Hypertension group knowledge recommendation method and system based on large language model multi-agent

Abstract

The invention belongs to the technical field of medical treatment and public health, and particularly relates to a hypertension group knowledge recommendation method and system based on a large language model multi-agent. Hypertension knowledge question-answer pair information, medical short video information and user personal health portraits are finely extracted by constructing a hypertension knowledge question-answer pair library; according to the method, video content quality evaluation and matching degree evaluation are performed, and user platform interaction data adjustment is combined, so that personalized and accurate knowledge recommendation services can be provided for hypertension users, the requirements of hypertension patients for health knowledge are met, the accuracy and effectiveness of knowledge recommendation are improved, and self-health management of the patients is facilitated.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical and health care, and particularly relates to a hypertension group knowledge recommendation method and system based on a large language model multi-agent. BACKGROUND

[0002] Under the background of global economic integration and technological innovation, big data technology has deeply penetrated into various aspects of human life, especially in the field of medical health. The industry is at a critical node of transformation and is gradually entering a new era driven by big data. First, medical institutions use big data analysis systems to deeply mine clinical databases, thereby developing more accurate and efficient treatment plans for patients. Second, the widespread deployment of electronic medical record systems has enabled the informatization of patient medical records. Through intelligent matching technology, individual cases are associated with medical teams to ensure timely and targeted diagnosis and treatment services. In addition, the combination of big data and cloud computing technology enables intelligent cloud management and configuration optimization of medical resources. Furthermore, through in-depth analysis of individual difference data, medical professionals identify patients' specific needs and provide personalized medical services, realizing a new medical model that is "people-oriented".

[0003] In the field of chronic disease management, such as hypertension, the public's demand for health knowledge continues to grow. However, there are many challenges in acquiring and applying hypertension health knowledge. On the one hand, the Internet is flooded with a large amount of information related to hypertension of varying quality, making it difficult for users to judge the professionalism and credibility of these information, and easily influenced by non-professional or misleading content. On the other hand, medical health knowledge exists in various forms such as text, audio, and video, with unstructured or semi-structured data dominating, making it difficult to manage and utilize systematically. For example, question and answer texts between doctors and patients contain rich implicit knowledge, and medical short videos attract users with their intuitive and vivid characteristics, but these contents are difficult to be effectively identified and extracted by machines. In addition, existing knowledge recommendation systems generally lack personalization and are difficult to provide precise and adaptive knowledge content according to the individual differences in the health status of hypertension patients, including blood pressure level, comorbidities, lifestyle, disease stage, etc. For example, hypertension patients with different risk levels have significant differences in the focus of required knowledge and intervention measures, and traditional generalized recommendation methods cannot meet this fine-grained demand. SUMMARY

[0004] The application provides a high blood pressure group knowledge recommendation method and system based on a large language model multi-agent, which can provide personalized and accurate knowledge recommendation services for high blood pressure users by constructing a high blood pressure knowledge question and answer pair library, finely extracting high blood pressure knowledge question and answer pair information, medical short video information, and user personal health portraits, performing video content quality evaluation and matching degree evaluation, and adjusting user platform interaction data.

[0005] A high blood pressure group knowledge recommendation method based on a large language model multi-agent, comprising: Based on high blood pressure knowledge, a high blood pressure knowledge question and answer pair library and a high blood pressure risk level determination system are constructed, and according to the high blood pressure knowledge question and answer pair library, knowledge question and answer pair texts are classified and structured knowledge question and answer pair information is extracted; Medical short videos of high blood pressure are collected and downloaded, video content, publisher type, video publishing time and interaction metadata are extracted, video quality evaluation standards are established according to the video content, publisher type and interaction metadata, and video quality scores are generated; Based on user personal health information and user platform interaction data, a user personal health portrait is constructed, and the high blood pressure risk level determination system is combined to determine the current user's high blood pressure risk level; The user personal health portrait is matched with the structured knowledge question and answer pair information and the video content in multiple dimensions to calculate the matching degree score; the multiple dimension matching includes semantic similarity matching and high blood pressure risk level alignment matching; Based on the video quality score, the matching degree score, the video publishing time and the user platform interaction data, an individualized video recommendation list is generated by sorting, and video recommendation optimization is performed according to user feedback behavior.

[0006] By constructing a high blood pressure knowledge question and answer pair library, finely extracting high blood pressure knowledge question and answer pair information, medical short video information and user personal health portraits, performing video content quality evaluation and matching degree evaluation, and adjusting user platform interaction data, personalized and accurate knowledge recommendation services can be provided for high blood pressure users, the needs of high blood pressure patients for health knowledge can be met, the accuracy and effectiveness of knowledge recommendation can be improved, and patient self-health management can be facilitated.

[0007] Further, the high blood pressure knowledge question and answer pair library and the high blood pressure risk level determination system are constructed based on high blood pressure knowledge, and according to the high blood pressure knowledge question and answer pair library, knowledge question and answer pair texts are classified and structured knowledge question and answer pair information is extracted, comprising: Based on the knowledge of hypertension, define the hypertension knowledge question pair, extract the entity type and the relationship type between entities, and construct the hypertension knowledge question pair library; the entity type includes: medication precautions, lifestyle intervention, symptoms, living habits, and curative effect; Based on the blood pressure level of the patient, the combined disease condition, and the living habits, comprehensive risk assessment is performed, the hypertension risk level is divided, and a hypertension risk level determination system is constructed; Based on the hypertension knowledge question pair library, the knowledge question pair text classification is performed, and the structured knowledge question pair information extraction is performed according to the knowledge question pair text classification result; the knowledge question pair text classification includes popular science question and answer type, clinical consultation and health management question and answer type, and the clinical consultation and health management question and answer type includes complication and risk type, health management type, psychological health type, disease condition judgment type, and treatment and drug type.

