Method and apparatus for generating medical reports based on evidence-based medicine

Abstract

Embodiments of the present specification relate to a method and device for generating a medical report based on evidence-based medicine, the method comprising: obtaining patient information and an original clinical question; inputting the patient information, the original clinical question and a first prompt word into a target model to generate a structured clinical question; the target model being a large language model in the medical field; retrieving a medical literature database according to the clinical question to obtain a plurality of medical literatures; inputting the target medical literature and a second prompt word into the target model to enable the target model to perform bias risk assessment on the target literature to generate a structured initial assessment result of the target literature; inputting the clinical question, each medical literature and its initial assessment result and a third prompt word into the target model to enable the target model to perform step-by-step reasoning based on a GRADE evaluation process in the third prompt word to generate a medical report; the medical report including a medical recommendation for the clinical question, a recommended strength of the medical recommendation and an evidence quality level.

Description

Technical Field

[0001] This specification relates to one or more embodiments in the field of machine learning, and more particularly to methods and apparatus for generating medical reports based on evidence-based medicine. Background Technology

[0002] With the widespread application of Artificial Intelligence (AI) in the medical field, Clinical Decision Support Systems (CDSS) are evolving from a traditional model based on fixed rules to intelligent and autonomous systems. However, existing medical AI systems generally rely on static knowledge bases or "black box" model outputs when providing medical advice, lacking the ability to dynamically acquire, assess the quality of, and make transparent inferences based on the latest clinical research evidence. Especially when dealing with complex clinical problems, they struggle to proactively search literature, judge the quality of evidence, and generate interpretable recommendations like human experts, affecting the credibility and safety of the recommendations. Meanwhile, the GRADE (Grading of Recommendations, Assessment, Development and Evaluation) framework in evidence-based medicine has been widely adopted internationally. It can systematically assess the quality of evidence and generate structured recommendations, but it currently relies mainly on manual operation, which is time-consuming, costly, and cannot be updated in real time.

[0003] Therefore, a method is needed to combine the GRADE methodology with artificial intelligence technology to automatically generate more rational medical recommendations. Summary of the Invention

[0004] This specification describes one or more embodiments of a method and apparatus for generating medical reports based on evidence-based medicine. Under the GRADE framework of evidence-based medicine, it utilizes a large language model (LMM) in the medical field to perform end-to-end reasoning on clinical questions in order to generate medical reports for those clinical questions.

[0005] Firstly, a method for generating medical reports based on evidence-based medicine is provided, including:

[0006] Obtain patient information and original clinical questions in natural language;

[0007] The patient information, original clinical question, and first prompt word are input into the target model, which then extracts the corresponding information from the input content based on the structured output template in the first prompt word, and generates a structured clinical question; the target model is a large language model in the medical field.

[0008] Based on the stated clinical problem, a medical literature database was searched to obtain several medical articles related to the stated clinical problem; each medical article contains research content related to the stated clinical problem.

[0009] Any target medical literature and a second prompt word are input into the target model, which then performs a bias risk assessment on the target literature and determines the initial evidence quality level based on the research type of the target literature, generating a structured initial assessment result for the target literature; the initial assessment result includes the bias risk level and the initial evidence quality level.

[0010] The clinical problem, various medical literature articles and their initial assessment results, and the third prompt word are input into the target model. The target model then uses the GRADE evaluation process in the third prompt word to perform step-by-step reasoning and generate a medical report. The medical report includes medical recommendations for the clinical problem, as well as the strength of recommendation and the level of evidence quality of the medical recommendations.

[0011] In some possible implementations, the first prompt word instructs the target model to transform the original clinical question and patient information according to the structured output template to generate a structured clinical question.

[0012] In some possible implementations, the first prompt word may further include a plurality of example samples, each example sample containing example input content and example output content; the first prompt word may also instruct the target model to perform inference using the plurality of example samples as examples.

[0013] In some possible implementations, the structured clinical problem includes several elements, such as patient characteristics, proposed interventions, control measures, and clinical outcomes of interest.

[0014] In some possible implementations, a medical literature database is searched based on the clinical problem to obtain several medical articles related to the clinical problem, including:

[0015] The clinical question is input into the target model, which is then instructed to perform synonym replacements on some or all of the content in the clinical question to obtain several synonymous clinical questions with the same meaning as the clinical question.

[0016] The clinical question and several synonymous clinical questions were searched in the medical literature database, and the search results were summarized to obtain several medical articles.

[0017] In some possible implementations, the second prompt word may further include several assessment scales; the second prompt word instructs the target model to use the corresponding assessment scale to assess the risk of bias in the target literature based on the research type of the target literature.

