Disclosure of Invention
The application provides a medical science popularization article recommending method and system based on user portraits, which are used for solving the problems of poor recommending accuracy and poor timeliness of the traditional Chinese medical science popularization article in the prior art.
In a first aspect, the present application provides a medical science popularization article recommendation method based on user portraits, including:
By analyzing the logical association of unstructured text and structured case data in the multi-source heterogeneous medical data, a dynamic knowledge graph containing dynamic dependency relations among medical concepts is constructed;
In the dynamic knowledge graph, carrying out multidimensional fusion on the acquired real-time diagnosis and treatment behavior data of the user and the history medical interaction records of the user and the medical concepts to generate a user portrait tag set taking the medical cognition level of the user as a layering reference;
Performing cross-scene feature conversion on the user medical preference features in the user portrait tag set from the non-real scene of the user history medical interaction records to the real scene of the user real-time diagnosis and treatment behavior data, and generating dynamic recommendation feature vectors which are updated synchronously with the medical knowledge evolution trend;
And dynamically adjusting the recommendation priority ordering of the medical science popularization articles according to the joint matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, and generating a personalized recommendation sequence matched with the real-time medical cognition level of the user.
Optionally, the step of performing cross-scene feature conversion on the user medical preference feature in the user portrait tag set from the non-real scene of the user history medical interaction record to the real scene of the user real-time diagnosis and treatment behavior data to generate a dynamic recommendation feature vector updated synchronously with the evolution trend of medical knowledge includes:
extracting preference components of unreal scenes of the user history medical interaction records according to the user medical preference characteristics in the user portrait tag set;
Constructing a nonlinear projection relation from the non-real scene to the real scene by analyzing an implicit association mode of an operation track of a user on a medical concept in the non-real scene and real-time diagnosis and treatment behavior data of the user in the real scene;
mapping the user medical preference feature to a feature space of the real scene through a preference component of the non-real scene based on the nonlinear projection relation, separating a scene noise component irrelevant to the user medical cognition level from the feature space, and reserving a core preference component which keeps stable transfer in cross-scene feature conversion;
And calculating the dynamic offset of the core preference component on the medical knowledge evolution trend by tracking the updated state change of the medical concept, and generating a dynamic recommendation feature vector containing the medical knowledge evolution trend by superposing the dynamic offset to the core preference component.
Optionally, the calculating the dynamic offset of the core preference component on the medical knowledge evolution trend by tracking the updated state change of the medical concept includes:
acquiring a change record of the association relation of the medical concepts in the dynamic knowledge graph, and extracting newly added dependency relation and failed dependency relation among the medical concepts to obtain an updated state mark of the medical knowledge evolution trend;
Based on the update status flag, determining a positive offset direction corresponding to the newly added dependency relationship and a negative offset direction corresponding to the failed dependency relationship in the core preference component;
And according to the superposition of the positive offset direction and the negative offset direction, and combining the proportional weight of the newly added dependency relationship and the failed dependency relationship, obtaining the dynamic offset of the core preference component on the evolution trend of medical knowledge.
Optionally, 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 failed dependency in the core preference component includes:
Mapping the newly added dependency relationship to a first adjustment parameter of the core preference component and mapping the failed dependency relationship to a second adjustment parameter of the core preference component based on the update status flag;
Generating a first direction vector set according to the product of the correlation strength between the medical concepts connected by each newly added dependency relationship and the first adjustment parameter, and generating a second direction vector set according to the product of the historical correlation strength between the medical concepts corresponding to each failed dependency relationship and the second adjustment parameter;
And superposing the direction vectors in the first direction vector set to obtain a positive offset direction, and superposing the direction vectors in the second direction vector set to obtain a negative offset direction.
Optionally, in the dynamic knowledge graph, the multi-dimensional fusion is performed on the acquired real-time diagnosis and treatment behavior data of the user and the user history medical interaction record with the medical concept, and the generating of the user portrait tag set with the user medical cognition level as a layering reference includes:
extracting an operation time stamp and an operation type identifier in the real-time diagnosis and treatment behavior data of the user, and behavior frequency and behavior duration in the historical medical interaction records of the user;
performing time sequence alignment on the operation time stamp 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 according to the operation type identifier to obtain a preliminary association set;
Calculating stability weights related to the medical concepts in the user history medical interaction records according to the accumulated distribution proportion of the behavior frequency and the behavior duration;
And carrying out hierarchical superposition on the medical concepts in the preliminary association set and the stability weight, and generating a user portrait tag set taking the user medical cognition level as a hierarchical reference by combining the hierarchical distribution proportion of the medical concepts in the dynamic knowledge graph.
Optionally, the aligning the operation timestamp with the time attribute of the medical concept in the dynamic knowledge graph in time sequence, and performing behavior matching with the category attribute of the medical concept according to the operation type identifier to obtain a preliminary association set, which includes:
performing time window overlapping comparison on the operation time stamp and the time attribute of the medical concept in the dynamic knowledge graph, and if the operation time stamp is in the effective time range of the time attribute, judging that the real-time diagnosis and treatment behavior of the user is aligned with the time sequence of the medical concept;
Matching the real-time diagnosis and treatment behavior of the user with the category association of the medical concept according to a preset mapping relation between the operation type identifier and the category attribute of the medical concept;
and combining the user real-time diagnosis and treatment behaviors and medical concepts with aligned time sequences and matched category correlations into behavior and concept correlation pairs, and summarizing the behavior and concept correlation pairs to obtain a preliminary correlation set.
Optionally, the dynamically adjusting the recommendation priority ranking of the medical science popularization articles according to the joint matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, and generating a personalized recommendation sequence matched with the real-time medical cognition level of the user, includes:
carrying out matching degree calculation on the association relation between the dynamic recommendation feature vector and the medical concept in the dynamic knowledge graph, and extracting a matching degree value directly associated with the medical concept in the dynamic recommendation feature vector;
Setting priority adjustment rules of different medical concept levels according to the range of the matching degree values;
based on the priority adjustment rule, performing priority weight distribution on medical science popularization articles associated with the real-time medical cognition level of the user in the dynamic knowledge graph;
And screening medical science popularization articles matched with the real-time medical cognition level of the user from the dynamic knowledge graph according to the priority weight distribution result, and sequencing the recommendation priorities from high to low according to the priority weight distribution proportion to generate a personalized recommendation sequence.
In a second aspect, the present application provides a medical science popularization article recommendation system based on user portraits, including:
the analysis module is used for constructing a dynamic knowledge graph containing dynamic dependency relations among medical concepts by analyzing logical association of unstructured text and structured case data in the multi-source heterogeneous medical data;
the fusion module is used for carrying out multidimensional fusion on the acquired real-time diagnosis and treatment behavior data of the user and the history medical interaction records of the user and the medical concepts in the dynamic knowledge map to generate a user portrait tag set taking the medical cognition level of the user as a layering reference;
The conversion module is used for carrying out cross-scene feature conversion on the user medical preference features in the user portrait tag set from the unreal scenes of the user history medical interaction records to the real scenes of the user real-time diagnosis and treatment behavior data, and generating dynamic recommendation feature vectors which are updated synchronously with the medical knowledge evolution trend;
And the adjustment module dynamically adjusts the recommendation priority ordering of the medical science popularization articles according to the joint matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph so as to generate a personalized recommendation sequence matched with the real-time medical cognition level of the user.
In a third aspect, an embodiment of the present application provides a computing device, including a processing component and a storage component, where the storage component stores one or more computer instructions, and the one or more computer instructions are used to be invoked and executed by the processing component to implement a medical science popularization article recommendation method based on a user portrait according to the first aspect.
In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program, where the computer program when executed by a computer implements a medical science popularization article recommendation method based on a user portrait according to the first aspect.
