A method and system for recommending medical popular science articles based on user portraits

Abstract

The present application provides a method and system for recommending medical popular science articles based on user portraits. Among them, by integrating multi-source medical data, analyzing its logical associations and constructing a dynamic knowledge graph, and then in the graph, integrating the user's real-time diagnosis and treatment behavior and historical medical interaction records to generate a user portrait label set stratified by medical cognitive level, reflecting the user's depth of understanding and preference for medical concepts, and further migrating the preference features in the portrait from the non-realistic scenario of historical interaction to the realistic scenario of real-time diagnosis and treatment, and generating a dynamic recommendation feature vector synchronized with the evolution of medical knowledge through cross-scenario conversion. Finally, according to the matching results between the feature vector and the concepts in the knowledge graph, the article recommendation priority is dynamically adjusted to generate a personalized recommendation sequence that fits the user's real-time cognitive level; the technical solution provided by this application can improve the accuracy and timeliness of the recommendation of medical popular science articles.

Description

Technical Field

[0001] The present application relates to the field of medical popular science technology, and in particular to a method and system for recommending medical popular science articles based on user portraits. Background Art

[0002] With rising health awareness, users are increasingly demanding multi-dimensional medical knowledge, including disease prevention and medication guidelines. Due to significant differences in health status, knowledge level, and information preferences among user groups, the traditional "one-size-fits-all" distribution model for popular science content is unable to meet personalized needs.

[0003] The current technical solution is to use machine learning algorithms to classify a large number of medical popular science articles and match and recommend them based on the characteristic data in the user portrait. This method first uses natural language processing technology to parse the content of medical popular science articles, extract key words and topic information, and then uses machine learning models to intelligently match them based on the attributes in the user portrait to determine which articles are most likely to meet the needs of specific users. This method reduces manual intervention through automated processes, improves recommendation efficiency and accuracy, and provides users with a more personalized reading experience.

[0004] However, the above-mentioned existing solutions have certain defects. On the one hand, due to the complexity and professionalism of the medical field, natural language processing technology may encounter the problem of insufficient understanding depth when parsing medical popular science articles, resulting in a deviation between the recommended articles and the actual needs of users; on the other hand, the construction of user portraits relies on multiple data sources. If the data is not collected comprehensively or updated in a timely manner, it will directly affect the accuracy and relevance of the recommendation results. In addition, privacy protection is also an issue that cannot be ignored in the implementation of this solution. How to protect the security of user personal information while ensuring the effectiveness of recommendations is an important challenge facing this solution. Summary of the Invention

[0005] The present application provides a method and system for recommending medical popular science articles based on user portraits, which is used to solve the problems of poor accuracy and timeliness in the recommendation of traditional Chinese medicine popular science articles in the prior art.

[0006] In a first aspect, this application provides a method for recommending medical popular science articles based on user portraits, comprising:

[0007] By analyzing the logical association between unstructured text and structured case data in multi-source heterogeneous medical data, a dynamic knowledge graph containing dynamic dependencies between medical concepts is constructed;

[0008] In the dynamic knowledge graph, the acquired real-time diagnosis and treatment behavior data of users, historical medical interaction records of users and medical concepts are integrated in multiple dimensions to generate a user portrait tag set based on the user's medical cognition level as a hierarchical benchmark;

[0009] Performing cross-scenario feature conversion on the user medical preference features in the user portrait tag set from the non-realistic scenario of the user's historical medical interaction records to the real-world scenario of the user's real-time diagnosis and treatment behavior data, and generating a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge;

[0010] According to the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, the recommendation priority ranking of medical popular science articles is dynamically adjusted to generate a personalized recommendation sequence that matches the user's real-time medical cognitive level.

[0011] Optionally, the cross-scenario feature conversion of the user medical preference features in the user portrait tag set from the non-realistic scenario of the user's historical medical interaction records to the real-world scenario of the user's real-time diagnosis and treatment behavior data is performed to generate a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge, including:

[0012] Extracting the preference components of the non-realistic scenarios of the user's historical medical interaction records based on the user's medical preference features in the user portrait tag set;

[0013] By analyzing the implicit correlation pattern between the user's operation trajectory of medical concepts in the non-realistic scenario and the user's real-time diagnosis and treatment behavior data in the real scenario, a nonlinear projection relationship from the non-realistic scenario to the real scenario is constructed;

[0014] Based on the nonlinear projection relationship, the user's medical preference features are mapped to the feature space of the real scene through the preference components of the non-real scene, and the scene noise components that are unrelated to the user's medical cognitive level are separated in the feature space, while retaining the core preference components that are stably transferred in the cross-scene feature conversion;

[0015] By tracking the update status changes of the medical concepts, the dynamic offset of the core preference component on the medical knowledge evolution trend is calculated, and the dynamic offset is superimposed on the core preference component to generate a dynamic recommendation feature vector containing the medical knowledge evolution trend.

[0016] Optionally, the calculating of the dynamic offset of the core preference component on the evolution trend of medical knowledge by tracking the update status change of the medical concept includes:

[0017] Obtain the change records of the association relationships of the medical concepts described in the dynamic knowledge graph, extract the newly added dependencies and invalid dependencies between medical concepts, and obtain the updated status mark of the evolution trend of medical knowledge;

[0018] Based on the update status flag, determining a positive offset direction corresponding to the newly added dependency and a negative offset direction corresponding to the invalidated dependency in the core preference component;

[0019] According to the superposition of the positive offset direction and the negative offset direction, and combined with the proportional weights of the newly added dependencies and the invalidated dependencies, the dynamic offset of the core preference component in the evolution trend of medical knowledge is obtained.

[0020] Optionally, determining, based on the update status flag, a positive offset direction corresponding to the newly added dependency and a negative offset direction corresponding to the invalidated dependency in the core preference component includes:

[0021] Based on the update status flag, mapping the newly added dependency to a first adjustment parameter of the core preference component, and mapping the invalidated dependency to a second adjustment parameter of the core preference component;

[0022] Generate a first direction vector set based on the product of the association strength between the medical concepts connected by each newly added dependency and the first adjustment parameter, and generate a second direction vector set based on the product of the historical association strength between the medical concepts corresponding to each invalidated dependency and the second adjustment parameter;

[0023] The direction vectors in the first direction vector set are superimposed to obtain a positive offset direction, and the direction vectors in the second direction vector set are superimposed to obtain a negative offset direction.

[0024] Optionally, in the dynamic knowledge graph, the acquired real-time diagnosis and treatment behavior data of the user, the user's historical medical interaction records and the medical concepts are multi-dimensionally integrated to generate a user portrait tag set based on the user's medical cognitive level as a hierarchical benchmark, including:

[0025] Extract the operation timestamp and operation type identifier from the user's real-time diagnosis and treatment behavior data, as well as the behavior frequency and behavior duration from the user's historical medical interaction records;

[0026] Performing temporal alignment on the operation timestamp and the time attribute of the medical concept in the dynamic knowledge graph, and performing behavior matching on the operation type identifier and the category attribute of the medical concept to obtain a preliminary association set;

[0027] Calculating the stability weight of the medical concept association in the user's historical medical interaction record according to the cumulative distribution ratio of the behavior frequency and the behavior duration;

[0028] The medical concepts in the preliminary association set are hierarchically superimposed with the stability weights, and combined with the hierarchical distribution ratio of medical concepts in the dynamic knowledge graph to generate a user portrait label set based on the user's medical cognitive level as the hierarchical benchmark.

[0029] Optionally, the operation timestamp is time-sequentially aligned with the time attribute of the medical concept in the dynamic knowledge graph, and behavior matching is performed based on the operation type identifier and the category attribute of the medical concept to obtain a preliminary association set, including:

[0030] Perform a time window overlap comparison on the operation timestamp and the time attribute of the medical concept in the dynamic knowledge graph. If the operation timestamp is within the effective time range of the time attribute, it is determined that the user's real-time diagnosis and treatment behavior is aligned with the time sequence of the medical concept.

[0031] Matching the user's real-time diagnosis and treatment behavior with the category association of the medical concept according to a preset mapping relationship between the operation type identifier and the category attribute of the medical concept;

[0032] The user's real-time diagnosis and treatment behaviors and medical concepts that are time-series aligned and have matching category associations are combined into behavior-concept association pairs, and each behavior-concept association pair is summarized to obtain a preliminary association set.

[0033] Optionally, dynamically adjusting the recommendation priority ranking of medical popular science articles based on the joint matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph to generate a personalized recommendation sequence that matches the user's real-time medical cognitive level includes:

[0034] Calculating the matching degree between the dynamic recommendation feature vector and the association relationship between the medical concepts in the dynamic knowledge graph, and extracting the matching degree value directly associated with the medical concepts in the dynamic recommendation feature vector;

[0035] According to the range of the matching degree value, setting priority adjustment rules for different medical concept levels;

[0036] Based on the priority adjustment rule, priority weights are assigned to medical popular science articles in the dynamic knowledge graph that are associated with the user's real-time medical cognition level;

[0037] According to the priority weight allocation result, medical popular science articles that match the user's real-time medical cognitive level are screened from the dynamic knowledge graph, and the recommendation priorities are sorted from high to low according to the priority weight allocation ratio to generate a personalized recommendation sequence.