[0008] By defining the hypertension knowledge question pair, constructing the hypertension knowledge question pair library, providing basic data for subsequent knowledge processing and recommendation, and classifying the knowledge question pair text, the knowledge is structured, which facilitates subsequent knowledge matching and recommendation, and improves the utilization efficiency.

[0009] Further, the comprehensive risk assessment based on the blood pressure level of the patient, the combined disease condition, and the living habits, the hypertension risk level is divided, and a hypertension risk level determination system is constructed, including: Obtain the blood pressure level score of the patient; the blood pressure level score is determined by the blood pressure value; Obtain the combined disease condition score of the patient; the combined disease condition score is determined by the prognosis influence degree of high-risk disease factors; Obtain the patient's living habit score; the patient's living habit score is determined by the harmful behavior and the degree of harm; Combine the patient's blood pressure level score, the patient's combined disease condition score, and the patient's living habit score to calculate the patient's comprehensive risk score; Based on the weighted calculation of the patient's comprehensive risk score, the hypertension risk level of the patient is determined according to the hypertension risk level division standard; the hypertension risk level includes low risk, medium risk, medium-high risk, high risk, high-very high risk, and very high risk.

[0010] By considering multiple factors to assess the patient's risk, the hypertension risk level of the patient is determined according to the hypertension risk level division standard, which accurately reflects the patient's hypertension condition and health risk, and provides a basis for personalized knowledge recommendation.

[0011] Further, the knowledge question pair text classification is performed based on the hypertension knowledge question pair library, and the structured knowledge question pair information extraction is performed according to the knowledge question pair text classification result, including: Based on the popular science question and answer type text information, the special groups and brief questions, brief replies and suggestion information are extracted to generate structured knowledge question and answer information. Based on the clinical consultation and health management question and answer type text information, a large language model is used to extract patient question and answer information to generate first-stage extracted information. Based on the clinical consultation and health management question and answer type text information, a one-prompt method is used to extract systematic patient question and answer information to generate second-stage extracted information, and a hypertension risk level determination system is used to determine the hypertension risk level of the patient. Based on the multi-model cooperation and cross-validation mechanism, the consistency of the first-stage extracted information and the second-stage extracted information is compared, and the optimal structured knowledge question and answer pair information is output.

[0012] Different processing methods are used to process the clinical consultation and health management question and answer type text information, and consistency comparison is used to more comprehensively and accurately extract knowledge information to provide data support for subsequent matching.

[0013] Further, the medical short video of hypertension is collected and downloaded, the video content, publisher type, video publishing time and interaction metadata are extracted, and the video quality evaluation standard is established according to the video content, publisher type and interaction metadata, and the video quality score is generated, including: Based on the preset prompt word template, the medical short video of hypertension is automatically recognized, downloaded and parsed, the video content, publisher type, video publishing time and interaction metadata are extracted, and the video content is converted into structured video text information; the interaction metadata includes the number of likes, the number of comments and the number of forwards; Based on the prompt word template, health behavior suggestions are extracted from the structured video text information to construct an intermediate semantic layer of knowledge recommendation; Based on the video interaction metadata, the video influence score is calculated; Based on the video publisher type, the video professional score is determined; Based on the video content, the DISCERN scale of medical information quality evaluation is combined to calculate the video content score; The video quality score is calculated by combining the video influence score, the video professional score and the video content score.

[0014] By extracting information from medical short videos and evaluating video quality based on different dimensions of information, high-quality videos can be screened, and the reliability and professionalism of recommended video content can be ensured, improving the quality of user knowledge acquisition.

[0015] Further, based on the video content, the DISCERN scale of medical information quality evaluation is combined to calculate the video content score, including: select a plurality of medical experts, and each medical short video is evaluated in multiple dimensions, and the DISCERN scale of medical information quality evaluation is combined to calculate the DISCERN score of each medical expert for the current medical short video; the multiple dimension evaluation includes definiteness evaluation, reliability evaluation, objectivity evaluation, sufficiency evaluation, and expressiveness evaluation; Fleiss's Kappa coefficient is introduced to test the cross-dimension consistency of the classification selection of the medical short video by the plurality of medical experts, to determine the average consistency of the current medical short video, and to generate a consistency quantization result; The consistency quantization result is compared with the preset threshold, and the video content score of the current medical short video is determined according to the comparison result; when the consistency quantization result reaches the preset threshold, the average of the DISCERN scores of each medical expert for the current medical short video is taken as the video content score of the current medical short video; when the consistency quantization result does not reach the preset threshold, the DISCERN score of the re-evaluation expert for the current medical short video is taken as the video content score of the current medical short video.

[0016] By introducing the multi-dimensional evaluation of medical experts on video content, the quality of video content can be evaluated professionally and comprehensively, and the reliability and accuracy of the video content score can be ensured according to the cross-dimension consistency test result, thereby improving the video quality evaluation standard.

[0017] Further, the multi-dimensional matching of the user's personal health portrait with the structured knowledge question and answer pair information and the video content, and the calculation of the matching degree score, includes: Integrate the structured knowledge question and answer pair information and the video content into knowledge items; Use a large language model to calculate the cosine similarity of the user's personal health portrait and the knowledge items, and filter and retain the knowledge items with a cosine similarity higher than a semantic similarity threshold as the semantic similarity matching result; Use a large language model to analyze the user's personal health portrait to identify the user's intent, and match the knowledge items corresponding to the high blood pressure risk level based on the user's intent, and take them as the risk level alignment matching result; Combine the semantic similarity matching result and the risk level alignment matching result to calculate the matching degree score.

[0018] By combining the user's personal health portrait and the knowledge items, the matching degree of the knowledge items with the user is measured, thereby improving the accuracy of the recommended video.

