A method and system for construction of a cardiac care knowledge base

Abstract

The present application relates to the technical field of data processing, and particularly relates to a construction method and system for a heart care knowledge base, comprising: collecting multi-source heterogeneous data and preprocessing to generate standardized knowledge metadata; constructing a heart care dynamic knowledge graph containing node confidence weight and timestamp attribute based on the knowledge metadata; collecting patient static and dynamic data to construct an individualized care portrait and generate a dynamic user state representation with timing characteristics; inputting the dynamic user state representation into a reasoning adaptation engine based on a graph neural network, reasoning and calculating the matching degree of candidate knowledge nodes on the graph, and screening to generate a personalized care knowledge set; collecting clinical feedback signals to dynamically update the confidence weight and engine parameters. The present application solves the problem of poor data correlation by constructing a semantic correlation graph, realizes personalized dynamic adaptation through dynamic state embedding, and guarantees the timeliness and scientificity of the knowledge base through a feedback loop.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for constructing a cardiac care knowledge base. Background Technology

[0002] With the development of medical informatization, a large amount of clinical data, nursing standards, and research literature have accumulated in the field of cardiac nursing. In order to systematize this scattered knowledge and facilitate its retrieval and application by medical staff, knowledge base technology has been introduced into the cardiac nursing management system.

[0003] Currently, existing methods for constructing cardiac nursing knowledge bases typically involve: collecting clinical guidelines and case data, classifying and storing the data in a relational database through manual annotation or keyword matching, forming a static knowledge entry base. When healthcare professionals need to query nursing protocols, they can search by entering keywords, and the system returns matching entries.

[0004] However, the aforementioned existing technologies have the following technical shortcomings: First, because cardiac care involves complex relationships across multiple dimensions such as etiology, symptoms, medication, and rehabilitation, traditional relational databases struggle to capture and express the deep semantic connections between these entities, resulting in low accuracy in knowledge retrieval and frequent instances of inaccurate or incomplete searches. Second, the knowledge bases generated by existing construction methods are mostly static and cannot dynamically adjust recommended nursing content based on the patient's real-time physiological indicators, making it difficult to meet the needs of personalized care. Finally, existing methods lack a dynamic evaluation mechanism for the timeliness and authority of knowledge sources, which can easily lead to the continued use of outdated nursing information, affecting the scientific nature of nursing decisions. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for constructing a cardiac care knowledge base, which solves the problems of poor correlation between multi-source heterogeneous data, inability to dynamically adjust nursing knowledge according to the real-time status of patients, and insufficient recommendation accuracy due to lagging knowledge updates in existing cardiac care knowledge base construction methods.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for constructing a cardiac care knowledge base, comprising the following steps: Collect multi-source heterogeneous data related to cardiac care, preprocess the multi-source heterogeneous data, and generate standardized knowledge metadata; Based on the standardized knowledge metadata, a dynamic knowledge graph for cardiac care is constructed, with entities in the cardiac care domain as nodes and semantic relationships between entities as edges; the nodes have initial confidence weights and timestamp attributes. Collect static attribute data and dynamic monitoring data of the target patient, construct a personalized nursing profile of the patient, and embed the dynamic monitoring data into the personalized nursing profile of the patient to generate a dynamic user status representation with time-series characteristics; The dynamic user state representation is input into a pre-trained reasoning adaptation engine, and reasoning is performed on the dynamic knowledge graph of cardiac care. The matching degree of each candidate knowledge node is calculated based on the dynamic user state representation, and knowledge nodes with matching degrees exceeding a preset threshold are selected to generate a personalized nursing knowledge set. Feedback signals to the personalized care knowledge set are collected, and the initial confidence weights and parameters of the inference adaptation engine are updated using the feedback signals.

[0007] The process involves preprocessing the multi-source heterogeneous data to generate standardized knowledge metadata, including: Natural language processing is performed on the unstructured and semi-structured data in the multi-source heterogeneous data. The natural language processing includes entity recognition based on a dictionary specific to the cardiac care domain, and relation extraction based on temporal relationships. The identified entities are mapped to a standard medical terminology set, the extracted relationships are mapped to a predefined relation ontology library, and each entity and relationship is marked with its data source and timestamp to generate the standardized knowledge metadata.

[0008] Specifically, based on the standardized knowledge metadata, a dynamic knowledge graph for cardiac care is constructed, using entities in the cardiac care domain as nodes and semantic relationships between entities as edges, including: The entities in the standardized knowledge metadata are transformed into nodes in the graph, and each node is assigned an initial confidence weight, and the creation time and the most recent verification time are recorded. The relationships in the standardized knowledge metadata are transformed into edges connecting nodes in a graph, and each edge is assigned a relationship type, relationship attribute, and initial confidence weight, and the creation time and the most recent verification time are recorded. The relational attributes include time constraint attributes.

[0009] This process involves collecting static attribute data and dynamic monitoring data of the target patient to construct an individualized patient care profile. The dynamic monitoring data is then embedded into this profile to generate a dynamic user state representation with temporal characteristics, including: Collect static attribute data and dynamic monitoring data of the target patients; A static profile containing demographic features and past medical history is constructed based on the aforementioned static attribute data; Based on the dynamic monitoring data, a real-time state vector is extracted, which includes the current value, trend characteristics, and fluctuation characteristics. The features of the static profile are fused with the real-time state vector, and time context information is injected to generate the dynamic user state representation through a time-series coding module.

[0010] The inference adaptation engine includes a graph coding module, a state graph interaction module, and a candidate node scoring module. The graph encoding module is used to encode the nodes and edges in the cardiac care dynamic knowledge graph into node embedding vectors and relation embedding vectors. The state graph interaction module is used to calculate the attention similarity between the dynamic user state representation and the node embedding vector; The candidate node scoring module is used to calculate the matching degree based on the attention similarity, the cumulative confidence of the path from the seed node to the candidate node, and the temporal adaptability of the candidate node.

[0011] The inference adaptation engine is obtained through a two-stage training process of pre-training and fine-tuning. The pre-training phase utilizes the structural information of the cardiac care dynamic knowledge graph to learn the node embedding vector and relation embedding vector by predicting the masked edges; The fine-tuning phase utilizes historical case data, taking the actually adopted nursing knowledge as positive examples, to optimize the calculation of the matching degree.

[0012] Before calculating the matching degree of each candidate knowledge node based on the dynamic user state representation, the method further includes: Seed nodes are matched from the cardiac care dynamic knowledge graph based on the key features in the dynamic user state representation. The seed node and its neighboring nodes within a preset hop count range are selected as candidate knowledge nodes.

[0013] The preset threshold is set between 0.6 and 0.75 and is dynamically adjusted according to the adoption rate of clinical scenarios or feedback signals.

[0014] The process includes collecting feedback signals from the personalized care knowledge set and using these feedback signals to update the initial confidence weights and the parameters of the inference adaptation engine, including: The feedback signals to the personalized nursing knowledge set are collected, wherein the feedback signals include data on the adoption behavior of medical staff in relation to the personalized nursing knowledge set and data on subsequent changes in the patient's physiological indicators. When the adoption behavior data is positive and the subsequent physiological indicator change data is positive, the confidence weight of the set nodes and edges is enhanced, and the parameters of the inference adaptation engine are adjusted positively. When the adoption behavior data is rejection or the subsequent physiological indicator change data is deterioration, the confidence weight of the set node and edge is reduced, and the inference adaptation engine is negatively adjusted.