According to the embodiment of the application, the dynamic knowledge graph containing the dynamic dependency relationship among medical concepts is constructed by analyzing the logical association of the unstructured text and the structured case data in the multi-source heterogeneous medical data, so that the semantic barriers of the multi-source medical data can be opened, a medical concept relationship network capable of being updated in real time is established, and an authoritative knowledge framework is provided for accurate recommendation. In the dynamic knowledge graph, the acquired real-time diagnosis and treatment behavior data of the user and the historical medical interaction records of the user are subjected to multidimensional fusion with the medical concepts to generate a user portrait tag set taking the medical cognition level of the user as a layering reference, and the step can fuse the real-time behavior and the historical interaction data of the user to realize multidimensional quantitative expression of the user demands. And performing cross-scene feature conversion from the unreal scene of the user history medical interaction record to the real scene of the user real-time diagnosis and treatment behavior data to generate a dynamic recommendation feature vector which is updated synchronously with the medical knowledge evolution trend. According to the combined matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, the recommendation priority ordering of the medical science popularization articles is dynamically adjusted, a personalized recommendation sequence matched with the real-time medical cognition level of the user is generated, and the recommendation result of dynamic adjustment of the content priority is output based on intelligent matching of the real-time cognition state of the user and the knowledge graph, so that triple adaptation of requirements, knowledge and scenes is realized.
Further, the method extracts behavior characteristics of the non-real scene of the user, digs implicit association of the behavior characteristics with diagnosis and treatment behaviors of the real scene, and builds a cross-scene mapping model. And (3) reserving the user cognition core characteristics through noise filtering, calculating the dynamic offset by combining with the medical knowledge evolution direction, and generating a recommended characteristic vector fusing the user cognition essence and the knowledge updating trend. The method solves the problem of semantic deviation between the behavior data of the virtual scene and the real demand scene, and enables the recommended features to respond to the evolution of medical knowledge in real time through a dynamic deviation mechanism, thereby guaranteeing the accuracy and professional timeliness of the recommended content under scene switching.
These and other aspects of the application will be more readily apparent from the following description of the embodiments.
Detailed Description
In order to enable those skilled in the art to better understand the present application, the following description will make clear and complete descriptions of the technical solutions according to the embodiments of the present application with reference to the accompanying drawings.
In some of the flows described in the specification and claims of the present application and in the foregoing figures, a plurality of operations occurring in a particular order are included, but it should be understood that the operations may be performed out of order or performed in parallel, with the order of operations such as 101, 102, etc., being merely used to distinguish between the various operations, the order of the operations themselves not representing any order of execution. In addition, the flows may include more or fewer operations, and the operations may be performed sequentially or in parallel. It should be noted that, the descriptions of "first" and "second" herein are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, and are not limited to the "first" and the "second" being different types.
Researchers find that the existing medical science popularization recommendation system is difficult to dynamically integrate medical knowledge updating and user multi-scene behavior difference, so that recommendation content deviates from user real-time cognitive requirements and authoritative medical progress. Based on the method, a dynamic knowledge map containing dynamic dependency relations among medical concepts is constructed, real-time diagnosis and treatment behavior data and historical medical interaction records of a user are combined to generate a label set reflecting medical cognition level and preference of the user, a dynamic recommendation feature vector synchronously updated with medical knowledge evolution trend is formed through cross-scene feature conversion, and finally recommendation priority is dynamically adjusted according to a matching result of the vector and the medical concepts in the knowledge map, so that a goal of providing a highly matched and personalized science popularization article recommendation sequence with the current medical cognition level for the user is achieved. The method not only improves the relevance and practicability of the information, but also promotes the effective propagation of medical knowledge and the improvement of personal health management capability.
The technical scheme of the application can be suitable for accurately matching recommended scenes of medical science popularization articles aiming at multi-dimensional health knowledge demands (such as disease prevention and medication guidelines) of users. According to the method, the dynamic knowledge graph is constructed, the user real-time and historical medical data are combined to generate personalized labels, the personalized labels are converted into dynamic recommendation feature vectors, and then the recommendation priority of science popularization articles is dynamically adjusted according to the medical concept matching result, so that accurate and personalized medical science popularization information recommendation is realized, information correlation and practicality are improved, and effective propagation of medical knowledge is promoted.
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to fall within the scope of the application.
Fig. 1 is a flowchart of a medical science popularization article recommending method based on user portrait according to an embodiment of the present application, and as shown in fig. 1, the method includes:
101. by analyzing the logical association of unstructured text and structured case data in the multi-source heterogeneous medical data, a dynamic knowledge graph containing dynamic dependency relations among medical concepts is constructed;
In this step, the multi-source heterogeneous medical data refers to medical data sources with different formats such as electronic medical records, image reports, scientific research documents and the like.
Unstructured text contains free text data such as handwriting records of doctors, medical forum discussions and the like, and structured case data refers to standardized data such as examination indexes, diagnostic codes and the like containing standard codes.
Dynamic dependency refers to the strength of the association between medical concepts such as disease and symptoms, drugs and side effects, etc. as new research evidence changes.
Medical concepts refer to structured case data that is patient diagnosis and treatment information organized according to certain standards and formats.
Dynamic knowledge graph refers to knowledge network constructed based on logical association between medical concepts (such as diseases and medicines), and its dependency relationship can be automatically adjusted along with data update.
In the embodiment of the application, the two-way long-short-term memory network in the natural language processing technology is adopted to carry out entity recognition on the unstructured text, and medical concepts such as disease names, medicine components and the like are extracted. The diagnostic path of the structured case data is then encoded using the iconic network, generating weighted medical concepts. And then logically associating the medical concepts extracted by the two types of data sources through a dynamic graph convolution network, and adjusting the node relation weight in the graph structure according to the time attenuation function and the evidence intensity when new case data is input. And finally forming a dynamic knowledge graph containing time dimension characteristics.
When a non-structural report of an image department and structural data of a clinical laboratory are integrated in a certain three-dimensional hospital, the system discovers a coronary calcification entity through text analysis, extracts a low-density lipoprotein elevation index from a structural inspection sheet, and establishes initial association according to an international cardiovascular guideline. And after the newly released clinical study proves that the correlation strength of the two is changed, the dynamic knowledge graph automatically updates the edge weight value.
102. In the dynamic knowledge graph, carrying out multidimensional fusion on the acquired real-time diagnosis and treatment behavior data of the user and the history medical interaction records of the user and the medical concepts to generate a user portrait tag set taking the medical cognition level of the user as a layering reference;
In this step, the real-time diagnosis and treatment behavior data of the user refer to an operation sequence (such as prescription making and inspection report inquiring) of the user in a real diagnosis and treatment scene.
The user history medical interaction record refers to a behavior track (such as disease simulation diagnosis exercise) of a user in non-real scenes such as virtual learning, simulation training and the like.
The user portrait tag set refers to tag groups based on medical cognition level (such as primary and advanced) hierarchical tags, and reflects the grasping degree of a user on medical concepts.
In the embodiment of the application, firstly, a medical concept trigger sequence (such as continuously inquiring 'insulin dosage adjustment' and 'blood sugar monitoring frequency') is extracted from real-time diagnosis and treatment behavior data of a user through a behavior pattern extraction algorithm (such as a time sequence pattern mining technology), and meanwhile, a simulated operation track sequence (such as a 'diabetes complication screening path' in virtual diagnosis) is extracted from a historical medical interaction record of the user. And then mapping the behavior sequence to a corresponding node of the dynamic knowledge graph, and carrying out multi-dimensional fusion on the operation frequency, the concept complexity and the path depth of the user by adopting a hierarchical clustering algorithm (such as a density-based spatial clustering method) to generate a medical cognition level hierarchical label (such as a base layer, an advanced layer and an expert layer) of the user. And finally, constructing a user portrait tag set according to the clustering result, wherein each tag is associated with a specific medical concept cluster and a cognitive depth index thereof in the map.
Continuing with the above example, each patient is finely classified based on its latest medical information and its past medical experiences. For example, some patients may be more concerned with the management of chronic diseases, while others are interested in the emergency treatment of acute disorders. Based on the information, corresponding labels are marked on different groups, so that detailed user figures are formed, and better understanding and meeting of the needs of patients are facilitated.
