the interest content generation prompt words include:

[0080] Medical role description, including descriptive information of the specified medical role;

[0081] Key information fragments, including one or more key information fragments, as the basis for generating interesting content;

[0082] Generate requirements, including:

[0083] Content relevance: the generated interesting content is highly relevant to the key information fragments provided, avoiding irrelevant or generalized information;

[0084] Role-specific: Based on the role's professional background, job responsibilities, common concerns, and typical behaviors, the generated content should be targeted and able to meet the specific needs of the role;

[0085] Practicality: the generated content should have practical application value and help the role better understand and apply relevant information in daily work;

[0086] Diversity, generating content of interest from multiple different perspectives, covering different aspects such as diagnosis, treatment, nursing, and drug management;

[0087] Generate rules, including:

[0088] Doctors, including:

[0089] Diagnosis-related: extract the parts related to disease diagnosis from the key information fragments and generate content about etiology analysis, diagnostic methods, auxiliary examinations, etc.

[0090] Treatment-related, generating information about the choice of treatment options, details of surgical procedures, and evaluation of the effectiveness of drug therapy;

[0091] Prognosis-related, generating predictions about patient prognosis, rehabilitation suggestions, long-term management, etc.;

[0092] Latest developments, generating content about the latest medical research, clinical guidelines, treatments, etc.;

[0093] Nurses, including:

[0094] Nursing-related: generating information on the focus of patient care, optimization of nursing processes, and improvement of nursing skills;

[0095] Monitoring-related, generating specifications for monitoring of patients’ vital signs, observing changes in their condition, and recording nursing care;

[0096] Safety-related: generating safeguards, risk prevention, and emergency plans for patient safety;

[0097] Health education, generating health education, life guidance, and psychological support for patients;

[0098] Patients, including:

[0099] Symptom description, generating detailed descriptions of symptoms, methods of self-observation, and symptom trends;

[0100] Treatment compliance, generating information on how to take medication correctly, follow up on time, and follow doctor's orders;

[0101] Lifestyle, generating suggestions on diet, exercise, daily routine, etc. to help patients improve their quality of life;

[0102] Psychological support, generating information on how to cope with the psychological stress caused by the disease and maintain a positive attitude;

[0103] Pharmacists, including:

[0104] Drug management, generating precautions on drug storage, distribution, and use;

[0105] Drug consultation, which generates content about the drug’s mechanism of action, side effects, interactions, etc., to help patients and healthcare professionals better understand the drug;

[0106] Efficacy monitoring: generating evaluation methods for drug efficacy, monitoring and treatment of adverse reactions;

[0107] Medication compliance, generating information on how to improve patients’ medication compliance and reduce missed or misplaced doses;

[0108] Nursing staff, including:

[0109] Daily life care, generating care suggestions on the patient's diet, cleaning, mobility, etc.;

[0110] Rehabilitation training: generating plans, implementation methods, and effect evaluations for rehabilitation training;

[0111] Psychological support, generating information on how to provide psychological support and relieve patients’ anxiety and fear;

[0112] Family support, which generates recommendations on how to communicate with patients’ families and provide family care;

[0113] Medical researchers, including:

[0114] Research design, generating information on the selection of research topics, the rationality of experimental design, and the methods of data collection;

[0115] Data analysis, generating information on data processing, statistical analysis, and interpretation of results;

[0116] Academic exchanges, generating information on how to write research reports, publish academic papers, and participate in academic conferences;

[0117] Transformation of research results: generating information about the application prospects of research results and the possibility of clinical transformation;

[0118] Hospital management staff, including:

[0119] Operation management, generating information about the hospital’s daily operations, resource allocation, and work efficiency improvement;

[0120] Quality control, generating information on the monitoring of medical service quality, the improvement of patient satisfaction, and the prevention of medical accidents;

[0121] Policy formulation, generating information on the formulation of hospital rules and regulations, implementation and adjustment of policies;

[0122] Emergency handling, generation of emergency plans, crisis management, communication and coordination for emergencies.

[0123] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, step S130 includes:

[0124] S131, encoding the original interest content, using the text encoding model and the vector database, encoding each original interest content into a 1×1024-dimensional feature vector, obtaining the original interest content feature vector, and saving it in the vector database;

[0125] S132, feature vector deduplication, each original interest content feature vector is compared with other original interest content feature vectors in the vector database using a feature vector similarity formula for similarity, and deduplication is performed. If the similarity is greater than or equal to a specified feature similarity threshold, it is discarded otherwise;

[0126] S133: generating interest content, saving the original interest content after deduplication to a vector database, and obtaining the interest content.

[0127] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the feature vector similarity formula is as follows:

[0128] ;

[0129] A and B represent the two feature vectors to be compared, and 1024 represents the maximum dimension of each feature vector. and Respectively represent the two feature vectors A and B to be compared in the The value of the dimension, It represents the sum of the corresponding elements of the eigenvectors A and B in each dimension. and Respectively represent the modulus lengths of the eigenvectors A and B to be compared, That is, feature vector similarity;

[0130] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the feature similarity threshold is specified to be equal to 0.5.

[0131] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the question type classification prompt words include preset question types and corresponding classification rules and classification requirements.

[0132] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the preset question types and corresponding classification rules include:

[0133] Diagnosis category, classification rules include:

[0134] Etiology analysis, including the cause, pathogenesis, and risk factors of the disease;

[0135] Auxiliary examinations, including imaging examinations, laboratory tests, pathological examinations and other diagnostic methods;

[0136] Diagnostic criteria, including diagnostic criteria, clinical manifestations, and differential diagnosis of the disease;

[0137] Early screening, including early detection of diseases, screening methods, and high-risk groups;

[0138] Treatment category, classification rules include:

[0139] Surgical treatment, including surgical procedures, postoperative care, and surgical complications;

[0140] Drug treatment, including drug selection, dosage, timing of medication, and drug interactions;

[0141] Physical therapy, including physical therapy, rehabilitation training, and physical therapy equipment;

[0142] Psychotherapy, including psychological counseling, psychological intervention, and emotion management;

[0143] Nursing category, classification rules include:

[0144] Daily care, including daily care, cleaning, and mobility of patients;

[0145] Condition monitoring, including monitoring of patients' vital signs, observation of condition changes, and nursing records;

[0146] Safety precautions, including patient safety measures, risk prevention, and emergency plans;

[0147] Health education, including health education, life guidance, and psychological support for patients;

[0148] Drug management category, classification rules include:

[0149] Drug selection, including the basis for drug selection, indications, and contraindications;

[0150] Medication use, including drug administration, dosage adjustment, and frequency of medication;

[0151] Drug efficacy, including evaluation of drug therapeutic effects and monitoring of adverse reactions;

[0152] Medication compliance, including how to improve patients’ medication compliance and reduce missed or mis-dosed doses;

[0153] Rehabilitation category, classification rules include:

[0154] Rehabilitation training, including planning, implementation methods, and effect evaluation of rehabilitation training;

[0155] Functional recovery, including functional recovery and improvement of patients’ quality of life;

[0156] Psychological rehabilitation, including patients’ psychological rehabilitation, emotional regulation, and social integration;

[0157] Long-term management, including the patient's long-term rehabilitation plan, follow-up arrangements, and health management;

[0158] Prevention category, classification rules include:

[0159] Disease prevention, including preventive measures for diseases, vaccinations, and healthy lifestyles;

[0160] Risk control, including disease risk factor control, early intervention, and health monitoring;

[0161] Public health, including public health policy, community health promotion, and population health management;

[0162] Health education category, classification rules include:

[0163] Disease knowledge, including the causes, symptoms, and treatments of diseases, helps patients and the public understand diseases;

[0164] Self-management, including how patients manage themselves, follow doctor's orders, and maintain a healthy lifestyle;

[0165] Mental health, including mental health knowledge, emotion management, and stress coping;

[0166] Family support, including how to communicate with family members, obtain family support, and face the disease together.

[0167] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the classification requirements include:

[0168] Accuracy, each interest content is correctly classified into the category most relevant to its topic, avoiding misclassification;

[0169] Comprehensiveness, all the content of interest is reasonably categorized without omission or duplication;

[0170] Fine-grained: Each preset question type is further subdivided into more specific subcategories to improve the precision of classification;

[0171] Explainability: Provide a brief explanation for each preset question type and subcategory, explaining why the content of interest belongs to that category, helping users understand the basis for the classification.

