A method and system for collecting vital sign monitoring data for a health online platform

Abstract

The present application relates to a kind of health online platform's sign monitoring data acquisition method and system, belong to health online platform data acquisition technical field, the present application is by the data field to be mapped as query node input knowledge graph, adopts graph neural network and carries out multi-hop inference, finds the most matched standard concept node in knowledge graph in semantics, forms mapping model, and utilizes mapping model output mapping result;Semantics mapping engine front deployment is positioned at data acquisition entrance, cooperates three-tier cache strategy, realizes that data flow is completed standardized online processing capability, fundamentally avoids the non-standard data into storage layer after cleaning management again pollution after management type data quality predicament, guarantees that platform data has consistency, standardization and usability from source head.Just by standardized data source semantic description model and dynamic registration mechanism, the access configuration time of new data source is shortened from the several days even several weeks required in traditional scheme.

Description

Technical Field

[0001] This invention relates to the field of technology, and in particular to a method and system for collecting vital sign monitoring data on an online health platform. Background Technology

[0002] With the rapid development of wearable devices, IoT technology, and artificial intelligence, online health platforms have become an important infrastructure for personal health management, chronic disease monitoring, and telemedicine. Currently, the sources of health data collection are increasingly diverse, encompassing various terminal devices such as smartwatches, dynamic electrocardiogram recorders, continuous glucose monitors, smart blood pressure monitors, body fat scales, and environmental sensors, enabling real-time acquisition of multimodal health data including heart rate, blood oxygen, blood pressure, blood glucose, body movement, sleep, and ambient temperature and humidity.

[0003] However, existing health data acquisition technologies generally adopt a "three-fixed" approach: fixed sampling frequency, fixed sensor combination, and fixed transmission strategy. This rigid acquisition strategy has significant technical drawbacks: on the one hand, during periods of stable user health, high-frequency acquisition leads to excessive device power consumption, redundant data storage, and wasted data transmission bandwidth, severely limiting the battery life of wearable devices and the efficiency of system resource utilization; on the other hand, when user health fluctuates or is in specific risk scenarios, fixed low-frequency sampling cannot capture the dynamic changes in key physiological parameters, resulting in the loss of important health risk characteristic signals and delaying the early detection of potential health risks.

[0004] Therefore, health data comes from diverse sources, and different data sources employ different data standards, coding systems, and semantic expressions—for example, the same clinical indicator may have different names, units, and value ranges in different systems. Existing data standardization schemes mostly rely on manually configuring mapping rules, which is inefficient and lacks scalability. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and system for collecting vital sign monitoring data on an online health platform.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for collecting vital sign monitoring data on an online health platform, comprising the following steps: Integrate international standards, clinical guidelines, medical literature, and expert knowledge to construct a health domain knowledge graph that includes entities, attributes, and relationships; For each data source accessed in the online health platform, a data source semantic description model is constructed based on the data source metadata, data field semantic descriptions, and the known mapping relationship between the data source's unique coding system and the standard coding system. The data field to be mapped is input into the knowledge graph as a query node. A graph neural network is used for multi-hop reasoning to find the standard concept node with the best semantic match in the knowledge graph, forming a mapping model. The mapping result is then output using the mapping model. A semantic mapping engine is deployed at the data collection entry point to perform real-time semantic parsing and standardized mapping on the collected raw data fields, and to convert heterogeneous data into a standard format before storing it in the online health platform. A confidence score is output for each mapping, and the semantic description model of the data source is continuously optimized based on the confidence score.

[0007] Furthermore, in the data collection methods for vital sign monitoring on online health platforms, international standards, clinical guidelines, medical literature, and expert knowledge are integrated to construct a health domain knowledge graph that includes entities, attributes, and relationships. Specifically: Integrate international standards, clinical guidelines, medical literature, and expert knowledge to define core entity types and relationship types; For different expressions of the same concept in international standards, clinical guidelines, medical literature and expert knowledge, UMLSCUI can achieve cross-terminology system normalization and complete terminology level alignment. Referring to the BiolinkModel standardized predicate system, relation types from different sources are mapped to a unified set of relation predicates to complete the structural layer alignment. Entity disambiguation is performed by combining the SapBERT model with the edit distance algorithm to complete the instance layer alignment. After alignment at the terminology layer, structure layer, and instance layer, a health domain knowledge graph containing entities, attributes, and relationships is constructed. At the same time, the semantic embedding of entities is generated using the SapBERT model, and synonymous entities are identified through vector similarity. Merge and simplify semantically equivalent or approximate relational paths, identify and retain key paths in knowledge propagation based on centrality indicators in graph theory, and remove edge noise.