103. Performing cross-scene feature conversion on the user medical preference features in the user portrait tag set from the non-real scene of the user history medical interaction records to the real scene of the user real-time diagnosis and treatment behavior data, and generating dynamic recommendation feature vectors which are updated synchronously with the medical knowledge evolution trend;
in this step, the unrealistic scene refers to an unrealistic application environment such as virtual learning and simulation operation in the user history medical interaction record.
The real scene refers to actual demand scenes such as real diagnosis and treatment, health management and the like of real-time diagnosis and treatment behaviors of users.
Cross-scene feature transformation refers to the process of migrating user preference features from a virtual environment to a real environment.
The dynamic recommendation feature vector refers to a mathematical representation that fuses user cognitive features with the evolution of medical knowledge.
In the embodiment of the application, firstly, a feature decoupling algorithm (such as an independent component analysis method) is used to separate unrealistic scenes (such as a cardio-pulmonary resuscitation process repeatedly operated in simulation training) from a user portrait tag set. Then, a cross-scene nonlinear mapping model (such as a feature migration framework based on an countermeasure generation network) is constructed, the non-real scene is subjected to cross-scene feature conversion to the real scene through a countermeasure training technology, and noise generated by free exploration of a simulation environment (such as nonsensical random jump operation) is filtered by adopting an attention mechanism. And finally, introducing a knowledge evolution perception module (such as a time sequence difference learning model), calculating the dynamic offset of the core preference component according to the update strength (such as the node weight change rate) of the medical concept in the knowledge graph, and generating a dynamic recommendation feature vector fusing the user cognition essence and the knowledge evolution trend.
Based on the previous cases, it was assumed that a patient showed great interest in reading an article about heart disease prevention. When this patient is admitted due to chest pain, the system will take into account their previous points of interest and provide personalized medical advice and support resources in combination with the latest guidelines for heart disease treatment.
104. And dynamically adjusting the recommendation priority ordering of the medical science popularization articles according to the joint matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, and generating a personalized recommendation sequence matched with the real-time medical cognition level of the user.
In this step, the joint matching result refers to the degree of association score of the dynamic recommendation feature vector and the medical concept in the knowledge graph.
The recommendation priority ranking refers to a rule for dynamically adjusting the article recommendation sequence according to the matching degree.
The personalized recommendation sequence refers to a group of recommendation item lists which are sequenced according to the priority after analysis according to the interests, the behavior habits and the current demand state of the user. In the medical field, it refers to a series of medical science popularization articles that are most suitable for users.
In the embodiment of the application, firstly, a multi-mode similarity calculation algorithm (such as an embedded vector matching method based on a graph neural network) is adopted to match a dynamic recommendation feature vector with an embedded representation of a medical concept node in a dynamic knowledge graph so as to generate an initial relevance score. And then, based on a real-time feedback mechanism (such as an online learning ordering algorithm), matching degree weights are dynamically adjusted according to implicit feedback data such as user clicks, stay time and the like, and recommendation priority ordering of medical science popularization articles is recalculated by combining initial association degree scores and incremental updating results (such as newly added 'gene editing treatment' nodes) of a knowledge graph. And finally, realizing personalized recommendation sequences through a streaming data processing engine, and ensuring strict synchronization of the recommendation sequences with the cognitive state and the medical progress of the user.
By combining the previous examples of the steps, the system provides not only heart disease-related science popularization articles, but also timely adjusts the recommendation sequence according to the disease progress of patients who show interest in heart disease prevention and later admission due to chest pain, and preferentially displays articles introducing acute myocardial infarction emergency measures, thereby greatly improving the relevance and practicability of information.
In summary, steps 101 to 104 integrate semantic association of multi-source medical data through dynamic knowledge graphs, build an authoritative knowledge framework updated in real time, and fuse multi-scene behaviors of users to generate cognition layered portraits, thereby accurately quantifying demand differences. And eliminating behavior deviation of virtual and real environments through the migration of scene characteristics, and generating dynamic recommendation characteristics by combining knowledge evolution trend. Finally, three-dimensional matching of medical science popularization content and user cognitive state, real-time diagnosis and treatment requirements and medical progress is realized, and recommendation accuracy, timeliness and scene adaptability are improved.
In order to solve the problem of semantic deviation between the user preference of the virtual scene and the real diagnosis and treatment requirement, a nonlinear projection relation is constructed by analyzing an implicit association mode of an operation track of the user in the non-real scene and real-time diagnosis and treatment behavior data in the real scene, and the preference characteristics are mapped to a characteristic space of the real scene. Then, by tracking the update state change of the medical concept, calculating the dynamic offset of the core preference component generated along with the evolution trend of the medical knowledge, and superposing the dynamic offset on the core preference component, and finally generating a dynamic recommendation feature vector capable of reflecting the latest medical knowledge evolution so as to realize accurate and personalized medical information recommendation. This procedure ensures that the user's preference profile can be efficiently transformed between different scenarios and updated in synchronization with the latest medical knowledge. In some embodiments, in step 103, performing cross-scene feature transformation from the non-real scene of the user history medical interaction record to the real scene of the user real-time diagnosis and treat behavior data to generate a dynamic recommendation feature vector updated synchronously with the medical knowledge evolution trend, where the step includes:
201. Extracting preference components of unreal scenes of the user history medical interaction records according to the user medical preference characteristics in the user portrait tag set;
In step 201, preference components of the unrealistic scene refer to behavior characteristics generated by the user in the unrealistic diagnosis and treatment scene such as simulated learning, virtual training and the like, including simulated operation path selection modes (such as step jump sequence in virtual diagnosis flow), knowledge node stay time length distribution (such as repeated consulting time length statistics of a certain disease mechanism), and jump logic chains (such as a continuous jump path from 'symptoms' to 'treatment scheme') of the cross medical concept. The user medical preference profile describes a particular interest or tendency of the user in the medical field, such as research into a certain disease, attention to a specific therapy, etc. These features originate from the user's history browsing records, search history, and other interactive behaviors.
In the embodiment of the application, firstly, the non-real scene operation fragment (such as a complete flow record in simulated diagnosis and treatment practice) is extracted from the historical medical interaction record of the user through a behavior sequence segmentation technology. Secondly, adopting a time sequence feature extraction model to encode path selection and jump logic in the operation fragment, and generating a preference component vector of the medical preference feature of the user. And finally, carrying out semantic alignment on the preference component vector and medical concept nodes in the dynamic knowledge graph (such as by a graph embedding matching algorithm) to form preference components of the non-real scene with semantic tags.
202. Constructing a nonlinear projection relation from the non-real scene to the real scene by analyzing an implicit association mode of an operation track of a user on a medical concept in the non-real scene and real-time diagnosis and treatment behavior data of the user in the real scene;
In step 202, a nonlinear projection relationship refers to a mapping rule from a non-real scene feature space to a real scene feature space, and is established by quantifying an implicit logical association (such as an association strength of "antibiotic selection simulation" and "infection therapy prescription") between a virtual operation track (such as a simulated diagnosis path) of a user and an actual diagnosis and treatment behavior (such as a prescription setting record). Implicit association patterns refer to the inherent relationships or laws that exist between data that are often not directly visible, but rather need to be mined through data analysis techniques (e.g., machine learning algorithms). In cross-scene feature conversion, this mode helps establish a mapping relationship between non-real scenes and real scenes.
In the embodiment of the application, firstly, an implicit association mode between an operation track of a user on a medical concept in a non-real scene and real-time diagnosis and treatment behavior data is analyzed through a machine learning algorithm (such as a random forest or a neural network). And secondly, constructing a framework of a generator and a discriminator according to the implicit association mode, mapping the non-real scene features to the real scene space by the generator, and optimizing a mapping rule by comparing the distribution difference of the generated features and the real behavior features by the discriminator. Finally, the nonlinear projection relation is adjusted through the countermeasure training iteration of the generator and the discriminator framework, so that the mapped features approach to the real scene distribution while keeping the cognitive essence of the user.