[0038] In a second aspect, this application provides a medical popular science article recommendation system based on user portraits, including:

[0039] The parsing module parses the logical associations between unstructured text and structured case data in multi-source heterogeneous medical data to construct a dynamic knowledge graph containing dynamic dependencies between medical concepts;

[0040] A fusion module, in the dynamic knowledge graph, multi-dimensionally integrates the acquired user real-time diagnosis and treatment behavior data, the user's historical medical interaction records, and the medical concepts to generate a user portrait tag set based on the user's medical cognition level as a hierarchical benchmark;

[0041] A conversion module converts the user's medical preference features in the user portrait tag set from the non-realistic scenario of the user's historical medical interaction records to the real-world scenario of the user's real-time diagnosis and treatment behavior data, thereby generating a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge;

[0042] The adjustment module dynamically adjusts the recommendation priority ranking of medical popular science articles based on the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, and generates a personalized recommendation sequence that matches the user's real-time medical cognitive level.

[0043] In a third aspect, an embodiment of the present application provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a method for recommending medical popular science articles based on user portraits as described in the first aspect above.

[0044] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a method for recommending medical popular science articles based on user portraits as described in the first aspect.

[0045] In an embodiment of the present application, by parsing the logical association between unstructured text and structured case data in multi-source heterogeneous medical data, a dynamic knowledge graph containing dynamic dependencies between medical concepts is constructed, which can break through the semantic barriers of multi-source medical data, establish a medical concept relationship network that can be updated in real time, and provide an authoritative knowledge framework for accurate recommendation. In the dynamic knowledge graph, the acquired user real-time diagnosis and treatment behavior data, the user's historical medical interaction records and the medical concepts are multi-dimensionally integrated to generate a user portrait tag set based on the user's medical cognitive level as a hierarchical benchmark. This step can integrate the user's real-time behavior and historical interaction data to achieve a multi-dimensional quantitative expression of user needs. The user's medical preference features in the user portrait tag set are converted from the non-realistic scenes of the user's historical medical interaction records to the real scenes of the user's real-time diagnosis and treatment behavior data through cross-scene feature conversion to generate a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge. This step can eliminate the behavioral pattern differences between virtual learning scenes and real diagnosis and treatment scenes, migrate user preference features to real demand scenes, and synchronize the dynamic evolution laws of medical knowledge. Based on the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, the recommendation priority ranking of medical popular science articles is dynamically adjusted to generate a personalized recommendation sequence that matches the user's real-time medical cognitive level. This step is based on the intelligent matching of the user's real-time cognitive state and the knowledge graph, and outputs the recommendation results with dynamically adjusted content priority, achieving triple adaptation of needs, knowledge and scenarios.

[0046] Furthermore, the method extracts the user's behavioral characteristics in non-realistic scenarios, explores their implicit associations with real-world diagnosis and treatment behaviors, and constructs a cross-scenario mapping model. The core cognitive features of the user are retained through noise filtering, and the dynamic offset is calculated in combination with the evolution direction of medical knowledge to generate a recommendation feature vector that integrates the user's cognitive essence and knowledge update trends. This method solves the problem of semantic deviation between virtual scenario behavior data and real-world demand scenarios. Through the dynamic offset mechanism, the recommendation features respond to the evolution of medical knowledge in real time, ensuring the accuracy and professional timeliness of the recommended content under scenario switching.

[0047] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0049] Figure 1A flowchart of a method for recommending medical popular science articles based on user portraits provided by this application is shown;

[0050] Figure 2 The following is a schematic diagram showing the structure of a medical popular science article recommendation system based on user portraits provided by the present application;

[0051] Figure 3 A schematic structural diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0052] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0053] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0054] Researchers have found that existing medical popular science recommendation systems have difficulty in dynamically integrating medical knowledge updates with differences in user multi-scenario behavior, resulting in recommended content deviating from users' real-time cognitive needs and authoritative medical progress. Based on this, a method for recommending medical popular science articles based on user portraits is provided. This method constructs a dynamic knowledge graph containing dynamic dependencies between medical concepts, and combines users' real-time diagnosis and treatment behavior data and historical medical interaction records to generate a set of labels that reflect the user's medical cognitive level and preferences. After cross-scenario feature conversion, a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge is formed. Finally, the recommendation priority is dynamically adjusted based on the matching results between this vector and the medical concepts in the knowledge graph, achieving the goal of providing users with a personalized sequence of popular science article recommendations that is highly matched with their current medical cognitive level. This method not only improves the relevance and practicality of information, but also promotes the effective dissemination of medical knowledge and the improvement of personal health management capabilities.

[0055] The technical solution of this application is applicable to scenarios where medical popular science articles are accurately matched and recommended to meet users' multi-dimensional health knowledge needs (such as disease prevention and medication guides). This method constructs a dynamic knowledge graph, combines users' real-time and historical medical data to generate personalized tags, and converts them into dynamic recommendation feature vectors. It then dynamically adjusts the recommendation priority of popular science articles based on the medical concept matching results, achieving accurate and personalized medical popular science information recommendations, improving the relevance and practicality of information, and promoting the effective dissemination of medical knowledge.

[0056] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0057] Figure 1 A flowchart of a method for recommending medical popular science articles based on user portraits is provided for the embodiment of this application, such as Figure 1 As shown, the method includes:

[0058] 101. By analyzing the logical association between unstructured text and structured case data in multi-source heterogeneous medical data, a dynamic knowledge graph containing dynamic dependencies between medical concepts is constructed;

[0059] In this step, multi-source heterogeneous medical data refers to medical data sources in different formats, such as electronic medical records, imaging reports, and scientific research documents.

[0060] Unstructured text includes free text data such as doctors' handwritten notes and medical forum discussions, while structured case data refers to standardized data such as examination indicators and diagnostic codes with standard codes.

[0061] Dynamic dependencies refer to the strength of associations between medical concepts such as diseases and symptoms, drugs and side effects, which change with new research evidence.

[0062] The medical concept refers to structured case data, which is patient diagnosis and treatment information organized according to certain standards and formats.

[0063] A dynamic knowledge graph refers to a knowledge network built based on the logical associations between medical concepts (such as diseases and drugs), and its dependencies can be automatically adjusted as data is updated.

[0064] In the embodiment of the present application, a bidirectional long short-term memory network in natural language processing technology is first used to perform entity recognition on unstructured text to extract medical concepts such as disease names and drug ingredients. Then, a graph attention network is used to encode the diagnostic path of structured case data to generate weighted medical concepts. The medical concepts extracted from the two types of data sources are then logically associated through a dynamic graph convolutional network. When new case data is input, the node relationship weights in the graph structure are dynamically adjusted according to the time decay function and the strength of evidence. Finally, a dynamic knowledge graph containing time dimension features is formed.

[0065] When a tertiary hospital integrated unstructured reports from its imaging department with structured data from its laboratory department, the system discovered the "coronary artery calcification" entity through text parsing. It also extracted the "elevated low-density lipoprotein" indicator from the structured test results, establishing an initial association based on international cardiovascular guidelines. When newly released clinical studies confirmed a change in the strength of the association, the dynamic knowledge graph automatically updated the edge weights.

[0066] 102. In the dynamic knowledge graph, the acquired real-time diagnosis and treatment behavior data of the user, the user's historical medical interaction records, and the medical concepts are integrated in multiple dimensions to generate a user portrait tag set based on the user's medical cognition level as a hierarchical basis;

[0067] In this step, the user's real-time diagnosis and treatment behavior data refers to the user's operation sequence in the actual diagnosis and treatment scenario (such as prescription writing and test report query).

[0068] User historical medical interaction records refer to the user's behavioral trajectory in non-realistic scenarios such as virtual learning and simulation training (such as disease simulation diagnosis exercises).

[0069] The user portrait tag set refers to a group of tags that are hierarchically labeled based on the level of medical cognition (such as elementary and advanced), reflecting the user's mastery of medical concepts.

[0070] In an embodiment of the present application, first, a medical concept trigger sequence (such as continuous query of "insulin dosage adjustment" and "blood glucose monitoring frequency") is extracted from the user's real-time diagnosis and treatment behavior data through a behavioral pattern extraction algorithm (such as time series pattern mining technology), and a simulated operation trajectory sequence (such as "diabetes complications screening path" in virtual diagnosis) is extracted from the user's historical medical interaction records. Then, the above-mentioned behavioral sequence is mapped to the corresponding node of the dynamic knowledge graph, and a hierarchical clustering algorithm (such as a density-based spatial clustering method) is used to perform multi-dimensional fusion of the user's operation frequency, concept complexity and path depth to generate hierarchical labels of the user's medical cognitive level (such as basic layer, advanced layer, expert layer). Finally, a user portrait label set is constructed based on the clustering results, in which each label is associated with a specific medical concept cluster in the graph and its cognitive depth index.