[0172] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the original question generates prompt words including:

[0173] Role description, including information on the specified role description;

[0174] Interest content;

[0175] Generate requirements, including:

[0176] Content relevance: the questions generated are highly relevant to the content of interest provided, avoiding irrelevant or general questions;

[0177] Role-specific: Based on the professional background, job responsibilities, common concerns and typical behaviors of the medical role, the questions generated should be targeted to meet the specific needs of the role;

[0178] Practicality: the questions generated should have practical application value, helping the role to better understand and apply relevant information in daily work;

[0179] Diversity, generating multiple questions from different angles, covering different aspects including diagnosis, treatment, nursing, and medication management;

[0180] Answerability: The generated questions have clear answers or solutions, which facilitates subsequent answers and discussions.

[0181] Optionally, in the role-driven medical text data synthesis method in any of the above embodiments, step S200 includes:

[0182] S210, semantic deduplication of the original question, using the text encoding model and the vector database to encode the original question into a database, and deduplication is performed to obtain the original generated question;

[0183] S220, relevance evaluation, based on the original generated question and the interest content, using the large language model to perform relevance discrimination and relevance screening to obtain the generated question;

[0184] S230, rewriting the generated question, rewriting the generated question using the large language model according to the role description and the question rewriting prompt words, so that the question conforms to the expression habits and information needs of the specific role;

[0185] S240, screening of generalized questions: using a large language model to evaluate the rewritten generation questions, screening out generation questions with generalization, retaining and saving generation questions that are not generalized, and obtaining targeted generation questions.

[0186] Optionally, in the role-driven medical text data synthesis method in any of the above embodiments, generating questions for rewriting includes tone rewriting and questioning style rewriting.

[0187] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the question rewriting prompt words include:

[0188] Role description, including information on the specified role description;

[0189] Generate questions, including already generated questions;

[0190] Rewrite requirements, including:

[0191] Adjust the tone of the question according to the role's professional background and communication habits;

[0192] Word choice: Choose professional terms or lay language that is appropriate to the role, ensuring that questions are both accurate and understandable; use more professional medical terms for doctors and researchers, and more lay language for patients and caregivers;

[0193] Structural optimization: Optimize the structure of questions to make them more concise and easy to understand and answer. Avoid lengthy or complex sentence structures and ensure that questions are clear and direct.

[0194] Personalization: adjust the content of questions to make them more targeted according to the specific needs and concerns of the role; doctors are more concerned about diagnosis and treatment plans, while patients are more concerned about symptoms and prognosis (the prediction of the possible development trend of a disease or condition);

[0195] Be polite and respectful, ensuring questions are asked in a polite and respectful manner, especially when involving patients or their families, and avoid using overly direct or blunt terms;

[0196] Rewrite the rules, including:

[0197] Doctors, including:

[0198] Professionalism, use medical terminology, maintain authority and rigor;

[0199] Be concise and clear. Questions should be short and punchy, and get to the point.

[0200] Pay attention to details, focusing on specific details such as diagnosis, treatment, and prognosis;

[0201] Nurses, including:

[0202] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0203] Operational: questions should be specific and actionable to help nurses better perform nursing tasks;

[0204] Safety, emphasizing the safety and health of patients;

[0205] Patients, including:

[0206] Be cordial and friendly. The tone should be cordial, friendly, and easy to understand;

[0207] Keep it simple and clear, avoid using complex medical terms, and try to explain in plain language;

[0208] Focus on yourself, and questions should revolve around the patient's symptoms, treatment, and prognosis;

[0209] Pharmacists, including:

[0210] Professionalism, use professional terminology in drug administration, maintain authority and accuracy;

[0211] Be detailed and comprehensive, and questions should cover the selection, use, and monitoring of medications;

[0212] Safety reminders, emphasizing the safety and precautions of drugs;

[0213] Nursing staff, including:

[0214] Operational: questions should be specific and actionable to help caregivers better perform daily care tasks;

[0215] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0216] Practicality: the questions should have practical application value and help nursing staff improve their work efficiency;

[0217] Medical researchers, including:

[0218] Scientific rigor, scientific and rigorous tone, focusing on the accuracy of research methods and data;

[0219] Detailed and comprehensive, with questions covering research design, data analysis, and interpretation of results;

[0220] Innovation: Encourage new research directions and innovative ideas.

[0221] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, step S210 includes:

[0222] S211, encoding the original question, using the text encoding model, encoding each original question into a 1×1024-dimensional feature vector, obtaining the original question feature vector, and saving it in the vector database;

[0223] S212, deduplication of original question feature vectors, comparing each original question feature vector with other original question feature vectors in the vector database using a feature vector similarity formula to perform deduplication processing, and discarding them if the similarity is greater than or equal to a specified question similarity threshold;

[0224] S213, the original generated question is obtained, and the original question after deduplication is saved to the vector database to obtain the original generated question.

[0225] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the question similarity threshold is specified to be equal to 0.5.

[0226] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, step S220 includes:

[0227] S221, relevance assessment, based on the original generated questions and interest content, combined with the relevance assessment prompt words, using the large language model to perform relevance assessment and output the comprehensive relevance;

[0228] S222, correlation screening, traversing all original generated questions, and discarding the original generated questions whose comprehensive correlation is less than a specified correlation threshold;

[0229] S223, obtaining the generated problem, saving the original generated problem after screening, and obtaining the generated problem.

[0230] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the relevance evaluation prompt words include:

[0231] Original generated questions, including providing one or more generated questions as a basis for relevance assessment;

[0232] Interest content, including providing one or more passages of interest content as reference text for assessment;

[0233] Assessment requirements, including:

[0234] Semantic similarity: evaluate the semantic similarity between the original generated question and the content of interest to ensure that the meanings expressed by the two are consistent or similar;

[0235] Topic consistency: evaluate whether the original generated questions and the content of interest revolve around the same topic to avoid disconnection between the original generated questions and the content;

[0236] Contextual relevance, evaluating whether the original generated question is closely related to the context of the content of interest, ensuring that the original generated question can be naturally derived from the content;

[0237] Practicality: evaluate whether the original generated questions have practical application value and whether they can help users better understand and apply the content of interest;

[0238] Answerability, assessing whether the original generated question has a clear answer or solution to facilitate subsequent answers and discussions.

[0239] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the value of the comprehensive correlation is a decimal of 0-1, and the larger the value of the comprehensive correlation is, the stronger the correlation is, 1 represents complete correlation, and 0 represents complete irrelevance.

[0240] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the specified relevance threshold is equal to 0.5.

[0241] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, step S240 includes:

[0242] S241, generalization question identification, using a large language model, based on the question generalization evaluation prompt words, evaluate the rewritten generated question to obtain the evaluation result, that is, whether it has generalization;

[0243] S242, screening of generalization problems, screening out the generation problems with generalization, and retaining the generation problems without generalization;

[0244] S243. Targeted generation problems are saved, generation problems that are not generalizable are saved, and targeted generation problems are obtained.

[0245] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the question generalization evaluation prompt words include:

[0246] Generate questions, including providing generative questions as a basis for assessment;

[0247] Assessment requirements, including:

[0248] Specificity: assess whether the question involves specific objects, situations or details, and avoid asking questions that are too broad;

[0249] Application scenario: whether the problem has a clear application scenario and can be applied in actual work;

[0250] Actionability: assess whether the problem has a clear answer or solution to facilitate subsequent responses and discussions;

[0251] Practical value: whether the problem has practical application value and can help users solve specific problems or obtain useful information;

[0252] Assessment methods, including:

[0253] Large language model evaluation: Use the large language model to analyze the generated questions to determine whether they are generalizable. Evaluation indicators include the breadth, depth, specificity, and application scenarios of the questions.

[0254] Keyword matching: Through keyword matching, we can identify whether the question contains specific objects, situations or details, for example, whether the question mentions a specific disease, drug or patient group;

[0255] Contextual relevance: assess whether the question is closely related to the content of interest, to avoid questions being out of touch with background information;

[0256] Manual labeling: Combined with the results of manual labeling, further adjust and optimize the evaluation results to ensure the accuracy of screening;

[0257] Filtering rules include:

[0258] The standard is retained. The problem has clear objects, situations or details, can be applied in actual work, and has practical value;

[0259] Screening criteria: questions that are too broad, lack specific content, or have no practical application scenarios will be screened out.