[0019] Further, the multi-dimensional matching of the user's personal health portrait with the structured knowledge question and answer pair information and video content also includes knowledge graph path association matching; when the construction form of the hypertension knowledge question and answer pair library is a graph form, the associated path of the user's personal health portrait and the structured knowledge question and answer pair information is mined based on graph embedding technology, and the shortest path or the highest weight associated path in the associated path is identified to obtain potential associated knowledge as the knowledge graph path association matching result.

[0020] Further, the generation of the personalized video recommendation list based on the video quality score, the matching degree score, the video publishing time and the user platform interaction data, and the optimization of video recommendation according to user feedback behavior include: determining video timeliness based on the video publishing time; determining user preferences by learning user implicit feedback using collaborative filtering based on user platform interaction data; combining the video quality score, the matching degree score, the video timeliness and the user preferences to calculate a video recommendation score, and sorting to generate a personalized video recommendation list with a recommendation reason; collecting user feedback behavior on recommended videos to optimize recommended videos.

[0021] The calculation of the video recommendation score by considering multiple factors, the generation of the personalized video recommendation list meeting user needs and user preferences, and the optimization of video recommendation according to user feedback behavior can improve the accuracy of video recommendation and thus improve user experience.

[0022] A system of a hypertension population knowledge recommendation method based on a large language model multi-agent, comprising: A knowledge acquisition module for constructing a hypertension knowledge question and answer pair library and a hypertension risk level determination system based on hypertension knowledge, and classifying and extracting structured knowledge question and answer pair information from knowledge question and answer pair texts according to the hypertension knowledge question and answer pair library; A video acquisition and quality evaluation module for collecting and downloading medical short videos of hypertension, extracting video content, publisher type, video publishing time and interaction metadata, and establishing video quality evaluation standards and generating video quality scores according to video content, publisher type and interaction metadata; A user information acquisition module for constructing a user's personal health portrait based on user personal health information and user platform interaction data, and determining the current user's hypertension risk level in combination with the hypertension risk level determination system; A matching and calculation module for multi-dimensional matching of the user's personal health portrait with the structured knowledge question and answer pair information and video content, and calculating a matching degree score; the multi-dimensional matching includes semantic similarity matching and hypertension risk level alignment matching; A recommendation and optimization module is used for sorting and generating a personalized video recommendation list based on the video quality score, matching degree score, video publishing time and user platform interaction data, and optimizing video recommendation according to user feedback behavior.

[0023] The present application has the following advantages: The present application can provide personalized and accurate knowledge recommendation services for hypertension users by constructing a hypertension knowledge question and answer pair database, fine-grained extraction of hypertension knowledge question and answer pair information, medical short video information and user personal health portraits, video content quality evaluation and matching degree evaluation, and adjustment of user platform interaction data, meet the needs of hypertension patients for health knowledge, improve the accuracy and effectiveness of knowledge recommendation, and help patients to manage their own health. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The flowchart of the present application is shown in the figure; Figure 2 The figure shows a schematic diagram of the framework of the hypertension group knowledge recommendation method; Figure 3 The figure shows a schematic diagram of the treatment inquiry entity in the hypertension knowledge question and answer pair; Figure 4 The figure shows a schematic diagram of the treatment suggestion entity in the hypertension knowledge question and answer pair; Figure 5 The figure shows a schematic diagram of the optimal structured knowledge question and answer pair information extraction framework; Figure 6 The figure shows a schematic diagram of the medical short video quality evaluation system process; Figure 7 The figure shows a schematic diagram of the framework of the personalized video recommendation list; Figure 8 The figure shows a schematic diagram of the system structure of the present application. DETAILED DESCRIPTION

[0025] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0026] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this disclosure, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number of aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using other structures and / or functionalities besides one or more of the aspects set forth herein.

[0027] In addition, specific details are provided in the following description to facilitate a thorough understanding of the examples, and those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0028] Example 1 Figure 1 This paper presents a knowledge recommendation method for hypertension patients based on a large language model and multi-agent approach. By constructing a hypertension knowledge question-and-answer pair database, it extracts hypertension knowledge question-and-answer pair information, medical short video information, and user personal health profiles in a refined manner. It then conducts video content quality assessment and matching degree assessment, and adjusts the method by combining user platform interaction data. This approach can provide personalized and accurate knowledge recommendation services for hypertension patients, meet their health knowledge needs, improve the accuracy and effectiveness of knowledge recommendations, and help patients manage their own health. Figure 2 The diagram shown illustrates the framework for a knowledge recommendation method for people with hypertension, which includes the following steps: S1: Based on hypertension knowledge, construct a hypertension knowledge question-and-answer pair database and a hypertension risk level determination system, and classify and extract structured knowledge question-and-answer pair information according to the hypertension knowledge question-and-answer pair database; S21: Based on hypertension knowledge, define hypertension knowledge question-answer pairs, extract knowledge entity types and relationship types between entities, and construct a hypertension knowledge question-answer pair library; In this embodiment, the entity types include treatment inquiry entities and treatment suggestion entities.

[0029] Among them, such as Figure 3 As shown, the treatment recommendations include patient gender, patient age, patient special population, patient hypertension type, patient family history, patient symptoms, patient comorbidities, patient blood pressure, patient lifestyle habits, patient medication (tablet name and dosage), patient medication effect, patient medication side effects, patient hypertension risk level, and patient mood. Among them, such as Figure 4As shown, the treatment recommendation entity includes doctor-recommended drugs (tablet name and usage), doctor's objective disease characterization description, doctor-recommended drug efficacy, doctor-recommended tablet side effects, doctor-recommended examination, doctor's medication precautions, and lifestyle interventions.