[0015] Secondly, the present invention provides a system for constructing a cardiac care knowledge base, applied to a method for constructing a cardiac care knowledge base as provided in the first aspect, comprising: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data related to cardiac care, preprocess the multi-source heterogeneous data, and generate standardized knowledge metadata. A dynamic knowledge graph construction module is used to construct a dynamic knowledge graph for cardiac care based on the standardized knowledge metadata, with entities in the cardiac care domain as nodes and semantic relationships between entities as edges; the nodes have initial confidence weights and timestamp attributes. The user profile and status embedding module is used to collect static attribute data and dynamic monitoring data of the target patient, construct a personalized nursing profile of the patient, and embed the dynamic monitoring data into the personalized nursing profile of the patient to generate a dynamic user status representation with time-series characteristics. The reasoning adaptation and knowledge generation module is used to input the dynamic user state representation into the pre-trained reasoning adaptation engine, perform reasoning on the dynamic knowledge graph of cardiac care, calculate the matching degree of each candidate knowledge node according to the dynamic user state representation, and filter out knowledge nodes with matching degree exceeding a preset threshold to generate a personalized nursing knowledge set. The dynamic update and feedback module is used to collect feedback signals on the personalized care knowledge set and use the feedback signals to update the initial confidence weights and the parameters of the inference adaptation engine.

[0016] This invention discloses a method and system for constructing a cardiac care knowledge base, comprising: collecting multi-source heterogeneous data and preprocessing it to generate standardized knowledge metadata; constructing a dynamic cardiac care knowledge graph containing node confidence weights and timestamp attributes based on the knowledge metadata; collecting static and dynamic patient data to construct individualized nursing profiles and generating dynamic user state representations with temporal features; inputting the dynamic user state representations into a graph neural network-based inference adaptation engine to infer and calculate the matching degree of candidate knowledge nodes on the graph, and filtering to generate a personalized nursing knowledge set; collecting clinical feedback signals and dynamically updating the confidence weights and engine parameters. This invention solves the problem of poor data correlation by constructing a semantic association graph, achieves personalized dynamic adaptation through dynamic state embedding, and ensures the timeliness and scientific rigor of the knowledge base through a feedback loop, making it suitable for clinical decision support in cardiac care. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0018] Figure 1 This is a schematic diagram illustrating the steps of a method for constructing a cardiac care knowledge base according to the first embodiment of the present invention.

[0019] Figure 2 This is a simplified flowchart illustrating a method for constructing a cardiac care knowledge base provided by the present invention.

[0020] Figure 3 This is a complete flowchart illustrating a method for constructing a cardiac care knowledge base provided by the present invention.

[0021] Figure 4 This is a structural schematic diagram of a system for constructing a cardiac care knowledge base according to a second embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram of the electronic device of the present invention.

[0023] In the diagram: 101 - Data Acquisition and Preprocessing Module, 102 - Dynamic Knowledge Graph Construction Module, 103 - User Profile and State Embedding Module, 104 - Reasoning Adaptation and Knowledge Generation Module, 105 - Dynamic Update and Feedback Module. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0025] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0026] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0027] The first embodiment of this application is as follows: Please see Figures 1-3 This invention provides a method for constructing a knowledge base for cardiac care, comprising the following steps: S1. Collect multi-source heterogeneous data related to cardiac care, preprocess the multi-source heterogeneous data, and generate standardized knowledge metadata.

[0028] Specifically, based on the characteristics of cardiac care knowledge, three core data sources were identified: Structured data sources primarily originate from electronic medical records (EMRs) and nursing records within the hospital information system. This data is stored in relational database tables and includes basic patient information, diagnostic records, medical orders, vital sign monitoring data, etc. Anonymized structured data is obtained through hospital data interfaces or by exporting from authorized databases.

[0029] Semi-structured data sources: These mainly refer to clinical practice guidelines, expert consensus documents, and nursing pathway documents for cardiac nursing published by authoritative medical institutions or academic organizations. These data are typically published in PDF or HTML format and have structural markers such as chapter titles and lists, but the main content is unstructured text. During collection, web crawlers are used to selectively retrieve guideline documents from the official websites of authoritative domestic and international cardiovascular societies (such as AHA, ESC, and the Chinese Society of Cardiology), and archive them according to publication date and version number.

[0030] Unstructured data sources include research literature, case reports, and summaries of nursing experience published in medical journals related to cardiac care. Full-text articles are obtained by searching academic databases such as PubMed and CNKI using keyword combinations (e.g., "heart failure nursing" and "coronary artery disease rehabilitation"). Furthermore, considering copyright issues, only open-access or authorized literature is collected.

[0031] For structured data, incremental data is extracted from the hospital's data warehouse periodically using ETL tools to ensure data real-time performance.

[0032] For semi-structured and unstructured data, a distributed crawler system is adopted for collection, and a crawling strategy is set, such as preferentially crawling guidelines and literature published in the past five years to ensure the timeliness of knowledge.

[0033] During the collection process, a data quality assessment mechanism is established: each document or record is scored for source credibility (for example, guidelines published by authoritative institutions are given a high score, while personal blogs or non-peer-reviewed articles are excluded), and the publication timestamp is recorded as the basis for subsequent knowledge timeliness assessment. Only the data that passes the initial quality screening will enter the preprocessing stage.

[0034] Since unstructured and semi-structured data account for the vast majority of knowledge sources and are difficult to process, the present invention designs a text processing pipeline specifically for the characteristics of the cardiac care field, including key steps such as cleaning, segmentation, entity recognition adapted to the field, and relationship extraction.

[0035] For the crawled PDF or HTML documents, first, parsing tools are used to extract the pure text content and remove noise information such as headers, footers, reference lists, and advertisement links. For text recognition errors caused by scanning, a correction algorithm based on a medical dictionary is used for correction, for example, correcting "coronary heart pain" to "coronary heart disease".

[0036] The cleaned text is uniformly converted into a pure text format encoded in UTF-8 and is initially segmented according to the document source and chapter structure to form text blocks in units of paragraphs for subsequent processing.

[0037] To improve the accuracy of subsequent entity recognition, the present invention pre-constructs a dynamically expanding dictionary for the cardiac care field. The construction of this dictionary is divided into two steps: [[ID=十六]] [[ID=十七]]Basic dictionary: Extract cardiac care-related entries from existing medical vocabularies (such as ICD-10 diagnostic codes, ATC drug classification systems), including disease names (such as "acute myocardial infarction"), symptom terms (such as "exertional dyspnea"), drug generic names (such as "aspirin"), nursing operations (such as "electrocardiogram monitoring", "postural drainage"), and rehabilitation indicators (such as "6-minute walk distance"). [[ID=十八]] [[ID=十九]]

[0038] [[ID=二十]]Expansion dictionary: Use word vector technology to train the collected corpus to discover clinically common abbreviations, acronyms, or synonymous expressions that are semantically similar to the basic entries (such as "MI" corresponding to "myocardial infarction"), and after expert review, add them to the dictionary. This dictionary is continuously updated during the preprocessing process to adapt to the dynamic changes in language use. [[ID=二十一]] [[ID=二十二]]

[0039] For semi-structured clinical guidelines, intelligent segmentation is performed using their inherent hierarchical structure (such as chapter titles and subheadings) to divide long texts into segments with thematic boundaries. For example, each drug paragraph under the "Drug Therapy" chapter in the guideline is segmented separately. For unstructured literature, segmentation is performed based on the natural boundaries of paragraphs, and rules are used to identify parts such as abstracts, methods, and results, facilitating subsequent differentiated processing according to content type.