[0260] Optionally, in the role-driven medical text data synthesis method in any of the above embodiments, step S300 includes:

[0261] S310, variant question and answer generation, through role description, combined with variant question and answer generation prompt words, using the large language model to rewrite the targeted generated questions and answer interest content, to generate variant questions and answers;

[0262] S320, scoring and screening, uses a large language model and a re-ranking model to score and screen variant questions and answers and synthesize medical text data.

[0263] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the variant question and answer generation prompt words include:

[0264] Role description, including information on the specified role description;

[0265] Generate targeted questions;

[0266] Corresponding document answers, including targeted generation of interesting content corresponding to the question;

[0267] Generate requirements, including:

[0268] Diversity: Generate multiple variant questions and answers from different angles, covering different scenarios, details, and expressions;

[0269] Targeted: Each variant is targeted to meet the specific needs of the role based on the role's professional background, job responsibilities, common concerns and typical behaviors;

[0270] Practicality: the generated variants have practical application value, helping the characters better understand and apply relevant information in their daily work;

[0271] Consistency: The variant questions and answers should keep the same core content as the original questions and answers, and avoid changing the core facts or conclusions;

[0272] Innovation: Encourage the generation of novel variant questions, propose new perspectives or application scenarios, and increase the richness and depth of data;

[0273] Generate rules, including:

[0274] Doctors, including:

[0275] Change of perspective, rewriting the question from different diagnostic, therapeutic, and prognostic perspectives;

[0276] Scenario expansion: introducing specific clinical cases or special situations to make the problem more realistic;

[0277] Details supplementation, adding more specific details, including the patient's medical history, test results, etc., to make the question more detailed;

[0278] Professional terminology: Use more professional medical terminology to maintain authority and rigor;

[0279] Nurses, including:

[0280] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0281] Operational: questions should be specific and actionable to help nurses better perform nursing tasks;

[0282] Safety, emphasizing the safety and health of patients and ensuring risk prevention during the nursing process;

[0283] Details supplement, adding more specific care details, such as monitoring frequency and care records;

[0284] Patients, including:

[0285] Be cordial and friendly. The tone should be cordial, friendly, and easy to understand;

[0286] Keep it simple and clear, avoid using complicated medical terms, and use layman's language to explain;

[0287] Focus on yourself, and questions should revolve around the patient's symptoms, treatment, and prognosis;

[0288] Detail supplementation, adding more details related to the patient's daily life, including diet and exercise;

[0289] Pharmacists, including:

[0290] Professionalism, use professional terminology in drug administration, maintain authority and accuracy;

[0291] Detailed and comprehensive, with questions covering the selection, use, and monitoring of medications;

[0292] Safety reminders, emphasizing the safety and precautions of drugs to ensure safe use of drugs;

[0293] Details supplement, adding more specific details about drug use, including dosage adjustments and adverse reactions;

[0294] Nursing staff, including:

[0295] Practicality: questions are specific and actionable, helping caregivers to better perform daily care tasks;

[0296] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0297] Practicality: the questions have practical application value and help nursing staff improve their work efficiency;

[0298] Supplement details and add more specific care details, including diet, cleaning, and rehabilitation training;

[0299] Medical researchers, including:

[0300] Scientific rigor, scientific and rigorous tone, focusing on the accuracy of research methods and data;

[0301] Detailed and comprehensive, with questions covering research design, data analysis, and interpretation of results;

[0302] Innovation: encourage new research directions and innovative ideas to increase the depth and breadth of research;

[0303] Details supplement, adding more specific details about the research design and implementation, including sample selection and experimental procedures;

[0304] Hospital management staff, including:

[0305] Decision-making oriented, decisive and methodical in tone, focusing on the feasibility and effectiveness of decisions;

[0306] A global perspective, where the problem is addressed from the perspective of the overall hospital operation, taking into account resource allocation, service quality, and policy formulation;

[0307] Risk management, emphasizing the importance of risk management and emergency plans to ensure the normal operation of the hospital;

[0308] Details are supplemented to add more specific details about hospital management, including process optimization, resource allocation, etc.

[0309] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, step S320 includes:

[0310] S321, based on the relevance calculation of the large language model, according to the variant question and answer and the interest content, combined with the variant question and answer content scoring prompt words, using the large language model to perform relevance judgment, and output a comprehensive score;

[0311] S322, relevance calculation based on the re-ranking model: input the variant question and answer and the interest content into the re-ranking model, perform relevance determination, and output the re-ranking relevance;

[0312] S323, correlation screening, calculate the average of the comprehensive score and the re-ranking correlation as the correlation, and discard the variant questions and answers whose correlation is less than the specified correlation threshold. The calculation formula is as follows:

[0313]

[0314] in Representative variant question and answer, Represents content of interest, Represents the scoring prompt words for variant question and answer content, represents the reordering association degree, represents the comprehensive score;

[0315] S324, generate and save questions in a targeted manner, save variant questions and answers and interest content retained after correlation screening, and synthesize medical text data.

[0316] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the variant question and answer content scoring prompt words include:

[0317] Variant problem;

[0318] Answer, the answer corresponding to the variant question;

[0319] Scoring requirements include:

[0320] Accuracy, assessing whether variant questions and answers are consistent with medical facts and clinical guidelines, and avoiding false or misleading information;

[0321] Relevance, assessing whether variant questions and answers are highly relevant to the content of interest and closely revolve around specific topics;

[0322] Operability: evaluate whether the answer has clear steps or solutions to facilitate its application in actual work;

[0323] Practical value: evaluate whether the variant questions and answers have practical application value and help users solve specific problems or obtain useful information;

[0324] Logical clarity, assessing whether the logic of questions and answers is clear, whether the expression is concise and clear, and whether long or complex sentence structures are avoided;

[0325] Professionalism, assessing whether the answer uses appropriate medical terminology, remains professional and authoritative, and is easy to understand;

[0326] Scoring methods include:

[0327] Large language model evaluation: using a large language model to encode variant questions and answers and calculate their scores on various dimensions, including accuracy, relevance, operability, practical value, logical clarity, and professionalism;

[0328] Manual labeling: Combined with the results of manual labeling, further adjust and optimize the scores to ensure the accuracy of the evaluation results;

[0329] Comprehensive score: The scores of each dimension are weighted averaged to obtain the final comprehensive score. The comprehensive score is a decimal between 0 and 1. The larger the comprehensive score, the stronger the correlation. 1 means complete correlation, and 0 means no correlation.

[0330] Filtering rules include:

[0331] Retention criteria: questions and answers with a comprehensive score greater than 0.5 are retained to ensure their excellent performance in terms of accuracy, relevance, operability and practical value;

[0332] Screening criteria: questions and answers with an overall score less than or equal to 0.5 will be screened out, especially those with obvious errors, irrelevant content or no practical application value.

[0333] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the reordering correlation value is a decimal of 0-1, and a larger reordering correlation value indicates a stronger correlation, 1 indicates complete correlation, and 0 indicates complete irrelevance.

[0334] Furthermore, in the role-driven medical text data synthesis method in the above embodiment, the specified relevance threshold is equal to 0.5.

[0335] The present invention introduces role description and interest content generation, enhances the adaptability to different medical role requirements, uses text encoding models and vector databases to achieve semantic deduplication and relevance evaluation, encodes and compares the original generated questions and interest content, removes semantically repeated content, uses a large language model to accurately evaluate relevance, improves data quality and diversity, and enriches the diversity and practicality of medical text data through variant question and answer generation. The present invention improves the personalization and pertinence of medical text data, enhances the quality and deduplication efficiency of medical text data, enriches the types of questions, optimizes the generated content, and comprehensively covers various information needs of medical text data.

[0336] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0337] Figure 1 is a flow chart illustrating a method for synthesizing medical text data based on role-driven according to an exemplary embodiment;

[0338] Figure 2 is a flow chart illustrating original question generation according to an exemplary embodiment;

[0339] Figure 3 is a flow chart illustrating original question screening according to an exemplary embodiment;

[0340] Figure 4 is a flowchart illustrating medical text data synthesis according to an exemplary embodiment. DETAILED DESCRIPTION

[0341] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0342] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is schematically exaggerated in some places in the drawings.