[0030] S22: Based on the patient's blood pressure level, the patient's comorbidities, and the patient's lifestyle, a comprehensive risk assessment is performed, the high blood pressure risk level is divided, and a high blood pressure risk level determination system is constructed. S221: Obtain the patient's blood pressure level score. In this embodiment, according to the authoritative standard of the definition of hypertension in the "Chinese Hypertension Prevention and Treatment Guidelines", the abnormal degree of systolic blood pressure (SBP) and diastolic blood pressure (DBP) is taken as the preliminary risk indicator. The blood pressure level score includes normal blood pressure (SBP < 120 mmHg and DBP < 80 mmHg) assigned 0 points, normal high value (SBP 120-139 mmHg or DBP 80-89 mmHg) assigned 1 point, 1st grade hypertension (SBP 140-159 mmHg or DBP 90-99 mmHg) assigned 2 points, 2nd grade hypertension (SBP 160-179 mmHg or DBP 100-109 mmHg) assigned 3 points, and 3rd grade hypertension (SBP ≥ 180 mmHg or DBP ≥ 110 mmHg) assigned 4 points.

[0031] S222: Obtain the patient's comorbidities score. In this embodiment, it is determined whether the patient has high-risk disease factors such as diabetes, chronic kidney disease, coronary heart disease, stroke, heart failure, peripheral arterial disease, and hyperlipidemia, and when the patient has one high-risk disease factor, different weight scores are assigned according to the influence of the high-risk disease factor on the prognosis of hypertension, and then the comorbidities score is determined.

[0032] S223: Obtain the patient's lifestyle score. In this embodiment, it is determined whether the patient has adverse behaviors such as smoking, excessive alcohol consumption, high-salt diet, lack of exercise, excessive mental stress, and obesity (BMI ≥ 28 kg / m²), and the patient's lifestyle score is set based on the harmfulness of the adverse behaviors, and when the patient has one adverse behavior, the patient's lifestyle score increases by 0.5-1 points.

[0033] S224: Combine the patient's blood pressure level score, the patient's comorbidities score, and the patient's lifestyle score to calculate the patient's comprehensive risk score. S225: Based on the weighted calculation of the patient's comprehensive risk score, and combined with the high blood pressure risk level division standard, the patient's high blood pressure risk level is determined. S2251: when the patient's comprehensive risk score is 0-2, define the hypertension risk level as low risk; S2252: when the patient's comprehensive risk score is 2.1-4, define the hypertension risk level as moderate risk; S2253: when the patient's comprehensive risk score is 4.1-6, define the hypertension risk level as moderate-high risk; S2254: when the patient's comprehensive risk score is 6.1-8, define the hypertension risk level as high risk; S2255: when the patient's comprehensive risk score is 8.1-10, define the hypertension risk level as high-very high risk; S2256: when the patient's comprehensive risk score is 10 or more, define the hypertension risk level as very high risk; S23: based on the knowledge question and answer pair library, perform knowledge question and answer pair text classification, and perform structured knowledge question and answer pair information extraction according to the knowledge question and answer pair text classification results; In this embodiment, the knowledge question and answer pair text classification includes popular science question and answer type, clinical consultation and health management question and answer type, and the clinical consultation and health management question and answer type includes complication and risk type, health management type, psychological health type, disease judgment type, and treatment and drug type.

[0034] S231: based on the popular science question and answer type text information, using a large language model to extract special population and brief questions, brief replies and suggestion information, and generate structured knowledge question and answer information; Specifically, special population and brief questions are extracted from the question text, and brief replies and suggestion information are extracted from the answer text, and the extracted information is integrated to form structured knowledge question and answer information.

[0035] S232: based on the text information of the clinical consultation and health management question and answer type, extract the optimal structured knowledge question and answer pair information; S2321: based on the text information of the clinical consultation and health management question and answer type, using a large language model to extract patient question and answer information, and generate first-stage extraction information; Specifically, based on the text information of the clinical consultation and health management question and answer type, the patient's personal basic condition and drug taking condition are extracted from the question text, and the question and answer pair is analyzed and graded based on the pre-set emotion level standard, and the drug treatment scheme, non-drug treatment suggestion and objective disease representation are extracted from the answer text, and the first-stage extraction information is integrated and output.

[0036] S2322: based on the text information of the clinical consultation and health management question and answer type, using one-prompt method to extract systematic patient question and answer information, generating second-stage extraction information, and determining the patient's hypertension risk level combined with the hypertension risk level determination system; Specifically, based on the clinical consultation and health management question and answer type text information, the patient's special population, patient blood pressure, patient lifestyle, patient comorbidities and other information are extracted from the patient's personal basic condition; the patient's medication (tablet name and usage), patient's medication effect are extracted from the medication taking condition; the patient's drug side effects are extracted from the drug treatment plan, and the second stage extraction information is integrated and output; and the patient's hypertension risk level is determined based on the hypertension risk level determination system.

[0037] S2323: Based on the multi-model cooperation and cross-validation mechanism, the consistency of the first stage extraction information and the second stage extraction information is compared, and the optimal structured knowledge question and answer pair information is integrated and output; Specifically, based on the multi-model cooperation and cross-validation mechanism, the consistency of the first stage extraction information and the second stage extraction information is compared according to the multi-dimensional standards of entity item content consistency, medical semantic integrity and logical rationality, that is, based on the same field, when the extraction information of the two is consistent, the consistent extraction information is retained; when the extraction information of the two is inconsistent, based on the original clinical consultation and health management question and answer type text information, re-extraction is performed, and according to the multi-dimensional standards of entity item content consistency, medical semantic integrity and logical rationality, the better extraction result is selected from the extraction results as the optimal structured knowledge question and answer pair information.

[0038] Figure 5 The figure shows an optimal structured knowledge question and answer pair information extraction framework diagram.

[0039] S2: Collect and download medical short videos of hypertension, extract video content, publisher type, video publishing time and interaction metadata, and establish video quality evaluation standards according to video content, publisher type and interaction metadata, and generate video quality score; S21: Based on the preset prompt word template, automatically identify, download and parse medical short videos of hypertension, extract video content, publisher type, video publishing time and interaction metadata, and convert the video content into structured video text information; In this embodiment, the interaction metadata includes the number of likes, the number of comments, and the number of forwards.