[0040] Based on word segmentation and part-of-speech tagging, an entity recognition method combining deep neural networks and dictionary rules is adopted: First, the text block is input into a BERT-based pre-trained language model to obtain a semantic vector representation of each character. This model has been fine-tuned on a massive medical corpus and is able to understand professional expressions in cardiac care texts.

[0041] Secondly, a multi-head attention mechanism is introduced, enabling the model to pay attention to the contextual information at different locations in the text, thereby accurately identifying nested entities (for example, in "heart failure with reduced left ventricular ejection fraction", "left ventricular ejection fraction" is identified as the inspection indicator entity and "heart failure" is identified as the disease entity).

[0042] Meanwhile, a dictionary of cardiac care is incorporated as external knowledge into the recognition process. Entries already in the dictionary are given higher probability weights during model decoding, effectively addressing the problem of low recognition rates for rare or novel entities. Using this method, seven categories of entities—diseases, symptoms, medications, nursing procedures, and examination indicators—are accurately extracted, and the location and confidence level of each entity in the original text are marked.

[0043] Building upon entity recognition, the semantic relationships between entities are further extracted, which is the core of constructing a knowledge graph. The relationship extraction module of this invention not only focuses on static relationships (such as "aspirin is used to treat coronary heart disease"), but also pays special attention to dynamic temporal relationships, laying the foundation for the subsequent construction of a dynamic knowledge graph.

[0044] Static relation extraction: A relation classifier based on a graph convolutional neural network is trained using predefined semantic relation types (such as "etiology", "clinical manifestations", "drug treatment", "nursing measures", "contraindications", etc.). This classifier takes a sentence containing two entities as input and predicts the relationship category between them. For example, it extracts a triple (heart failure, clinical manifestations, dyspnea) from the sentence "Patients with heart failure often experience dyspnea".

[0045] Temporal Relationship Mining: Much knowledge in cardiac care possesses temporal attributes, such as "closely monitor blood pressure within 24 hours post-surgery" and "monitor INR values ​​weekly while taking warfarin." This invention extracts temporally constrained relationships through dependency parsing and temporal expression identification (e.g., "three days post-surgery," "once daily"). Specifically, after identifying the relationship, the presence of words or phrases indicating time, frequency, or periodicity within the sentence is detected, and these are extracted as attributes of the relationship. For example, from "patients with acute myocardial infarction need emergency PCI within 12 hours of onset," the following are extracted: (acute myocardial infarction, treatment measures, emergency PCI), with the temporal attribute "within 12 hours of onset." This relationship with temporal labels provides a crucial basis for dynamically adjusting knowledge based on the patient's real-time status.

[0046] The extracted entities and relations still suffer from inconsistent naming (e.g., the same drug may have a brand name and a generic name). To address this, all entities are mapped to a standard medical terminology set (such as SNOMED CT or the Unified Medical Language System UMLS), and relation types are mapped to a predefined relation ontology. Finally, standardized knowledge metadata is generated, containing entity ID, entity type, relation type, target entity ID, time attribute (if present), confidence score, data source, and timestamp. Each piece of knowledge metadata constitutes a basic unit of the knowledge graph.

[0047] Knowledge metadata from different data sources may conflict or be redundant. For example, guidelines from different years may offer different recommendations for the same nursing procedure. This invention employs a fusion strategy based on confidence level and timeliness: For each piece of knowledge metadata, a comprehensive confidence score is calculated based on its source authority and publication time. Source authority is determined using a pre-defined institutional rating scale (e.g., 0.9 for AHA guidelines and 0.5 for a case report); the more recent the publication time, the higher the confidence score.

[0048] When multiple knowledge elements describe the same entity relationship but conflict in content, the one with the highest confidence is retained, and the others are used as alternative additional contextual information.

[0049] The merged knowledge metadata is ultimately stored in the knowledge graph database, while retaining its timestamp and confidence weight for optimization during the dynamic update process.

[0050] Through the aforementioned refined data collection and preprocessing process, this invention transforms scattered, multi-source, heterogeneous cardiac care data into high-quality, semantically rich, and time-series-informed standardized knowledge metadata, providing a solid data foundation for the subsequent construction of dynamic knowledge graphs and inference adaptation engines. The core innovation of this process lies in: constructing a dynamically expanding dictionary specifically for the cardiac care field, introducing a time-series relationship mining mechanism, and ensuring the accuracy and timeliness of the knowledge through multi-source fusion and confidence assessment.

[0051] S2. Based on the standardized knowledge metadata, construct a dynamic knowledge graph for cardiac care, with entities in the cardiac care domain as nodes and semantic relationships between entities as edges; the nodes have initial confidence weights and timestamp attributes.

[0052] Specifically, the dynamic knowledge graph for cardiac care constructed in this invention adopts a directed attribute graph model. In this model, nodes represent core entities in the cardiac care domain, and edges represent semantic relationships between entities. Each node and edge contains a set of attributes to describe its identity, type, confidence level, and time information. This structure can intuitively map the intricate network of relationships in cardiac care knowledge; for example, a disease may be associated with multiple symptoms, correspond to multiple medications, and depend on specific nursing procedures.

[0053] Entity nodes are the basic units of a knowledge graph, corresponding to the various cardiac care concepts identified and standardized in step 1. Based on the actual needs of cardiac care, this invention defines the following five core entities: Disease entities: covering various heart diseases and their subtypes, such as "acute myocardial infarction", "chronic heart failure", "atrial fibrillation", etc.

[0054] Symptoms: Describe the clinical manifestations of the disease, such as "chest pain", "dyspnea", "edema", etc.

[0055] Drug entities: This includes commonly used drugs for treating heart disease, such as "aspirin", "warfarin", and "metoprolol", and distinguishes between generic names and brand names, mapping them uniformly to standard names.

[0056] Nursing procedure entities: refer to specific nursing interventions, such as "electrocardiographic monitoring", "postural care", "dietary guidance", "rehabilitation training", etc.

[0057] Rehabilitation indicators are quantitative indicators or states used to assess a patient's condition, such as "left ventricular ejection fraction", "New York heart function classification", and "blood pressure range".

[0058] For each standardized entity output from step 1, the system creates a corresponding node in the knowledge graph. Each node contains the following key attributes: Entity unique identifier: Employ globally unique encoding, such as a concept unique identifier based on the Unified Medical Language System (UMLS) or a custom UUID, to ensure that each entity can be accurately referenced.

[0059] Entity name: The standard name of the entity, such as "acute ST-segment elevation myocardial infarction".

[0060] Entity type: Marks which of the five categories the node belongs to, facilitating subsequent retrieval and reasoning by category.

[0061] Initial confidence weight: This is a value between 0 and 1, used to measure the credibility of the entity itself. The calculation of the initial confidence weight considers two factors: Source authority score: The score is based on the data source from which the entity originated in Step 1. For example, an entity from the latest American Heart Association (AHA) guidelines is assigned a high authority score of 0.95, an entity from a meta-analysis in an authoritative journal is assigned 0.90, and an entity from a single case report is assigned 0.60. If the same entity appears in multiple sources, the highest authority score is used.