[0018] In some possible implementations, the initial evaluation results may also include the following: research type, research results, and research value.

[0019] In some possible implementations, the target model performs step-by-step reasoning based on the GRADE evaluation process in the third cue word, including:

[0020] The target model adjusts the initial evidence quality level of each medical article based on its individual risk of bias, as well as the consistency, directness, precision, and publication bias among the articles, to obtain the evidence quality level of the medical recommendations.

[0021] Based on various medical literature articles, the target model determines the favorable and unfavorable outcomes of interventions for the stated clinical problem.

[0022] The target model summarizes the research conclusions of various medical literature articles to determine medical recommendations;

[0023] The target model determines the strength of the medical recommendations based on the evidence quality level of each medical article, the favorable and unfavorable outcomes, the obtained patient preference model, and the resource cost of the medical recommendations.

[0024] In some possible implementations, the medical report may also include the following: the logical chain in the stepwise reasoning process, the medical literature and initial assessment results, and uncertainty warnings.

[0025] In some possible implementations, the method further includes, before inputting the clinical question, the medical literature and their initial assessment results, and the third prompt word into the target model:

[0026] The initial evaluation results of each medical article are input into the target model, which is then instructed to semantically align the technical terms in each initial evaluation result in order to update the initial evaluation results of each medical article.

[0027] In some possible implementations, it also includes:

[0028] The medical knowledge graph is updated based on the initial evaluation results.

[0029] Secondly, a device for generating medical reports based on evidence-based medicine is provided, comprising:

[0030] The acquisition unit is configured to acquire patient information and raw clinical questions in natural language.

[0031] The structured transformation unit is configured to input the patient information, the original clinical question, and the first prompt word into the target model, so that the target model extracts the corresponding information from the input content based on the structured output template in the first prompt word and generates a structured clinical question; the target model is a large language model in the medical field.

[0032] The retrieval unit is configured to search a medical literature database based on the clinical problem and obtain several medical articles related to the clinical problem; each medical article contains research content related to the clinical problem.

[0033] The initial evaluation unit is configured to input any target medical literature and a second prompt word into the target model, enabling the target model to perform a bias risk assessment on the target literature, determine the initial evidence quality level based on the research type of the target literature, and generate a structured initial evaluation result for the target literature; the initial evaluation result includes the bias risk level and the initial evidence quality level.

[0034] The report generation unit is configured to input the clinical problem, each medical literature article and its initial assessment results, and the third prompt word into the target model, so that the target model performs step-by-step reasoning based on the GRADE evaluation process in the third prompt word to generate a medical report; the medical report includes medical recommendations for the clinical problem, as well as the recommendation strength and evidence quality level of the medical recommendations.

[0035] Thirdly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of the first aspect.

[0036] Fourthly, a computing device is provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method of the first aspect.

[0037] The method and apparatus for generating medical reports based on evidence-based medicine, as proposed in the embodiments of this specification, combine the GRADE framework of evidence-based medicine with a large language model in the medical field. This achieves automated generation and dynamic updating of clinical decision recommendations, provides high-quality, low-risk evidence-based recommendations, and the output results possess high interpretability, meet medical safety requirements, support individualized decision support in precision medicine scenarios, and provide an architectural path for AI-driven intelligent guide systems.

[0038] The method described in this specification integrates problem understanding, literature retrieval, evidence extraction, quality assessment, and recommendation generation using a large language model, forming an end-to-end autonomous reasoning loop. Simultaneously, it extracts structured evidence from original literature using the large language model and stores it in a medical knowledge graph, supporting the correlation analysis and continuous updating of multi-source evidence. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the various embodiments disclosed in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only a few embodiments disclosed in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This illustration shows a scenario of generating medical reports based on evidence-based medicine according to one embodiment;

[0041] Figure 2 A flowchart illustrating a method for generating medical reports based on evidence-based medicine according to one embodiment is shown.

[0042] Figure 3 A schematic block diagram of an apparatus for generating medical reports based on evidence-based medicine, according to one embodiment, is shown. Detailed Implementation

[0043] The solution provided in this specification will now be described with reference to the accompanying drawings.

[0044] As mentioned earlier, existing medical AI systems struggle to access the latest knowledge when providing medical advice, and their conclusions often lack supporting evidence. Specifically, traditional rule-based or knowledge-based clinical decision support systems rely on manually constructed knowledge bases and pre-defined logical rules to generate recommendations. While these systems offer relatively high interpretability, their knowledge updates are lagging, making it impossible to dynamically acquire the latest research evidence, and they struggle to address complex, emerging, or individualized clinical problems. Deep learning-based medical AI models (such as black-box models used for diagnosis or prognosis prediction) utilize massive amounts of data to train models for reasoning, possessing a certain generalization ability. However, their decision-making process lacks transparency, making it impossible to trace the specific evidence upon which recommendations are based. Furthermore, they do not assess the scientific quality of input data or external literature, making them susceptible to low-quality or biased research and posing potential medical risks.