[0008] Furthermore, in the data collection method for vital sign monitoring on the online health platform, for each data source accessed by the online health platform, a data source semantic description model is constructed based on the data source metadata, data field semantic descriptions, and the known mapping relationship between the data source's unique coding system and the standard coding system. Specifically, this includes: For each data source connected to the online health platform, a metadata layer is constructed based on the data source name, the type of preset device, the version number, and the access time. Based on each data field in the data source, record the field name, data type, semantic label, value range, unit, and sampling frequency to construct a field semantic layer, and record the known partial mapping relationship between the data source's unique local encoding system and the standard encoding system to form an encoding mapping layer; Record the scene information of data collection and the relevant context information of the population characteristics to form a context layer, and establish a standardized semantic description model based on the meta-information layer, field semantic layer, encoding mapping layer and context layer; When a new data source is connected to the platform, the data source submits a DSSP through a standardized interface, inserts the field description information in the DSSP as a new node into the knowledge graph, establishes the data source entity, and associates it with the data source node through relevant relation types, and automatically finds the best standard concept mapping for each field of the current data source. After the mapping results are evaluated for confidence, if the confidence level meets the standard, the predetermined relationship type is confirmed and established; otherwise, it is transferred to the manual review queue.

[0009] Furthermore, in the data collection method for vital sign monitoring on the online health platform, the data field to be mapped is input into the knowledge graph as a query node. A graph neural network is used for multi-hop reasoning to find the semantically most matching standard concept node in the knowledge graph, forming a mapping model. The mapping model is then used to output the mapping result, specifically: A mapping model is constructed, and the text describing the data source fields to be mapped is encoded into semantic vectors using the mapping model. The cosine similarity between semantic vectors is calculated, and the Top-K candidate standard concepts are recalled in the knowledge graph based on the cosine similarity of the semantic vectors. The first phase recall of a preset number of candidate concepts and their multi-hop neighborhood subgraphs in the knowledge graph are extracted, a candidate subgraph is constructed, a graph isomorphism transformer model is used to sort the candidate concepts, and the semantic vector, node type encoding, node degree and other features of the candidate nodes are fused into the initial node representation. The aggregation strategy of graph isomorphic network is adopted to aggregate the neighborhood information of each node in multiple rounds to capture the structural role information of the node. Self-attention calculation is performed on all candidate subgraph nodes in the global scope to capture the long-range dependencies between nodes. The semantic and structural features are adaptively weighted and fused using a gated aggregator to generate the final representation vector of each candidate concept. The matching score between the representation of the field to be mapped and the representation of each candidate concept is calculated. The candidate concepts with a preset ranking and their matching scores are output. The candidate concepts with a preset ranking and their matching scores are then scored a second time, and the final mapping result and confidence score are output. Finally, the mapping model is trained by comparing positive and negative samples.

[0010] Furthermore, in the data collection method for vital sign monitoring on the online health platform, a semantic mapping engine is deployed at the data collection entry point to perform real-time semantic parsing and standardized mapping on the collected raw data fields. This converts heterogeneous data into a standardized format before storing it on the online health platform. Specifically: A semantic mapping engine is deployed at the data acquisition entry point. The semantic mapping engine is used to parse the field names, data types and values ​​in the incoming data packets, and to query the local LRU cache. If the field already has a mapping record, it is returned directly. For fields that are not cached, the mapping model is invoked for real-time inference, and the original data fields are converted into standardized concept codes and standardized values ​​based on the mapping results. The standardized data is then encapsulated into standardized health events in a unified format. The system collects standardized health events in a unified format and distinguishes between high-frequency mapping pairs and full mapping pairs based on the historical number of each standardized health event. High-frequency mapping pairs are cached locally, while low-frequency mapping pairs are cached in a distributed manner. At the same time, the knowledge graph is used to store authoritative records of all mapping relationships.

[0011] Furthermore, in the data collection method for vital sign monitoring on the online health platform, a confidence score is output for each mapping, and the semantic description model of the data source is continuously optimized based on the confidence score, specifically as follows: For each mapping output, a comprehensive confidence score is calculated by considering factors such as model confidence, contextual consistency, historical consistency, and knowledge graph density, and hierarchical processing is performed based on the comprehensive confidence score. Manually confirmed mapping results are added to the training data as positive samples, and manually corrected mappings are added to the training dataset as hard negative samples or corrected samples to adjust the decision boundary of the model. Meanwhile, the mapping model is incrementally retrained regularly using the expanded training set, the semantic description model of the data source is continuously optimized, and the concept addition process of the knowledge graph is triggered when the case is marked as a new concept, the new concept and its position and relationship in the knowledge graph are defined, and the dynamic expansion of the knowledge graph is completed.

[0012] Furthermore, the methods for collecting vital sign monitoring data on online health platforms also include: All standardized health events are organized according to user and time dimensions to form a time-series event sequence for each user. In addition to the defined standard fields, each event is given derived time-series features such as event interval, trend label, and anomaly marker. For heterogeneous data from different data sources, we implement time granularity unification, missing value handling, time alignment window, and cross-modal time alignment for time-series relationship reasoning. We also map each standardized health event to a node in a knowledge graph and establish the association between events and standard concepts. The knowledge graph embedding method is used to embed event nodes and concept nodes into a unified low-dimensional vector space to construct a vector index, and the vector index is used to guide fast retrieval based on semantic similarity.