[0062] Timeliness coefficient: Calculated based on the entity's most recent verification time (i.e., the latest publication time of the knowledge corresponding to the entity in step 1). The closer to the current time, the higher the timeliness coefficient. For example, the timeliness coefficient is set to 1.0 for the past five years, 0.8 for five to ten years, and 0.5 for more than ten years.

[0063] The initial confidence weight is derived by multiplying the authority score by the timeliness coefficient. For example, an entity called "acute myocardial infarction" derived from the 2022 AHA guidelines would have an initial confidence weight of 0.95 × 1.0 = 0.95. This weight will be dynamically adjusted in the feedback loop of subsequent step 5.

[0064] Timestamp: Records two points in time—creation time (the time when the node was first added to the graph) and most recent verification time (the time when the knowledge source on which the node is based was most recently confirmed to be valid). The latter will be updated in subsequent updates as new guidelines are released or feedback loops are confirmed.

[0065] In this way, all entities extracted from multi-source data are transformed into nodes carrying quality labels, laying a reliable foundation for subsequent association construction and reasoning.

[0066] Semantic relation edges are used to connect entity nodes and express the inherent relationships between them. Based on the characteristics of knowledge in the field of cardiac care, this invention defines the following core relation types, each corresponding to key logic in clinical nursing: Etiological relationship: This indicates the connection between a disease and its cause, such as the causal relationship between "hypertension" and "coronary heart disease", or the causal relationship between "infection" and "infective endocarditis".

[0067] Clinical manifestation relationship: connects disease and symptoms, expressing the symptoms that a certain disease usually presents, such as the "clinical manifestation" relationship between "heart failure" and "dyspnea".

[0068] Medication relationship: Connects disease and drug, expressing the commonly used or recommended drugs for treating a certain disease, such as the "medication" relationship between "atrial fibrillation" and "warfarin". This relationship can be supplemented with attributes such as usage and dosage.

[0069] Nursing dependency relationship: connects disease or symptom with nursing procedures, expressing the specific nursing measures that should be taken for a certain disease or symptom, such as the nursing dependency relationship between "acute myocardial infarction" and "bed rest", or the nursing dependency relationship between "edema" and "elevation of the lower limbs".

[0070] Contraindication: Expresses a mutually exclusive relationship between certain drugs, procedures, or conditions, such as the contraindication between "warfarin" and "aspirin" under specific circumstances, or the contraindication between "severe bleeding tendency" and "thrombolytic therapy".

[0071] Monitoring Relationship: Connects disease or nursing procedures with rehabilitation indicators, expressing the indicators that need to be monitored and their target range, such as the "monitoring" relationship between "heart failure" and "daily weight", and includes the frequency attribute of "daily measurement".

[0072] For each relation triple (head entity, relation type, tail entity) extracted in step 1, the system creates a directed edge between the corresponding two nodes (if the relation has directionality, such as a cause-effect relation, it usually points from the cause to the disease). Each edge also contains the following attributes: Relationship type: Clearly indicate which of the above semantic relationships it belongs to.

[0073] Relationship Attributes: For relationships with time constraints or additional conditions, attach these conditions as attributes to the edges. For example, for the "Nursing Dependence" relationship from "Acute Myocardial Infarction" to "Emergency PCI", the attribute "Time Window: Within 12 Hours of Onset" can be attached; for the "Monitoring" relationship from "Warfarin" to "INR Monitoring", the attribute "Frequency: Once a Week" can be attached.

[0074] Initial Confidence Weights: Similar to nodes, each relation edge also needs a confidence weight, representing the degree of credibility of the relation. Its calculation is also based on source authority and timeliness, but it additionally considers the amount of supporting evidence appearing in the text supporting the relation. For example, if the same relation is mentioned in multiple authoritative guidelines, its confidence weight will be increased through a weighted average.

[0075] Timestamp: Also records the creation time and most recent verification time of the relationship, ensuring the timeliness of the relationship is traceable.

[0076] In this way, previously isolated knowledge nodes are connected into a network by rich relational edges, forming a semantic network covering diseases, symptoms, drugs, nursing procedures, and rehabilitation indicators. This network not only reflects static classification knowledge, but more importantly, it captures the dynamic conditions and temporal requirements in clinical decision-making through attribute-based relational edges.

[0077] The core innovation of this invention's knowledge graph lies in its "dynamic nature," which is reflected in the following two aspects: (1) Timestamp-driven knowledge evolution Each node and edge carries its creation time and most recent verification time. As medical knowledge is constantly updated, older knowledge may be revised or obsolete. The system periodically scans knowledge sources (such as newly published clinical guidelines) to extract newly added or changed knowledge metadata. For newly added knowledge, a new node or edge is created directly and assigned the latest timestamp; for new knowledge that conflicts with older knowledge, version management is performed by comparing confidence weights and timestamps—if the new knowledge has a higher confidence level or is more recent, the older knowledge is marked as "pending review" or its weight is reduced, rather than being deleted directly, thus preserving the trajectory of knowledge evolution. This design allows the knowledge graph to continuously absorb the latest research results while retaining historical reference information.

[0078] (2) Dynamic adjustment of confidence weights The initial confidence weights serve only as the starting point for graph construction. In the feedback loop of subsequent step 5, the system iteratively optimizes the confidence weights of relevant nodes and edges based on feedback signals from real clinical adoption behavior and patient outcomes. For example, if a nursing recommendation is frequently adopted in clinical practice and results in a good patient outcome, the confidence weight of that relationship edge will be positively enhanced; conversely, if a piece of knowledge is frequently ignored or leads to adverse outcomes, its weight will be reduced. This mechanism enables the knowledge graph to continuously self-correct, improving the practicality and accuracy of the knowledge.

[0079] To support efficient subsequent reasoning queries, the constructed knowledge graph needs to be stored in a graph-supporting database. This invention preferably uses a native graph database (such as Neo4j or JanusGraph) for persistence. During storage, necessary indexes are created for nodes and edges, such as by entity type, entity name, and confidence weight range, to accelerate subsequent graph traversal and pattern matching operations. Simultaneously, to support time-series queries (such as "querying high-confidence nursing procedures for acute myocardial infarction verified in the last five years"), the database needs to support range retrieval of timestamp attributes.

[0080] S3. Collect static attribute data and dynamic monitoring data of the target patient, construct an individualized nursing profile of the patient, and embed the dynamic monitoring data into the individualized nursing profile of the patient to generate a dynamic user status representation with time-series characteristics.

[0081] Specifically, patient data collection is the foundation for building personalized nursing profiles, and its accuracy, real-time performance, and security directly affect the effectiveness of subsequent reasoning. This invention employs a tiered data collection strategy, balancing data comprehensiveness with the feasibility of data acquisition.

[0082] Static attribute data refers to information that is relatively stable and does not change frequently in the short term, mainly including demographic characteristics and past medical history.

[0083] Data source: Primarily extracted from the hospital's electronic medical records (EMR). Basic patient information is obtained once upon admission or initial registration for nursing services through an authorized data interface.

[0084] The data collected includes basic information such as the patient's age, gender, height, and weight, as well as past medical history (e.g., "5-year history of hypertension", "type 2 diabetes"), surgical history (e.g., "post-coronary artery bypass grafting"), and allergy history (e.g., "aspirin allergy"). This data is stored in a structured field format.