[0045] The evidence-based GRADE approach relies heavily on manual operation, has low automation, high cost, and cannot be updated.

[0046] To overcome the above problems, this specification proposes a method for generating medical reports based on evidence-based medicine.

[0047] First, let me briefly introduce the GRADE methodology. Developed by the international GRADE working group, the GRADE methodology is a systematic, transparent, and reproducible evidence-based medicine approach used to formulate healthcare recommendations and assess the quality of evidence and determine the strength of recommendations. Its goal is to address the problems of inconsistent evaluation standards and opaque conclusions in past medical and public health guidelines. Evidence-based medicine (EBM) is a medical practice approach that combines available scientific research evidence, clinical expertise, and patient values, emphasizing that medical decisions should be based on high-quality research data rather than experiential intuition.

[0048] GRADE categorizes the quality of evidence in medical literature into four levels: High, Moderate, Low, and Very Low. It also classifies the strength of recommendation of the final medical conclusion as strong or weak, emphasizing that decision-making should consider not only the credibility of the evidence itself, but also the balance of benefits and risks, patient values ​​and preferences, and resource availability.

[0049] Figure 1 This illustration shows a scenario for generating medical reports based on evidence-based medicine, according to one embodiment. Figure 1 As shown, firstly, the patient information and the original clinical questions in natural language form input by the user (e.g., a doctor or medical professional) are fed into a large language model in the medical field, which is then instructed to convert them into structured clinical questions in PICO form.

[0050] The large language model in the medical field can be obtained by fine-tuning a general pre-trained large language model using a training set containing medical knowledge. PICO is a standardized structural format for constructing clinical questions, containing four elements: P (Patient): patient characteristics, I (Intervention): proposed intervention, C (Comparison): control measures, and O (Outcome): the clinical outcome of interest; used to guide literature retrieval and evidence extraction.

[0051] For example, in a specific case, the original clinical question in natural language could be: "A female, aged xx, has undergone wedge resection of a ground-glass nodule in her right lung. The lesion has been stable on imaging follow-up for 2 years. She has a family history of the disease. In the absence of a final pathological description of the histological subtype, how can an individualized postoperative monitoring strategy be developed to balance the risk of recurrence with the possibility of overtreatment?" Meanwhile, patient information can be retrieved from a patient database, including basic patient information, disease information, examination results, treatment process and results, medications used, etc.

[0052] Correspondingly, the structured clinical questions in the form of PICO output by the large language model could be: "P: Positive for xxx disease, has received yyy treatment, experienced asymptomatic progression after treatment, and is in stable condition. I: Continue the current yyy treatment if progression occurs, instead of immediately switching to another zzz treatment regimen. C: Immediately switch to zzz treatment regimen. O: Progression-free survival, time to next first-line systemic treatment, overall survival, toxicity, and quality of life."

[0053] Then, return to Figure 1 After obtaining the structured clinical question, the question is entered into a medical literature database for retrieval, yielding several medical articles related to the question. Each article contains research content relevant to the stated clinical question.

[0054] Next, the bias risk assessment is performed on each of the retrieved medical articles using a large language model, and the initial evidence quality level of each medical article is determined according to the criteria in GRADE, generating the initial assessment results for each medical article.

[0055] Specifically, any medical article and an assessment scale for bias risk evaluation are input into the large language model. The large language model then extracts relevant content from the medical article based on the questions in the assessment scale, and determines the bias risk level of the medical article based on the questions and the extracted content. Simultaneously, the large language model determines the initial quality of evidence level based on the research type within the medical article.

[0056] Risk of bias is a systematic error in the design or implementation of a study that may cause results to deviate from the true value, such as insufficient randomization or lack of blinding. The risk of bias level is the primary downgrade factor in the GRADE assessment. It includes three levels: High, Low, and Unclear. Study types include randomized controlled trials (RCTs), observational studies, and other types (such as case reports and expert opinions). When the study type is an RCT, the initial quality of evidence is high; when the study type is an observational study, the initial quality of evidence is low; and when the study type is other, the initial quality of evidence is very low.

[0057] Furthermore, it enables the large language model to extract content from the medical literature to determine the research type, research results, and research value of the medical literature, and adds it to the initial evaluation results of the medical literature to obtain structured initial evaluation results.