[0013] A second aspect of the present invention provides a vital sign monitoring data acquisition system for an online health platform, comprising a memory and a processor. The memory includes a vital sign monitoring data acquisition method program for the online health platform. When the processor executes the vital sign monitoring data acquisition method program for the online health platform, it implements the steps of the vital sign monitoring data acquisition method for the online health platform as described in any one of the claims.

[0014] This invention addresses the shortcomings of the prior art and has the following beneficial effects: This invention, by deploying a semantic mapping engine at the data acquisition entry point and employing a three-layer caching strategy, achieves standardized online processing capabilities as data flows in. This fundamentally avoids the "contaminate first, then clean up" data quality dilemma of non-standard data entering the storage layer and then undergoing cleaning and treatment, ensuring that platform data possesses consistency, standardization, and availability from the source. Furthermore, through a standardized data source semantic description model and dynamic registration mechanism, the configuration time for new data sources is shortened from several days or even weeks required by traditional solutions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0016] Figure 1 The overall flowchart of the vital sign monitoring data collection method of the online health platform is shown; Figure 2 The system block diagram of the vital signs monitoring data acquisition system of the online health platform is shown. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0019] like Figure 1As shown, the first aspect of the present invention provides a method for collecting vital sign monitoring data on an online health platform, comprising the following steps: Integrate international standards, clinical guidelines, medical literature, and expert knowledge to construct a health domain knowledge graph that includes entities, attributes, and relationships; For each data source accessed in the online health platform, a data source semantic description model is constructed based on the data source metadata, data field semantic descriptions, and the known mapping relationship between the data source's unique coding system and the standard coding system. The data field to be mapped is input into the knowledge graph as a query node. A graph neural network is used for multi-hop reasoning to find the standard concept node with the best semantic match in the knowledge graph, forming a mapping model. The mapping model is then used to output the mapping result. A semantic mapping engine is deployed at the data collection entry point to perform real-time semantic parsing and standardized mapping on the collected raw data fields, and to convert heterogeneous data into a standard format before storing it in the online health platform. A confidence score is output for each mapping, and the semantic description model of the data source is continuously optimized based on the confidence score.

[0020] It should be noted that this invention, by deploying the semantic mapping engine at the data acquisition entry point and employing a three-layer caching strategy, achieves standardized online processing capabilities as data flows in. This fundamentally avoids the "contaminate first, then clean up" data quality dilemma of non-standard data entering the storage layer and then undergoing cleaning and treatment, ensuring that platform data possesses consistency, standardization, and usability from the source. Furthermore, through a standardized Data Source Semantic Description Model (DSSP) and dynamic registration mechanism, the configuration time for accessing new data sources is shortened from several days or even weeks required by traditional solutions. More importantly, the "map once, reuse everywhere" knowledge transfer mechanism allows established mapping knowledge to be automatically reused across different data sources. The more data sources the system accesses, the faster the adaptation speed to new data sources, demonstrating a significant scale effect.

[0021] Furthermore, in the data collection methods for vital sign monitoring on online health platforms, international standards, clinical guidelines, medical literature, and expert knowledge are integrated to construct a health domain knowledge graph that includes entities, attributes, and relationships. Specifically: Integrate international standards, clinical guidelines, medical literature, and expert knowledge to define core entity types and relationship types; For different expressions of the same concept in international standards, clinical guidelines, medical literature and expert knowledge, UMLSCUI can achieve cross-terminology system normalization and complete terminology level alignment. Referring to the BiolinkModel standardized predicate system, relation types from different sources are mapped to a unified set of relation predicates to complete the structural layer alignment. Entity disambiguation is performed by combining the SapBERT model with the edit distance algorithm to complete the instance layer alignment. After alignment at the terminology layer, structure layer, and instance layer, a health domain knowledge graph containing entities, attributes, and relationships is constructed. At the same time, the semantic embedding of entities is generated using the SapBERT model, and synonymous entities are identified through vector similarity. Merge and simplify semantically equivalent or approximate relational paths, identify and retain key paths in knowledge propagation based on centrality indicators in graph theory, and remove edge noise.

[0022] It should be noted that entity types include diseases (e.g., type 2 diabetes, hypertension), symptoms (e.g., polydipsia, polyuria, dizziness), examination items (e.g., fasting blood glucose, electrocardiogram), test indicators (e.g., blood glucose level, systolic blood pressure), drugs (e.g., metformin, amlodipine), anatomical structures (e.g., pancreas, cardiovascular system), standard codes (e.g., SNOMEDCT:44054006), and data source structures (e.g., "FBG", "fasting blood glucose"), etc.

[0023] It should be noted that the relation types adopt a hierarchical design, including: Semantic relationships: such as "disease-symptom-symptom", "drug-treatment-disease", "gene-association-disease", etc.