[0040] S22: Based on the prompt word template, extract health behavior recommendations from the structured video text information to build an intermediate semantic layer of knowledge recommendation; Specifically, based on the prompt word template, extract health behavior recommendations such as exercise recommendations, psychological adjustment methods, diet management plans and common cognitive errors from the structured video text information to build an intermediate semantic layer of knowledge recommendation.

[0041] S23: Based on the video interaction metadata, calculate the video influence score; Among them, the video interaction metadata based on the number of likes, comments, and shares is normalized, and its calculation expression is as follows: ; ; ; In the formula, Indicates the normalized number of... The number of likes for each medical short video; Indicates the first The number of likes for each medical short video; This represents the maximum number of likes for all videos. This represents the minimum number of likes for all videos. Indicates the normalized number of... The number of comments on each medical short video; Indicates the first The number of comments on each medical short video; This represents the maximum number of comments across all videos. This represents the minimum number of comments for all videos. Indicates the normalized number of... The number of reposts for each medical short video; Indicates the first The number of reposts for each medical short video; This represents the maximum number of times all videos have been forwarded. This represents the minimum number of times all videos have been forwarded. The expression for calculating the video influence score is as follows: ; In the formula, Indicates the first The video influence score of each medical short video; Indicates the weight of the number of likes; Indicates the weight of the number of comments; Indicates the weight of the number of forwards; S24: Determine the video's professionalism score based on the type of video publisher; The video publishers include medical professionals, professional institutions, and individual users. A video professionalism score is assigned to each type of video publisher, with the score ranging from 0 to 1.

[0042] In this embodiment, the video professionalism score for medical workers is set to 1.0; the video professionalism score for professional institutions is set to 0.8; and the video professionalism score for individual users is set to 0.3.

[0043] S25: Calculate the video content score based on the video content and the DISCERN scale for assessing the quality of medical information; S251: Select multiple medical experts to conduct multi-dimensional evaluations of the video content of each medical short video, and combine the DISCERN scale for medical information quality assessment to calculate the DISCERN score of each medical expert for the current medical short video. In this embodiment, the multi-dimensional evaluation includes clarity assessment, reliability assessment, objectivity assessment, adequacy assessment, and expressiveness assessment. Clarity assessment determines whether the medical short video clearly states its objective; reliability assessment determines whether the medical short video mentions reliable information sources; objectivity assessment determines whether the medical short video contains advertisements, exaggerations, or misleading information; adequacy assessment determines whether the medical short video systematically explains the treatment or management plan for hypertension; and expressiveness assessment determines whether the language of the medical short video is easy to understand and suitable for the general public.

[0044] In this embodiment, five medical experts are selected to evaluate the content of each medical short video based on five dimensions, with each dimension having a scoring standard of 0, 0.5, and 1 point. The scores from the five dimensions are summed to obtain the medical expert's DISCERN score for the current medical short video, and the calculation expression is as follows: The formula for calculating the overall DISCERN score of current medical short videos by a single medical expert is as follows: ; In the formula, Indicates the first The first medical expert in the The DISCERN rating of a medical short video; Indicates the first The first medical expert in the The first medical short video Scoring across multiple dimensions; Indicates the number of evaluation dimensions. , Indicates the total number of evaluation dimensions; Indicates the number of medical experts. , Indicates the total number of medical experts; S252: Introduce Fleiss's Kappa coefficient to conduct a cross-dimensional consistency test on the rating category selection of medical short videos by multiple medical experts, determine the average consistency of the current medical short videos, and generate consistency quantification results; Among them, Fleiss's Kappa coefficient is used to quantify the inter-group consistency level of multiple medical experts on a discrete scoring task, and the overall average consistency coefficient of the current medical short videos is calculated and used as the consistency quantification result. The expression is as follows: ; In the formula, represents the overall consistency coefficient of the mth medical short video; represents the actual consistency degree of the mth medical short video in the nth dimension, and its expression is: represents the actual consistency degree of the mth medical short video in the nth dimension, and its expression is: represents the evaluation category, represents the total number of evaluation categories, represents the number of times that the nth evaluation category of the mth medical short video in the nth dimension is selected by the experts; represents the number of times that the nth evaluation category of the mth medical short video in the nth dimension is selected by the experts; represents the expected consistency degree of the mth medical short video in the nth dimension, and its expression is: S253: Compare the consistency quantization result with a preset threshold value, and determine the video content score of the current medical short video according to the comparison result; S2531: When the consistency quantization result reaches the preset threshold value, the average DISCERN score of the current medical short video given by each medical expert is taken as the video content score of the current medical short video; When the consistency quantization result reaches the preset threshold value, i.e. , the calculation expression of the video content score of the current medical short video is: ; In the formula, represents the final DISCERN total score of the mth medical short video; represents the total number of experts participating in the final score; represents the preset threshold value, in this embodiment, ; S2532: When the consistency quantization result does not reach the preset threshold value, the DISCERN score of the current medical short video given by the re-evaluation expert is taken as the video content score of the current medical short video; When the consistency quantization result does not reach the preset threshold value, i.e. , the calculation expression of the video content score of the current medical short video is: ; In the formula, represents the final DISCERN total score of the mth medical short video; represents the total number of experts participating in the final score; represents the preset threshold value, in this embodiment, ; represents the final DISCERN total score of the mth medical short video; represents the total number of experts participating in the final score; represents the preset threshold value, in this embodiment, ​​​​​DISCERN score of a medical short video; S26: combine the video influence score, the video professionalism score, and the video content score to calculate a video quality score by weighting; Based on the video influence score, the video professionalism score, and the video content score, the video quality score of the current medical short video is calculated after normalization processing, and the expression is: ; In the formula, represents the final video quality score of the i-th medical short video; represents the normalized influence score of the i-th medical short video; represents the weight of the normalized influence score of the i-th medical short video; represents the video professionalism score; represents the weight of the video professionalism score; represents the normalized video content score of the i-th medical short video; represents the weight of the normalized video content score of the i-th medical short video; ; ; Figure 6 As shown in FIG. 1, it is a schematic diagram of the medical short video quality evaluation system.