[0343] The inventors have designed a role-driven method for synthesizing medical text data, such as Figure 1 As shown, the following steps are included:

[0344] S100, original question generation, parsing the collected medical information documents, extracting key information fragments, generating role descriptions based on the occupations and concerns of the preset medical roles, generating original interest content for the key information fragments based on the role descriptions and performing semantic deduplication, generating interest content and classifying it, and using a large language model to generate original questions. The large language model uses Qwen2.5-72B-Instruct; Figure 2 As shown, specifically including:

[0345] S110, medical information document parsing, using a document parser to split the collected medical information documents and extract key information fragments; specifically including:

[0346] S111, document slicing, slicing the collected medical information documents according to line breaks and serial number codes to improve the efficiency and accuracy of subsequent processing;

[0347] S112, extracting key information fragments. For the sliced ​​text, calculate the string length, take 512 characters as the upper limit, process slice by slice, and store the excess part as a new slice. All slice texts constitute the key information fragment.

[0348] S120, generating interest content, generating role descriptions according to the occupations and concerns of the preset medical roles, and generating original interest content for key information fragments based on the role descriptions; specifically including:

[0349] S121, role description generation, based on the occupation and focus of the preset medical role, combined with the medical role description to generate prompt words, using the large language model to generate role descriptions, the preset medical roles include doctors, nurses, and patients; the prompt words generated by the medical role description include:

[0350] Role name, including:

[0351] Doctor

[0352] Nurse

[0353] Patient

[0354] Pharmacist

[0355] Caregiver

[0356] Medical Researcher

[0357] Hospital Administrator

[0358] Professional background, including:

[0359] Educational background, including Doctor of Medicine, Bachelor of Nursing, Master of Pharmacy;

[0360] Professional qualifications, including practicing physician certificate, registered nurse certificate, and pharmacist qualification certificate;

[0361] Work experience, including years of clinical experience, specialist training experience, and participation in scientific research projects;

[0362] Job responsibilities include:

[0363] Physicians, including those responsible for diagnosing diseases, formulating treatment plans, performing surgeries, and supervising residents;

[0364] Nurses, including those responsible for patient care, executing doctor's orders, monitoring patient conditions, and assisting doctors in diagnosis and treatment;

[0365] Patients, including providing medical history, describing symptoms, following doctor's orders, and cooperating with treatment;

[0366] Pharmacists, including those responsible for drug dispensing, drug consultation, drug management, and drug efficacy monitoring;

[0367] Nursing staff, including those responsible for daily care, life care, rehabilitation training, and psychological support;

[0368] Medical researchers, including those responsible for conducting clinical trials, data analysis, writing research reports, and publishing academic papers;

[0369] Hospital management personnel, including those responsible for hospital operations, resource management, quality control, and policy making;

[0370] Common concerns include:

[0371] Doctors, including the accuracy of disease diagnosis, treatment effectiveness, patient prognosis, and the latest medical advances;

[0372] Nurses, including patient comfort, quality of care, patient safety, and optimization of nursing processes;

[0373] patients, including the causes of illness, treatment, recovery time, cost issues, and quality of life;

[0374] Pharmacists, including drug safety, effectiveness, drug interactions, and patient medication compliance;

[0375] caregivers, including patients’ daily living needs, recovery progress, mental health, and family support;

[0376] Medical researchers, including the selection of research topics, rationality of experimental design, reliability of data, and application value of research results;

[0377] Hospital managers, including hospital operational efficiency, quality of medical services, patient satisfaction, and medical cost control;

[0378] Typical behaviors include:

[0379] Doctors, including communicating with patients, reviewing medical records, prescribing, conducting consultations, and attending academic conferences;

[0380] Nurses, including recording patients’ vital signs, administering medications, changing dressings, accompanying patients, and conducting health education;

[0381] Patients, including taking medication on time, regular follow-up visits, recording symptom changes, and maintaining good living habits;

[0382] Pharmacists, including reviewing prescriptions, dispensing medications, answering patients’ medication questions, and conducting adverse drug reaction monitoring;

[0383] Nursing staff, including helping patients eat, bathe, turn over, conduct rehabilitation training, and provide psychological support;

[0384] Medical researchers, including designing experimental plans, collecting and analyzing data, writing research reports, and applying for scientific research funds;

[0385] Hospital management, including conducting management meetings, reviewing medical records, developing hospital rules and regulations, and handling emergencies;

[0386] Personalized needs, including:

[0387] Doctors, including those who want to obtain the latest medical knowledge, advanced diagnosis and treatment techniques, and efficient case management systems;

[0388] Nurses, including those who want to improve their nursing skills, improve their work environment, and gain more career development opportunities;

[0389] Patients, including those who want clearer explanations, more considerate care, and more convenient medical services;

[0390] Pharmacists, including those who want to learn the latest drug information, master more drug management skills, and improve patients' medication safety;

[0391] Nursing staff, including the desire for more professional training, better welfare benefits, and more social recognition;

[0392] Medical researchers, including those who hope to obtain more scientific research resources, broader cooperation opportunities, and higher academic influence;

[0393] Hospital managers, including those who hope to improve the hospital's management level, optimize resource allocation, and enhance the hospital's social reputation.

[0394] S122, generating interest content, based on the role description, combined with the interest content to generate prompt words, using the large language model to generate original interest content for key information fragments; the interest content generation prompt words include:

[0395] Medical role description, including descriptive information of the specified medical role;

[0396] Key information fragments, including one or more key information fragments, as the basis for generating interesting content;

[0397] Generate requirements, including:

[0398] Content relevance: the generated interesting content is highly relevant to the key information fragments provided, avoiding irrelevant or generalized information;

[0399] Role-specific: Based on the role's professional background, job responsibilities, common concerns, and typical behaviors, the generated content should be targeted and able to meet the specific needs of the role;

[0400] Practicality: the generated content should have practical application value and help the role better understand and apply relevant information in daily work;

[0401] Diversity, generating content of interest from multiple different perspectives, covering different aspects such as diagnosis, treatment, nursing, and drug management;

[0402] Generate rules, including:

[0403] Doctors, including:

[0404] Diagnosis-related: extract the parts related to disease diagnosis from the key information fragments and generate content about etiology analysis, diagnostic methods, auxiliary examinations, etc.

[0405] Treatment-related, generating information about the choice of treatment options, details of surgical procedures, and evaluation of the effectiveness of drug therapy;

[0406] Prognosis-related, generating predictions about patient prognosis, rehabilitation suggestions, long-term management, etc.;

[0407] Latest developments, generating content about the latest medical research, clinical guidelines, treatments, etc.;

[0408] Nurses, including:

[0409] Nursing-related: generating information on the focus of patient care, optimization of nursing processes, and improvement of nursing skills;

[0410] Monitoring-related, generating specifications for monitoring of patients’ vital signs, observing changes in their condition, and recording nursing care;

[0411] Safety-related: generating safeguards, risk prevention, and emergency plans for patient safety;

[0412] Health education, generating health education, life guidance, and psychological support for patients;

[0413] Patients, including:

[0414] Symptom description, generating detailed descriptions of symptoms, methods of self-observation, and symptom trends;

[0415] Treatment compliance, generating information on how to take medication correctly, follow up on time, and follow doctor's orders;

[0416] Lifestyle, generating suggestions on diet, exercise, daily routine, etc. to help patients improve their quality of life;

[0417] Psychological support, generating information on how to cope with the psychological stress caused by the disease and maintain a positive attitude;

[0418] Pharmacists, including:

[0419] Drug management, generating precautions on drug storage, distribution, and use;

[0420] Drug consultation, which generates content about the drug’s mechanism of action, side effects, interactions, etc., to help patients and healthcare professionals better understand the drug;

[0421] Efficacy monitoring: generating evaluation methods for drug efficacy, monitoring and treatment of adverse reactions;

[0422] Medication compliance, generating information on how to improve patients’ medication compliance and reduce missed or misplaced doses;

[0423] Nursing staff, including:

[0424] Daily life care, generating care suggestions on the patient's diet, cleaning, mobility, etc.;

[0425] Rehabilitation training: generating plans, implementation methods, and effect evaluations for rehabilitation training;

[0426] Psychological support, generating information on how to provide psychological support and relieve patients’ anxiety and fear;

[0427] Family support, which generates recommendations on how to communicate with patients’ families and provide family care;

[0428] Medical researchers, including:

[0429] Research design, generating information on the selection of research topics, the rationality of experimental design, and the methods of data collection;

[0430] Data analysis, generating information on data processing, statistical analysis, and interpretation of results;

[0431] Academic exchanges, generating information on how to write research reports, publish academic papers, and participate in academic conferences;

[0432] Transformation of research results: generating information about the application prospects of research results and the possibility of clinical transformation;

[0433] Hospital management staff, including:

[0434] Operation management, generating information about the hospital’s daily operations, resource allocation, and work efficiency improvement;

[0435] Quality control, generating information on the monitoring of medical service quality, the improvement of patient satisfaction, and the prevention of medical accidents;

[0436] Policy formulation, generating information on the formulation of hospital rules and regulations, implementation and adjustment of policies;

[0437] Emergency handling, generation of emergency plans, crisis management, communication and coordination for emergencies.