203. Mapping the user medical preference feature to a feature space of the real scene through a preference component of the non-real scene based on the nonlinear projection relation, separating a scene noise component irrelevant to the user medical cognition level from the feature space, and reserving a core preference component which keeps stable transfer in cross-scene feature conversion;
In step 203, the scene noise component refers to invalid features introduced by the non-realistic scene operational degrees of freedom (e.g., random jump, purposeless browsing). The core preference component refers to the cognitive characteristics (e.g., persistent focus on specific treatment logic) that the user stably delivers across the scene. Feature space refers to a multi-dimensional abstract space for representing characteristics of data, where each dimension represents a particular attribute or feature. In the medical field, feature space may contain information about the age, sex, medical history, etc. of the patient, as well as more complex features extracted based on such information, such as disease risk scores, etc. The user medical cognitive level refers to the user's understanding and mastering of medical knowledge, including but not limited to, knowledge of the disease, understanding of the treatment regimen, and general knowledge of health management, etc. The level may be assessed by a variety of data such as learning records, interactive behavior, etc. of the user.
In the embodiment of the application, firstly, the mapped non-real scene features are input into an attention weight calculation model, and the contribution degree of each feature dimension to the feature space behavior prediction of the real scene (such as the prediction weight of the feature of 'simulation path selection' to 'actual prescription') is analyzed. Second, low contribution features (e.g., random jump operations) are filtered based on a contribution threshold (e.g., a preset weight score), leaving high contribution features as core preference components. And finally, embedding the core preference component into a real scene feature space through a feature reconstruction technology to form a denoised core preference component.
204. And calculating the dynamic offset of the core preference component on the medical knowledge evolution trend by tracking the updated state change of the medical concept, and generating a dynamic recommendation feature vector containing the medical knowledge evolution trend by superposing the dynamic offset to the core preference component.
In step 204, the dynamic offset refers to the directionality adjustment amount of the core preference component generated by the evolution of medical knowledge (such as the release of new diagnosis and treatment guidelines and the update of medication contraindications), and is used for synchronizing the cognition of the user and the update of authoritative knowledge. The evolution trend of medical knowledge refers to the change and development direction of knowledge over time in the medical field. Such trends may be captured by tracking updated state changes of medical concepts reflecting advances or adjustments in medical theory, practice, and technology. The dynamic recommendation feature vector is a data structure combining the user preference and the latest development condition of medical knowledge and is used in a personalized recommendation system to improve the relevance and timeliness of the recommended content.
In the embodiment of the application, firstly, the update state of the medical concept is monitored, and the text analysis technology is used for tracking the related change. Then, the dynamic offset of the core preference component under the new trend is calculated, which generally requires a method of combining time series analysis to evaluate the trend of the preference over time. Finally, this dynamic offset is added to the original core preference component, resulting in a directional adjustment that contains the latest medical knowledge trend. In this way, it is ensured that the health information recommended to the user is always up-to-date and meets the current cognitive level and preferences of the user.
The following is a specific example:
In a trimethyl hospital, the method is adopted for optimization in order to improve the service experience of patients and the accuracy of personalized medical information recommendation. First, by analyzing browsing history, interactive records and online course participation on a patient's health education platform, interest preferences of the patient for specific medical topics such as heart disease prevention are identified, and the preference intensity in each field is calculated to form a preference component set. Next, machine learning algorithms are utilized to mine potential links between these preferences and behavioral data at the time of actual patient visits. For example, the actual diagnosis and treatment behaviors of the patient after reading the heart disease prevention data are analyzed, including selection of physical examination items, drug use and the like, a nonlinear projection model is constructed, and theoretical interests are accurately mapped to the actual diagnosis and treatment requirements. On this basis, signal processing techniques such as principal component analysis are applied to distinguish between the stable core preference component and the scene noise component. By removing those fluctuating factors caused by specific circumstances, the core preference features that truly reflect the long-term interests and needs of the patient are preserved. And finally, combining the latest medical research results, monitoring the updated state of the heart disease prevention related knowledge, calculating the development trend of the core preference component along with time, and dynamically adjusting the preference characteristics of the patient. In this way, the system is able to provide personalized healthcare advice to the patient based on the latest medical evidence, such as recommending the latest guidelines for heart disease prevention or treatment methods, ensuring that the recommended content is always up-to-date and meets the current cognitive level and preferences of the patient.
In summary, steps 201 to 204 eliminate random noise in the virtual environment through cross-scene feature conversion, retain user cognitive essential features, and dynamically correct the recommendation direction by combining with the medical knowledge evolution trend, so as to achieve accurate synchronization of recommendation content, real-time diagnosis and treatment requirements of users and update of authoritative knowledge. Finally, the problem of semantic deviation caused by scene splitting of the traditional recommendation system is solved, the accuracy of cross-scene recommendation is improved, and the recommendation result is ensured to have professional timeliness and personalized adaptation capability.
In order to further improve the accuracy of dynamic offset calculation of the core preference component on the evolution trend of medical knowledge, an updated state marking system is constructed by tracking new and invalid records of concept association relations in the knowledge graph. And designing a calculation model of the positive and negative offset directions, and quantifying the influence of knowledge evolution on the core preference of the user by combining the proportional weights. The method creatively converts the knowledge evolution trend into computable vector space motion, realizes dynamic calibration of preference components through superposition, and solves the problem of error accumulation caused by knowledge update lag of the traditional recommendation system. In some embodiments, the calculating the dynamic offset of the core preference component on the medical knowledge evolution trend by tracking the updated state change of the medical concept in step 303 includes:
301. Acquiring a change record of the association relation of the medical concepts in the dynamic knowledge graph, and extracting newly added dependency relation and failed dependency relation among the medical concepts to obtain an updated state mark of the medical knowledge evolution trend;
In step 301, the association change record refers to a time sequence change dataset of connection relations between concept nodes in the medical science in the dynamic knowledge graph, and includes attributes such as relation establishment time and failure time. The newly added dependency relationship refers to the semantic association between the newly generated concepts, and the failed dependency relationship refers to the old connection which is no longer established through authority verification. The update status flag is an identification or tag that reflects the change in the association between particular medical concepts. Such labels are typically generated based on newly added dependencies and failed dependencies between medical concepts recorded in the dynamic knowledge-graph.
In the embodiment of the application, firstly, the version management function of the dynamic knowledge graph is utilized to compare the graph structure snapshots in adjacent time periods, and the association relation change record among the nodes is identified. And then scanning the newly added semantic association edges by adopting a graph structure difference analysis algorithm and adopting a sub-graph matching technology, and simultaneously positioning the invalid association edges exceeding the validity period based on a timestamp filtering mechanism. And finally, marking the detected newly-added dependency relationship as a positive evolution event, marking the failed dependency relationship as a negative evolution event, and recording the time attribute and the influence range parameter of each event to form an updated state mark with the evolution trend of medical knowledge.
302. Based on the update status flag, determining a positive offset direction corresponding to the newly added dependency relationship and a negative offset direction corresponding to the failed dependency relationship in the core preference component;
In step 302, the forward bias direction represents an enhanced action vector of semantic expansion of the medical concept on the core preference due to the newly added association. The negative-going offset direction represents the weakening vector of action of the concept dimension contraction induced by the old association failure on the core preference.
In the embodiment of the application, firstly, a state marker application graph representation learning technology is updated, and medical concepts and association relations thereof are mapped to a low-dimensional vector space, so that concepts with similar semantics have adjacency in the vector space. And then, aiming at the newly added dependency relationship, calculating vector direction difference of concept nodes at two ends of the relationship, and adopting cosine similarity variation as a forward shift direction. And finally, extracting the association strength attenuation value of the relationship in the history vector model as a negative offset direction for the failure dependency relationship.
303. And according to the superposition of the positive offset direction and the negative offset direction, and combining the proportional weight of the newly added dependency relationship and the failed dependency relationship, obtaining the dynamic offset of the core preference component on the evolution trend of medical knowledge.