[0024] Temporal relationships: Introduce timeline marking standards to distinguish the version evolution relationships of knowledge (such as "diabetes diagnostic criteria - 1999 version" and "2023 version").

[0025] Evidence Relationship: Mark the type of evidence source (guidelines, literature, expert consensus) and the level of evidence for knowledge.

[0026] It should be noted that this method, based on the knowledge distillation mechanism of evidence levels and network centrality, effectively controls the graph size while preserving core knowledge, thus improving query and reasoning efficiency. This method constructs a multi-layered health knowledge graph that integrates international standard terminology, clinical guidelines, and real-world data. In particular, by incorporating data source fields as independent entity types into the graph and establishing dynamic mapping relationships with standard concepts, the knowledge graph becomes not only a static medical knowledge base but also a "living" mapping knowledge base, providing rich semantic context for subsequent intelligent mapping.

[0027] Furthermore, in the data collection method for vital sign monitoring on the online health platform, for each data source accessed by the online health platform, a data source semantic description model is constructed based on the data source metadata, data field semantic descriptions, and the known mapping relationship between the data source's unique coding system and the standard coding system. Specifically, this includes: For each data source connected to the online health platform, a metadata layer is constructed based on the data source name, the type of preset device, the version number, and the access time. Based on each data field in the data source, record the field name, data type, semantic label, value range, unit, and sampling frequency to construct a field semantic layer, and record the known partial mapping relationship between the data source's unique local encoding system and the standard encoding system to form an encoding mapping layer; Record the contextual information of the data collection scenario and the relevant contextual information of the population characteristics to form a context layer, and establish a standardized semantic description model based on the meta-information layer, field semantic layer, encoding mapping layer and context layer; When a new data source is connected to the platform, the data source submits a DSSP through a standardized interface, inserts the field description information in the DSSP as a new node into the knowledge graph, establishes the data source entity, and associates it with the data source node through relevant relation types, and automatically finds the best standard concept mapping for each field of the current data source. It should be noted that DSSP is a standardized semantic description model, described in JSON-LD or RDF format.

[0028] After the mapping results are evaluated for confidence, if the confidence level meets the standard, the predetermined relationship type is confirmed and established; otherwise, it is transferred to the manual review queue.

[0029] It should be noted that this method proposes a standardized Data Source Semantic Description Model (DSSP) and a dynamic registration mechanism, transforming the access of new data sources from "manual configuration of mapping tables" to "automatic semantic registration and mapping," significantly reducing the marginal cost of system expansion. Simultaneously, by incorporating data source fields as first-class entities into the knowledge graph, the accumulation and reuse of mapping knowledge are achieved—when multiple data sources contain semantically similar fields, existing mapping relationships can be automatically reused through reasoning within the knowledge graph.

[0030] Furthermore, in the data collection method for vital sign monitoring on the online health platform, the data field to be mapped is input into the knowledge graph as a query node. A graph neural network is used for multi-hop reasoning to find the standard concept node with the best semantic match in the knowledge graph, forming a mapping model. The mapping model is then used to output the mapping result, specifically: A mapping model is constructed. For the data source field description text to be mapped, it is encoded into a semantic vector. The cosine similarity between semantic vectors is calculated. Based on the cosine similarity of semantic vectors, the Top-50 candidate standard concepts are recalled in the knowledge graph. The first phase recall of a preset number of candidate concepts and their multi-hop neighborhood subgraphs in the knowledge graph are extracted, a candidate subgraph is constructed, a graph isomorphism transformer model is used to sort the candidate concepts, and the semantic vector, node type encoding, node degree and other features of the candidate nodes are fused into the initial node representation. The aggregation strategy of graph isomorphic network is adopted to aggregate the neighborhood information of each node in multiple rounds to capture the structural role information of the node. Self-attention calculation is performed on all candidate subgraph nodes in the global scope to capture the long-range dependencies between nodes. The semantic and structural features are adaptively weighted and fused using a gated aggregator to generate the final representation vector of each candidate concept. The matching score between the representation of the field to be mapped and the representation of each candidate concept is calculated. The candidate concepts with a preset ranking and their matching scores are output. The candidate concepts with a preset ranking and their matching scores are then scored a second time, and the final mapping result and confidence score are output. Finally, the mapping model is trained by comparing positive and negative samples.

[0031] It should be noted that for the Top-3 candidate mappings output after fine ranking, a verification mechanism based on contrastive learning is introduced: 1. Positive sample construction: extract positive sample pairs (field descriptions, standard concepts) from existing confirmed mapping relationships; 2. Negative sample construction: construct difficult negative samples through random replacement, synonym replacement, etc.; 3. Contrastive learning training: adopt supervised contrastive loss function to narrow the representation distance of positive sample pairs and widen the representation distance of negative sample pairs, thereby enhancing the model's ability to discriminate subtle semantic differences.