[0071] Continuing with the above example, each patient can be categorized in detail based on their latest medical history and past medical experience. For example, some patients may be more interested in managing chronic conditions, while others are interested in emergency treatment for acute illnesses. Based on this information, different groups can be labeled accordingly, forming detailed user profiles to better understand and meet patient needs.

[0072] 103. Perform cross-scenario feature conversion on the user medical preference features in the user portrait tag set from the non-realistic scenario of the user's historical medical interaction records to the real-world scenario of the user's real-time diagnosis and treatment behavior data, and generate a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge;

[0073] In this step, non-realistic scenarios refer to non-actual application environments such as virtual learning and simulated operations in the user's historical medical interaction records.

[0074] Real-life scenarios refer to actual demand scenarios such as real diagnosis and treatment, health management, etc. in which users engage in real-time diagnosis and treatment behaviors.

[0075] Cross-scenario feature transfer refers to the process of transferring user preference features from a virtual environment to a real environment.

[0076] The dynamic recommendation feature vector refers to the mathematical representation that integrates user cognitive characteristics and the evolution of medical knowledge.

[0077] In an embodiment of the present application, first, a feature decoupling algorithm (such as an independent component analysis method) is used to separate non-realistic scenarios (such as the "cardiopulmonary resuscitation process" repeatedly performed in simulation training) from the user portrait label set. Then, a cross-scenario nonlinear mapping model is constructed (such as a feature transfer framework based on a generative adversarial network), and cross-scenario feature conversion of non-realistic scenarios to realistic scenarios is performed through adversarial training technology, and an attention mechanism is used to filter the noise generated by free exploration of the simulated environment (such as meaningless random jump operations). Finally, a knowledge evolution perception module (such as a temporal difference learning model) is introduced to calculate the dynamic offset of the core preference component based on the update intensity of medical concepts in the knowledge graph (such as the rate of change of node weights), and generate a dynamic recommendation feature vector that integrates the user's cognitive nature and the knowledge evolution trend.

[0078] Building on the previous example, let's say a patient shows a keen interest in reading an article about heart disease prevention. When this patient is admitted to the hospital with chest pain, the system will take this prior interest into account and combine it with the latest heart disease treatment guidelines to provide personalized medical advice and support resources.

[0079] 104. Based on the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, dynamically adjust the recommendation priority ranking of medical popular science articles to generate a personalized recommendation sequence that matches the user's real-time medical cognitive level.

[0080] In this step, the joint matching result refers to the association score between the dynamic recommendation feature vector and the medical concept in the knowledge graph.

[0081] Recommendation priority sorting refers to the rules for dynamically adjusting the order of article recommendations based on the degree of matching.

[0082] A personalized recommendation sequence is a prioritized list of recommended items based on a user's interests, behavior, and current needs. In the healthcare field, this refers to a list of popular medical articles that are most suitable for the user.

[0083] In the embodiments of the present application, a multimodal similarity calculation algorithm (such as an embedding vector matching method based on a graph neural network) is first used to match the dynamic recommendation feature vector with the embedded representation of the medical concept node in the dynamic knowledge graph to generate an initial relevance score. Then, based on a real-time feedback mechanism (such as an online learning ranking algorithm), the matching weight is dynamically adjusted according to implicit feedback data such as user clicks and dwell time. The recommendation priority ranking of medical popular science articles is recalculated based on the initial relevance score and the incremental update results of the knowledge graph (such as the newly added "gene editing therapy" node). Finally, a personalized recommendation sequence is implemented through a streaming data processing engine to ensure that the recommendation sequence is strictly synchronized with the user's cognitive state and medical progress.

[0084] Combining the examples in the previous steps, for patients who showed interest in heart disease prevention and were later hospitalized due to chest pain, the system not only provided popular science articles related to heart disease, but also adjusted the recommendation order in a timely manner according to the progression of their condition, giving priority to articles that introduce first aid measures for acute myocardial infarction, greatly improving the relevance and practicality of the information.

[0085] In summary, steps 101 through 104 integrate the semantic associations of multi-source medical data through a dynamic knowledge graph, constructing a real-time, authoritative knowledge framework. This integrates user behavior across multiple scenarios to generate a layered cognitive profile, accurately quantifying differences in demand. Cross-scenario feature transfer eliminates behavioral biases between virtual and real environments, and dynamic recommendation features are generated based on knowledge evolution trends. Ultimately, this achieves a three-dimensional match between medical popular science content and user cognitive status, real-time diagnostic and treatment needs, and medical advances, improving recommendation accuracy, timeliness, and scenario adaptability.

[0086] In order to solve the problem of semantic deviation between user preferences in virtual scenes and real medical needs, by analyzing the implicit association pattern of the user's operation trajectory in the non-realistic scene and the real-time medical behavior data in the real scene, a nonlinear projection relationship is constructed to map these preference features to the feature space of the real scene. Subsequently, by tracking the update state changes of medical concepts, the dynamic offset of the core preference component generated by the evolution trend of medical knowledge is calculated, and it is superimposed on the core preference component, and finally a dynamic recommendation feature vector that can reflect the latest evolution of medical knowledge is generated to achieve accurate and personalized medical information recommendation. This process ensures that the user's preference features can be effectively converted between different scenes and kept synchronized with the latest medical knowledge. In some embodiments, the user medical preference features in the user portrait tag set described in step 103 are converted from the non-realistic scene of the user's historical medical interaction record to the real scene of the user's real-time medical behavior data to perform cross-scene feature conversion to generate a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge, including:

[0087] 201. Extracting a preference component of a non-realistic scenario from the user's historical medical interaction record based on the user's medical preference features in the user portrait tag set;

[0088] In step 201, the preference component for non-realistic scenarios refers to the behavioral characteristics of users in non-realistic diagnosis and treatment scenarios such as simulated learning and virtual training. This includes simulated operation path selection patterns (e.g., the order of step jumps in a virtual diagnosis process), the distribution of dwell time on knowledge nodes (e.g., the duration of repeated references to a disease mechanism), and the logical chain of jumps across medical concepts (e.g., a coherent jump path from "symptoms" to "treatment plan"). User medical preference characteristics describe a user's specific interests or tendencies in the medical field, such as a focus on research on a particular disease or a particular treatment. These characteristics are derived from the user's browsing history, search history, and other interactive behaviors.

[0089] In this embodiment, behavioral sequence segmentation technology is first used to extract non-realistic scenario operation fragments (e.g., complete process records of simulated diagnosis and treatment exercises) from the user's historical medical interaction records. Second, a temporal feature extraction model is used to encode the path selection and jump logic in the operation fragments, generating a preference component vector representing the user's medical preference features. Finally, the preference component vector is semantically aligned with the medical concept nodes in the dynamic knowledge graph (e.g., using a graph embedding matching algorithm) to form a semantically labeled non-realistic scenario preference component.

[0090] 202. Constructing a nonlinear projection relationship from the non-realistic scene to the real scene by analyzing the implicit correlation pattern between the user's operation trajectory on the medical concept in the non-realistic scene and the user's real-time diagnosis and treatment behavior data in the real scene;

[0091] In step 202, the nonlinear projection relationship refers to the mapping rules from the feature space of the non-realistic scenario to the feature space of the real scenario. This is established by quantifying the implicit logical associations (e.g., the strength of the association between "antibiotic selection simulation" and "infection treatment prescription") between the user's virtual operation trajectory (e.g., simulated diagnostic path) and actual diagnosis and treatment behavior (e.g., prescription records). Implicit association patterns refer to inherent connections or regularities between data. These connections are often not directly visible but require data analysis techniques (e.g., machine learning algorithms) to discover. In cross-scenario feature conversion, this pattern helps establish the mapping relationship between the non-realistic scenario and the real scenario.

[0092] In an embodiment of the present application, first, a machine learning algorithm (such as a random forest or neural network) is used to analyze the implicit association pattern between the user's operation trajectory of medical concepts in non-realistic scenarios and real-time diagnosis and treatment behavior data. Secondly, a generator and discriminator framework is constructed based on the implicit association pattern. The generator maps the non-realistic scene features to the real scene space, and the discriminator optimizes the mapping rules by comparing the distribution differences between the generated features and the real behavior features. Finally, the nonlinear projection relationship is iteratively adjusted through adversarial training of the generator and discriminator frameworks, so that the mapped features approximate the real scene distribution while retaining the user's cognitive nature.