[0058] For example, continuing the above example, the structured initial assessment result of a medical article could be: "Document metadata: xxx. Study type: observational study. Study outcome: median Extended-PFS 6.7 months; consistent trends among multiple drivers. GRADE initial evidence quality level: low. Reasons for downgrading: retrospective, heterogeneous. Study value: supports the universal strategy of oligoprogressive local therapy + maintenance TKI."

[0059] After obtaining the initial evaluation results of each article, optionally, the initial evaluation results can be converted into triples using a large language model, and existing medical knowledge graphs can be created or updated.

[0060] Next, return to Figure 1 After obtaining the initial evaluation results of each article, the aforementioned structured clinical question, the content of each article and the initial evaluation results, as well as the GRADE framework document, are input into the large language model. This allows the model to perform step-by-step reasoning based on the evaluation process in the GRADE document, generating a medical report. The medical report includes the following items: medical recommendations for the clinical question, the strength of the recommendation and the level of evidence quality, the logical chain in the step-by-step reasoning process, the limitations of the medical recommendation, the medical article and the initial evaluation results, and uncertainty warnings.

[0061] Using pre-defined prompts, the large language model reasons step-by-step based on the evaluation process in the GRADE document, arriving at the final medical report. This medical report is highly readable and clearly structured, designed for medical professionals.

[0062] For example, continuing the above example, a medical report could be: "Continuing the current yyy treatment regimen can be considered instead of a mandatory change," recommendation strength: weak, evidence quality level: low, logical chain: In xxx condition, the relevant research results of the initial yyy treatment are... The relevant research results of the replacement treatment regimen zzz are... Overall strength of conclusion: GRADE is low to very low overall (mainly observational, single-arm, and indirect evidence). However, consistent effect direction, strong clinical rationale, toxicity, and quality of life benefits (reducing premature medication changes) support its classification as "weak (conditional recommendation)."

[0063] pass Figure 1 The process demonstrated allows large language models in the medical field to be deeply involved in each step of the GRADE reasoning process, realizing an automated end-to-end method for generating medical reports.

[0064] The following describes the specific implementation steps of the above-mentioned method for generating medical reports based on evidence-based medicine, with reference to specific embodiments.

[0065] Figure 2 A flowchart illustrating a method for generating medical reports based on evidence-based medicine according to one embodiment is provided. The subject executing the method can be any platform, server, or device cluster with computing and processing capabilities (hereinafter referred to as electronic equipment). Figure 2 As shown, the method is performed by an electronic device and includes at least steps S202 to S210.

[0066] The specific execution process of each of the above steps is described below.

[0067] First, in step S202, patient information and original clinical questions in natural language are obtained.

[0068] Patient information can be obtained from a patient database, specifically from the patient's Electronic Health Record (EHR), which may include basic patient information, disease information, allergy history, family medical history, examination results, treatment process and results, medications used, etc. The original clinical questions can be directly entered by the doctor or relevant personnel, and contain information related to the treatment of the patient's disease.

[0069] Then, in step S204, the patient information, the original clinical question, and the first prompt word are input into the target model, so that the target model extracts the corresponding information from the input content according to the structured output template in the first prompt word and generates a structured clinical question; the target model is a large language model in the medical field.

[0070] The medical language model can be an existing model, such as the ClinicalBERT-Large model or the Med-PaLM model; or it can be fine-tuned using a training set containing medical knowledge based on a pre-trained general language model. There are no restrictions here.

[0071] The first prompt may include the structured output template and indicative content for the task assigned to the target model. The target model is instructed to extract patient information from the input content and relevant content from the original clinical question based on the structured output template, and then fill these into the structured output template to generate a structured clinical question.

[0072] In one embodiment, the first prompt word in step S204 instructs the target model to transform the original clinical question and patient information according to the structured output template to generate a structured clinical question.

[0073] In one embodiment, the structured output template is a PICO template.

[0074] In a more specific embodiment, the structured clinical question includes several elements, such as patient characteristics, proposed intervention, control measures, and clinical outcomes of interest.

[0075] For example, in a specific example, the first prompt could be: "Please convert the patient information and the original clinical question into a structured clinical question based on the following structured output template. The structured output template is: [Structured Output Template], the patient information is: [Patient Information], and the original clinical question is: [Original Clinical Question]."

[0076] The [Structured Output Template], [Patient Information], and [Original Clinical Question] sections are used to fill in the corresponding specific content. The structured output template can be a PICO template.

[0077] In a more specific embodiment, few-shot prompting can also be used to guide the target model to accurately identify elements in the structured output template and integrate individual patient information for contextualized modeling.