[0045] S3: Based on the user personal health information and the user platform interaction data, a user personal health portrait is constructed, and the high blood pressure risk level of the current user is determined by combining the high blood pressure risk level determination system; In this embodiment, the user personal health information is the personal health information input by the user, which includes gender, age, special population, family history, symptoms, comorbidities, blood pressure, lifestyle, patient medication (tablet name and usage), medication effect, and medication side effects. The user platform interaction data includes user historical questions, historical browsing records, user likes, collections, sharing records, and user viewing time.

[0046] Based on the user personal health portrait and the high blood pressure risk level determination system, the current high blood pressure risk level of the user is automatically updated.

[0047] S4: Multi-dimensional matching of the user personal health portrait, structured knowledge question and answer pair information, and video content is performed to calculate a matching degree score; S41: The structured knowledge question and answer pair information and the video content are integrated into a knowledge item; ​​​​In the embodiment, the information and the video content are integrated to form the knowledge entry in combination with the structured knowledge question and answer. The knowledge entry specifically includes a knowledge question and answer question, a knowledge question and answer answer abstract, a video script main body, and a video tag.

[0048] S42: A large language model is used to calculate the cosine similarity between the user personal health portrait and the knowledge entry, and the knowledge entry with a cosine similarity higher than a semantic similarity threshold is reserved as a semantic similarity matching result; In the embodiment, a pre-trained large language model BERT is used to calculate the cosine similarity between the keywords and phrases in the user personal health portrait and the embedding vectors of the knowledge entries, and compared with the semantic similarity threshold. The knowledge entry with a cosine similarity higher than the semantic similarity threshold is reserved as a semantic similarity matching result. By selecting the knowledge entry with a high similarity, the content theme can be more relevant.

[0049] S43: A large language model is used to analyze the user personal health portrait, perform user intent recognition, and match the knowledge entry corresponding to the high blood pressure risk level based on the user intent, and the knowledge entry is taken as a risk level alignment matching result; In the embodiment, the user's historical questions or actively input personal health information is analyzed by a large language model, and a multi-classification or sequence labeling task is used to identify its deep health needs and perform knowledge question and answer text classification to identify the specific intent of the user.

[0050] Based on the current high blood pressure risk level of the user, the knowledge content corresponding to the high blood pressure risk level is preferentially recommended, such as “healthy lifestyle for preventing high blood pressure” for low-risk users and “high blood pressure medication precautions” for high-risk users. When a boundary condition occurs or the user wants to improve health literacy, appropriate knowledge content of adjacent high blood pressure risk levels is recommended, but the recommendation of the current high blood pressure risk level of the user is still the main one. At the same time, based on different high blood pressure risk levels, the focus of the recommended knowledge content is adjusted, such as health education and lifestyle intervention for low-risk users, early intervention and monitoring for medium-risk users, and drug management, complication prevention and management, and specialist diagnosis and treatment recommendations for high-risk and very high-risk users.

[0051] S44: When the construction form of the high blood pressure knowledge question and answer library is a graph form, based on graph embedding technology, the association path between the user personal health portrait and the structured knowledge question and answer information is mined, knowledge graph path association matching is performed, and the shortest path or the highest weight association path in the association path is identified to obtain potential associated knowledge, which is taken as a knowledge graph path association matching result; In this embodiment, when the construction form of the hypertension knowledge question and answer pair library is a graph form, the graph embedding technology TransE is introduced to find the association path between the entities in the user's personal health portrait and the entities in the structured knowledge question and answer pair information, identify the shortest path or the highest weight association path in the association path, recommend potential associated knowledge, and take it as the knowledge graph path association matching result. For example, the user's personal health portrait involves "diabetes", and through the knowledge graph path association matching, the knowledge of "dietary recommendations for hypertension combined with diabetes" is obtained.

[0052] S45: Combine the semantic similarity matching result, the knowledge graph path association matching result, and the risk level alignment matching result to calculate the matching degree score.

[0053] The calculation expression of the matching degree score is: ; In the formula, represents the matching degree score; represents the semantic similarity matching result, which is obtained by cosine similarity; represents the semantic similarity matching result weight; represents the knowledge graph path association matching result, which is obtained by graph association degree; represents the knowledge graph path association matching result weight; represents the risk level alignment matching result, that is, when completely aligned , when partially aligned , and when completely not aligned ; represents the risk level alignment matching result weight; S5: Based on the video quality score, the matching degree score, the video publishing time, and the user platform interaction data, sort and generate a personalized video recommendation list, and optimize the video recommendation according to the user feedback behavior.

[0054] S51: Determine the video timeliness based on the video publishing time. In this embodiment, for health information with strong timeliness (guideline updates, new drug developments), the newer knowledge content is preferentially recommended; for basic popular science knowledge, the timeliness weight is appropriately reduced according to the time decay function. The expression of the time decay function is: ; In the formula, represents the video timeliness; represents the time difference between the video publishing time and the current time; represents the decay coefficient.

[0055] S52: learning user implicit feedback by collaborative filtering based on user platform interaction data, and determining user preference; In this embodiment, based on user platform interaction data, the user's preference for specific types, forms or thematic content is learned, and user implicit feedback is learned by collaborative filtering, and higher weight is given to the knowledge content preferred by the user.

[0056] S53: combining video quality score, matching degree score, video timeliness, and user preference to calculate video recommendation score, sorting to generate personalized video recommendation list, and attaching recommendation reason; The expression of the video recommendation score is: ; In the formula, represents the video recommendation score; represents the video quality score; represents the user preference; represents the matching degree score weight; represents the video quality score weight; represents the video timeliness weight; represents the user preference weight, and .