[0085] Data collection method: ETL tools are used to extract new or updated static data from the EMR database periodically (e.g., once a day), and the data is anonymized to remove direct identifiers such as names and ID numbers and replace them with unique patient codes to ensure privacy and security.

[0086] Dynamic monitoring data refers to indicators that change continuously over time and reflect the patient's current physiological state, and is the key to driving personalized dynamic care.

[0087] Data sources: mainly from bedside monitoring equipment, wearable devices (such as smart bracelets and dynamic electrocardiographs), and bedside measurement records from nursing staff.

[0088] Data collected includes: core indicators such as real-time heart rate, blood pressure (systolic / diastolic), blood oxygen saturation, respiratory rate, and body temperature. Depending on the specific needs of cardiac care, more specialized indicators may also be collected, such as cardiac output, central venous pressure (for critically ill patients), or daily activity levels (for patients in the recovery phase).

[0089] Data collection method: High-frequency real-time data: For continuous monitoring indicators such as heart rate, blood pressure, and blood oxygen saturation, data is transmitted in real time to the central monitoring system or IoT gateway via the data output interface of the medical device (such as RS232, Bluetooth, HL7 protocol). The system receives the data at a fixed frequency (such as once per minute or triggered by each measurement) and adds a precise timestamp.

[0090] Low-frequency measurement data: For indicators measured at regular intervals every day (such as weight and urine output), nurses can input the data into a mobile nursing terminal or upload it automatically via a smart scale. The system collects this data in an event-driven manner.

[0091] Data preprocessing: Raw monitoring data often contains noise or outliers. Real-time cleaning is performed at the acquisition end or access layer, such as using median filtering to remove transient interference, setting a reasonable physiological range (e.g., heart rate 30-200 beats / min) to filter obviously erroneous data, and marking missing values ​​to ensure the quality of data entering the profile.

[0092] A patient's personalized care profile is a multi-dimensional, structured data model used to comprehensively describe a patient's health status. Its construction process includes the creation of a static profile and the integration of dynamic states.

[0093] Based on the collected static attribute data, the system first constructs a basic profile, which includes the following core modules: The demographic characteristics module transforms age, sex, height, weight, and other factors into standardized characteristics. For example, age is categorized into "youth," "middle-aged," and "elderly" based on medical classifications; weight is combined with height to calculate the Body Mass Index (BMI) and categorized into "underweight," "normal," "overweight," and "obese" ranges.

[0094] The past medical history module maps the patient's historical diagnoses and surgical records to standard disease and surgical entities in the knowledge graph from step 2. For example, "coronary artery disease" is mapped to the "coronary atherosclerotic heart disease" node in the knowledge graph, and the diagnosis time is recorded. Multiple medical histories may be related, and the system will retain their original diagnostic sequence and the attending physician's diagnostic conclusion.

[0095] Basic risk labels: Based on the above information, some basic risk labels are generated through predefined rules, such as "elderly woman", "history of hypertension", and "history of hyperlipidemia", to provide preliminary context for subsequent reasoning.

[0096] Static profiles are stored in individual patient files as key-value pairs or structured documents and are updated as new information is entered (such as newly diagnosed diseases), but the update frequency is low.

[0097] Dynamic monitoring data is continuously arriving streaming data that requires processing before it can be integrated into the profile. This invention uses a sliding window technique to construct a real-time state vector: For each dynamic metric (such as heart rate), the system maintains a fixed-length time series window (e.g., the past hour). Whenever new data arrives, the window slides forward, discarding the oldest data and adding the latest data.

[0098] Statistical features are extracted from each time series window to form an instantaneous description of the indicator's state. The extracted features include: Current value: The value of the most recent measurement.

[0099] Trend characteristics: such as the slope of change over the past 15 minutes and the degree of deviation from the base value.

[0100] Fluctuation characteristics: such as the variance over the past hour, and the difference between the maximum and minimum values.

[0101] Event characteristics: Whether an over-limit alarm occurs (such as heart rate exceeding 120 beats / min) or a specific event occurs (such as premature beats).

[0102] The above characteristics of all indicators are combined into a multi-dimensional vector called the "real-time state vector". This vector is updated at fixed time intervals (e.g., every 5 minutes) and reflects the patient's physiological dynamics at the current moment.

[0103] This invention effectively fuses real-time state vectors with static profiles to generate a dynamic user state representation that reflects both the patient's inherent characteristics and captures current transient changes. Since real-time state vectors and static profile features differ in dimensionality and semantics, alignment and fusion are necessary. This invention employs a hierarchical fusion strategy: Base layer fusion: The categorical features (such as age segmentation and gender) in the static image are one-hot encoded or embedded encoded to transform them into numerical feature vectors. At the same time, the continuous values ​​in the real-time state vector are normalized to make them distributed on the same order of magnitude.

[0104] Temporal context injection: Considering that many decisions in cardiac care depend on temporal context (e.g., "day after surgery", "how long after medication"), the system calculates the time offset of the patient's current moment relative to key events (e.g., surgery date, admission date) and adds it as an additional feature to the real-time state vector. For example, generating the temporal feature of "day 3 after surgery".

[0105] Feature concatenation: The processed static feature vector, real-time state vector, and temporal context features are concatenated into a long vector to form the original fused features.

[0106] Simple vector concatenation cannot fully capture the complex interactions between static and dynamic features, nor can it reflect the temporal patterns in dynamic data. Therefore, this invention introduces a lightweight temporal coding module to deeply abstract the fused features and generate the final dynamic user state representation.

[0107] This module is built upon the encoder portion of a recurrent neural network or Transformer, but its specific network structure will not be elaborated here. Instead, its function will be described: Capturing historical dependencies: The module receives a fused feature sequence from multiple consecutive time steps as input and, through its internal memory unit, can capture patterns in the evolution of the patient's state over time. For example, it can identify trends such as "heart rate has been rising continuously over the past half hour," rather than just the current heart rate value.

[0108] Strengthening key features: Through an attention mechanism, the module automatically learns which static features (such as age) and dynamic features (such as decreased blood oxygen saturation) are more important to nursing decisions at the current moment and assigns them higher weights. For example, for an elderly patient with heart failure, a slight decrease in blood oxygen saturation may be more clinically significant than a small fluctuation in heart rate, and the module will highlight this feature in the representation.

[0109] Output state representation: The module outputs a fixed-dimensional vector for the current moment, namely the "dynamic user state representation". This vector is a high-dimensional semantic representation that implicitly encodes the patient's identity, medical history, real-time physiological parameters, and recent trends.

[0110] The generated dynamic user state representation has the following characteristics: Comprehensive: It integrates static demographics, medical history information, and multimodal real-time monitoring data to comprehensively reflect the patient's condition.

[0111] Time sensitivity: Through sliding window and time-series coding, it incorporates recent trends and can distinguish between "stable mild anomalies" and "rapidly deteriorating mild anomalies".

[0112] Dynamic updates: As new monitoring data arrives, the system will be periodically refreshed to ensure it always reflects the patient's latest condition. The update frequency can be adjusted according to clinical needs (e.g., updates every minute for ICU patients, and every hour for general ward patients).

[0113] This representation will serve as the query input for the inference adaptation engine in step 4. The engine will calculate the semantic matching degree between this representation and candidate knowledge nodes in the knowledge graph, and then select the most suitable nursing recommendations for the current patient's condition. For example, for a patient whose dynamic user condition representation shows prominent features of "rapid increase in heart rate and decrease in blood oxygen," the engine will prioritize recommending nursing measures related to acute heart failure exacerbation; while for a stable patient in the recovery period, it may recommend rehabilitation training guidance.