In step 303, the proportional weight is comprehensively calculated by the indexes such as confidence level, academic impact factor, clinical verification level and the like of the newly added dependency and the invalid dependency, and is used for quantifying the contribution degree of different change events to the offset. Dynamic offsets refer to adjustment values that result from changes in user preference characteristics according to new medical concepts, treatment methods, or health advice as medical knowledge is updated and developed.
In the embodiment of the application, a multidimensional weight evaluation system is firstly established, characteristic data such as academic source credibility, clinical evidence number, time freshness and the like of knowledge change events are collected, and contribution degree weight coefficients of all the characteristics are calculated through an entropy weight method. And multiplying the positive and negative offset vectors obtained in the step 302 with the comprehensive weights of the corresponding events respectively to realize the proportional weights of the influences of different evolution events. And finally, carrying out synthesis operation on the weighted positive and negative vector groups by adopting a vector space linear superposition principle, eliminating dimension differences through normalization processing, and finally outputting standardized dynamic offset.
The following is a specific example:
In a certain trimethyl hospital, in the knowledge evolution scene in the cardiovascular disease treatment field, the dynamic knowledge graph is newly added with the association (weight 0.8) of the gene mutation and the antiplatelet drug reactivity, and simultaneously fails to be the old association (weight 0.6) of the beta receptor blocker and the asthma tabu. Step 301 extracts the two change events and marks the state. Step 302 calculates a positive offset vector (0.12,0.08) and a negative offset vector (-0.09,0.05) associated with the respiratory system, respectively, of the gene therapy related concept. Step 303, performing weighted superposition according to the weights to finally obtain dynamic offset (0.042,0.094) and guiding the recommendation system to adjust to the accurate medical direction.
In summary, steps 301 to 303 effectively solve the problem of recommendation deviation caused by knowledge update lag in the medical field of the traditional recommendation system by establishing a dynamic mapping mechanism of knowledge evolution and user preference. The system can automatically capture the fine granularity change of the medical concept relationship, and precisely quantize the influence direction and strength of knowledge evolution on the user core preference by combining a multidimensional weight evaluation system. Compared with a static model, the method remarkably improves the adaptability of the recommendation result to the front medical progress and the filtering efficiency of the recommendation result to outdated knowledge, ensures that the recommendation content always accords with the latest medical consensus, and simultaneously keeps the stable evolution of the personalized features. The graph difference analysis and vector space modeling method adopted in the technical implementation process provides an interpretable computing framework for processing complex knowledge evolution.
In order to precisely quantify the dynamic influence of medical knowledge evolution on user preference, a direction vector is generated through the product of the correlation strength and the adjustment parameter, and a vector space superposition technology is adopted to synthesize the net influence direction. The scheme breakthroughly converts discrete knowledge change events into continuous vector operation, and utilizes the historical association strength to keep the attenuation effect of the failure relationship, so that preference adjustment is ensured to respond to the latest medical progress and is compatible with the historical cognitive inertia. In some embodiments, determining a positive offset direction corresponding to the newly added dependency and a negative offset direction corresponding to the failed dependency in the core preference component based on the updated status flag in step 302 includes:
401. mapping the newly added dependency relationship to a first adjustment parameter of the core preference component and mapping the failed dependency relationship to a second adjustment parameter of the core preference component based on the update status flag;
In step 401, the first adjustment parameter is a quantization coefficient that reflects the influence degree of the newly added dependency relationship on the core preference, and includes dimensions such as knowledge authority, evidence level, and the like. The second adjustment parameter is an attenuation coefficient for representing the attenuation degree of the failed dependency relationship to the core preference, and comprises elements such as historical reference frequency, failure duration and the like.
In the embodiment of the application, a knowledge evolution influence evaluation model is firstly constructed by updating a state mark, characteristics such as a periodical influence factor (for example, a periodical coefficient is 5.2) of a source of the knowledge evolution influence evaluation model, the number of multi-center clinical tests (for example, 3 III-phase tests) and the like are extracted aiming at a newly added dependency relationship, and the characteristics are normalized to a first adjustment parameter in a 0-1 interval through an S-shaped function. And for the failed dependency relationship, calculating a time attenuation curve (such as an exponential attenuation factor of 0.3) of the number of times of introduction in the last five years, and generating a second adjustment parameter by adopting a linear regression model in combination with the failure confirmation time (such as 18 months of failure).
402. Generating a first direction vector set according to the product of the correlation strength between the medical concepts connected by each newly added dependency relationship and the first adjustment parameter, and generating a second direction vector set according to the product of the historical correlation strength between the medical concepts corresponding to each failed dependency relationship and the second adjustment parameter;
In step 402, the association strength refers to the tightness of semantic connections between medical concepts, and is represented by the relationship weights of the knowledge-graph. The historical association strength refers to the average influence of the failure relationship in the past period. The direction vector is a weighted multidimensional space displacement quantity and characterizes the direction and the amplitude of the adjustment of the preference component. The first direction vector set is a set composed of a series of vectors generated by multiplying the correlation strength between medical concepts connected by the newly added dependency relationship in the dynamic knowledge graph and corresponding adjustment parameters (namely, first adjustment parameters). The second direction vector set is a set composed of a series of vectors generated by multiplying the historical association strength between the corresponding medical concepts in the dynamic knowledge graph by the corresponding adjustment parameters (namely, the second adjustment parameters) of the failed dependency relationship.
In the embodiment of the application, a current weight value (such as the association strength of 'gene detection and targeted drug' of 0.78) of a new relationship is firstly obtained from a dynamic knowledge graph, and multiplied by a first adjustment parameter to obtain a first direction vector set. For failure dependency, the historical version library is called to obtain the average weight (such as the historical intensity of 'traditional chemotherapy and immunosuppression' of 0.65) in three years before failure, and the average weight is multiplied by a second adjustment parameter to be used as a second direction vector set. And then converting scalar products into displacement vectors in a vector space through a graph embedding projection technology, and retaining topological relation features of concept nodes.
403. And superposing the direction vectors in the first direction vector set to obtain a positive offset direction, and superposing the direction vectors in the second direction vector set to obtain a negative offset direction.
In step 403, the vector superposition is algebraically summing the coordinate components of the plurality of direction vectors by means of a spatial vector synthesis rule. The positive offset direction is the net influence direction generated by the newly added relation group, and the negative offset direction is the net weakening direction caused by the failure relation group. The negative bias direction refers to an indication of a change in user preference characteristics toward reduced interest or demand due to a failure or weakening of a dependency between certain medical concepts.
In the embodiment of the application, first, coordinate decomposition is carried out on all vectors in a first direction vector set, and each axial component (such as X axis+0.34 and Y axis-0.12) is obtained by respectively superposing and summing according to dimensions. The second set of direction vectors is processed in the same way to obtain the respective axial components (e.g., X-0.25, y + 0.08). The magnitude difference of each axial component is eliminated through vector modular length normalization processing, and finally a positive offset direction vector (0.94, -0.33) and a negative offset direction vector (-0.95,0.32) of unit length are formed.
The following is a specific example:
In the diabetes diagnosis and treatment knowledge updating scene of a certain trimethyl hospital, the dynamic knowledge graph is newly added with the relation between intestinal flora detection and insulin resistance (authority coefficient 0.9 and association strength 0.8), and the old relation between the failure of metformin and vitamin B12 (history strength 0.7 and attenuation coefficient 0.6). Step 401 calculates a first adjustment parameter of 0.72 (0.9x0.8) and a second adjustment parameter of 0.42 (0.7x0.6). Step 402 maps the new relationship to a direction vector (0.58,0.15) and the failure relationship to (-0.36,0.22). Step 403, after stacking other related vectors, finally generates a positive offset direction (0.62,0.31) and a negative offset direction (-0.55,0.28), and guides the recommendation system to strengthen pushing of microbiome treatment related content and weaken outdated drug side effect warning.