[0078] In this embodiment, the first prompt word also includes several example samples, each of which contains example input content and example output content; the first prompt word also instructs the target model to perform inference using the several example samples as examples.

[0079] The example samples are used to show the target model the correct output content, so that the target model can better understand the semantic relationships between the input content, the structured output template, and the output content.

[0080] For example, continuing from the example above, the first prompt can also include: "The following are examples of multiple input and output contents: [Input 1]-[Output 1]; [Input 2]-[Output 2]...", where [Input 1], [Output 1], [Input 2], [Output 2], etc. are used to fill in specific example input contents and example output contents.

[0081] After obtaining structured clinical questions, return to Figure 2 In step S206, a medical literature database is searched according to the clinical problem to obtain several medical articles related to the clinical problem; each medical article contains research content related to the clinical problem.

[0082] Medical literature databases such as PubMed and the Cochrane Library can contain various types of medical-related literature, such as medical research papers, case reports, expert opinions, etc. Each medical article contains one or more research topics, including those related to the stated clinical problem.

[0083] The medical literature retrieved in step S206 can be the full text of the medical literature, or it can be the abstract of the medical literature and its metadata (title, author, publication year, DOI, research design, etc.), without any limitation.

[0084] In one embodiment, since different forms of expression exist for the same concept in medicine, such as abbreviations, aliases, etc., parts of the clinical question can be replaced with synonyms to expand and generate several synonymous clinical questions with the same meaning, thereby improving the retrieval recall rate.

[0085] In this embodiment, step S206 specifically includes:

[0086] The clinical question is input into the target model, which is then instructed to perform synonym replacements on some or all of the content in the clinical question to obtain several synonymous clinical questions with the same meaning as the clinical question.

[0087] The clinical question and several synonymous clinical questions were searched in the medical literature database, and the search results were summarized to obtain several medical articles.

[0088] In another embodiment, a target model can also be used to transform structured clinical questions into retrieval expressions (such as Boolean logic expressions) for exploring medical literature databases. In this embodiment, step S206 includes:

[0089] The clinical question is input into the target model, which then instructs the target model to convert the clinical question into a corresponding database retrieval expression.

[0090] By searching the medical literature database using the aforementioned retrieval formula, several medical articles related to the clinical problem were obtained.

[0091] In another embodiment, the medical literature retrieved in step S206 can be sorted according to a preset sorting rule (e.g., prioritizing the return of high-quality studies such as systematic reviews, meta-analyses, and RCTs) to update the list of medical literature composed of the medical literature.

[0092] After obtaining several medical articles related to the clinical problem in step S206, in step S208, any target medical article and the second prompt word are input into the target model, so that the target model performs a bias risk assessment on the target article and determines the initial evidence quality level according to the research type of the target article, generating a structured initial assessment result of the target article; the initial assessment result includes the bias risk level and the initial evidence quality level.

[0093] Risk of bias assessment is used to determine the risk of bias level of experiments in the target literature. The risk of bias level is determined by evaluating the experimental process using assessment scales. The initial quality of evidence level is determined using a pre-defined correspondence based on the research type of the target literature.

[0094] In one embodiment, step S208 further includes several assessment scales in the second prompt word; the second prompt word instructs the target model to use the corresponding assessment scale to assess the risk of bias in the target literature according to the research type of the target literature.

[0095] The target model first extracts information related to research methodology from the target literature (such as randomization methods, blinding implementation, loss to follow-up rate, intention-to-treat analysis, etc.) to determine the research type of the target literature. Then, based on the research type, it uses corresponding assessment scales (such as the Cochrane RoB 2.0, ROBINS-I, or NEWCASTLE-OTTAWA scales) to assess the risk of bias in the target literature.

[0096] The assessment scale contains multiple questions, each with several options. Participants answer the questions based on information from medical literature, and the combined answers determine the risk level of bias in the medical literature.

[0097] Specifically, the target model extracts content related to the questions in the assessment scale from the target literature, answers the corresponding questions based on the relevant content, and then integrates the answers to each question to determine the risk level of bias.

[0098] In one example, the second prompt could be: "First, based on the methodological content of the medical literature, determine the research type of the medical literature. Then, according to the research type, use the corresponding assessment scale to assess the risk of bias in the medical literature and determine the risk level of bias. You need to first extract the relevant content from the medical literature based on the questions in the assessment scale, then answer each question, and finally determine the risk level of bias. The medical literature is: [Medical Literature], the assessment scale is: [Assessment Scale], and the correspondence between the research type and the assessment scale is: [Correspondence]."