[0057] According to the video recommendation score from high to low, the user is recommended.

[0058] In this embodiment, the sorted knowledge entries are recommended in a structured list form to directly present the summary of the structured knowledge question and answer pair information, the title and brief description of the medical short video, and mark the video quality score, video professionalism and other key information; at the same time, the recommendation reason is generated for each recommended knowledge content to explain the adaptability of the knowledge entry to the current user.

[0059] In addition, in actual application, according to the user preference, various forms of content such as text, picture, video link can be flexibly used to further improve the user experience.

[0060] Figure 7 The frame schematic diagram of the personalized video recommendation list.

[0061] S54: collecting user feedback behavior on the recommended video, and optimizing the recommended video.

[0062] According to the feedback behavior of the user on the recommended video, such as click rate, stay time, collection, sharing, and user active evaluation, the recommended video is continuously optimized to improve the accuracy of the recommended knowledge content and user satisfaction.

[0063] Embodiment 2 Based on the same design concept, such as Figure 8As shown, the embodiment provides a high blood pressure population knowledge recommendation system based on a large language model multi-agent, which includes a knowledge acquisition module, a video acquisition and quality evaluation module, a user information acquisition module, a matching and calculation module, a recommendation and optimization module.

[0064] Specifically, the knowledge acquisition module is configured to construct a high blood pressure knowledge question and answer pair library and a high blood pressure risk level determination system based on high blood pressure knowledge, and classify knowledge question and answer pair texts and extract structured knowledge question and answer pair information according to the high blood pressure knowledge question and answer pair library. Specifically, the video acquisition and quality evaluation module is configured to collect and download medical short videos of high blood pressure, extract video content, publisher type, video publishing time and interactive metadata, and establish video quality evaluation standards and generate video quality scores according to the video content, publisher type and interactive metadata. Specifically, the user information acquisition module is configured to construct a user personal health portrait based on user personal health information and user platform interaction data, and determine the current user's high blood pressure risk level in combination with the high blood pressure risk level determination system. Specifically, the matching and calculation module is configured to perform multi-dimensional matching of the user personal health portrait, structured knowledge question and answer pair information and video content, and calculate a matching score; the multi-dimensional matching includes semantic similarity matching and high blood pressure risk level alignment matching. Specifically, the recommendation and optimization module is configured to sort and generate a personalized video recommendation list based on the video quality score, the matching score, the video publishing time and the user platform interaction data, and optimize video recommendation according to user feedback behavior.

[0065] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0066] Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

Claims

1. A method for recommending hypertension population knowledge based on a large language model multi-agent, characterized in that, The method comprises the following steps: Based on the knowledge of hypertension, build a hypertension knowledge question and answer pair library and a hypertension risk level judgment system, and according to the hypertension knowledge question and answer pair library, classify and structure the knowledge question and answer pair information extraction of the text; Collect and download medical short videos of hypertension, extract video content, publisher type, video release time and interactive metadata, and establish video quality evaluation standard according to video content, publisher type and interactive metadata, and generate video quality score; Based on the user's personal health information and the user's platform interaction data, build the user's personal health portrait, and determine the current user's hypertension risk level combined with the hypertension risk level judgment system; Multi-dimensional matching of user personal health portrait, structured knowledge question and answer pair information and video content is carried out to calculate the matching score; the multi-dimensional matching includes semantic similarity matching and hypertension risk level alignment matching; Based on the video quality score, matching score, video release time and user platform interaction data, sort and generate a personalized video recommendation list, and optimize the video recommendation according to the user feedback behavior.

2. The method according to claim 1, wherein, Based on the knowledge of hypertension, build a hypertension knowledge question and answer pair library and a hypertension risk level judgment system, and according to the hypertension knowledge question and answer pair library, classify and structure the knowledge question and answer pair information extraction of the text, including: Based on the knowledge of hypertension, define the hypertension knowledge question and answer pair, extract the knowledge entity type and the relationship type between entities, and build the hypertension knowledge question and answer pair library; the entity type includes: medication precautions, lifestyle intervention, symptoms, living habits, and therapeutic effect; Based on the patient's blood pressure level, combined disease condition and living habits, comprehensive risk assessment is carried out to divide the hypertension risk level and build the hypertension risk level judgment system; Based on the hypertension knowledge question and answer pair library, the knowledge question and answer pair text classification is carried out, and the structured knowledge question and answer pair information extraction is carried out according to the knowledge question and answer pair text classification result; the knowledge question and answer pair text classification includes popular science question and answer type, clinical consultation and health management question and answer type, and the clinical consultation and health management question and answer type includes complication and risk type, health management type, psychological health type, disease judgment type and treatment and drug type.

3. The method according to claim 2, wherein, Based on the patient's blood pressure level, combined disease condition and living habits, comprehensive risk assessment is carried out to divide the hypertension risk level and build the hypertension risk level judgment system, including: Obtain the patient's blood pressure level score; the blood pressure level score is determined by blood pressure value; Obtain the patient's combined disease condition score; the combined disease condition score is determined by the prognosis influence degree of high-risk disease factors; Obtain the patient's living habit score; the patient's living habit score is determined by the harmful behavior and the degree of harm; Combine the patient's blood pressure level score, the patient's combined disease condition score and the patient's living habit score to calculate the patient's comprehensive risk score; Based on the weighted calculation of the patient's comprehensive risk score, determine the patient's hypertension risk level combined with the hypertension risk level division standard; the hypertension risk level includes low risk, medium risk, medium-high risk, high risk, high-very high risk and very high risk.