In summary, steps 401 to 403 effectively solve the key problem that the traditional recommendation system is difficult to adapt to the dynamic evolution of the medical knowledge by establishing the quantitative mapping mechanism of the knowledge evolution event and the preference space. The system can accurately analyze the differentiated influence of the newly added knowledge and the obsolete knowledge on the preference of the user, and a vector space modeling method is adopted to convert discrete knowledge change events into continuous directivity adjustment. Compared with rule-based empirical adjustment, the method realizes the calculability and the interpretability of evolution of preference components, ensures that a recommendation strategy follows the progress of leading-edge medicine in time, and avoids preference drift instability caused by knowledge updating. In the technical implementation process, multidimensional influence factor evaluation and vector space synthesis algorithm are fused, and a reliable mathematical modeling framework is provided for processing complex knowledge evolution.
In order to construct a personalized recommendation system dynamically adapting to medical knowledge evolution, a cognitive tag is generated by combining a knowledge hierarchy through time alignment of real-time behaviors and stability analysis of historical behaviors. The innovation point is to establish dynamic mapping of behavior data and knowledge version, adopt cumulative distribution to quantify long-term cognition precipitation, reflect professional architecture through hierarchical superposition, and realize conversion from discrete behavior to structured cognition level. In some embodiments, in step 102, the step of performing multidimensional fusion on the acquired real-time diagnosis and treatment behavior data of the user and the user history medical interaction record and the medical concept in the dynamic knowledge graph to generate a user portrait tag set with the user medical cognition level as a hierarchical reference includes:
501. Extracting an operation time stamp and an operation type identifier in the real-time diagnosis and treatment behavior data of the user, and behavior frequency and behavior duration in the historical medical interaction records of the user;
in step 501, the operation timestamp refers to specific time information of performing diagnosis and treatment actions (such as consulting literature and prescribing) by the user. The operation type identifier is a classification code (such as diagnosis class A01 and medication class B02) for distinguishing diagnosis and treatment behavior types. The frequency of behavior represents statistics of the number of user accesses to a particular medical concept. The duration of the action refers to the difference in start and stop times of a single interactive action.
In the embodiment of the application, firstly, the real-time diagnosis and treatment line data of the user is collected as a log through a medical information system interface, and an operation time stamp (such as '2023-08-20-14:30:00') and an operation type identifier are extracted by using a regular expression. And then, the historical medical interaction records of the user within three months of the user are called from the historical database, the behavior frequency of each medical concept is calculated by adopting a sliding window statistical method (e.g. 8 times of access in 'hypertension' month), and the behavior duration is calculated by using the timestamp difference value (e.g. 25 minutes of single document reading duration).
502. Performing time sequence alignment on the operation time stamp 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 according to the operation type identifier to obtain a preliminary association set;
In step 502, the time alignment is to match the user behavior occurrence time with the time period of validity of the medical concept in the knowledge graph. Behavior matching is to calculate the related operation type and the medical concept belonging field through semantic similarity. Category attributes refer to features or labels used in dynamic knowledge-graph to describe and distinguish different medical concepts. These attributes help to clarify the specific classification to which each medical concept belongs. The preliminary association set is an intermediate data set containing a mapping relationship between user behaviors and medical concepts.
In the embodiment of the application, a dynamic time warping algorithm is firstly applied to align the time sequence of the user operation time stamp and the time attribute of the medical concept in the knowledge graph. And then constructing a medical behavior ontology library, adopting cosine similarity to calculate the matching degree of operation type identification and medical concept category attribute (such as 'drug treatment scheme'), and screening behavior and concept pairs with similarity threshold exceeding 0.7. And finally integrating the time sequence alignment and behavior matching results to form a preliminary association set containing time effectiveness marks.
503. Calculating stability weights related to the medical concepts in the user history medical interaction records according to the accumulated distribution proportion of the behavior frequency and the behavior duration;
In step 503, the duration of the action refers to the length of time it takes for the user to perform a particular activity or operation. In the medical health field, it may refer to a specific duration of time that a user is engaged in a health interaction. The cumulative distribution ratio is the distribution concentration of the user's historical behavior in the time dimension. The stability weight reflects the persistence and regularity of the user's attention to a particular medical concept.
In the embodiment of the application, firstly, time slicing processing is carried out on the user history medical interaction record within three years of the user, and the variation coefficient of the behavior frequency and the behavior duration of each medical concept in each time slice (such as a quarter) is calculated. And then, giving higher weight to recent behavior data by adopting an exponential smoothing method, and calculating the cumulative distribution proportion of the behavior distribution by using the coefficient of the kene. Finally, the frequency stability (accounting for 60% of the weight) and the duration stability (accounting for 40% of the weight) are linearly combined to generate the stability weight in the range of 0-1.
504. And carrying out hierarchical superposition on the medical concepts in the preliminary association set and the stability weight, and generating a user portrait tag set taking the user medical cognition level as a hierarchical reference by combining the hierarchical distribution proportion of the medical concepts in the dynamic knowledge graph.
In step 504, the hierarchical overlay is a weighted fusion of the preliminary associations and stability weights according to a medical concept hierarchy. The hierarchical distribution proportion refers to the hierarchical weight of the medical concept in the discipline system (such as 30% for the basic medical layer and 70% for the clinical medical layer).
In the embodiment of the application, a medical concept hierarchy tree is firstly constructed, and the hierarchy where each node is located is determined according to the upper and lower relationship of medical concepts in the dynamic knowledge graph (for example, "coronary heart disease" belongs to a three-level clinical concept). And multiplying each medical concept in the preliminary association set by the stability weight of the medical concept, and combining the hierarchical distribution proportion of the medical concepts in the dynamic knowledge graph. And finally, mapping the weighted concept association degree to primary, middle and high-level medical cognitive level labels by adopting a hierarchical clustering algorithm to form a user portrait label set containing a hierarchical benchmark of weight values.
The following is a specific example:
In a cardiovascular department application scenario in a certain three-dimensional hospital, a user looks at a heart failure guideline of a certain version (the operation type is guideline review), and a history record shows continuous attention to natriuretic peptide detection (12 visits in months) for nearly two years and average study time of 38 minutes. Step 501 extracts the user's operation time stamp, guide reference identifier, and calculates the behavior frequency stability weight of "natriuretic peptide test" of 0.85. Step 502 aligns the guideline review time with the period of effectiveness of the knowledge-graph center decline treatment guideline, matching to the "natriuretic peptide monitoring" concept. Step 503 calculates the stability weight of the "natriuretic peptide test" concept of 0.76. Step 504 superimposes the "natriuretic peptide monitoring" with the stability weight, combined with its distribution ratio at the diagnostic standard layer of 0.6, to finally generate a layered tag set containing "high level cardiac marker cognition" (weight 0.68).
In summary, steps 501 to 504 solve the problem of static user portrayal and single dimension in the medical field innovatively by deeply fusing the real-time behavior data and the space-time characteristics of the knowledge graph. The system can dynamically capture the space-time correlation of diagnosis and treatment behaviors and medical knowledge evolution of the user, and combines the analysis of the stability of the historical behaviors and the distribution of concept level weights to construct a multi-layer cognitive portrait with time sensitivity and subject structural characteristics. Compared with the traditional method, the scheme remarkably improves the professionality and timeliness of the user tag system, accurately reflects the instant knowledge demands of users, deeply describes the long-term professional cognitive structure of the users, and provides accurate cognitive level reference standard for personalized medical services. The space-time alignment algorithm and the hierarchical fusion mechanism adopted in the technical implementation provide a reliable computing framework for processing complex medical behavior data.
In order to accurately correlate user behaviors with a dynamically evolving medical knowledge system, concept effectiveness is ensured through time window comparison, and category correlation is ensured by utilizing semantic mapping. The method creatively merges a dynamic programming algorithm with deep semantic matching, and solves the problems of age dislocation and semantic gap in the relation between behaviors and concepts. By generating correlation pairs with intensity values, a reliable data basis for space-time two-dimensional verification is provided for subsequent analysis. In some embodiments, in step 502, the aligning the operation timestamp with the time attribute of the medical concept in the dynamic knowledge graph, and performing behavior matching with the category attribute of the medical concept according to the operation type identifier, to obtain a preliminary association set, includes:
601. Performing time window overlapping comparison on the operation time stamp and the time attribute of the medical concept in the dynamic knowledge graph, and if the operation time stamp is in the effective time range of the time attribute, judging that the real-time diagnosis and treatment behavior of the user is aligned with the time sequence of the medical concept;
In step 601, the time window overlap alignment refers to verifying the intersection of the user behavior occurrence period and the knowledge-graph concept validity period. The validation time range is the start-stop time period in which the medical concept in the knowledge graph is accepted by the academic community. The operation time stamp is used to mark the exact time that the user performs certain medical related activities (e.g., views a particular medical science popularization article, receives a certain treatment, etc.). The real-time diagnosis and treatment behavior of the user describes activities or behaviors related to medical care, including diagnosis, diagnosis or treatment, health information consulting and the like, which are currently performed by the user.