[0099] The fields [Medical Literature], [Assessment Scales], and [Correspondence] should be filled with the corresponding content. Medical literature and assessment scales can be directly entered text or links to relevant content; there are no restrictions here. [Correspondence] can be in text form, for example: "RCT - High; Observational Studies - Low; Other Types - Very Low".

[0100] In another embodiment, if the target model has already learned the aforementioned assessment scales during the training phase, the second prompt does not need to include the assessment scales. Instead, it directly instructs the target model to use the corresponding assessment scale to assess the risk of bias in the target document based on its research type.

[0101] In one embodiment, the initial evaluation results output by the target model in step S208 include, in addition to the bias risk level and initial evidence quality level of the target literature, the following items: research type, research results, and research value.

[0102] The research results and research value can be obtained by extracting relevant paragraphs from the target literature using the target model.

[0103] According to step S208, each medical article can be evaluated separately to obtain the initial evaluation results corresponding to each article.

[0104] In some possible implementations, also because different forms of expression exist for the same concept in medicine, such as abbreviations, aliases, etc., the method further includes the following before step S210:

[0105] The initial evaluation results of each medical article are input into the target model, which is then instructed to semantically align the technical terms in each initial evaluation result in order to update the initial evaluation results of each medical article.

[0106] Finally, in step S210, the clinical problem, each medical literature article and its initial assessment results, and the third prompt word are input into the target model, so that the target model performs step-by-step reasoning based on the GRADE evaluation process in the third prompt word to generate a medical report; the medical report includes medical recommendations for the clinical problem, as well as the recommendation strength and evidence quality level of the medical recommendations.

[0107] The third prompt may include the GRADE evaluation process and indicative content regarding the tasks assigned to the target model. It instructs the target model to perform step-by-step reasoning according to the GRADE evaluation process to generate a medical report.

[0108] For example, in one instance, the third prompt could include: "Based on the steps of the GRADE evaluation process, integrate the research content from various medical literatures to generate a medical recommendation for a clinical problem, along with the strength of recommendation and the quality of evidence level of that recommendation. Clinical problem: [Clinical problem], List of medical literatures: [List of medical literatures], GRADE evaluation process: [GRADE]."

[0109] Among them, [Clinical Question] is used to fill in the specific content of the clinical question, [Medical Literature List] is used to fill in each medical literature and its initial assessment results, and [GRADE] is used to fill in the content of GRADE documents or their links.

[0110] The target model can proceed step by step according to the reasoning steps in the GRADE document to obtain the final result.

[0111] In one embodiment, the target model in step S210 performs step-by-step reasoning based on the GRADE evaluation process in the third cue word, including:

[0112] According to the GRADE evaluation process, the target model first integrates the initial evidence quality levels of each medical article to determine the overall evidence quality level.

[0113] Then, based on the instructions in the GRADE document, the target model judges the consistency, directness, precision, and publication bias among the various documents. Combining this with the risk level of bias of each medical document determined in step S208, the overall evidence quality level is adjusted based on the above five aspects to obtain the evidence quality level of the final medical recommendation.

[0114] Next, the target model, based on each medical literature article, determines the favorable and unfavorable outcomes of interventions in the clinical problem.

[0115] Then, the target model summarizes the research conclusions of each medical article to determine medical recommendations.

[0116] Finally, the target model determines the strength of the medical recommendation based on the evidence quality level of each medical article, the favorable and unfavorable outcomes, the patient preference model obtained from external sources, and the resource cost of the medical recommendation determined based on the medical articles or external resources.

[0117] In a more specific embodiment, the third prompt in step S210 also instructs the target model to output a list of supporting evidence for the medical recommendation (including literature citations, effect size, quality rating), a reasoning path diagram (a logical chain showing the entire process from the level of evidence to the basis for evaluation, the trade-offs, and the strength of the recommendation), and an uncertainty prompt due to insufficient research content in existing medical literature (such as "more follow-up data are needed to support the long-term effects on renal function").

[0118] In this embodiment, the medical report also includes the following items: the logical chain in the step-by-step reasoning process, the medical literature and initial assessment results, and uncertainty warnings.

[0119] The above describes the process of automatically generating medical reports end-to-end based on the GRADE methodology and large language model.

[0120] In some possible implementations, the structured initial evaluation results of the target documents identified in step S208 can also be converted into knowledge points in the form of triples to generate a corresponding medical knowledge graph, or to update an existing medical knowledge graph.

[0121] In this embodiment, the method further includes:

[0122] The medical knowledge graph is updated based on the initial evaluation results.

[0123] In one embodiment, the structured initial evaluation results are input into the target model, which is then instructed to convert each initial evaluation result into knowledge points in the form of triples and update the medical knowledge graph based on each knowledge point.