[0032] It should be noted that this method proposes a three-stage mapping architecture of "semantic coarse screening - graph neural network fine ranking - contrastive learning verification", which combines semantic matching with structural reasoning, and solves the shortcomings of traditional methods that rely solely on text similarity and ignore the structural relationships between concepts. Moreover, the graph isomorphic transformer (GIT) is applied to the health data standardization mapping scenario, making full use of the rich hierarchical structure and relational information in the medical ontology for mapping decisions. Finally, a contrastive learning mechanism is introduced for mapping verification, and the model's ability to distinguish easily confused concepts is improved through adversarial training with hard negative samples.

[0033] Furthermore, in the data collection method for vital sign monitoring on the online health platform, a semantic mapping engine is deployed at the data collection entry point to perform real-time semantic parsing and standardized mapping on the collected raw data fields. This converts heterogeneous data into a standardized format before storing it on the online health platform. Specifically: Deploy a semantic mapping engine at the data acquisition entry point. Use the semantic mapping engine to parse the field names, data types, and values ​​in the incoming data packets, and query the local LRU cache. If the field already has a mapping record, return it directly. For fields that are not cached, the mapping model is invoked for real-time inference, and the original data fields are converted into standardized concept codes and standardized values ​​based on the mapping results. The standardized data is then encapsulated into standardized health events in a unified format. The system collects standardized health events in a unified format and distinguishes between high-frequency mapping pairs and full mapping pairs based on the historical number of each standardized health event. High-frequency mapping pairs are cached locally, while low-frequency mapping pairs are cached in a distributed manner. At the same time, authoritative records of all mapping relationships are stored using a knowledge graph.

[0034] It's important to note that this method elevates semantic mapping from offline batch processing to online real-time processing at the data acquisition entry point, achieving a paradigm shift of "standardization upon data acquisition." Through a combination of lightweight model inference and a three-layer caching strategy, latency is controlled to the millisecond level while ensuring mapping accuracy, preventing standardization from becoming a bottleneck in the data acquisition pipeline. Furthermore, by deploying the semantic mapping engine at the data acquisition entry point, coupled with the three-layer caching strategy, online standardization is achieved as data flows in. This fundamentally avoids the "contaminate first, then clean up" data quality dilemma of non-standard data entering the storage layer and then undergoing cleaning and treatment, ensuring that platform data possesses consistency, standardization, and usability from the source.

[0035] Furthermore, in the data collection method for vital sign monitoring on the online health platform, a confidence score is output for each mapping, and the semantic description model of the data source is continuously optimized based on the confidence score. Specifically: For each mapping output, a comprehensive confidence score is calculated by considering factors such as model confidence, contextual consistency, historical consistency, and knowledge graph density, and hierarchical processing is performed based on the comprehensive confidence score. It should be noted that the overall confidence level is: ; For model confidence, For context consistency, For the sake of historical consistency, represents the knowledge graph density. Wherein, These are weighting coefficients, and the sum of the four is 1, for example, 0.1, 0.2, 0.3, and 0.4 respectively.

[0036] It should be noted that the model confidence level It comes from the matching score of the output; context consistency This involves checking whether the mapping results are consistent with the data source context (such as data source type and demographic characteristics). Historical consistency is also crucial. It checks whether the historical mapping results of the same field in the data source are consistent; knowledge graph density. It checks the neighborhood richness of the target standard concept in the knowledge graph (the richer the neighborhood, the clearer the concept definition, and the higher the mapping credibility).

[0037] Based on the overall confidence level The value is processed hierarchically: For example, when S ≥ 0.92, automatic confirmation occurs, indicating that the mapping result takes effect directly and is added to the training set as a positive sample. When S falls within the range of 0.75 ≤ S < 0.92, automatic confirmation plus sampling review occurs, the mapping takes effect, but 10% is randomly selected for manual review. When S falls within the range of 0.50 ≤ S < 0.75, manual review occurs, the mapping result is temporarily ineffective, and the result is pushed to the review queue. When S < 0.50, the result is marked as unmappingable, pushed to expert processing, and the knowledge graph completion process is triggered simultaneously.

[0038] Manually confirmed mapping results are added to the training data as positive samples, and manually corrected mappings are added to the training dataset as hard negative samples or corrected samples to adjust the decision boundary of the model. Meanwhile, the mapping model is incrementally retrained regularly using the expanded training set, the semantic description model of the data source is continuously optimized, and the concept addition process of the knowledge graph is triggered when the case is marked as a new concept, the new concept and its position and relationship in the knowledge graph are defined, and the dynamic expansion of the knowledge graph is completed.

[0039] It should be noted that this solution directly transforms the results of manual review into structured supervision signals (positive samples, hard negative samples, and corrected samples) for model training, achieving continuous learning and evolution under "human-machine collaboration." In particular, by linking the "new concept" scenario with the knowledge graph completion process, the system gains the ability to autonomously adapt to new medical concepts.