[0093] 203. Based on the nonlinear projection relationship, the user's medical preference features are mapped to the feature space of the real scene through the preference components of the non-real scene, and the scene noise components unrelated to the user's medical cognitive level are separated in the feature space, while retaining the core preference components that are stably transferred in the cross-scene feature conversion;

[0094] In step 203, the scene noise component refers to the invalid features introduced by the non-realistic scene operation freedom (such as random jumps and aimless browsing). The core preference component refers to the cognitive characteristics that users stably transmit across scenarios (such as sustained attention to specific treatment logic). Feature space refers to a multidimensional abstract space used to represent data characteristics, where each dimension represents a specific attribute or feature. In the medical field, the feature space can contain information such as the patient's age, gender, medical history, and more complex features extracted based on this information, such as disease risk scores. The user's medical cognitive level refers to the user's understanding and mastery of medical knowledge, including but not limited to understanding of diseases, understanding of treatment plans, and basic knowledge of health management. This level can be evaluated through a variety of data such as the user's learning records and interactive behaviors.

[0095] In the embodiments of the present application, first, the mapped non-realistic scene features are input into the attention weight calculation model to analyze the contribution of each feature dimension to the prediction of the feature space behavior of the real scene (such as the prediction weight of the "simulated path selection" feature to the "actual prescription issuance"). Secondly, based on the contribution threshold (such as a preset weight quantile), low-contribution features (such as random jump operations) are filtered, and high-contribution features are retained as core preference components. Finally, the core preference components are embedded in the real scene feature space through feature reconstruction technology to form a denoised core preference component.

[0096] 204. By tracking the update status changes of the medical concepts, the dynamic offset of the core preference component on the evolution trend of medical knowledge is calculated, and the dynamic offset is superimposed on the core preference component to generate a dynamic recommendation feature vector containing the evolution trend of medical knowledge.

[0097] In step 204, the dynamic offset refers to the directional adjustment of the core preference component due to the evolution of medical knowledge (such as the release of new treatment guidelines or updates to drug contraindications). It is used to synchronize user cognition with authoritative knowledge updates. Medical knowledge evolution trends refer to the changes and development direction of knowledge within the medical field over time. This trend can be captured by tracking the update status of medical concepts, reflecting the progress or adjustments in medical theory, practice, and technology. The dynamic recommendation feature vector is a data structure that combines user preferences and the latest developments in medical knowledge. It is used in personalized recommendation systems to improve the relevance and timeliness of recommendations.

[0098] In the embodiment of the present application, first, the update status of medical concepts is monitored and relevant changes are tracked using text analysis technology. Then, the dynamic offset of the core preference component under the new trend is calculated, which usually requires combining time series analysis methods to evaluate the development trend of preferences over time. Finally, this dynamic offset is added to the original core preference component to generate a directional adjustment that includes the latest medical knowledge trends. In this way, it can be ensured that the health information recommended to the user is always the latest and in line with the user's current cognitive level and preferences.

[0099] Here's a specific example:

[0100] To improve the patient experience and the accuracy of personalized medical information recommendations, a tertiary hospital employed the aforementioned optimization method. First, by analyzing patients' browsing history, interaction records, and online course participation on a health education platform, the method identified patients' interests and preferences for specific medical topics, such as heart disease prevention. The strength of preference within each area was calculated to form a set of preference components. Next, machine learning algorithms were used to identify potential connections between these preferences and patients' actual medical behavior data. For example, after reading information on heart disease prevention, patients' actual medical behavior, including selection of physical examination items and medication use, was analyzed. A nonlinear projection model was constructed to accurately map theoretical interests to actual medical needs. Furthermore, signal processing techniques such as principal component analysis were applied to distinguish stable core preference components from contextual noise components. By removing fluctuations caused by specific contexts, the method retained core preference features that truly reflected patients' long-term interests and needs. Finally, incorporating the latest medical research findings, the updated state of heart disease prevention knowledge was monitored, and the trends of core preference components over time were calculated to dynamically adjust patient preference features. In this way, the system can provide patients with personalized medical service recommendations based on the latest medical evidence, such as recommending the latest heart disease prevention guidelines or treatment methods, ensuring that the recommendations are always up to date and in line with the patient's current level of knowledge and preferences.

[0101] In summary, steps 201 to 204 eliminate random noise in the virtual environment through cross-scenario feature conversion, preserving the essential cognitive characteristics of users. Dynamically modifying recommendation directions based on the evolution of medical knowledge ensures precise synchronization of recommended content with users' real-time diagnosis and treatment needs and authoritative knowledge updates. Ultimately, this solves the semantic bias problem caused by scenario fragmentation in traditional recommendation systems, improves the accuracy of cross-scenario recommendations, and ensures that recommendation results are both professional, timely, and personalized.

[0102] In order to further improve the accuracy of dynamic offset calculation of core preference components in the evolution trend of medical knowledge, an update status marking system is constructed by tracking the new and invalid records of concept associations in the knowledge graph. Calculation models for positive and negative offset directions are designed, and the impact of knowledge evolution on user core preferences is quantified in combination with proportional weights. This method innovatively converts the knowledge evolution trend into a computable vector space movement, realizes the dynamic calibration of preference components through superposition, and solves the error accumulation problem caused by the lag in knowledge update in traditional recommendation systems. In some embodiments, the calculation of the dynamic offset of the core preference component in the evolution trend of medical knowledge by tracking the update status changes of the medical concepts in step 303 includes:

[0103] 301. Obtaining the change records of the association relationships of the medical concepts described in the dynamic knowledge graph, extracting the newly added dependencies and invalidated dependencies between the medical concepts, and obtaining an update status mark of the evolution trend of medical knowledge;

[0104] In step 301, the association relationship change record refers to a time-series change dataset of the connection relationship between medical concept nodes in the dynamic knowledge graph, including attributes such as relationship establishment time and expiration time. New dependency refers to the latest semantic association between concepts, and expired dependency refers to an old connection that is no longer valid after authoritative verification. The update status mark is an identifier or label used to reflect the changes in the association relationship between specific medical concepts. This mark is usually generated based on the new dependency and expired dependency between medical concepts recorded in the dynamic knowledge graph.

[0105] In the embodiment of the present application, the version management function of the dynamic knowledge graph is first used to compare the graph structure snapshots in adjacent time periods to identify the change records of the association relationships between nodes. Then, through the graph structure difference analysis algorithm, the subgraph matching technology is used to scan the newly added semantic association edges, and at the same time, the invalid association edges that have exceeded the validity period are located based on the timestamp filtering mechanism. Finally, the detected new dependencies are marked as "positive evolution events", and the invalid dependencies are marked as "negative evolution events", and the time attributes and impact range parameters of each event are recorded to form an update status mark with the evolution trend of medical knowledge.

[0106] 302. Determine, based on the update status flag, a positive offset direction corresponding to the newly added dependency and a negative offset direction corresponding to the invalidated dependency in the core preference component;

[0107] In step 302, the positive offset direction represents the vector of the enhancement of the core preference caused by the semantic expansion of the medical concept due to the newly added associations, while the negative offset direction represents the vector of the weakening of the core preference caused by the shrinkage of the concept dimension due to the invalidation of the old associations.

[0108] In the embodiment of the present application, the state mark is first updated using graph representation learning technology to map medical concepts and their associations into a low-dimensional vector space, so that semantically similar concepts have proximity in the vector space. Then, for the newly added dependency relationship, the vector direction difference between the concept nodes at both ends of the relationship is calculated, and the change in cosine similarity is used as the positive offset direction. Finally, for the failed dependency relationship, the attenuation value of the association strength of the relationship in the historical vector model is extracted as the negative offset direction.

[0109] 303. According to the superposition of the positive offset direction and the negative offset direction, and in combination with the proportional weights of the newly added dependency and the invalid dependency, the dynamic offset of the core preference component in the evolution trend of medical knowledge is obtained.

[0110] In step 303, proportional weights are calculated using a combination of indicators such as the confidence level, academic impact factor, and clinical validation level of the newly added and invalidated dependencies to quantify the contribution of different change events to the offset. Dynamic offsets are adjustments made to user preferences based on new medical concepts, treatments, or health recommendations as medical knowledge evolves and develops.

[0111] In the embodiment of the present application, a multi-dimensional weight evaluation system is first established to collect characteristic data such as the credibility of the academic source, the number of clinical evidence, and the time freshness of the knowledge change event, and the contribution weight coefficient of each feature is calculated by the entropy weight method. Then, the positive and negative offset vectors obtained in step 302 are multiplied by the comprehensive weight of the corresponding event respectively to achieve the proportional weight of the influence of different evolutionary events. Finally, the linear superposition principle of vector space is used to perform a synthetic operation on the weighted positive and negative vector groups, and the dimensional difference is eliminated by normalization processing, and finally the standardized dynamic offset is output.

[0112] Here's a specific example:

[0113] In a knowledge evolution scenario for cardiovascular disease treatment at a tertiary hospital, a new association (weight 0.8) was added to the dynamic knowledge graph for "gene mutation and antiplatelet drug responsiveness," while the old association (weight 0.6) for "β-blockers and asthma contraindications" was invalidated. Step 301 extracts these two change events and marks their status. Step 302 calculates the positive offset vector (0.12, 0.08) for concepts related to gene therapy and the negative offset vector (-0.09, 0.05) for concepts related to the respiratory system. Step 303 adds these values ​​together based on the weights, ultimately yielding a dynamic offset (0.042, 0.094), guiding the recommendation system toward precision medicine.