[0124] This specification's embodiments combine the GRADE methodology and large language models to propose a method for generating medical reports based on evidence-based medicine, effectively overcoming several shortcomings of related technologies. By actively accessing real-time updated medical databases, it overcomes the problem of static knowledge, achieving dynamic acquisition and real-time updating of research evidence, and avoiding reliance on outdated manual knowledge bases.

[0125] By using a large language model for GRADE evidence assessment, retrieved medical studies are automatically scored across five dimensions: risk of bias, consistency, directness, precision, and publication bias. This ensures that only highly credible evidence is used in recommendation generation, thus addressing the problem of missing evidence quality assessment.

[0126] By using a large language model to output a complete decision chain that includes "level of evidence - assessment basis - cost-benefit trade-off - recommendation strength", the recommendation process becomes traceable and auditable, significantly enhancing clinical trust and improving the transparency and interpretability of medical reports.

[0127] By combining large language models, knowledge graphs, and rule-based reasoning, it completes end-to-end processing from natural language questions to structured recommendations, significantly reducing the need for manual intervention, supporting the rapid iteration and personalized application of clinical guidelines, and achieving synergy between automation and intelligence.

[0128] According to another embodiment, an apparatus for generating medical reports based on evidence-based medicine is also provided. Figure 3 A schematic block diagram of an apparatus for generating medical reports based on evidence-based medicine, according to one embodiment, is shown. This apparatus can be deployed in any device, platform, or cluster of devices with computing and processing capabilities. Figure 3 As shown, the device 300 includes:

[0129] Acquisition unit 302 is configured to acquire patient information and original clinical questions in natural language form;

[0130] The structured conversion unit 304 is configured to input the patient information, the original clinical question, and the first prompt word into the target model, so that the target model extracts the corresponding information from the input content based on the structured output template in the first prompt word and generates a structured clinical question; the target model is a large language model in the medical field.

[0131] The retrieval unit 306 is configured to retrieve a medical literature database based on the clinical problem and obtain several medical articles related to the clinical problem; each medical article contains research content related to the clinical problem.

[0132] The initial evaluation unit 308 is configured to input any target medical literature and a second prompt word into the target model, so that the target model performs a bias risk assessment on the target literature, determines the initial evidence quality level according to the research type of the target literature, and generates a structured initial evaluation result of the target literature; the initial evaluation result includes the bias risk level and the initial evidence quality level.

[0133] The report generation unit 310 is configured to input the clinical problem, each medical literature article and its initial assessment results, and the third prompt word into the target model, so that the target model performs step-by-step reasoning based on the GRADE evaluation process in the third prompt word to generate a medical report; the medical report includes medical recommendations for the clinical problem, as well as the recommendation strength and evidence quality level of the medical recommendations.

[0134] In some possible implementations, the device 300 further includes:

[0135] The alignment unit is configured to input the initial evaluation results of each medical article into the target model, and instruct the target model to perform semantic alignment of the technical terms in each initial evaluation result in order to update the initial evaluation results of each medical article.

[0136] In some possible implementations, the device 300 further includes:

[0137] The knowledge graph update unit is configured to update the medical knowledge graph based on each initial evaluation result.

[0138] According to another embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed in a computer, causes the computer to perform the methods described in any of the above embodiments.

[0139] According to another embodiment, a computing device is also provided, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method described in any of the above embodiments.

[0140] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0141] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0142] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0143] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0144] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating medical reports based on evidence-based medicine, comprising: Obtain patient information and raw clinical questions in natural language; the patient information includes basic patient information, disease information, examination results, treatment process and results, and medications used; the raw clinical questions include content related to the treatment of the patient's disease. The patient information, original clinical question, and first prompt word are input into the target model. The target model extracts the patient information and corresponding content from the original clinical question from the input content based on the structured output template in the first prompt word, fills the structured output template, and integrates the patient information to perform contextualized modeling to generate a structured clinical question. The target model is a large language model in the medical field; the structured clinical questions include: patient characteristics, proposed interventions, control measures, and clinical outcomes of interest. A search of medical literature databases was conducted based on the stated clinical problem, yielding several medical articles related to the clinical problem; each medical article contained research content related to the stated clinical problem. The target model is input into a set of medical literature articles, including any target article and a second prompt word. The target model then performs a bias risk assessment on the target article and determines an initial evidence quality level based on the research type of the target article, generating a structured initial assessment result. This initial assessment result includes a bias risk level and an initial evidence quality level. The second prompt word also includes several assessment scales, each containing multiple questions and several options. The second prompt word instructs the target model to answer each question in the assessment scale based on the content of the target article, and the bias risk level of the target article is determined by combining the answers to all questions. The clinical problem, various medical literature articles and their initial assessment results, and the third cue word are input into the target model. The target model then uses the GRADE evaluation process in the third cue word and the patient preference model obtained from an external source to perform step-by-step reasoning based on the thought chain, and generates a medical report. The medical report includes medical recommendations for treating the patient's disease in response to the clinical problem, as well as the recommendation strength and evidence quality level of the medical recommendations.