[0040] Furthermore, the methods for collecting vital sign monitoring data on online health platforms also include: All standardized health events are organized according to user and time dimensions to form a time-series event sequence for each user. In addition to the defined standard fields, each event is given derived time-series features such as event interval, trend label, and anomaly marker. For heterogeneous data from different data sources, we implement time granularity unification, missing value handling, time alignment window, and cross-modal time alignment for time-series relationship reasoning. We also map each standardized health event to a node in a knowledge graph and establish the association between events and standard concepts. We employ a knowledge graph embedding method to embed event nodes and concept nodes into a unified low-dimensional vector space, construct a vector index, and use vector guidance to perform fast retrieval based on semantic similarity.

[0041] It should be noted that for heterogeneous data from different data sources (such as heart rate data from wearable devices collected every second, and blood lipid data from health check centers collected quarterly), cross-modal time-series alignment is implemented, for example: Uniform time granularity: unify the timestamps of all events to UTC standard time, with precision down to the second level; Missing value handling: For sparse data (such as quarterly physical examinations), reasonable estimation is made through digital twin models or statistical imputation methods; Event Alignment Window: Defines an alignment window (e.g., in days) that aggregates multimodal events within the same time window into a "health status snapshot"; Temporal relation reasoning: Using causal relationships in knowledge graphs (such as "hypertension → increased risk of coronary heart disease"), we can perform causal chain reasoning on events at different points in time.

[0042] This method encapsulates standardized data into Standard Health Events (SHEs), providing a unified data abstraction for subsequent data analysis and enabling joint analysis across data sources. Furthermore, through cross-modal temporal alignment and temporal relationship reasoning, it organizes discrete and heterogeneous health events into logically related health state evolution sequences, achieving a leap from "data standardization" to "knowledge temporalization".

[0043] Because the aforementioned technical solutions require independent policy learning for each user, training from scratch demands a large amount of interactive data and a long convergence time. However, in real-world health monitoring scenarios, new users urgently need to quickly acquire personalized data collection strategies after joining the platform, and cannot afford a lengthy policy learning process. How to achieve rapid transfer and adaptation of data collection strategy knowledge to new users based on the data collection experience accumulated by the existing user base is a critical technical problem that urgently needs to be solved. Therefore, this method also includes: Each user's health data collection strategy optimization is defined as an independent meta-task. The meta-task distribution consists of the historical data of existing users in the platform. Each meta-task includes the user's state space distribution, state transition rules, and reward function characteristics. A meta-training set is constructed based on the meta-task. The initial parameters of the acquisition strategy decision network are trained using a model-independent meta-learning algorithm. For each sampled meta-task, starting from the initial parameters, K-step gradient updates are performed on each eye task to obtain task-specific parameters. On all sampling tasks, minimize the sum of the losses of updating task-specific parameters on the corresponding tasks, and update the initial parameters. During the meta-training process, simultaneously learn a task embedding network to map each user's meta-task into a low-dimensional task embedding vector. When a new user joins, the task embedding vector is calculated, and the M most similar historical meta-tasks are retrieved in the task embedding space. The optimized parameters corresponding to the M most similar historical tasks are then weighted and initialized. Gradient fine-tuning is performed on the actual collected data of new users to obtain personalized collection strategies. After each new user completes the collection strategy learning, the corresponding task data is anonymized and added to the meta-training set as a new meta-task. For completely new users with no historical data available, the initial parameters are directly used as the initial parameters of the collection strategy network, and the strategy is initialized in conjunction with general demographic prior knowledge.

[0044] It should be noted that for entirely new users with no historical data available (cold start scenario), meta-parameters are directly used as the initial parameters of the data collection strategy network. Combined with general demographic prior knowledge (age, gender, body mass index, etc.), this is used for strategy initialization, ensuring a reasonable data collection strategy from the very first collection. This method introduces meta-reinforcement learning into the cross-user transfer scenario of health data collection strategies, achieving a leap from "independent training for each user" to "group knowledge sharing and rapid transfer." Furthermore, it proposes a task similarity-aware meta-knowledge retrieval and weighted initialization mechanism, making strategy initialization for new users more accurate and efficient. Finally, it solves the problem of rapid deployment of collection strategies in cold start scenarios, enabling new users to obtain a reasonable collection strategy from their first use.