[0438] S130, semantic deduplication of the interest content, using a text encoding model and a vector database to encode the original interest content and remove semantically repeated content, obtain the original interest content after deduplication, obtain the interest content and save it to the vector database, the text encoding model uses BGE M3-Embedding, and the vector database uses Milvus; specifically including:

[0439] S131, encoding the original interest content, using the text encoding model and the vector database, encoding each original interest content into a 1×1024-dimensional feature vector, obtaining the original interest content feature vector, and saving it in the vector database;

[0440] S132, feature vector deduplication, each original interest content feature vector is compared with other original interest content feature vectors in the vector database using the feature vector similarity formula to perform deduplication processing, if the similarity is greater than or equal to the specified feature similarity threshold (0.5), otherwise it is discarded; the feature vector similarity formula is as follows:

[0441] ;

[0442] A and B represent the two feature vectors to be compared, and 1024 represents the maximum dimension of each feature vector. and Respectively represent the two feature vectors A and B to be compared in the The value of the dimension, It represents the sum of the corresponding elements of the eigenvectors A and B in each dimension. and Respectively represent the modulus lengths of the eigenvectors A and B to be compared, That is, feature vector similarity;

[0443] S133: generating interest content, saving the original interest content after deduplication to the vector database to obtain the interest content.

[0444] S140, classifying the interest content, using a large language model to classify the interest content according to the preset question type and the question type classification prompt words; the question type classification prompt words include the preset question type and the corresponding classification rules and classification requirements:

[0445] The preset question types and corresponding classification rules include:

[0446] Diagnosis category, classification rules include:

[0447] Etiology analysis, including the cause, pathogenesis, and risk factors of the disease;

[0448] Auxiliary examinations, including imaging examinations, laboratory tests, pathological examinations and other diagnostic methods;

[0449] Diagnostic criteria, including diagnostic criteria, clinical manifestations, and differential diagnosis of the disease;

[0450] Early screening, including early detection of diseases, screening methods, and high-risk groups;

[0451] Treatment category, classification rules include:

[0452] Surgical treatment, including surgical procedures, postoperative care, and surgical complications;

[0453] Drug treatment, including drug selection, dosage, timing of medication, and drug interactions;

[0454] Physical therapy, including physical therapy, rehabilitation training, and physical therapy equipment;

[0455] Psychotherapy, including psychological counseling, psychological intervention, and emotion management;

[0456] Nursing category, classification rules include:

[0457] Daily care, including daily care, cleaning, and mobility of patients;

[0458] Condition monitoring, including monitoring of patients' vital signs, observation of condition changes, and nursing records;

[0459] Safety precautions, including patient safety measures, risk prevention, and emergency plans;

[0460] Health education, including health education, life guidance, and psychological support for patients;

[0461] Drug management category, classification rules include:

[0462] Drug selection, including the basis for drug selection, indications, and contraindications;

[0463] Medication use, including drug administration, dosage adjustment, and frequency of medication;

[0464] Drug efficacy, including evaluation of drug therapeutic effects and monitoring of adverse reactions;

[0465] Medication compliance, including how to improve patients’ medication compliance and reduce missed or mis-dosed doses;

[0466] Rehabilitation category, classification rules include:

[0467] Rehabilitation training, including planning, implementation methods, and effect evaluation of rehabilitation training;

[0468] Functional recovery, including functional recovery and improvement of patients’ quality of life;

[0469] Psychological rehabilitation, including patients’ psychological rehabilitation, emotional regulation, and social integration;

[0470] Long-term management, including the patient's long-term rehabilitation plan, follow-up arrangements, and health management;

[0471] Prevention category, classification rules include:

[0472] Disease prevention, including preventive measures for diseases, vaccinations, and healthy lifestyles;

[0473] Risk control, including disease risk factor control, early intervention, and health monitoring;

[0474] Public health, including public health policy, community health promotion, and population health management;

[0475] Health education category, classification rules include:

[0476] Disease knowledge, including the causes, symptoms, and treatments of diseases, helps patients and the public understand diseases;

[0477] Self-management, including how patients manage themselves, follow doctor's orders, and maintain a healthy lifestyle;

[0478] Mental health, including mental health knowledge, emotion management, and stress coping;

[0479] Family support, including how to communicate with family members, obtain family support, and face the disease together;

[0480] Classification requirements include:

[0481] Accuracy, each interest content is correctly classified into the category most relevant to its topic, avoiding misclassification;

[0482] Comprehensiveness, all the content of interest is reasonably categorized without omission or duplication;

[0483] Fine-grained: Each preset question type is further subdivided into more specific subcategories to improve the precision of classification;

[0484] Explainability: Provide a brief explanation for each preset question type and subcategory, explaining why the content of interest belongs to that category, helping users understand the basis for the classification.

[0485] S150, generating the original question, generating prompt words based on the interest content classification and role description, combined with the original question, and using the large language model to generate the original question; the prompt words generated from the original question include:

[0486] Role description, including information on the specified role description;

[0487] Interest content;

[0488] Generate requirements, including:

[0489] Content relevance: the questions generated are highly relevant to the content of interest provided, avoiding irrelevant or general questions;

[0490] Role-specific: Based on the professional background, job responsibilities, common concerns and typical behaviors of the medical role, the questions generated should be targeted to meet the specific needs of the role;

[0491] Practicality: the questions generated should have practical application value, helping the role to better understand and apply relevant information in daily work;

[0492] Diversity, generating multiple questions from different angles, covering different aspects including diagnosis, treatment, nursing, and medication management;

[0493] Answerability: The generated questions have clear answers or solutions, which facilitates subsequent answers and discussions.

[0494] S200, screening the original questions, performing semantic deduplication on the original questions to obtain the original generated questions, evaluating and screening the relevance between the original generated questions and the content of interest to obtain the generated questions, rewriting the generated questions using the large language model and screening the generalized questions to obtain the targeted generated questions; Figure 3 As shown, specifically including:

[0495] S210: semantic deduplication of the original question, using the text encoding model and the vector database to encode the original question into the database, and deduplication to obtain the original generated question; specifically including:

[0496] S211, encoding the original question, using the text encoding model, encoding each original question into a 1×1024-dimensional feature vector, obtaining the original question feature vector, and saving it in the vector database;

[0497] S212, deduplication of original question feature vectors, comparing the similarity of each original question feature vector with other original question feature vectors in the vector database for deduplication, and discarding them if the similarity is greater than or equal to the specified question similarity threshold (0.5);

[0498] S213, the original generated question is obtained, and the original question after deduplication is saved to the vector database to obtain the original generated question.

[0499] S220, relevance assessment, based on the original generated question and the interest content, using the large language model to perform relevance discrimination and relevance screening to obtain the generated question; specifically including:

[0500] S221, relevance assessment, based on the original generated questions and interest content, combined with the relevance assessment prompt words, use the large language model to perform relevance assessment, and output the comprehensive relevance. The value of the comprehensive relevance is a decimal between 0 and 1. The larger the value of the comprehensive relevance, the stronger the relevance. 1 means complete relevance, and 0 means complete irrelevance. The relevance assessment prompt words include:

[0501] Original generated questions, including providing one or more generated questions as a basis for relevance assessment;

[0502] Interest content, including providing one or more passages of interest content as reference text for assessment;

[0503] Assessment requirements, including:

[0504] Semantic similarity: evaluate the semantic similarity between the original generated question and the content of interest to ensure that the meanings expressed by the two are consistent or similar;

[0505] Topic consistency: evaluate whether the original generated questions and the content of interest revolve around the same topic to avoid disconnection between the original generated questions and the content;

[0506] Contextual relevance, evaluating whether the original generated question is closely related to the context of the content of interest, ensuring that the original generated question can be naturally derived from the content;

[0507] Practicality: evaluate whether the original generated questions have practical application value and whether they can help users better understand and apply the content of interest;

[0508] Answerability, assessing whether the original generated question has a clear answer or solution to facilitate subsequent answers and discussions.