2. The method according to claim 1, wherein, The first prompt word instructs the target model to transform the original clinical question and patient information into a structured clinical question based on the structured output template.

3. The method according to claim 2, wherein, The first prompt word also includes several example samples, each of which contains example input content and example output content; the first prompt word also instructs the target model to perform inference using the several example samples as examples.

4. The method according to claim 1, wherein, A search of medical literature databases based on the stated clinical problem yielded several medical articles related to the clinical problem, including: The clinical question is input into the target model, which is then instructed to perform synonym replacements on some or all of the content in the clinical question to obtain several synonymous clinical questions with the same meaning as the clinical question. The clinical question and several synonymous clinical questions were searched in the medical literature database, and the search results were summarized to obtain several medical articles.

5. The method according to claim 1, wherein, The second prompt word instructs the target model to use the corresponding assessment scale to assess the risk of bias in the target literature based on the research type of the target literature.

6. The method according to claim 1, wherein, The initial evaluation results also include the following items: research type, research results, and research value.

7. The method according to claim 1, wherein, The target model, based on the GRADE evaluation process in the third cue word, performs step-by-step reasoning through a thought process, including: The target model adjusts the initial evidence quality level of each medical article based on its individual risk of bias, as well as the consistency, directness, precision, and publication bias among the articles, to obtain the evidence quality level of the medical recommendations. Based on various medical literature articles, the target model determines the favorable and unfavorable outcomes of interventions for the stated clinical problem. The target model summarizes the research conclusions of various medical literature articles to determine medical recommendations; The target model determines the strength of the medical recommendations based on the evidence quality level of each medical article, the favorable and unfavorable outcomes, the obtained patient preference model, and the resource cost of the medical recommendations.

8. The method according to claim 1, wherein, The medical report also includes the following items: the logical chain in the step-by-step reasoning process, the medical literature and initial assessment results, and uncertainty warnings.

9. The method according to claim 1, further comprising, before inputting the clinical problem, each medical literature article and its initial assessment results, and the third prompt word into the target model: The initial evaluation results of each medical article are input into the target model, which is then instructed to semantically align the technical terms in each initial evaluation result in order to update the initial evaluation results of each medical article.

10. The method according to claim 1, further comprising: The medical knowledge graph is updated based on the initial evaluation results.

11. A device for generating medical reports based on evidence-based medicine, comprising: The acquisition unit is configured to acquire patient information and raw clinical questions in natural language; the patient information includes basic patient information, disease information, examination results, treatment process and results, and medications used; the raw clinical questions include content related to the treatment of the patient's disease. The structured transformation unit is configured to input the patient information, the original clinical question, and the first prompt word into the target model, so that the target model extracts the patient information and the corresponding content in the original clinical question from the input content according to the structured output template in the first prompt word, fills the structured output template, and integrates the patient information to perform contextualized modeling to generate a structured clinical question. The target model is a large language model in the medical field; the structured clinical questions include: patient characteristics, proposed interventions, control measures, and clinical outcomes of interest. The retrieval unit is configured to search a medical literature database based on the clinical problem and obtain several medical articles related to the clinical problem; each medical article contains research content related to the clinical problem. The initial evaluation unit is configured to input any target document and a second prompt word from several medical articles into the target model. The target model then performs a bias risk assessment on the target document and determines an initial evidence quality level based on the research type of the target document, generating a structured initial evaluation result. The initial evaluation result includes a bias risk level and an initial evidence quality level. The second prompt word also includes several evaluation scales, each containing multiple questions and several options. The second prompt word instructs the target model to answer each question in the evaluation scale based on the content of the target document, and the overall bias risk level of the target document is determined by combining the answers to all questions. The report generation unit is configured to input the clinical problem, various medical literature articles and their initial assessment results, and the third prompt word into the target model, so that the target model performs step-by-step reasoning based on the GRADE evaluation process in the third prompt word and the patient preference model obtained from the outside, and generates a medical report; the medical report includes medical recommendations for treating the patient's disease in response to the clinical problem, as well as the recommendation strength and evidence quality level of the medical recommendations.

12. A computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method of any one of claims 1-10.

13. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and when the processor executes the executable code, it implements the method of any one of claims 1-10.

Images