[0045] In health monitoring, certain key health events (such as arrhythmia, hypoglycemia, and sudden rise in blood pressure) occur infrequently, resulting in extremely scarce relevant data samples. However, these scarce events are precisely the most valuable information for health risk assessment and disease early warning. Traditional data collection methods employ a uniform collection strategy for all data types, failing to specifically enhance the collection capability for scarce health events. How to achieve targeted enhancement of scarce key health event collection while ensuring routine monitoring efficiency is a key bottleneck in improving the effectiveness of health monitoring. Therefore, this method also includes: Based on clinical guidelines, medical literature and historical data from the platform, rare health events are extracted. For each type of rare event, the trend, fluctuation range and cross-correlation of various physiological indicators are extracted in the period before the event occurs. For each type of rare health event, a conditional generative adversarial network is trained using a small number of existing real precursor signal samples. The generator takes a random noise vector and an event type condition vector as input to generate a realistic precursor signal sequence. The discriminator distinguishes between generated precursor signals and real precursor signals. Through adversarial training, the generator learns to produce synthetic samples that are consistent with the distribution of real precursor signals, thus expanding the training data of scarce events. Using the synthetic dataset expanded by the generator, a lightweight precursor event detection model is pre-trained. The lightweight precursor event detection model is deployed on edge nodes to continuously scan the real-time data stream and identify whether there are signals that match the known scarce event precursor patterns. When the lightweight precursor event detection model identifies a possible precursor signal, it triggers an active learning mechanism to mark the data of that period as high-value candidate samples, increases the acquisition frequency of relevant sensors, activates backup sensors to acquire richer multimodal data, and stores the acquired enhanced data in a high-priority buffer for priority uploading to the platform for in-depth analysis. After the collected data is analyzed by cloud-based expert models, if it is confirmed as a real rare health event, the complete precursor signal sequence corresponding to the event is added as a positive sample to the real sample library for retraining the generative adversarial network and precursor detection model; if it is confirmed as a false positive, it is added as a negative sample to the training set to suppress similar false triggers.

[0046] It should be noted that applying generative adversarial networks to synthesize and enhance precursor signals of scarce health events fundamentally solves the problem of insufficient scarce event samples. Moreover, when the lightweight precursor event detection model identifies a possible precursor signal, it triggers an active learning mechanism to mark the data of that period as high-value candidate samples, increases the acquisition frequency of relevant sensors, activates backup sensors to collect richer multimodal data, and stores the collected enhanced data in a high-priority buffer for priority uploading to the platform for in-depth analysis, thereby improving the rationality of data collection.

[0047] In reality, making independent acquisition decisions based on a single sensor or a single data modality ignores the collaborative relationships between sensors and data modalities. In fact, complex physiological coupling relationships exist between different health indicators—for example, heart rate and blood oxygenation changes are highly correlated, blood pressure and heart rate fluctuations are linked, and exercise state and ECG signal quality influence each other. How to explicitly model the collaborative relationships between sensors in the acquisition strategy optimization to achieve a synergistic acquisition effect of "1+1>2" is an important direction not addressed by existing technologies. Therefore, this method also includes: Each sensor / data modality is defined as a node in the graph. The node features include sensor type, power consumption characteristics, data sampling rate range, data accuracy, and privacy sensitivity attributes. Physiological coupling edges, statistical correlation edges, resource competition edges, and complementary enhancement edges are constructed. It should be noted that physiological coupling edges are established based on medical knowledge when there is a known coupling relationship between two physiological indicators (such as between heart rate and blood pressure, or between blood glucose and heart rate variability); statistical correlation edges are established based on historical data to calculate the statistical correlation between the data collected by two sensors, and when the correlation coefficient exceeds a threshold; resource competition edges are established when two sensors share the same computing or communication resources; and complementary enhancement edges are established when the combined use of data from two sensors can improve detection accuracy.

[0048] For each sensor node, the attention coefficient between each node and its neighboring nodes is calculated, and the attention weight is normalized. Through multi-layer graph attention stacking, the representation of each node gradually integrates the information of its multi-hop neighbors and updates the node representation. Based on the updated node representation, an independent policy head outputs acquisition decisions for each sensor. At each decision moment, the current state of all sensors is encoded as the initial feature of the node. Through multiple rounds of information transmission via a graph attention network, the decision of each sensor incorporates the state information of its associated sensors. Each sensor outputs acquisition actions in parallel according to the updated representation. After executing the acquisition action, the graph neural network is continuously trained by collecting comprehensive data information gain, system energy consumption, and privacy risks. When a sensor fails or the data quality deteriorates significantly, the graph neural network automatically identifies the affected associated sensors and dynamically reconfigures the acquisition strategy.

[0049] It should be noted that the acquisition strategy includes: neighboring nodes of the faulty sensor perceive the abnormal state of the node through a graph attention mechanism; related sensors automatically adjust the acquisition strategy to partially compensate for the missing information of the faulty sensor (e.g., when the heart rate sensor fails, the blood oxygen sensor and accelerometer work together to increase the sampling rate and indirectly estimate heart rate information through other modalities); and a "acquisition strategy reconfiguration scheme" is generated through a graph neural network to achieve self-healing of the acquisition strategy under sensor failure.

[0050] It should be noted that introducing graph neural networks into the optimization of multi-sensor collaborative acquisition strategies explicitly models the physiological coupling, statistical correlation, and resource competition relationships between sensors, which can further optimize data acquisition. This allows for the explicit modeling of collaborative relationships between sensors in the acquisition strategy optimization, achieving a collaborative acquisition effect of "1+1>2".