[0114] In summary, steps 301 to 303 effectively address the recommendation bias problem of traditional recommendation systems in the medical field caused by lagging knowledge updates by establishing a dynamic mapping mechanism between knowledge evolution and user preferences. The system can automatically capture fine-grained changes in medical concept relationships and, combined with a multidimensional weight evaluation system, accurately quantify the direction and intensity of the impact of knowledge evolution on user core preferences. Compared with static models, this method significantly improves the adaptability of recommendation results to cutting-edge medical advances and the efficiency of filtering out outdated knowledge, ensuring that recommended content always conforms to the latest medical consensus while maintaining the stable evolution of personalized features. The graph difference analysis and vector space modeling methods used in the technical implementation process provide an interpretable computational framework for processing complex knowledge evolution.

[0115] In order to accurately quantify the dynamic impact of the evolution of medical knowledge on user preferences, a direction vector is generated by multiplying the association strength and the adjustment parameter, and the vector space superposition technology is used to synthesize the net impact direction. This solution breakthroughly transforms discrete knowledge change events into continuous vector operations, and uses the historical association strength to retain the attenuation effect of the invalid relationship, ensuring that preference adjustments are both responsive to the latest medical advances and compatible with historical cognitive inertia. In some embodiments, the determination of the positive offset direction corresponding to the newly added dependency and the negative offset direction corresponding to the invalid dependency in the core preference component based on the update status mark in step 302 includes:

[0116] 401. Map the newly added dependency to a first adjustment parameter of the core preference component based on the update status flag, and map the invalidated dependency to a second adjustment parameter of the core preference component;

[0117] In step 401, the first adjustment parameter is a quantitative coefficient reflecting the impact of the newly added dependency on the core preference, including dimensions such as knowledge authority and evidence level. The second adjustment parameter is an attenuation coefficient representing the degree to which the expired dependency weakens the core preference, including factors such as historical citation frequency and expiration time.

[0118] In this embodiment, a knowledge evolution impact assessment model is first constructed by updating status markers. For newly added dependencies, features such as the source journal impact factor (e.g., a journal with a coefficient of 5.2) and the number of multicenter clinical trials (e.g., three Phase III trials) are extracted. These features are then normalized to a first adjustment parameter in the 0-1 range using an S-shaped function. For expired dependencies, a time decay curve (e.g., an exponential decay factor of 0.3) is calculated for the number of citations over the past five years. This is combined with the expiration confirmation time (e.g., 18 months after expiration) to generate a second adjustment parameter using a linear regression model.

[0119] 402. Generate a first set of direction vectors based on the product of the association strength between the medical concepts connected by each newly added dependency and the first adjustment parameter, and generate a second set of direction vectors based on the product of the historical association strength between the medical concepts corresponding to each invalidated dependency and the second adjustment parameter;

[0120] In step 402, the association strength refers to the closeness of the semantic connection between medical concepts, which is represented by the relationship weight of the knowledge graph. The historical association strength refers to the average influence of the invalid relationship in the past period. The direction vector is a weighted multidimensional spatial displacement that represents the direction and magnitude of the adjustment of the preference component. The first direction vector set is a set consisting of a series of vectors generated by multiplying the association strength between the medical concepts connected by the newly added dependency in the dynamic knowledge graph by the corresponding adjustment parameter (i.e., the first adjustment parameter). The second direction vector set is a set consisting of a series of vectors generated by multiplying the historical association strength between the medical concepts corresponding to the invalid dependency in the dynamic knowledge graph by the corresponding adjustment parameter (i.e., the second adjustment parameter).

[0121] In this embodiment, the current weight of each newly added relationship (e.g., the correlation strength of "genetic testing and targeted drugs" is 0.78) is first obtained from the dynamic knowledge graph and multiplied by the first adjustment parameter to obtain a first set of direction vectors. For expired dependencies, the historical version library is retrieved to obtain the average weight of each relationship over the three years prior to expiration (e.g., the historical strength of "traditional chemotherapy and immunosuppression" is 0.65). This value is then multiplied by the second adjustment parameter to form a second set of direction vectors. Then, using graph embedding and projection techniques, the scalar product is converted into a displacement vector in the vector space, preserving the topological relationship characteristics of the concept nodes.

[0122] 403. Superimpose the direction vectors in the first direction vector set to obtain a positive offset direction, and superimpose the direction vectors in the second direction vector set to obtain a negative offset direction.

[0123] In step 403, vector superposition is performed by algebraically summing the coordinate components of multiple directional vectors using the spatial vector composition rule. A positive offset indicates the net impact of a newly added relationship group, while a negative offset indicates the net weakening caused by an invalid relationship group. A negative offset indicates a shift in user preference characteristics toward a reduced interest or demand due to the invalidation or weakening of dependencies between certain medical concepts.

[0124] In the embodiment of the present application, all vectors in the first direction vector set are first decomposed into coordinates, and then summed up by dimension to obtain each axial component (e.g., +0.34 on the X axis and -0.12 on the Y axis). The same method is used to process the second direction vector set to obtain each axial component (e.g., -0.25 on the X axis and +0.08 on the Y axis). Vector modulus normalization is then performed to eliminate differences in the magnitude of each axial component, ultimately forming a positive offset direction vector (0.94, -0.33) and a negative offset direction vector (-0.95, 0.32) of unit length.

[0125] Here's a specific example:

[0126] In a diabetes diagnosis and treatment knowledge update scenario at a tertiary hospital, the dynamic knowledge graph added a new association between "gut microbiome testing and insulin resistance" (authority coefficient 0.9, association strength 0.8) and invalidated the old association between "metformin and vitamin B12 deficiency" (historical strength 0.7, decay coefficient 0.6). Step 401 calculated the first adjustment parameter 0.72 (0.9 × 0.8) and the second adjustment parameter 0.42 (0.7 × 0.6). Step 402 mapped the new relationship into a direction vector (0.58, 0.15) and the invalidated relationship into (-0.36, 0.22). Step 403, after superimposing other correlation vectors, ultimately generated a positive offset direction (0.62, 0.31) and a negative offset direction (-0.55, 0.28). This guides the recommendation system to enhance the push of content related to microbiome treatment while de-emphasizing outdated drug side effect warnings.

[0127] In summary, steps 401 to 403 effectively address the key challenge of traditional recommendation systems in adapting to the dynamic evolution of medical knowledge by establishing a quantitative mapping mechanism between knowledge evolution events and preference space. The system accurately analyzes the differential impact of new and obsolete knowledge on user preferences, using vector space modeling to transform discrete knowledge change events into continuous directional adjustments. Compared to rule-based empirical adjustments, this method achieves computability and interpretability of preference component evolution, ensuring that recommendation strategies keep pace with cutting-edge medical advances while avoiding preference drift and instability caused by knowledge updates. The technical implementation process integrates multidimensional influencing factor evaluation and vector space synthesis algorithms, providing a reliable mathematical modeling framework for handling complex knowledge evolution.

[0128] In order to build a personalized recommendation system that dynamically adapts to the evolution of medical knowledge, cognitive labels are generated by combining the time alignment of real-time behavior and the stability analysis of historical behavior with the knowledge hierarchy. The innovation lies in establishing a dynamic mapping between behavioral data and knowledge versions, using cumulative distribution to quantify long-term cognitive precipitation, and reflecting the professional architecture through hierarchical superposition to achieve the transformation from discrete behavior to structured cognitive level. In some embodiments, in the dynamic knowledge graph described in step 102, the acquired user real-time diagnosis and treatment behavior data, the user's historical medical interaction records and the medical concepts are multi-dimensionally integrated to generate a user portrait label set based on the user's medical cognitive level as the hierarchical benchmark, including:

[0129] 501. Extract the operation timestamp and operation type identifier from the user's real-time diagnosis and treatment behavior data, as well as the behavior frequency and behavior duration from the user's historical medical interaction records;

[0130] In step 501, the operation timestamp refers to the specific time when the user performed a medical action (e.g., consulting literature, prescribing). The operation type identifier is a classification code that distinguishes the type of medical action (e.g., A01 for diagnosis, B02 for medication). The action frequency represents the number of times a user accesses a specific medical concept. The action duration refers to the difference between the start and end times of a single interaction.

[0131] In this embodiment, real-time user medical records are first collected as logs through a medical information system interface. Regular expressions are used to extract the operation timestamp (e.g., "2023-08-20 14:30:00") and operation type identifier. Three months of historical user medical interaction records are then retrieved from a historical database. A sliding window statistical method is used to calculate the frequency of each medical concept (e.g., "hypertension" has an average of 8 visits per month). The duration of each interaction is then calculated using timestamp differences (e.g., a single document reading session lasts 25 minutes).