In the embodiment of the application, firstly, time attribute metadata of a target medical concept is extracted from a dynamic knowledge graph version library, wherein the time attribute metadata comprises a concept effective start time (such as 1 month and 1 day of a certain year) and an end time (marked as permanently effective if the concept is not invalid). And then, adopting a dynamic programming algorithm to carry out window overlapping comparison on an operation time stamp (such as 09:30 of 15 days of 3 months of a certain year) and the concept effective starting time, and generating time sequence alignment when the operation time is more than or equal to the effective starting time and the operation time is less than or equal to the effective ending time.
602. Matching the real-time diagnosis and treatment behavior of the user with the category association of the medical concept according to a preset mapping relation between the operation type identifier and the category attribute of the medical concept;
In step 602, the preset mapping relationship is a classification correspondence rule base established based on the medical behavior ontology, and includes a semantic association matrix of operation types (such as laboratory report interpretation) and concept categories (such as test index analysis). The category association refers to a matching relationship established between a specific behavior type (such as clicking, reading, sharing, etc.) of a user and category attributes of a specific medical concept in a dynamic knowledge graph.
In the embodiment of the application, a mapping knowledge base of medical behaviors and concept categories is firstly constructed, and a preset mapping relation (such as image examination reservation corresponding to image diagnosis standard) between operation type identifiers and category attributes of medical concepts is screened in the mapping knowledge base by adopting an expert labeling mode. Then, a semantic similarity calculation model is introduced, text semantic matching degree of operation type identification and concept category and a preset mapping relation (such as a 'laboratory detection index') is analyzed through a bi-directional encoder characterization technology, and matching pairs with similarity threshold exceeding 0.75 are screened. And finally generating category association with a confidence value.
603. And combining the user real-time diagnosis and treatment behaviors and medical concepts with aligned time sequences and matched category correlations into behavior and concept correlation pairs, and summarizing the behavior and concept correlation pairs to obtain a preliminary correlation set.
In step 603, the behavior and concept association pair is a valid matching unit that is subjected to space-time double verification, and includes a time stamp alignment proof, a category matching proof and an association strength value. The preliminary association set is a data set formed by combining real-time diagnosis and treatment behaviors of the user after time sequence alignment and category association matching with corresponding medical concepts.
In the embodiment of the present application, first, a cartesian product operation is performed on the time sequence alignment result of step 601 and the category matching result of step 602, so that a combination of time effectiveness and category matching performance is maintained. Behavior and concept association pairs are then calculated based on the confidence level (e.g., 0.85) and the time window overlap level (e.g., fully contained by 1.0, partially overlapped by 0.6) of the mapping relationship. And finally summarizing all behavior and concept association pairs meeting the conditions, and arranging the behavior and concept association pairs according to the descending order of association strength to form a preliminary association set with weight identifiers.
The following is a specific example:
In a hypertension management scenario of a certain trimethyl hospital, a user executes a dynamic blood pressure monitoring report interpretation operation (time stamp 05.20:00) on the 5 th month and 20 th day of a certain year, and the effective time of a 24-hour dynamic blood pressure diagnosis standard concept in a knowledge graph is 11 months to 12 months of a certain year. Step 601 compares the time window to confirm that the operation time is within the effective range. Step 602 matches the "report interpretation" operation type with the "diagnostic criteria" concept category (similarity 0.89) according to the preset map. Step 603 generates an association pair (dynamic blood pressure monitoring report interpretation, 24-hour dynamic blood pressure diagnosis standard, association strength of 0.93) and forms a preliminary association set together with other effective associations.
In summary, steps 601 to 603 effectively solve the problem of space-time misalignment between medical behaviors and knowledge concepts through a dual verification mechanism. The system creatively combines time validity verification with semantic association matching, so that not only is knowledge concepts associated with user behaviors ensured to be in a current valid state, but also semantic consistency of operation intentions and concept categories is ensured. Compared with a single-dimension matching method, the method has the advantages that the timeliness accuracy and professional relevance of the association result are remarkably improved, and the recommendation of outdated knowledge or semantic deviation content is avoided. The dynamic programming time comparison and semantic depth matching algorithm adopted in the technical implementation provides a reliable computing framework for processing complex medical space-time data, and supports and builds a knowledge service system which accurately reflects the real-time requirements of users.
In order to dynamically adapt to the medical knowledge evolution and the user cognitive change, the scheme calculates and matches the core requirement through vector similarity, sets a hierarchical adjustment rule to balance key coverage and knowledge expansion, fuses time attenuation and hot spot reinforcement double factors, adopts a diversity control strategy, and breaks through the homogenization limitation of traditional recommendation. The method realizes collaborative optimization of personalized demand matching, knowledge timeliness maintenance and content ecological balance. In some embodiments, in step 104, dynamically adjusting the recommendation priority ranking of the medical science popularization articles according to the joint matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph, and generating a personalized recommendation sequence matching the real-time medical cognitive level of the user includes:
701. Carrying out matching degree calculation on the association relation between the dynamic recommendation feature vector and the medical concept in the dynamic knowledge graph, and extracting a matching degree value directly associated with the medical concept in the dynamic recommendation feature vector;
In step 701, the matching degree calculation is to evaluate the association compactness of the dynamic recommendation feature vector and the medical concept in the knowledge graph through a vector space similarity measurement method. The matching degree value is the normalized similarity score, and the range is between 0 and 1.
In the embodiment of the application, firstly, a graph embedding technology is adopted to convert medical concepts in a dynamic knowledge graph into vector representations to form a concept vector library. And then, calculating the similarity score of the association relation between the dynamic recommendation feature vector (such as [0.34, -0.12, 0.78 ]) of the user and the medical concept in the dynamic knowledge graph by using a cosine similarity algorithm. Finally, mapping similarity scores to a 0-1 interval through maximum and minimum value normalization processing, and extracting directly-related matching degree values exceeding a threshold value of 0.6.
702. Setting priority adjustment rules of different medical concept levels according to the range of the matching degree values;
In step 702, the priority adjustment rule is a weight allocation policy according to the matching degree interval division, and includes three mechanisms of a core concept reinforcement rule, an associated concept expansion rule and an edge concept suppression rule. The medical concept hierarchy is a way to organize rich medical information (such as diseases, symptoms, treatment methods, drugs, etc.) by dividing these concepts into different layers according to their nature, relevance or importance.
In the embodiment of the application, a three-level interval of matching degree values is firstly set, high matching degree (0.8-1.0) corresponds to a core concept layer, medium matching degree (0.6-0.8) corresponds to an associated concept layer, and low matching degree (< 0.6) corresponds to an edge concept layer. Then constructing a hierarchical weight coefficient matrix through a medical concept hierarchy, giving 3 times of basic weight to a core concept layer, giving 1.5 times of basic weight to an associated concept layer, and giving 0.3 times of attenuation weight to an edge concept layer. And finally, generating a priority adjustment rule containing parameters such as a weight coefficient, an upper limit of recommended frequency and the like according to the three-level interval of the matching degree value.
703. Based on the priority adjustment rule, performing priority weight distribution on medical science popularization articles associated with the real-time medical cognition level of the user in the dynamic knowledge graph;
In step 703, the priority weight assignment is a comprehensive calculation process combining the concept-level weight and the article relevance, and includes two dynamic parameters, namely a time decay factor and a hot spot enhancement factor. The medical science popularization paper refers to a paper of popular medical knowledge, and generally covers the contents of prevention, diagnosis, treatment methods and the like of diseases, and aims to improve public health awareness.