4. The method according to claim 2, wherein, The knowledge question pair based on hypertension is classified, and the structured knowledge question pair information is extracted according to the knowledge question pair text classification result, including: Based on the popular science question and answer type text information, the special population and brief question, brief reply and suggestion information are extracted, and the structured knowledge question information is generated; Based on the clinical consultation and health management question and answer type text information, a large language model is used for patient question and answer information extraction to generate first-stage extraction information; Based on the clinical consultation and health management question and answer type text information, a one-prompt method is used for systematic patient question and answer information extraction to generate second-stage extraction information, and the hypertension risk level determination system is used to determine the hypertension risk level of the patient; Based on the multi-model cooperation and cross-validation mechanism, the consistency of the first-stage extraction information and the second-stage extraction information is compared, and the optimal structured knowledge question pair information is output.

5. The method according to claim 1, wherein, The medical short video of hypertension is collected and downloaded, the video content, publisher type, video publishing time and interactive metadata are extracted, and the video quality evaluation standard is established according to the video content, publisher type and interactive metadata, and the video quality score is generated, including: Based on the preset prompt word template, the medical short video of hypertension is automatically recognized, downloaded and parsed, the video content, publisher type, video publishing time and interactive metadata are extracted, and the video content is converted into structured video text information; the interactive metadata includes the number of likes, the number of comments and the number of forwards; Based on the prompt word template, health behavior recommendations are extracted from the structured video text information to build an intermediate semantic layer of knowledge recommendation; Based on the video interactive metadata, the video influence score is calculated; Based on the video publisher type, the video professional score is determined; Based on the video content, combined with the DISCERN scale of medical information quality evaluation, the video content score is calculated; Combined with the video influence score, the video professional score and the video content score, the video quality score is calculated by weighting.

6. The method according to claim 5, wherein, Based on the video content, combined with the DISCERN scale of medical information quality evaluation, the video content score is calculated, including: Select a plurality of medical experts to perform multi-dimensional evaluation on the video content of each medical short video, and calculate the DISCERN score of each medical expert for the current medical short video combined with the DISCERN scale of medical information quality evaluation; the multi-dimensional evaluation includes definiteness evaluation, reliability evaluation, objectivity evaluation, sufficiency evaluation and expressiveness evaluation; Fleiss's Kappa coefficient is introduced to perform cross-dimensional consistency test on the classification selection of the medical short video by multiple medical experts to determine the average consistency of the current medical short video, and generate a consistency quantization result; The consistency quantification result is compared with a preset threshold value, and a video content score of the current medical short video is determined according to a comparison result; when the consistency quantification result reaches the preset threshold value, the average DISCERN score of the current medical short video given by each medical expert is taken as the video content score of the current medical short video; when the consistency quantification result does not reach the preset threshold value, the DISCERN score of the current medical short video given by the re-evaluation expert is taken as the video content score of the current medical short video.

7. The method according to claim 1, wherein, The multi-dimensional matching of the user's personal health portrait with the structured knowledge question and answer pair information and the video content, and the calculation of the matching degree score include: The structured knowledge question and answer pair information and the video content are integrated into knowledge entries; The cosine similarity of the user's personal health portrait and the knowledge entries is calculated by using a large language model, and the knowledge entries with a cosine similarity higher than a semantic similarity threshold value are screened and reserved as the semantic similarity matching result; The user's personal health portrait is analyzed by using the large language model to identify the user's intention, and the knowledge entries corresponding to the high blood pressure risk level are matched based on the user's intention as the risk level alignment matching result; The semantic similarity matching result and the risk level alignment matching result are combined to calculate the matching degree score.

8. The method according to claim 7, wherein, The multi-dimensional matching of the user's personal health portrait with the structured knowledge question and answer pair information and the video content further includes knowledge graph path association matching; when the construction form of the high blood pressure knowledge question and answer pair library is a graph form, the associated paths of the user's personal health portrait and the structured knowledge question and answer pair information are mined based on graph embedding technology, and the shortest path or the associated path with the highest weight in the associated paths is identified to obtain potential associated knowledge as the knowledge graph path association matching result.

9. The method according to claim 1, wherein, The personalized video recommendation list is generated by sorting based on the video quality score, the matching degree score, the video publishing time, and the user platform interaction data, and the video recommendation is optimized according to the user feedback behavior, including: The video timeliness is determined based on the video publishing time; The user preference is determined by using collaborative filtering to learn the user's implicit feedback based on the user platform interaction data; The video recommendation score is calculated by combining the video quality score, the matching degree score, the video timeliness, and the user preference, the personalized video recommendation list is generated by sorting, and the recommendation reason is attached; The feedback behavior of the user to the recommended video is collected to optimize the recommended video.

10. A system for implementing the large language model-based multi-agent high blood pressure population knowledge recommendation method of claim 1, characterized in that, It includes: A knowledge acquisition module is configured to construct a high blood pressure knowledge question and answer pair library and a high blood pressure risk level determination system based on high blood pressure knowledge, and classify knowledge question and answer pair texts and extract structured knowledge question and answer pair information according to the high blood pressure knowledge question and answer pair library; A video acquisition and quality evaluation module is configured to collect and download medical short videos of high blood pressure, extract video content, publisher type, video publishing time, and interaction metadata, establish a video quality evaluation standard according to the video content, the publisher type, and the interaction metadata, and generate a video quality score; A user information acquisition module is configured to construct a user's personal health portrait based on the user's personal health information and user platform interaction data, and determine the high blood pressure risk level of the current user in combination with the high blood pressure risk level determination system. A matching and computing module is configured to perform multi-dimensional matching of the user's personal health profile with structured knowledge question-answer pair information and video content, and to calculate a matching score; the multi-dimensional matching includes semantic similarity matching and high blood pressure risk level alignment matching; A recommendation and optimization module is configured to sort and generate a personalized video recommendation list based on the video quality score, the matching score, the video publishing time, and the user platform interaction data, and to optimize the video recommendation according to the user feedback behavior.

Images