[0114] S4. Input the dynamic user state representation into the pre-trained reasoning adaptation engine, perform reasoning on the dynamic knowledge graph of cardiac care, calculate the matching degree of each candidate knowledge node according to the dynamic user state representation, and filter out knowledge nodes with matching degree exceeding a preset threshold to generate a personalized nursing knowledge set.

[0115] Specifically, the reasoning adaptation engine is an intelligent reasoning model built on a graph neural network. Its core function is to receive dynamic user state representations as queries, perform semantic matching and path reasoning on a dynamic knowledge graph of cardiac nursing, and finally output a set of nursing knowledge nodes most relevant to the current patient's condition. The engine is designed to understand the deeper meaning of the patient's state and, by utilizing the rich entity associations and temporal attributes in the knowledge graph, simulate the reasoning process of clinical experts to generate truly personalized nursing recommendations.

[0116] The inference adaptation engine adopts an encoder-decoder architecture based on graph neural networks, which mainly consists of three core modules: graph encoding module, state-graph interaction module, and candidate node scoring module.

[0117] (1) Graph Coding Module This module is responsible for converting the nodes and edges in the dynamic knowledge graph of cardiac care into computer-processable vector representations, i.e., graph embedding. Specifically: For each entity node in the graph (such as disease, drug, nursing procedure, etc.), the graph coding module generates a vector that reflects the semantics and contextual relationships of the node by aggregating information from its neighboring nodes. This vector is called the "node embedding vector". For example, the embedding vector of the "aspirin" node will incorporate information from the "coronary heart disease" node connected by its "medication relationship" and the "bleeding tendency" node connected by its "contraindication relationship", thereby representing its multi-dimensional clinical meaning in the vector space.

[0118] At the same time, for different types of edges (semantic relations), the module also learns a "relation embedding vector" for each relation type, which is used to distinguish the importance of different relations when aggregating neighbor information.

[0119] The graph encoding module propagates information layer by layer through multi-layer graph convolution or graph attention mechanisms, ensuring that the final embedding of each node not only includes its own attributes but also incorporates the topological structure information of its multi-hop neighbors. This process is completed during the engine training phase. The trained module can encode the entire knowledge graph into a high-dimensional vector space, where semantically similar nodes are closer together in the space.

[0120] (2) Status-Graph Interaction Module This module is responsible for interacting with the dynamic user state representation and the node embedding vectors generated by the graph coding module, achieving the alignment and fusion of patient state and graph knowledge. Its working principle is as follows: First, the dynamic user state representation (a high-dimensional vector) is mapped to the same semantic space as the node embedding vector through a linear transformation layer to obtain the "query vector".

[0121] Then, an attention mechanism is used to calculate the relevance score between the query vector and the embedding vector of each candidate node in the graph. Specifically, for each node in the graph, the similarity between its embedding vector and the query vector is calculated. This similarity reflects the degree to which the concept represented by the node matches the patient's current state. For example, if the patient's state representation highlights "dyspnea" and "edema," then the embedding vector of the "heart failure" node will show a high similarity to the query vector.

[0122] The attention mechanism also considers the path relationships between nodes. For example, if the "heart failure" node is connected to the "fluid restriction" node through "care dependency," then when the patient's state highly matches "heart failure," the engine will pass attention through the graph structure, indirectly improving the relevance score of the "fluid restriction" node. This allows the engine to perform multi-hop reasoning and discover potential connections between the patient's state and indirectly related care knowledge.

[0123] (3) Candidate node scoring module This module uses the attention score output by the state-graph interaction module to comprehensively score each candidate knowledge node and generate the final matching degree. The scoring comprehensively considers the following factors: Attention similarity: the direct similarity between node embeddings and query vectors.

[0124] Path cumulative confidence: The product of the confidence weights of all relation edges traversed from the core node most relevant to the patient's state (such as the disease node) to the current candidate node. This ensures that the recommended knowledge is not only semantically relevant, but also that the reasoning path it relies on is itself highly reliable.

[0125] Node self-confidence: The initial confidence weight of the candidate node itself, which comes from the construction in step 2.

[0126] Temporal fit: Whether the temporal attribute associated with the candidate node (e.g., "within 24 hours post-surgery") matches the patient's current temporal context (e.g., "day 3 post-surgery"). If they match, points are awarded; if they do not match, points may be reduced or the candidate may be excluded.

[0127] Ultimately, each candidate knowledge node receives a matching score between 0 and 1, representing the applicability of the knowledge to the current patient.

[0128] The inference adaptation engine needs to be fully trained to have accurate reasoning capabilities. The training process is divided into two stages: pre-training and fine-tuning.

[0129] (1) Pre-training phase The goal of pre-training is to enable the graph encoding module to learn high-quality node embedding vectors that reflect the semantic structure of the knowledge graph. A graph autoencoder-based approach is employed. Training data: The structure of the cardiac care dynamic knowledge graph itself, including all nodes and edges.

[0130] Training task: Randomly mask a portion of the edges in the graph (e.g., randomly remove some "medication relationships"), and have the model predict whether the masked edges exist. The model needs to determine the most likely relationship type between two nodes based on their embedding vectors. By minimizing the prediction error, the model learns how to cluster similar nodes together and how to infer missing relationships based on their neighbors.

[0131] Unsupervised characteristics: No manual annotation is required in the pre-training stage. It fully utilizes the structural information of the graph, enabling node embeddings to capture the inherent connections of domain knowledge.

[0132] (2) Fine-tuning stage The goal of fine-tuning is to enable the entire engine (including the state-graph interaction module and the scoring module) to learn to accurately recommend nursing knowledge based on dynamic user state representations.

[0133] Training data: Historical case data are collected. Each training sample includes: a patient's dynamic user state at a certain point in time, the nursing knowledge actually adopted by medical staff at that moment (as a positive example), and other candidate knowledge that was not adopted (as a negative example). Positive examples can be obtained by retrospectively analyzing nursing records in electronic medical records, while negative examples are obtained by randomly sampling knowledge that is irrelevant to the patient's state from the graph.

[0134] Training task: Input the patient state representation into the engine, and the engine calculates a matching score for all candidate knowledge nodes. The training objective is to maximize the matching score of positive nodes and minimize the matching score of negative nodes. Optimization is achieved using a ranking loss function (such as margin ranking loss) so that the score of positive examples is at least a certain threshold higher than that of negative examples.

[0135] Feedback signal integration: During the training process, the clinical adoption feedback collected in step 5 can be combined to use the actual adoption behavior as a reinforcement signal to further adjust the model parameters and make the engine closer to the real clinical decision-making mode.

[0136] Through the two-stage training described above, the inference adaptation engine can combine abstract graph structures with specific patient states to achieve accurate personalized inference.

[0137] In actual reasoning, it is impossible for the engine to score every node in the entire knowledge graph individually, as the graph can be very large and the computational cost would be too high. Therefore, this invention designs a two-stage screening strategy: first, it quickly narrows down the candidate range, and then it performs a refined scoring.