In the embodiment of the application, firstly, the association concept and the matching degree value of each medical science popularization article in the dynamic knowledge graph are extracted, and are converted into the hierarchical weight according to the rule of step 702. Then introducing a time decay function to reduce the article weight exceeding the validity period (such as the article weight being released for more than 2 years by 0.7), and superposing the real-time popularity coefficient (such as the new article weight being 3 months by 1.2) provided by the hot spot monitoring module. And finally, calculating the article comprehensive priority weight distribution through a linear weighting formula, and reserving the two-bit precision after the decimal point.
704. And screening medical science popularization articles matched with the real-time medical cognition level of the user from the dynamic knowledge graph according to the priority weight distribution result, and sequencing the recommendation priorities from high to low according to the priority weight distribution proportion to generate a personalized recommendation sequence.
In step 704, the recommendation prioritization is based on the descending ranking result of the weight values, and the recommendation sequence is dynamically maintained by using an adaptive queue management mechanism. The real-time medical cognition level of a user refers to the understanding degree of medical knowledge, interest points and grasp of specific medical concepts at the current moment of the user. Such cognitive levels are derived based on real-time diagnosis and treatment performance data (e.g., recently accepted treatments, reviewed health information, etc.) of the user and historical medical interaction records dynamic assessment. The personalized recommendation sequence refers to an ordered list of a series of recommendation contents customized according to specific requirements and preferences of users, and particularly refers to medical science popularization articles which are screened from dynamic knowledge maps according to priority weight distribution results and matched with real-time medical cognition levels of the users.
In the embodiment of the application, a candidate article pool is established through a priority weight distribution result, and article items of medical science popularization articles with weight values lower than 0.5 are filtered. And then, ranking the articles entering the surrounding according to the weight value by adopting a heap ranking algorithm to generate an initial recommendation sequence. Finally, introducing a diversity control module, and implementing an interval arrangement strategy (such as that articles with the same disease type are separated by more than 3 positions) on articles with the same category in the initial recommendation sequence to form a final personalized recommendation sequence.
The following is a specific example:
In a heart failure patient educational scenario in a certain trimethyl hospital, the user dynamically recommends a degree of matching of the feature vector with the concept of "natriuretic peptide monitoring" of 0.92 (core layer) and the concept of matching of "diuretic use" of 0.75 (associated layer). Step 701, extracting matching degree values of the two concepts, step 702, giving a weight which is 3 times that of a core layer and a weight which is 1.5 times that of an associated layer, step 703, calculating the weight of a journal (associated core layer and release time of 3 months) to obtain a weight value of 2.76, and the weight of the other journal (associated layer and release time of 18 months) is 0.9, step 704, sequencing, preferentially recommending natriuretic peptide related articles, and inserting heart failure diet management in the third position to ensure diversity.
In summary, steps 701 to 704 effectively solve the dual problems of insufficient individuation degree and lack of knowledge timeliness of the medical recommendation system through a multi-level weight dynamic allocation mechanism. The system innovatively carries out fusion calculation on the real-time cognitive characteristics, knowledge evolution state and content quality factors of the user, and realizes three key breakthroughs, namely, a quantized mapping model of concept matching degree and content priority is established, and the cognitive demands of the user are accurately reflected. And secondly, introducing time attenuation and hot spot reinforcement double-factor dynamic adjustment, and balancing knowledge accuracy and frontier. Thirdly, a diversity control strategy is adopted, so that homogenization of a recommended result is avoided. Compared with the traditional static recommendation method, the dynamic fitness of the recommendation content and the cognitive state of the user is remarkably improved, and intelligent knowledge service adaptation capability is provided for scenes such as patient education, clinical decision support and the like.
Fig. 2 is a schematic structural diagram of a medical science popularization article recommendation system based on user portrait according to an embodiment of the present application, where, as shown in fig. 2, the system includes:
The analysis module 21 is used for constructing a dynamic knowledge graph containing dynamic dependency relations among medical concepts by analyzing the logical association of unstructured text and structured case data in the multi-source heterogeneous medical data;
the fusion module 22 is used for carrying out multidimensional fusion on the acquired real-time diagnosis and treatment behavior data of the user and the history medical interaction records of the user and the medical concepts in the dynamic knowledge graph to generate a user portrait tag set taking the medical cognition level of the user as a layering reference;
The conversion module 23 is used for performing cross-scene feature conversion from the unreal scene of the user history medical interaction records to the real scene of the user real-time diagnosis and treatment behavior data to generate a dynamic recommendation feature vector which is updated synchronously with the medical knowledge evolution trend;
and the adjusting module 24 dynamically adjusts the recommendation priority ordering of the medical science popularization articles according to the joint matching result between the dynamic recommendation feature vector and the medical concepts in the dynamic knowledge graph so as to generate a personalized recommendation sequence matched with the real-time medical cognition level of the user.
The medical science popularization article recommending system based on the user portrait shown in fig. 2 can execute the medical science popularization article recommending method based on the user portrait shown in the embodiment shown in fig. 1, and the implementation principle and technical effects are not repeated. The specific manner in which the various modules and units perform operations in the user portrayal-based medical science popularization article recommendation system in the above embodiments has been described in detail in connection with the embodiments of the method and will not be described in detail herein.
In one possible design, a user portrayal-based medical science popularization article recommendation system of the embodiment of FIG. 2 may be implemented as a computing device, as shown in FIG. 3, which may include a storage component 31 and a processing component 32;
The storage component 31 stores one or more computer instructions for execution by the processing component 32.
The processing component 32 is configured to implement a user portrayal-based medical science popularization article recommendation method according to the embodiment of fig. 1.
Wherein the processing component 32 may include one or more processors to execute computer instructions to perform all or part of the steps of the methods described above. 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 elements for executing the methods described above.
The storage component 31 is configured to store various types of data to support operations at the terminal. The memory component may be implemented by any type or combination of volatile or nonvolatile memory devices 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 or optical disk.
Of course, the computing device may necessarily include other components as well, such as input/output interfaces, display components, communication components, and the like.
The input/output interface provides an interface between the processing component and a peripheral interface module, which may be an output device, an input device, etc.
The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, and the like.
The computing device may be a physical device or an elastic computing host provided by the cloud computing platform, and at this time, the computing device may be a cloud server, and the processing component, the storage component, and the like may be a base server resource rented or purchased from the cloud computing platform.
The embodiment of the application also provides a computer storage medium which stores a computer program, and the computer program can realize the medical science popularization article recommending method based on the user portrait in the embodiment shown in the figure 1 when being executed by a computer.
It will be clear to those skilled in the art that, for convenience and brevity of description, specific working procedures of the above-described systems, apparatuses and units may refer to corresponding procedures in the foregoing method embodiments, which are not repeated herein.
The apparatus embodiments described above are merely illustrative, wherein the elements illustrated as separate elements may or may not be physically separate, and the elements shown as elements may or may not be physical elements, may be located in one place, or may be distributed over a plurality of network elements. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art will understand and implement the present invention without undue burden.
From the above description of the embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented by means of software plus necessary general hardware platforms, or of course may be implemented by means of hardware. Based on this understanding, the foregoing technical solution may be embodied essentially or in a part contributing to the prior art in the form of a software product, which may be stored in a computer readable storage medium, such as ROM/RAM, a magnetic disk, an optical disk, etc., including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute the method described in the respective embodiments or some parts of the embodiments.
It should be noted that the above-mentioned embodiments are merely for illustrating the technical solution of the present application, and not for limiting the same, and although the present application has been described in detail with reference to the above-mentioned embodiments, it should be understood by those skilled in the art that the technical solution described in the above-mentioned embodiments may be modified or some technical features may be equivalently replaced, and these modifications or substitutions do not make the essence of the corresponding technical solution deviate from the spirit and scope of the technical solution of the embodiments of the present application.