[0132] 502. Align the operation timestamp with the time attribute of the medical concept in the dynamic knowledge graph, and perform behavior matching based on the operation type identifier and the category attribute of the medical concept to obtain a preliminary association set;

[0133] In step 502, temporal alignment involves matching the time of user behavior with the effective time period of medical concepts in the knowledge graph. Behavior matching involves associating the operation type with the domain to which the medical concept belongs through semantic similarity calculation. Category attributes refer to the features or labels used to describe and distinguish different medical concepts in the dynamic knowledge graph. These attributes help clarify the specific classification to which each medical concept belongs. The preliminary association set is an intermediate dataset containing the mapping relationship between user behavior and medical concepts.

[0134] In this embodiment, a dynamic time warping algorithm is first used to align user operation timestamps with the temporal attributes of medical concepts in the knowledge graph. A medical behavior ontology is then constructed, and cosine similarity is used to calculate the matching degree between operation type identifiers and medical concept category attributes (e.g., "drug treatment plan"). Behavior and concept pairs with a similarity threshold exceeding 0.7 are selected. Finally, the temporal alignment and behavior matching results are integrated to form a preliminary association set containing a time validity marker.

[0135] 503. Calculate the stability weight of the medical concept association in the user's historical medical interaction record based on the cumulative distribution ratio of the behavior frequency and the behavior duration;

[0136] In step 503, behavior duration refers to the length of time a user spends performing a specific activity or operation. In the healthcare field, this can refer to the specific duration of a user's participation in a health interaction. The cumulative distribution ratio is the distribution concentration of a user's historical behavior over time. The stability weight reflects the persistence and regularity of a user's attention to a specific medical concept.

[0137] In this embodiment, the user's historical medical interaction records over a three-year period are first time-sliced, and the coefficient of variation of the behavior frequency and behavior duration for each medical concept within each time slice (e.g., quarter) is calculated. Exponential smoothing is then used to assign higher weights to recent behavior data, and the cumulative distribution ratio of the behavior distribution is calculated using the Gini coefficient. Finally, a linear combination of frequency stability (60% weight) and duration stability (40% weight) is performed to generate a stability weight ranging from 0 to 1.

[0138] 504. The medical concepts in the preliminary association set are hierarchically superimposed with the stability weights, and combined with the hierarchical distribution ratio of the medical concepts in the dynamic knowledge graph to generate a user portrait label set based on the user's medical cognitive level as the hierarchical basis.

[0139] In step 504, hierarchical superposition involves weighted fusion of the preliminary associations and stability weights according to the medical concept hierarchy. The hierarchical distribution ratio refers to the hierarchical weight of the medical concepts in the knowledge graph within the discipline system (e.g., 30% for basic medicine and 70% for clinical medicine).

[0140] In this embodiment, a medical concept hierarchy tree is first constructed. The hierarchy of each node is determined based on the hierarchical relationships of medical concepts in the dynamic knowledge graph (e.g., "coronary heart disease" belongs to a third-level clinical concept). Each medical concept in the preliminary association set is then multiplied by its stability weight, and the weighted distribution of these concepts in the hierarchy of the dynamic knowledge graph is then combined. Finally, a hierarchical clustering algorithm is used to map the weighted concept associations to labels at the elementary, intermediate, and advanced levels of medical cognition, forming a user profile label set with a hierarchical basis that includes weighted values.

[0141] Here's a specific example:

[0142] In a cardiology application scenario at a tertiary hospital, a user is viewing a certain version of the heart failure guidelines in real time (operation type: guideline review). Historical records show that over the past two years, they have consistently focused on "natriuretic peptide testing" (average 12 visits per month), spending an average of 38 minutes per session. Step 501 extracts the user's operation timestamp and guideline review identifier, and calculates a stability weight of 0.85 for the behavior frequency of "natriuretic peptide testing." Step 502 aligns the guideline review time with the effective period of the heart failure treatment guidelines in the knowledge graph, matching the concept of "natriuretic peptide monitoring." Step 503 calculates a stability weight of 0.76 for the concept of "natriuretic peptide monitoring." Step 504 superimposes the stability weight of "natriuretic peptide monitoring" and combines it with its distribution ratio of 0.6 in the diagnostic criteria layer to generate a stratified label set containing "high-level cardiac marker awareness" (weight 0.68).

[0143] In summary, steps 501 to 504 innovatively solve the problem of static and single-dimensional user portraits in the medical field by deeply integrating real-time behavioral data with the spatiotemporal characteristics of the knowledge graph. The system can dynamically capture the spatiotemporal correlation between user diagnosis and treatment behavior and the evolution of medical knowledge, and combine historical behavioral stability analysis and concept hierarchy weight distribution to construct a multi-layered cognitive portrait with time sensitivity and disciplinary structure characteristics. Compared with traditional methods, this solution significantly improves the professionalism and timeliness of the user labeling system, accurately reflecting the user's immediate knowledge needs, and deeply portraying their long-term professional cognitive structure, providing a precise cognitive level reference benchmark for personalized medical services. The spatiotemporal alignment algorithm and hierarchical fusion mechanism used in the technical implementation provide a reliable computational framework for processing complex medical behavior data.

[0144] In order to accurately associate user behavior with the dynamically evolving medical knowledge system, the validity of concepts is ensured by time window comparison, and category relevance is guaranteed by semantic mapping. This method innovatively integrates dynamic programming algorithms with deep semantic matching to solve the problems of time misalignment and semantic gap in the association between behavior and concepts. By generating association pairs with intensity values, a reliable data foundation for spatiotemporal dual-dimensional verification is provided for subsequent analysis. In some embodiments, in step 502, the operation timestamp is aligned with the time attribute of the medical concept in the dynamic knowledge graph, and behavior matching is performed based on the operation type identifier and the category attribute of the medical concept to obtain a preliminary association set, including:

[0145] 601. Perform a time window overlap comparison on the operation timestamp and the time attribute of the medical concept in the dynamic knowledge graph. If the operation timestamp is within the effective time range of the time attribute, it is determined that the user's real-time diagnosis and treatment behavior is aligned with the time sequence of the medical concept.

[0146] In step 601, time window overlap comparison verifies the intersection of the time period during which the user's behavior occurred and the validity period of the knowledge graph concept. The validity period is the period from the beginning to the end of the period when the medical concept in the knowledge graph is recognized by the academic community. Operation timestamps are used to mark the exact time when the user performs certain medical-related activities (such as viewing a specific medical popular science article or receiving a certain treatment). Real-time user medical behavior describes the user's current healthcare-related activities or behaviors, including visiting a doctor, receiving a diagnosis or treatment, and accessing health information.

[0147] In this embodiment, we first extract the time attribute metadata of the target medical concept from the dynamic knowledge graph version library, including the concept's effective start time (e.g., January 1st of a certain year) and end time (marked as permanently valid if not expired). Then, using a dynamic programming algorithm, we perform a window overlap comparison between the operation timestamp (e.g., 09:30 on March 15th of a certain year) and the concept's effective start time. When the conditions "operation time ≥ effective start time" and "operation time ≤ effective end time" are met, a time alignment is generated.

[0148] 602. Match the user's real-time diagnosis and treatment behavior with the category association of the medical concept according to the preset mapping relationship between the operation type identifier and the category attribute of the medical concept;

[0149] In step 602, the pre-set mapping relationship is a classification rule library established based on the medical behavior ontology, which contains a semantic association matrix between operation types (such as laboratory report interpretation) and concept categories (such as test indicator analysis). Category association refers to the matching relationship established between specific user behavior types (such as clicks, reading, sharing, etc.) and the category attributes of specific medical concepts in the dynamic knowledge graph.

[0150] In this embodiment, a knowledge base mapping medical actions to concept categories is first constructed. Expert annotation is then used to identify pre-defined mapping relationships between operation type identifiers and medical concept category attributes within the mapping knowledge base (e.g., "imaging examination appointment" corresponds to "imaging diagnostic criteria"). A semantic similarity calculation model is then introduced, using bidirectional encoder representation technology to analyze the textual semantic matching between operation type identifiers, concept categories, and pre-defined mapping relationships (e.g., "laboratory test indicators"). Matches with a similarity threshold exceeding 0.75 are selected. Finally, category associations with confidence values ​​are generated.

[0151] 603. Combining the user's real-time diagnosis and treatment behaviors and medical concepts that are time-series aligned and have matching category associations into behavior-concept association pairs, and summarizing each behavior-concept association pair to obtain a preliminary association set.

[0152] In step 603, the behavior-concept association pair is a valid matching unit that has undergone both temporal and spatial verification, including timestamp alignment proof, category matching evidence, and association strength value. The preliminary association set is a dataset composed of the user's real-time diagnosis and treatment behaviors and the corresponding medical concepts after temporal alignment and category association matching.

[0153] In this embodiment, a Cartesian product operation is first performed on the temporal alignment results of step 601 and the category matching results of step 602, retaining combinations that meet both temporal validity and category matching requirements. Then, behavior-concept association pairs are calculated based on the confidence level of the mapping relationship (e.g., 0.85) and the degree of time window overlap (e.g., 1.0 for complete inclusion and 0.6 for partial overlap). Finally, all eligible behavior-concept association pairs are summarized and sorted in descending order of association strength to form a preliminary set of weighted associations.