[0138] (1) Preliminary screening Based on key features in dynamic user state representation, entities directly related to the patient's current condition are quickly located from the knowledge graph as "seed nodes." Specific methods include: Entity matching: Matches the entities explicitly contained in the patient's status representation (such as diseases in past medical history and descriptions of current symptoms) with the entity names in the graph, directly extracting the corresponding nodes. For example, if the patient's status representation contains "history of hypertension", then the "hypertension" node in the graph is selected as the seed node.

[0139] Tag retrieval: Utilizing risk tags (such as "elderly" or "post-operative") implicit in the state representation, nodes or related subgraphs with these tags are retrieved in the atlas. For example, all nursing operation nodes related to "elderly heart failure" can be retrieved.

[0140] Scope limitation: Based on the patient's department (such as cardiology ICU, general ward) or treatment stage, a set of related knowledge subgraphs (such as "ICU nursing" "rehabilitation nursing") are predefined to limit the scope of reasoning to these subgraphs, thereby reducing the amount of computation.

[0141] After initial screening, a set of hundreds of candidate knowledge nodes was obtained, which are macroscopically correlated with the patient's condition.

[0142] (2) Generation of refined candidate sets The seed nodes obtained from the initial screening, along with their neighbor nodes within two to three hops in the knowledge graph, are included in the candidate set. This is because knowledge directly related to patients is often indirectly connected through multi-hop relationships. For example, the seed node is "heart failure," and its two-hop neighbors include "daily weight monitoring" connected through "care dependency" and "diuretics" connected through "medication relationship." These nodes are then collectively considered as candidate knowledge nodes and input into the subsequent scoring module.

[0143] For each knowledge node in the candidate set, the engine performs the following steps to calculate its matching degree: Step A: Obtain node embedding Extract the pre-trained embedding vector of the node from the graph coding module. This vector contains the semantics of the node itself and its topological information in the graph.

[0144] Step B: Calculate state-node similarity A basic similarity score is obtained by calculating the dot product or cosine similarity between the dynamic user state representation (query vector) and the embedding vector of the node. The higher the score, the more closely the concept represented by the node matches the patient's current state.

[0145] Step C: Calculate path confidence Starting from the seed node, the engine uses a graph attention mechanism to calculate the cumulative confidence of all possible paths from the seed node to the current candidate node. Specifically, the engine calculates the product of the confidence weights of all relation edges on each path and takes the maximum value or weighted sum as the path confidence score. This step ensures that the recommended knowledge is supported by a reliable reasoning chain.

[0146] Step D: Timing Constraint Check Check if candidate nodes and their associated edges have time attributes (such as "immediately after surgery" or "once daily"). If so, obtain the patient's current time context (such as "day after surgery" or "current time") and determine if the time condition is met. If it is met, award a time-series fit bonus; if it is not met, deduct points based on the degree of mismatch, or even eliminate the node entirely (for example, if an operation requires execution within "24 hours after surgery," but the patient has been post-surgery for 72 hours, then this knowledge should not be recommended).

[0147] Step E: Overall Scoring The basic similarity score, path confidence score, node self-confidence weight, and temporal adaptation bonus are weighted and fused to obtain the final matching score. The fusion weights can be learned through a fine-tuning stage to ensure the model achieves optimal performance on the validation set.

[0148] To filter out truly applicable nursing recommendations from candidate knowledge nodes, a matching threshold needs to be set. Only nodes with a matching degree exceeding this threshold will be output to healthcare professionals.

[0149] The preset threshold is set based on the following criteria: Initialization based on experience: The initial threshold was set between 0.6 and 0.7, based on the minimum standard of "acceptable" recommended knowledge set by clinical experts. This range was based on simulation testing to ensure that the selected knowledge was sufficient in quantity but not excessively irrelevant.

[0150] Dynamic adjustment: The threshold can be fine-tuned according to the actual application scenario. For example, in high-risk environments such as the ICU, in order not to miss any potentially important prompts, the threshold can be appropriately lowered (e.g., 0.5) to expand the recommendation range, with medical staff making the final judgment; in general wards or during the recovery period, in order to reduce information overload, the threshold can be increased (e.g., 0.75) to push only knowledge with high confidence.

[0151] Learning optimization: In the feedback loop of step 5, the system can automatically adjust the threshold based on the adoption rate of medical staff. If the adoption rate is too low (indicating that too much irrelevant knowledge has been recommended), the threshold will be automatically increased; if key knowledge is found to have been missed (through post-analysis), the threshold will be appropriately decreased.

[0152] Typically, the preset threshold is kept at around 0.65 as a benchmark value for balancing recommendation accuracy and recall.

[0153] S5. Collect feedback signals on the personalized nursing knowledge set, and use the feedback signals to update the initial confidence weights and the parameters of the inference adaptation engine.

[0154] Specifically, the personalized nursing knowledge set is pushed to the terminal device, and data on the adoption behavior of medical staff and the subsequent changes in the patient's physiological indicators are collected. Based on the adoption behavior data and the changes in physiological indicators, a feedback signal is generated, and the initial confidence weight in step S2 and the inference adaptation engine parameters in step S4 are updated using the feedback signal to form a closed-loop optimization. When the adoption behavior data is positive and the subsequent changes in physiological indicators are improving, the confidence weight of the set nodes and edges is increased, and the parameters of the inference adaptation engine are adjusted positively. When the adoption behavior data is negative or the subsequent changes in physiological indicators are worsening, the confidence weight of the set nodes and edges is decreased, and the parameters of the inference adaptation engine are adjusted negatively.

[0155] The method for constructing a cardiac care knowledge base provided by this invention has the following beneficial effects: 1. Improve knowledge relevance and retrieval accuracy: By constructing a dynamic knowledge graph for cardiac care, multi-source heterogeneous data is transformed into a semantically related entity network, overcoming the shortcomings of traditional relational databases in expressing complex relationships. This enables knowledge retrieval to be based on semantic relationships for reasoning, significantly improving the accuracy and recall of knowledge matching.

[0156] 2. Achieve personalized dynamic nursing plan recommendations: By embedding real-time dynamic monitoring data of patients into user profiles and combining it with dynamic reasoning on the knowledge graph by the reasoning and adaptation engine, the generated knowledge can respond in real time to changes in the patient's physiological state, realizing the leap from "static query" to "dynamic adaptation" and providing patients with accurate personalized nursing decision support.

[0157] 3. Construct a knowledge self-optimization mechanism: By introducing a feedback loop based on clinical adoption behavior and patient outcomes, the confidence weights of the knowledge graph and the parameters of the inference engine are dynamically adjusted, enabling the knowledge base to automatically optimize according to actual clinical effects, eliminating outdated or ineffective nursing information, and ensuring the timeliness and scientific nature of the knowledge base.

[0158] The second embodiment of this application is as follows: Please see Figure 4 This invention provides a system for constructing a cardiac care knowledge base, applied to a method for constructing a cardiac care knowledge base as provided in the first embodiment, comprising: The data acquisition and preprocessing module 101 is used to acquire multi-source heterogeneous data related to cardiac care, preprocess the multi-source heterogeneous data, and generate standardized knowledge metadata. The dynamic knowledge graph construction module 102 is used to construct a dynamic knowledge graph for cardiac care based on the standardized knowledge metadata, with entities in the cardiac care domain as nodes and semantic relationships between entities as edges; the nodes have initial confidence weights and timestamp attributes. The user profile and state embedding module 103 is used to collect static attribute data and dynamic monitoring data of the target patient, construct a personalized nursing profile of the patient, and embed the dynamic monitoring data into the personalized nursing profile of the patient to generate a dynamic user state representation with time-series characteristics. The reasoning adaptation and knowledge generation module 104 is used to input the dynamic user state representation into the pre-trained reasoning adaptation engine, perform reasoning on the dynamic knowledge graph of cardiac care, calculate the matching degree of each candidate knowledge node according to the dynamic user state representation, and filter out knowledge nodes with matching degree exceeding a preset threshold to generate a personalized nursing knowledge set. The dynamic update and feedback module 105 is used to collect feedback signals on the personalized nursing knowledge set and use the feedback signals to update the initial confidence weights and the parameters of the inference adaptation engine.