[0509] S222, correlation screening, traversing all original generated questions, and discarding the original generated questions whose comprehensive correlation is less than the specified correlation threshold (0.5).

[0510] S223, obtaining the generated problem, saving the original generated problem after screening, and obtaining the generated problem.

[0511] S230, rewrite the generated question. According to the role description and the question rewriting prompt words, the generated question is rewritten using the large language model, including tone rewriting and questioning method rewriting, so that the question conforms to the expression habits and information needs of the specific role; the question rewriting prompt words include:

[0512] Role description, including information on the specified role description;

[0513] Generate questions, including already generated questions;

[0514] Rewrite requirements, including:

[0515] Adjust the tone of the question according to the role's professional background and communication habits;

[0516] Word choice: Choose professional terms or lay language that is appropriate to the role, ensuring that questions are both accurate and understandable; use more professional medical terms for doctors and researchers, and more lay language for patients and caregivers;

[0517] Structural optimization: Optimize the structure of questions to make them more concise and easy to understand and answer. Avoid lengthy or complex sentence structures and ensure that questions are clear and direct.

[0518] Personalization: adjust the content of questions to make them more targeted according to the specific needs and concerns of the role; doctors are more concerned about diagnosis and treatment plans, while patients are more concerned about symptoms and prognosis (the prediction of the possible development trend of a disease or condition);

[0519] Be polite and respectful, ensuring questions are asked in a polite and respectful manner, especially when involving patients or their families, and avoid using overly direct or blunt terms;

[0520] Rewrite the rules, including:

[0521] Doctors, including:

[0522] Professionalism, use medical terminology, maintain authority and rigor;

[0523] Be concise and clear. Questions should be short and punchy, and get to the point.

[0524] Pay attention to details, focusing on specific details such as diagnosis, treatment, and prognosis;

[0525] Nurses, including:

[0526] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0527] Operational: questions should be specific and actionable to help nurses better perform nursing tasks;

[0528] Safety, emphasizing the safety and health of patients;

[0529] Patients, including:

[0530] Be cordial and friendly. The tone should be cordial, friendly, and easy to understand;

[0531] Keep it simple and clear, avoid using complex medical terms, and try to explain in plain language;

[0532] Focus on yourself, and questions should revolve around the patient's symptoms, treatment, and prognosis;

[0533] Pharmacists, including:

[0534] Professionalism, use professional terminology in drug administration, maintain authority and accuracy;

[0535] Be detailed and comprehensive, and questions should cover the selection, use, and monitoring of medications;

[0536] Safety reminders, emphasizing the safety and precautions of drugs;

[0537] Nursing staff, including:

[0538] Operational: questions should be specific and actionable to help caregivers better perform daily care tasks;

[0539] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0540] Practicality: the questions should have practical application value and help nursing staff improve their work efficiency;

[0541] Medical researchers, including:

[0542] Scientific rigor, scientific and rigorous tone, focusing on the accuracy of research methods and data;

[0543] Detailed and comprehensive, with questions covering research design, data analysis, and interpretation of results;

[0544] Innovation: Encourage new research directions and innovative ideas.

[0545] S240, screening of generalized questions: using a large language model to evaluate the rewritten generation questions, screening out generation questions with generalization, retaining and saving generation questions without generalization, and obtaining targeted generation questions; including:

[0546] S241, generalization question identification, using the large language model, based on the question generalization evaluation prompt words, evaluate the rewritten generated question to obtain the evaluation result, that is, whether it is generalized; the question generalization evaluation prompt words include:

[0547] Generate questions, including providing generative questions as a basis for assessment;

[0548] Assessment requirements, including:

[0549] Specificity: assess whether the question involves specific objects, situations or details, and avoid asking questions that are too broad;

[0550] Application scenario: whether the problem has a clear application scenario and can be applied in actual work;

[0551] Actionability: assess whether the problem has a clear answer or solution to facilitate subsequent responses and discussions;

[0552] Practical value: evaluate whether the problem has practical application value and can help users solve specific problems or obtain useful information.

[0553] Assessment methods, including:

[0554] Large language model evaluation: Use the large language model to analyze the generated questions to determine whether they are generalizable. Evaluation indicators include the breadth, depth, specificity, and application scenarios of the questions.

[0555] Keyword matching: Through keyword matching, we can identify whether the question contains specific objects, situations or details, for example, whether the question mentions a specific disease, drug or patient group;

[0556] Contextual relevance: assess whether the question is closely related to the content of interest, to avoid questions being out of touch with background information;

[0557] Manual labeling: Combined with the results of manual labeling, further adjust and optimize the evaluation results to ensure the accuracy of screening;

[0558] Filtering rules include:

[0559] The standard is retained. The problem has clear objects, situations or details, can be applied in actual work, and has practical value;

[0560] Screening criteria: questions that are too broad, lack specific content, or have no practical application scenarios will be screened out.

[0561] S242. Screening of generalization problems: Screening out generation problems with generalization and retaining generation problems without generalization.

[0562] S243, saving the targeted generated problems, saving the generated problems that are not generalizable, and obtaining the targeted generated problems.

[0563] S300, medical text data synthesis, through role description, using the large language model to rewrite the targeted generated questions and interest content, generate variant questions and answers, use the large language model and reranking model to score and filter the variant questions and answers, synthesize medical text data, and use BGE-Rerank-Large as the reranking model; Figure 4 As shown, specifically including:

[0564] S310, variant question and answer generation, through role description, combined with variant question and answer generation prompt words, using the large language model to rewrite the targeted generated questions and answer interest content, to generate variant questions and answers; variant question and answer generation prompt words include:

[0565] Role description, including information on the specified role description;

[0566] Generate targeted questions;

[0567] Corresponding document answers, including targeted generation of interesting content corresponding to the question;

[0568] Generate requirements, including:

[0569] Diversity: Generate multiple variant questions and answers from different angles, covering different scenarios, details, and expressions;

[0570] Targeted: Each variant is targeted to meet the specific needs of the role based on the role's professional background, job responsibilities, common concerns and typical behaviors;

[0571] Practicality: the generated variants have practical application value, helping the characters better understand and apply relevant information in their daily work;

[0572] Consistency: The variant questions and answers should keep the same core content as the original questions and answers, and avoid changing the core facts or conclusions;

[0573] Innovation: Encourage the generation of novel variant questions, propose new perspectives or application scenarios, and increase the richness and depth of data;

[0574] Generate rules, including:

[0575] Doctors, including:

[0576] Change of perspective, rewriting the question from different diagnostic, therapeutic, and prognostic perspectives;

[0577] Scenario expansion: introducing specific clinical cases or special situations to make the problem more realistic;

[0578] Details supplementation, adding more specific details, including the patient's medical history, test results, etc., to make the question more detailed;

[0579] Professional terminology: Use more professional medical terminology to maintain authority and rigor;

[0580] Nurses, including:

[0581] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0582] Operational: questions should be specific and actionable to help nurses better perform nursing tasks;

[0583] Safety, emphasizing the safety and health of patients and ensuring risk prevention during the nursing process;

[0584] Details supplement, adding more specific care details, such as monitoring frequency and care records;

[0585] Patients, including:

[0586] Be cordial and friendly. The tone should be cordial, friendly, and easy to understand;

[0587] Keep it simple and clear, avoid using complicated medical terms, and use layman's language to explain;

[0588] Focus on yourself, and questions should revolve around the patient's symptoms, treatment, and prognosis;

[0589] Detail supplementation, adding more details related to the patient's daily life, including diet and exercise;

[0590] Pharmacists, including:

[0591] Professionalism, use professional terminology in drug administration, maintain authority and accuracy;

[0592] Detailed and comprehensive, with questions covering the selection, use, and monitoring of medications;

[0593] Safety reminders, emphasizing the safety and precautions of drugs to ensure safe use of drugs;

[0594] Details supplement, adding more specific details about drug use, including dosage adjustments and adverse reactions;

[0595] Nursing staff, including:

[0596] Practicality: questions are specific and actionable, helping caregivers to better perform daily care tasks;

[0597] Caring: The tone should be warm and caring, reflecting concern for the patient;

[0598] Practicality: the questions have practical application value and help nursing staff improve their work efficiency;

[0599] Supplement details and add more specific care details, including diet, cleaning, and rehabilitation training;

[0600] Medical researchers, including:

[0601] Scientific rigor, scientific and rigorous tone, focusing on the accuracy of research methods and data;

[0602] Detailed and comprehensive, with questions covering research design, data analysis, and interpretation of results;

[0603] Innovation: encourage new research directions and innovative ideas to increase the depth and breadth of research;

[0604] Details supplement, adding more specific details about the research design and implementation, including sample selection and experimental procedures;

[0605] Hospital management staff, including:

[0606] Decision-making oriented, decisive and methodical in tone, focusing on the feasibility and effectiveness of decisions;

[0607] A global perspective, where the problem is addressed from the perspective of the overall hospital operation, taking into account resource allocation, service quality, and policy formulation;

[0608] Risk management, emphasizing the importance of risk management and emergency plans to ensure the normal operation of the hospital;

[0609] Details are supplemented to add more specific details about hospital management, including process optimization, resource allocation, etc.

[0610] S320, scoring and screening, using a large language model and a re-ranking model to score and screen variant questions and answers, and synthesize medical text data; specifically including:

[0611] S321, based on the relevance calculation of the large language model, according to the variant question and answer and the interest content, combined with the variant question and answer content scoring prompt words, the large language model is used to perform relevance judgment and output a comprehensive score; the variant question and answer content scoring prompt words include:

[0612] Variant problem;

[0613] Answer, the answer corresponding to the variant question;

[0614] Scoring requirements include:

[0615] Accuracy, assessing whether variant questions and answers are consistent with medical facts and clinical guidelines, and avoiding false or misleading information;

[0616] Relevance, assessing whether variant questions and answers are highly relevant to the content of interest and closely revolve around specific topics;

[0617] Operability: evaluate whether the answer has clear steps or solutions to facilitate its application in actual work;

[0618] Practical value: evaluate whether the variant questions and answers have practical application value and help users solve specific problems or obtain useful information;

[0619] Logical clarity, assessing whether the logic of questions and answers is clear, whether the expression is concise and clear, and whether long or complex sentence structures are avoided;

[0620] Professionalism, assessing whether the answer uses appropriate medical terminology, remains professional and authoritative, and is easy to understand;

[0621] Scoring methods include:

[0622] Large language model evaluation: using a large language model to encode variant questions and answers and calculate their scores on various dimensions, including accuracy, relevance, operability, practical value, logical clarity, and professionalism;

[0623] Manual labeling: Combined with the results of manual labeling, further adjust and optimize the scores to ensure the accuracy of the evaluation results;

[0624] Comprehensive score: The scores of each dimension are weighted averaged to obtain the final comprehensive score. The comprehensive score is a decimal between 0 and 1. The larger the comprehensive score, the stronger the correlation. 1 means complete correlation, and 0 means no correlation.

[0625] Filtering rules include:

[0626] Retention criteria: questions and answers with a comprehensive score greater than 0.5 are retained to ensure their excellent performance in terms of accuracy, relevance, operability and practical value;

[0627] Screening criteria: questions and answers with an overall score less than or equal to 0.5 will be screened out, especially those with obvious errors, irrelevant content or no practical application value.

[0628] S322, relevance calculation based on re-ranking model: input variant questions and answers and interest content into the re-ranking model, perform relevance judgment, and output re-ranking relevance, where the re-ranking relevance is a decimal value between 0 and 1. The larger the re-ranking relevance, the stronger the relevance. 1 indicates complete relevance, and 0 indicates complete irrelevance.

[0629] S323, correlation screening, calculate the average of the comprehensive score and the re-ranking correlation as the correlation, and discard the variant questions and answers with a correlation less than the specified correlation threshold (0.5). The calculation formula is as follows:

[0630]

[0631] in Representative variant question and answer, Represents content of interest, Represents the scoring prompt words for variant question and answer content, represents the reordering association degree, represents the comprehensive score;

[0632] S324, generate and save questions in a targeted manner, save variant questions and answers and interest content retained after correlation screening, and synthesize medical text data.

[0633] To verify the technical effect of this application, the applicant conducted an experiment. Medical information was collected using public data sets. Medical text data was synthesized through the above steps S100 to S300. The details are as follows:

[0634] The collected medical information documents are as follows:

[0635] A. Common diseases and symptoms

[0636] 1. Hypertension

[0637] Description: Hypertension is a condition in which the pressure of blood on the artery walls increases. Long-term, uncontrolled high blood pressure can lead to serious complications such as heart disease and stroke.

[0638] Common symptoms: There are usually no obvious symptoms, but they may include headache, shortness of breath, chest pain, or blurred vision.

[0639] Self-assessment: If your home blood pressure is often above 140 / 90 mmHg, you should consult your doctor.

[0640] Emergencies: Seek emergency medical care immediately if you experience severe headache, severe dizziness, loss of consciousness, or chest pain.

[0641] Prevention: Maintain a healthy weight, reduce salt intake, limit alcohol consumption, exercise regularly, and avoid smoking.

[0642] 2. Diabetes Mellitus

[0643] Description: Diabetes is a chronic disease in which the body is unable to effectively use insulin to regulate blood sugar levels.

[0644] Common symptoms: Frequent urination, thirst, extreme hunger, unexplained weight loss, fatigue, blurred vision, slow wound healing.

[0645] Self-assessment: People with a family history of diabetes should have their blood sugar tested regularly; a fasting blood sugar ≥7.0 mmol / L (126 mg / dL) may be a sign of diabetes.

[0646] Emergency: If the patient is confused, breathing is deep and rapid, and there is a fruity smell in the breath, it may be diabetic ketoacidosis and requires emergency treatment.

[0647] Prevention: Reduce your risk of disease by eating a healthy diet, getting regular exercise, and maintaining a normal weight.

[0648] B. Treatment Options

[0649] 1. Drug treatment

[0650] Antihypertensive drugs: such as ACE inhibitors, beta-blockers, calcium channel blockers, etc., used to lower blood pressure.

[0651] Antidiabetic drugs: such as metformin, sulfonylureas, GLP-1 receptor agonists, etc., help control blood sugar levels.

[0652] 2. Non-drug treatment

[0653] Physical therapy: For people with arthritis, physical therapy can help relieve pain and improve mobility.

[0654] Psychotherapy: Cognitive behavioral therapy (CBT) can be effective in treating depression and other mental health problems.

[0655] 3. Alternative Therapies

[0656] Acupuncture: In traditional Chinese medicine, acupuncture is used to treat a variety of conditions, including chronic pain and stress management.

[0657] C. Introduction of hospital departments

[0658] 1. Internal Medicine

[0659] Department Function: Internal Medicine physicians focus on the diagnosis and non-surgical treatment of adult illnesses, particularly complex multi-system disorders.

[0660] Expert team: We have experts in multiple sub-specialties such as cardiovascular disease, endocrinology, and gastroenterology.

[0661] Appointment process: You can make an appointment through the hospital’s official website or by phone. Some hospitals also provide online consultation services.

[0662] 2. Surgery

[0663] Department Function: Surgery is responsible for all types of surgical procedures, from simple appendectomies to complex heart transplants.

[0664] Expert team: including experts in general surgery, neurosurgery, orthopedics and other fields.

[0665] Appointment process: You need to be referred by the internal medicine department or other departments first, and then the surgery time will be arranged according to your condition.

[0666] D. Patient education

[0667] 1. Preoperative and postoperative care

[0668] Preoperative preparation: Follow the instructions given by the doctor, such as stopping taking certain medications, fasting, etc.

[0669] Postoperative recovery: pay attention to rest, take painkillers as prescribed by the doctor, change dressings regularly, and prevent infection.

[0670] 2. Chronic disease management

[0671] Diabetes management: monitor blood sugar, take medication or insulin injections on time, and attend diabetes education classes.

[0672] Hypertension management: Take antihypertensive medication regularly, measure blood pressure regularly, and adjust your lifestyle to maintain stable blood pressure.

[0673] 3. Nutrition and diet

[0674] Low-salt diet: Reducing the intake of processed foods and high-sodium condiments can help control high blood pressure.