[0154] Here's a specific example:

[0155] In a hypertension management scenario at a tertiary hospital, a user performed the "Ambulatory Blood Pressure Monitoring Report Interpretation" operation on May 20th of a certain year (timestamp 05:20 14:00). The "24-Hour Ambulatory Blood Pressure Diagnostic Criteria" concept in the knowledge graph is valid from November to December of that year. Step 601 confirms that the operation time is within the valid range through time window comparison. Step 602 matches the "Report Interpretation" operation type with the "Diagnostic Criteria" concept category based on the preset mapping (similarity 0.89). Step 603 generates an association pair (Ambulatory Blood Pressure Monitoring Report Interpretation, 24-Hour Ambulatory Blood Pressure Diagnostic Criteria, association strength 0.93) and combines this with other valid associations to form a preliminary association set.

[0156] In summary, steps 601 to 603 effectively solve the problem of spatiotemporal misalignment between medical behavior and knowledge concepts through a dual verification mechanism. The system innovatively combines time validity verification with semantic association matching, ensuring that the knowledge concepts associated with user behavior are in the current valid state and that the semantic consistency between the operation intention and the concept category is guaranteed. Compared with the single-dimensional matching method, this solution significantly improves the timeliness accuracy and professional relevance of the association results, avoiding the recommendation of outdated knowledge or semantically biased content. The dynamic planning time comparison and semantic deep matching algorithms used in the technical implementation provide a reliable computing framework for processing complex medical spatiotemporal data, supporting the construction of a knowledge service system that accurately reflects the real-time needs of users.

[0157] In order to dynamically adapt to the evolution of medical knowledge and changes in user cognition, the solution matches core needs through vector similarity calculation, sets hierarchical adjustment rules to balance key coverage and knowledge expansion, and integrates the dual factors of time decay and hot spot reinforcement, adopts a diversity control strategy, and breaks through the homogeneity limitations of traditional recommendations. This method achieves the coordinated optimization of personalized demand matching, knowledge timeliness maintenance, and content ecological balance. In some embodiments, in step 104, according to the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, the recommendation priority ranking of medical popular science articles is dynamically adjusted to generate a personalized recommendation sequence that matches the user's real-time medical cognitive level, including:

[0158] 701. Calculate the matching degree between the dynamic recommendation feature vector and the association relationship between the medical concepts in the dynamic knowledge graph, and extract the matching degree value directly associated with the medical concept in the dynamic recommendation feature vector;

[0159] In step 701, the matching degree is calculated by using a vector space similarity metric to evaluate the closeness of the association between the dynamic recommendation feature vector and the medical concept in the knowledge graph. The matching degree value is a normalized similarity score ranging from 0 to 1.

[0160] In this embodiment, graph embedding technology is first used to convert medical concepts in a dynamic knowledge graph into vector representations, forming a concept vector library. A cosine similarity algorithm is then used to calculate the similarity score between the user's dynamic recommendation feature vector (e.g., [0.34, -0.12, 0.78]) and the association between the medical concepts in the dynamic knowledge graph. Finally, the similarity score is mapped to the 0-1 range through maximum-minimum normalization, and the matching degree values ​​for directly associated relationships exceeding a threshold of 0.6 are extracted.

[0161] 702. Setting priority adjustment rules for different medical concept levels according to the range of the matching degree values;

[0162] In step 702, the priority adjustment rule is a weighted allocation strategy based on the matching degree intervals. It includes a three-tier mechanism: core concept reinforcement rules, associated concept expansion rules, and marginal concept suppression rules. The medical concept hierarchy is a way to organize abundant medical information (such as diseases, symptoms, treatments, and medications) by dividing these concepts into different levels based on factors such as their nature, relevance, or importance.

[0163] In this embodiment, a three-level range of matching values ​​is first set: a high matching value (0.8-1.0) corresponds to the core concept layer, a medium matching value (0.6-0.8) corresponds to the associated concept layer, and a low matching value (<0.6) corresponds to the edge concept layer. A hierarchical weight coefficient matrix is ​​then constructed using the medical concept hierarchy, with the core concept layer assigned 3 times the basic weight, the associated concept layer assigned 1.5 times the basic weight, and the edge concept layer assigned 0.3 times the attenuation weight. Finally, a priority adjustment rule is generated based on the three-level range of matching values, including parameters such as the weight coefficient and the upper limit of the recommendation frequency.

[0164] 703. Based on the priority adjustment rule, assign priority weights to medical popular science articles in the dynamic knowledge graph that are associated with the user's real-time medical knowledge level;

[0165] In step 703, priority weight assignment is a comprehensive calculation process that combines concept hierarchy weights with article relevance, including two dynamic parameters: a time decay factor and a hotspot enhancement factor. Medical popular science articles are articles that popularize medical knowledge, typically covering topics such as disease prevention, diagnosis, and treatment methods, with the goal of raising public health awareness.

[0166] In this embodiment of the present application, the associated concepts and matching values ​​of each medical popular science article in the dynamic knowledge graph are first extracted and converted into hierarchical weights according to the rules of step 702. Then, a time decay function is introduced to reduce the weight of articles that have exceeded their validity period (e.g., the weight of articles published more than 2 years ago is multiplied by 0.7), and the real-time popularity coefficient provided by the hotspot monitoring module is superimposed (e.g., the weight of new articles published 3 months ago is multiplied by 1.2). Finally, the comprehensive priority weight distribution of the articles is calculated using a linear weighting formula, retaining two decimal places.

[0167] 704. Based on the priority weight allocation result, medical popular science articles that match the user's real-time medical cognitive level are screened from the dynamic knowledge graph, and the recommendation priorities are sorted from high to low according to the priority weight allocation ratio to generate a personalized recommendation sequence.

[0168] In step 704, recommendation priority sorting is performed based on the results of descending order of weight values, and an adaptive queue management mechanism is used to dynamically maintain the recommendation sequence. The user's real-time medical cognition level refers to the user's current understanding of medical knowledge, points of interest, and mastery of specific medical concepts. This cognition level is derived from a dynamic assessment of the user's real-time diagnosis and treatment behavior data (such as recent treatments, health information consulted, etc.) and historical medical interaction records. A personalized recommendation sequence refers to an ordered list of recommended content customized according to the user's specific needs and preferences, specifically medical popular science articles that match the user's real-time medical cognition level and are screened from the dynamic knowledge graph based on the priority weight allocation results.

[0169] In this embodiment, a pool of candidate articles is first established based on the priority weighting results, filtering out medical popular science articles with weights below 0.5. A heap sort algorithm is then used to prioritize the selected articles by weight, generating an initial recommendation sequence. Finally, a diversity control module is introduced to implement a staggered arrangement strategy for articles of the same category in the initial recommendation sequence (e.g., articles of the same disease type are separated by at least three positions), forming the final personalized recommendation sequence.

[0170] Here's a specific example:

[0171] In a heart failure patient education scenario at a tertiary hospital, the user dynamic recommendation feature vector matched the concept of "natriuretic peptide monitoring" at a 0.92 degree (core layer) and the concept of "diuretic use" at a 0.75 degree (association layer). Step 701 extracts the matching values ​​for these two concepts; Step 702 assigns a weight of 3 times to the core layer and 1.5 times to the association layer; Step 703 calculates the weight of a journal (association layer, published 3 months ago) to 2.76, while another journal (association layer, published 18 months ago) receives a weight of 0.9; Step 704 prioritizes natriuretic peptide-related articles after sorting, and inserts "Dietary Management of Heart Failure" in the third position to ensure diversity.

[0172] In summary, steps 701 to 704 effectively solve the dual problems of insufficient personalization and lack of knowledge timeliness in the medical recommendation system through a multi-level dynamic weight allocation mechanism. The system innovatively integrates the user's real-time cognitive characteristics, knowledge evolution status and content quality factors for calculation, achieving three key breakthroughs. The first is to establish a quantitative mapping model of concept matching and content priority to accurately reflect user cognitive needs. The second is to introduce dynamic adjustment of the two factors of time decay and hot spot reinforcement to balance the accuracy and cutting-edge nature of knowledge. The third is to adopt a diversity control strategy to avoid the homogenization of recommendation results. Compared with traditional static recommendation methods, this solution significantly improves the dynamic fit between recommended content and user cognitive status, providing intelligent knowledge service adaptation capabilities for scenarios such as patient education and clinical decision support.