[0159] Regarding the system in the above embodiments, the specific ways in which each module performs operations have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0160] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0161] Accordingly, this application also provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the method for constructing a cardiac care knowledge base as described above. Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, used in a system for constructing a cardiac care knowledge base according to an embodiment of the present invention. (Except for...) Figure 5 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0162] Accordingly, this application also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the method for constructing a cardiac care knowledge base as described above. The computer-readable storage medium can be an internal storage unit of any data-processing device as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data-processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data-processing device, and can also be used to temporarily store data that has been output or will be output.

[0163] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0164] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for constructing a knowledge base for cardiac care, characterized in that, Includes the following steps: Collect multi-source heterogeneous data related to cardiac care, preprocess the multi-source heterogeneous data, and generate standardized knowledge metadata; Based on the standardized knowledge metadata, a dynamic knowledge graph for cardiac care is constructed, with entities in the field of cardiac care as nodes and semantic relationships between entities as edges. The node has an initial confidence weight and a timestamp attribute; Collect static attribute data and dynamic monitoring data of the target patient, construct a personalized nursing profile of the patient, and embed the dynamic monitoring data into the personalized nursing profile of the patient to generate a dynamic user status representation with time-series characteristics; The dynamic user state representation is input into a pre-trained reasoning adaptation engine, and reasoning is performed on the dynamic knowledge graph of cardiac care. The matching degree of each candidate knowledge node is calculated based on the dynamic user state representation, and knowledge nodes with matching degrees exceeding a preset threshold are selected to generate a personalized nursing knowledge set. Feedback signals to the personalized care knowledge set are collected, and the initial confidence weights and parameters of the inference adaptation engine are updated using the feedback signals.

2. The method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, The multi-source heterogeneous data is preprocessed to generate standardized knowledge metadata, including: Natural language processing is performed on the unstructured and semi-structured data in the multi-source heterogeneous data. The natural language processing includes entity recognition based on a dictionary specific to the cardiac care domain, and relation extraction based on temporal relationships. The identified entities are mapped to a standard medical terminology set, the extracted relationships are mapped to a predefined relation ontology library, and each entity and relationship is marked with its data source and timestamp to generate the standardized knowledge metadata.

3. The method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, Based on the standardized knowledge metadata, a dynamic knowledge graph for cardiac care is constructed, using entities in the cardiac care domain as nodes and semantic relationships between entities as edges, including: The entities in the standardized knowledge metadata are transformed into nodes in the graph, and each node is assigned an initial confidence weight, and the creation time and the most recent verification time are recorded. The relationships in the standardized knowledge metadata are transformed into edges connecting nodes in a graph, and each edge is assigned a relationship type, relationship attribute, and initial confidence weight, and the creation time and the most recent verification time are recorded. The relational attributes include time constraint attributes.

4. The method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, Collect static attribute data and dynamic monitoring data of the target patient to construct a personalized patient care profile. Embed the dynamic monitoring data into the personalized patient care profile to generate a dynamic user status representation with time-series characteristics, including: Collect static attribute data and dynamic monitoring data of the target patients; A static profile containing demographic features and past medical history is constructed based on the aforementioned static attribute data; Based on the dynamic monitoring data, a real-time state vector is extracted, which includes the current value, trend characteristics, and fluctuation characteristics. The features of the static profile are fused with the real-time state vector, and time context information is injected to generate the dynamic user state representation through a time-series coding module.

5. The method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, The inference adaptation engine includes a graph coding module, a state graph interaction module, and a candidate node scoring module. The graph encoding module is used to encode the nodes and edges in the cardiac care dynamic knowledge graph into node embedding vectors and relation embedding vectors. The state graph interaction module is used to calculate the attention similarity between the dynamic user state representation and the node embedding vector; The candidate node scoring module is used to calculate the matching degree based on the attention similarity, the cumulative confidence of the path from the seed node to the candidate node, and the temporal adaptability of the candidate node.

6. The method for constructing a cardiac care knowledge base as described in claim 5, characterized in that, The inference adaptation engine is obtained through a two-stage training process of pre-training and fine-tuning. The pre-training phase utilizes the structural information of the cardiac care dynamic knowledge graph to learn the node embedding vector and relation embedding vector by predicting the masked edges; The fine-tuning phase utilizes historical case data, taking the actually adopted nursing knowledge as positive examples, to optimize the calculation of the matching degree.

7. The method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, Before calculating the matching degree of each candidate knowledge node based on the dynamic user state representation, the method further includes: Seed nodes are matched from the cardiac care dynamic knowledge graph based on the key features in the dynamic user state representation. The seed node and its neighboring nodes within a preset hop count range are selected as candidate knowledge nodes.

8. The method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, The preset threshold is set between 0.6 and 0.75 and is dynamically adjusted according to the adoption rate of clinical scenarios or feedback signals.

9. The method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, Collecting feedback signals on the personalized care knowledge set, and using the feedback signals to update the initial confidence weights and the parameters of the inference adaptation engine, including: The feedback signals to the personalized nursing knowledge set are collected, wherein the feedback signals include data on the adoption behavior of medical staff in relation to the personalized nursing knowledge set and data on subsequent changes in the patient's physiological indicators. When the adoption behavior data is positive and the subsequent physiological indicator change data is positive, the confidence weight of the set nodes and edges is enhanced, and the parameters of the inference adaptation engine are adjusted positively. When the adoption behavior data is rejection or the subsequent physiological indicator change data is deterioration, the confidence weight of the set node and edge is reduced, and the inference adaptation engine is negatively adjusted.

10. A system for constructing a cardiac care knowledge base, applied to the method for constructing a cardiac care knowledge base as described in claim 1, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous data related to cardiac care, preprocess the multi-source heterogeneous data, and generate standardized knowledge metadata. The dynamic knowledge graph construction module is used to construct a dynamic knowledge graph for cardiac care based on the standardized knowledge metadata, with entities in the cardiac care domain as nodes and semantic relationships between entities as edges. The node has an initial confidence weight and a timestamp attribute; The user profile and status embedding module is used to collect static attribute data and dynamic monitoring data of the target patient, construct a personalized nursing profile of the patient, and embed the dynamic monitoring data into the personalized nursing profile of the patient to generate a dynamic user status representation with time-series characteristics. The reasoning adaptation and knowledge generation module is used to input the dynamic user state representation into the pre-trained reasoning adaptation engine, perform reasoning on the dynamic knowledge graph of cardiac care, calculate the matching degree of each candidate knowledge node according to the dynamic user state representation, and filter out knowledge nodes with matching degree exceeding a preset threshold to generate a personalized nursing knowledge set. The dynamic update and feedback module is used to collect feedback signals on the personalized care knowledge set and use the feedback signals to update the initial confidence weights and the parameters of the inference adaptation engine.

Images