[0675] Low-carbohydrate diet: Appropriately reducing carbohydrate intake can help diabetic patients control blood sugar.

[0676] E. Health tools

[0677] 1. Symptom Checker

[0678] Interactive tools: Users can answer a series of questions based on their symptoms, and the tool will provide preliminary health guidance or recommend medical treatment.

[0679] 2. Medication reminder

[0680] Set reminders: Users can set reminders on their phones to ensure they take their medications on time and avoid missing or duplicating doses.

[0681] 3. Health Records

[0682] Track data: Record health indicators such as blood pressure, blood sugar, weight, etc., generate charts to observe trends and share with doctors.

[0683] The synthesis results of medical text data (in question-and-answer format) are as follows, where Q is the question and A is the answer:

[0684] Q: My blood pressure has been high recently. Do I really have high blood pressure?

[0685] A: If your blood pressure is often higher than 140 / 90 mmHg when you measure it at home, you may have high blood pressure. But don't worry too much. Record the results of several measurements at different time points and then go to the hospital to see a doctor. The doctor may ask you to do some additional tests to confirm the diagnosis and give you treatment recommendations that are suitable for you.

[0686] Q: I sometimes feel dizzy. Is it related to my high blood pressure?

[0687] A: Dizziness can indeed be a symptom of high blood pressure, but it can also be caused by other reasons, such as low blood sugar, anemia, or inner ear problems. If you often feel dizzy, it is best to go to the hospital for a comprehensive examination. The doctor will help you find the specific cause. At the same time, remember to continue to monitor your blood pressure. If you find that your blood pressure is abnormal when you feel dizzy, you should contact your doctor immediately.

[0688] Q: I heard that exercise is good for health, but I have high blood pressure. Can I exercise?

[0689] A: Yes, moderate exercise is very beneficial for people with high blood pressure. It can help lower blood pressure and improve cardiovascular health. It is recommended to choose low to moderate intensity activities, such as walking, swimming, or cycling. Before starting a new exercise program, it is best to consult a doctor to make sure the exercise you choose is suitable for you. Remember to increase the amount of exercise slowly and avoid strenuous exercise, especially in hot or cold weather.

[0690] Q: I just had surgery. What should I pay attention to after the surgery?

[0691] A: Postoperative care is very important and directly affects the speed and effect of recovery. You should:

[0692] Follow your doctor's orders: Take your medications, including pain relievers and other necessary medications, exactly as directed by your doctor.

[0693] Get enough rest: Give your body enough time to heal and avoid overexerting yourself.

[0694] Wound care: Keep the surgical incision clean and dry, change the dressing as recommended by the doctor, and observe for signs of infection such as redness, swelling, and exudation.

[0695] Gradually resume activities: Gradually increase your daily activities according to your doctor's advice, but avoid heavy lifting or strenuous exercise until your doctor says it's okay.

[0696] Regular follow-up: Return to the hospital for follow-up at the scheduled time to make sure everything is recovering as planned.

[0697] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A role-driven medical text data synthesis method, characterized in that: The steps include: S100, generating original questions, parsing collected medical information documents, extracting key information fragments, generating role descriptions according to the occupations and concerns of preset medical roles, generating original interest content for the key information fragments based on the role descriptions and performing semantic deduplication, generating interest content and classifying it, and generating original questions using a large language model; including: S110, parsing medical information documents, using a document parser to split the collected medical information documents and extract key information fragments; S120, generating interest content, generating a role description according to the occupation and focus of the preset medical role, and generating original interest content for the key information fragment based on the role description; S130, semantically deduplicating the interest content, using a text encoding model and a vector database to encode the original interest content and remove semantically repeated content, obtain the deduplicated original interest content, obtain the interest content and save it to the vector database; S140, classifying the interest content, using the large language model to classify the interest content according to the preset question type and the question type classification prompt word; S150, generating an original question, generating prompt words according to the interest content classification and the role description in combination with the original question, and using the large language model to generate the original question; S200, screening original questions, performing semantic deduplication on the original questions to obtain original generated questions, evaluating and screening the relevance between the original generated questions and the interest content to obtain generated questions, rewriting the generated questions using the large language model and screening generalized questions to obtain targeted generated questions; S300, medical text data synthesis, through the role description, using the large language model to rewrite the targeted generated questions and the interest content to generate variant questions and answers, using the large language model and the re-ranking model to score and screen the variant questions and answers to synthesize medical text data.

2. The role-driven medical text data synthesis method according to claim 1, characterized in that: The step S110 includes: S111, document slicing, slicing the collected medical information document according to line breaks and serial number codes to improve the efficiency and accuracy of subsequent processing; S112, extracting key information fragments. For the sliced ​​text, calculate the string length, take 512 characters as the upper limit, process slice by slice, and store the excess part as a new slice. All slice texts constitute the key information fragment.

3. The role-driven medical text data synthesis method according to claim 1, characterized in that: The step S120 includes: S121, generating a role description, generating prompt words based on the occupation and focus of the preset medical role in combination with the medical role description, and generating a role description using the large language model; S122, generating interest content, based on the role description and in combination with interest content generation prompt words, and using the large language model to generate original interest content for the key information fragment.

4. The role-driven medical text data synthesis method according to claim 1, characterized in that: The step S130 includes: S131, encoding the original interest content, using the text encoding model and the vector database, encoding each original interest content into a 1×1024-dimensional feature vector, obtaining an original interest content feature vector, and saving it in the vector database; S132, feature vector deduplication, comparing each original interest content feature vector with other original interest content feature vectors in the vector database using a feature vector similarity formula, and performing deduplication processing, if the similarity is greater than or equal to a specified feature similarity threshold, otherwise discard; S133: generating interest content, saving the original interest content after deduplication to the vector database, and obtaining the interest content.

5. The role-driven medical text data synthesis method according to claim 1, characterized in that: The step S200 includes: S210, semantically de-duplicating the original question, encoding the original question into a database using the text encoding model and the vector database, and de-duplicating the original question to obtain the original generated question; S220, relevance evaluation, based on the original generated question and the interest content, using the large language model to perform relevance discrimination and relevance screening to obtain a generated question; S230, rewriting the generated question, rewriting the generated question using the large language model according to the role description and the question rewriting prompt words; S240, screening of generalized questions: using the large language model to evaluate the rewritten generation questions, screening out generation questions with generalization, retaining and saving generation questions without generalization, and obtaining targeted generation questions.

6. The role-driven medical text data synthesis method according to claim 5, characterized in that: The step S210 includes: S211, encoding the original question, using the text encoding model to encode each original question into a 1×1024-dimensional feature vector, obtaining the original question feature vector, and saving it in the vector database; S212, deduplication of original question feature vectors, comparing each original question feature vector with other original question feature vectors in the vector database using a feature vector similarity formula to perform deduplication processing, and discarding them if the similarity is greater than or equal to a specified question similarity threshold; S213, the original generated question is obtained, and the original question after deduplication is saved to the vector database to obtain the original generated question.

7. The role-driven medical text data synthesis method according to claim 5, characterized in that: The step S220 includes: S221, relevance assessment, based on the original generated question and the interest content, combined with the relevance assessment prompt words, using the large language model to perform relevance assessment and output a comprehensive relevance; S222, correlation screening, traversing all the original generated questions, and discarding the original generated questions whose comprehensive correlation is less than the specified correlation threshold; S223, obtaining the generated problem, saving the original generated problem after screening, and obtaining the generated problem.

8. The role-driven medical text data synthesis method according to claim 5, characterized in that: The step S240 includes: S241, generalization question identification, using the large language model, based on the question generalization evaluation prompt words, the rewritten generated question is evaluated to obtain an evaluation result, that is, whether it has generalization; S242, screening of generalization problems, screening out the generation problems with generalization, and retaining the generation problems without generalization; S243. Targeted generation problems are saved, generation problems that are not generalizable are saved, and targeted generation problems are obtained.

9. The role-driven medical text data synthesis method according to claim 5, characterized in that: The step S300 includes: S310, generating variant questions and answers, by combining the role description with the variant question and answer generation prompt words, and using the large language model to rewrite the targeted generated questions and answer interest content to generate variant questions and answers; S320, scoring and screening, using the large language model and the re-ranking model to score and screen the variant questions and answers, and synthesize medical text data.

Images