[0173] Figure 2 The present application provides a schematic diagram of a user-portrait-based medical science article recommendation system. Figure 2 As shown, the system includes:

[0174] Parsing module 21, which builds a dynamic knowledge graph containing dynamic dependencies between medical concepts by parsing the logical association between unstructured text and structured case data in multi-source heterogeneous medical data;

[0175] A fusion module 22 performs multi-dimensional fusion of the acquired user real-time diagnosis and treatment behavior data, the user's historical medical interaction records, and the medical concepts in the dynamic knowledge graph to generate a user portrait tag set based on the user's medical cognition level as a hierarchical basis;

[0176] A conversion module 23 converts the user's medical preference features in the user portrait tag set from the non-realistic scenario of the user's historical medical interaction records to the real-world scenario of the user's real-time diagnosis and treatment behavior data, thereby generating a dynamic recommendation feature vector that is updated synchronously with the evolution trend of medical knowledge;

[0177] The adjustment module 24 dynamically adjusts the recommendation priority ranking of medical popular science articles based on the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, and generates a personalized recommendation sequence that matches the user's real-time medical cognitive level.

[0178] Figure 2 The medical popular science article recommendation system based on user portrait can be executed Figure 1 The implementation principle and technical effects of the user-profile-based medical popular science article recommendation method described in the illustrated embodiment will not be elaborated on here. The specific manner in which each module and unit performs operations in the user-profile-based medical popular science article recommendation system in the above embodiment has been described in detail in the embodiments of the method and will not be elaborated on here.

[0179] In one possible design, Figure 2 The medical popular science article recommendation system based on user portraits of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0180] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0181] The processing component 32 is used for the above Figure 1 The embodiment provides a method for recommending medical popular science articles based on user portraits.

[0182] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0183] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0184] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0185] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0186] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0187] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0188] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is a method for recommending medical popular science articles based on user portraits.

[0189] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0190] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0191] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for recommending medical popular science articles based on user portraits, characterized in that: include: By analyzing the logical association between unstructured text and structured case data in multi-source heterogeneous medical data, a dynamic knowledge graph containing dynamic dependencies between medical concepts is constructed; In the dynamic knowledge graph, the acquired real-time diagnosis and treatment behavior data of users, historical medical interaction records of users and medical concepts are integrated in multiple dimensions to generate a user portrait tag set based on the user's medical cognition level as a hierarchical benchmark; Extracting the preference components of the non-realistic scenarios of the user's historical medical interaction records based on the user's medical preference features in the user portrait tag set; By analyzing the implicit correlation pattern between the user's operation trajectory on the medical concept in the non-realistic scene and the user's real-time diagnosis and treatment behavior data in the real scene, a nonlinear projection relationship from the non-realistic scene to the real scene is constructed; based on the nonlinear projection relationship, the user's medical preference characteristics are mapped to the feature space of the real scene through the preference components of the non-realistic scene, and the scene noise components that are irrelevant to the user's medical cognitive level are separated in the feature space, and the core preference components that maintain stable transmission in the cross-scene feature conversion are retained; by tracking the update state changes of the medical concepts, the dynamic offset of the core preference components on the medical knowledge evolution trend is calculated, and the dynamic offset is superimposed on the core preference components to generate a dynamic recommendation feature vector containing the medical knowledge evolution trend; According to the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, the recommendation priority ranking of medical popular science articles is dynamically adjusted to generate a personalized recommendation sequence that matches the user's real-time medical cognitive level.

2. The method according to claim 1, characterized in that The step of calculating the dynamic offset of the core preference component on the evolution trend of medical knowledge by tracking the update status change of the medical concept includes: Obtain the change records of the association relationships of the medical concepts described in the dynamic knowledge graph, extract the newly added dependencies and invalid dependencies between medical concepts, and obtain the updated status mark of the evolution trend of medical knowledge; Based on the update status flag, determining a positive offset direction corresponding to the newly added dependency and a negative offset direction corresponding to the invalidated dependency in the core preference component; According to the superposition of the positive offset direction and the negative offset direction, and combined with the proportional weights of the newly added dependencies and the invalidated dependencies, the dynamic offset of the core preference component in the evolution trend of medical knowledge is obtained.

3. The method according to claim 2, characterized in that The determining, based on the update status flag, a positive offset direction corresponding to the newly added dependency and a negative offset direction corresponding to the invalidated dependency in the core preference component includes: Based on the update status flag, mapping the newly added dependency to a first adjustment parameter of the core preference component, and mapping the invalidated dependency to a second adjustment parameter of the core preference component; Generate a first direction vector set based on the product of the association strength between the medical concepts connected by each newly added dependency and the first adjustment parameter, and generate a second direction vector set based on the product of the historical association strength between the medical concepts corresponding to each invalidated dependency and the second adjustment parameter; The direction vectors in the first direction vector set are superimposed to obtain a positive offset direction, and the direction vectors in the second direction vector set are superimposed to obtain a negative offset direction.

4. The method according to claim 1, wherein In the dynamic knowledge graph, the acquired real-time diagnosis and treatment behavior data of the user, the user's historical medical interaction records and the medical concepts are integrated in multiple dimensions to generate a user portrait tag set based on the user's medical cognition level as a hierarchical benchmark, including: Extract the operation timestamp and operation type identifier from the user's real-time diagnosis and treatment behavior data, as well as the behavior frequency and behavior duration from the user's historical medical interaction records; Performing temporal alignment on the operation timestamp and the time attribute of the medical concept in the dynamic knowledge graph, and performing behavior matching on the operation type identifier and the category attribute of the medical concept to obtain a preliminary association set; Calculating the stability weight of the medical concept association in the user's historical medical interaction record according to the cumulative distribution ratio of the behavior frequency and the behavior duration; The medical concepts in the preliminary association set are hierarchically superimposed with the stability weights, and combined with the hierarchical distribution ratio of medical concepts in the dynamic knowledge graph to generate a user portrait label set based on the user's medical cognitive level as the hierarchical benchmark.

5. The method according to claim 4, characterized in that The operation timestamp is aligned with the time attribute of the medical concept in the dynamic knowledge graph, and behavior matching is performed according to the operation type identifier and the category attribute of the medical concept to obtain a preliminary association set, including: Perform a time window overlap comparison on the operation timestamp and the time attribute of the medical concept in the dynamic knowledge graph. If the operation timestamp is within the effective time range of the time attribute, it is determined that the user's real-time diagnosis and treatment behavior is aligned with the time sequence of the medical concept. Matching the user's real-time diagnosis and treatment behavior with the category association of the medical concept according to a preset mapping relationship between the operation type identifier and the category attribute of the medical concept; The user's real-time diagnosis and treatment behaviors and medical concepts that are time-series aligned and have matching category associations are combined into behavior-concept association pairs, and each behavior-concept association pair is summarized to obtain a preliminary association set.

6. The method according to claim 1, characterized in that The method dynamically adjusts the recommendation priority ranking of medical popular science articles based on the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph to generate a personalized recommendation sequence that matches the user's real-time medical cognition level, including: Calculating the matching degree between the dynamic recommendation feature vector and the association relationship between the medical concepts in the dynamic knowledge graph, and extracting the matching degree value directly associated with the medical concepts in the dynamic recommendation feature vector; According to the range of the matching degree value, setting priority adjustment rules for different medical concept levels; Based on the priority adjustment rule, priority weights are assigned to medical popular science articles in the dynamic knowledge graph that are associated with the user's real-time medical cognition level; According to the priority weight allocation result, medical popular science articles that match the user's real-time medical cognitive level are screened from the dynamic knowledge graph, and the recommendation priorities are sorted from high to low according to the priority weight allocation ratio to generate a personalized recommendation sequence.

7. A medical popular science article recommendation system based on user portraits, characterized in that: include: The parsing module parses the logical associations between unstructured text and structured case data in multi-source heterogeneous medical data to construct a dynamic knowledge graph containing dynamic dependencies between medical concepts; A fusion module, in the dynamic knowledge graph, multi-dimensionally integrates the acquired user real-time diagnosis and treatment behavior data, the user's historical medical interaction records, and the medical concepts to generate a user portrait tag set based on the user's medical cognition level as a hierarchical benchmark; a conversion module, extracting the preference component of the non-realistic scenario of the user's historical medical interaction record based on the user's medical preference features in the user portrait tag set; By analyzing the implicit correlation pattern between the user's operation trajectory on the medical concept in the non-realistic scene and the user's real-time diagnosis and treatment behavior data in the real scene, a nonlinear projection relationship from the non-realistic scene to the real scene is constructed; based on the nonlinear projection relationship, the user's medical preference characteristics are mapped to the feature space of the real scene through the preference components of the non-realistic scene, and the scene noise components that are irrelevant to the user's medical cognitive level are separated in the feature space, and the core preference components that maintain stable transmission in the cross-scene feature conversion are retained; by tracking the update state changes of the medical concepts, the dynamic offset of the core preference components on the medical knowledge evolution trend is calculated, and the dynamic offset is superimposed on the core preference components to generate a dynamic recommendation feature vector containing the medical knowledge evolution trend; The adjustment module dynamically adjusts the recommendation priority ranking of medical popular science articles based on the joint matching results between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, and generates a personalized recommendation sequence that matches the user's real-time medical cognitive level.

8. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a medical popular science article recommendation method based on user portraits as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, it implements the method for recommending medical popular science articles based on user portraits as described in any one of claims 1 to 6.

Images