[0051] like Figure 2 As shown, the second aspect of the present invention provides a vital sign monitoring data acquisition system for an online health platform, including a memory and a processor. The memory includes a vital sign monitoring data acquisition method program for the online health platform. When the vital sign monitoring data acquisition method program for the online health platform is executed by the processor, it implements the steps of any one of the vital sign monitoring data acquisition methods for the online health platform.

[0052] In a preferred embodiment of the present invention, the process of constructing a health domain knowledge graph is as follows: (I) Data Sources and Integration Medical knowledge is collected from the following sources: International standards: SNOMED CT, LOINC, ICD-10, UMLS, and other standard terminology systems; Clinical guidelines: Chinese guidelines for the prevention and treatment of type 2 diabetes (2023 edition), Chinese guidelines for the prevention and treatment of hypertension (2022 edition), ACC / AHA guidelines for the prevention of cardiovascular disease, etc.; Medical literature: High-quality clinical research literature from the past 10 years in the PubMed / MEDLINE database; Expert knowledge: Through collaboration with clinical experts from the departments of endocrinology, cardiology, and geriatrics in top-tier hospitals, the key relationships in the atlas are verified and supplemented.

[0053] (II) Definition of Entity and Relationship The core entity types defined include: diseases, symptoms, examination items, test indicators, drugs, anatomical structures, standard codes, and data source structures.

[0054] Examples of defined relationship types: Disease - Symptoms - Symptoms, Drug - Treatment - Disease, Gene - Association - Disease, Examination - Measurement - Indicator, Indicator - Normal Range - Range Value, Disease - Risk Factor - Factor.

[0055] (III) Hierarchical Alignment and Entity Disambiguation Example of terminology level alignment: Source A: Concept "DM" (abbreviation).

[0056] Source B: Concept "Diabetes Mellitus" (full English name).

[0057] Source C: Concept "diabetes" (Chinese).

[0058] Source D: Concept "Type 2 Diabetes" (subtype) → is uniformly identified as the "Diabetes" concept family through UMLS CUI (such as C0011849), while retaining the hierarchical relationship.

[0059] Example of structural layer alignment: Source A uses the relationship "treats" to indicate drug treatment for diseases. Source B uses the relation "indicated_for" to indicate drug indications. A unified mapping to gene-association-disease relationships Example of instance layer alignment: The SapBERT model is used to generate semantic embeddings of entities. For the entities "FBG", "Fasting Glucose" and "Fasting Blood Glucose", they are identified as synonymous entities through vector similarity (similarity > 0.92) and merged.

[0060] (IV) Path Simplification and Knowledge Distillation Based on centrality indicators in graph theory (betweenness centrality, closeness centrality, PageRank, etc.), key paths in knowledge propagation are identified and retained. For example, the path "diabetes → complications → nephropathy" has high betweenness centrality and is retained; while noise paths (such as relationships that only appear in a single literature and have not been clinically verified) are pruned.

[0061] (V) Graph Scale and Storage The constructed knowledge graph contains approximately 500,000 entity nodes and 3,000,000 relational edges, covering a complete knowledge system of common chronic diseases (diabetes, hypertension, coronary heart disease, etc.) and their related examination indicators, drugs, symptoms, etc. The graph is stored in the Neo4j graph database, and an Elasticsearch index is established to support efficient semantic retrieval.

[0062] In a preferred embodiment of the present invention, taking accessing a new smart bracelet device as an example, the construction and dynamic registration process of DSSP is described: The new data source submits the DSSP to the platform through a standardized API interface; The platform parses the DSSP and inserts device information and field information as new entity nodes into the knowledge graph; For the field "hr", the system detects that there is an existing known mapping relationship hr → LOINC:8867-4, which is directly confirmed; For the field "steps", although there is no direct known mapping, it is matched to the standard concept "LOINC:55423-8 (step count)" through inference by the mapping model in step S3, and the confidence score , which is automatically confirmed; After the registration is completed, the data source can start reporting standardized data to the platform.

[0063] In a preferred embodiment of the present invention, taking mapping the data source field description "morning fasting fingertip blood glucose" to a standard concept as an example, the specific working process of the mapping model is described: (I) Semantic Coarse Screening Stage Text to be mapped: "morning fasting fingertip blood glucose".

[0064] A pre-trained medical BERT model (such as BioBERT or ClinicalBERT) is used to encode into a 768-dimensional semantic vector: ; Calculate cosine similarity in the standard concept vector library of the knowledge graph to recall the Top-50 candidate concepts: LOINC:2339-0 (glucose^fasting^fingerprint blood): similarity 0.89.

[0065] LOINC:2341-6 (glucose^fasting^venous blood): similarity 0.87.

[0066] LOINC:14771-0 (glucose^2-hour postprandial^fingerprint): similarity 0.72.

[0067] SNOMED:166717003 (abnormal fasting blood glucose): similarity 0.68.

[0068] (II) Fine Ranking Stage of Graph Neural Network Extracting multi-hop neighborhood subgraphs (within 2 hops) of the Top-50 candidate concepts from a knowledge graph. For example, the neighborhood of LOINC:2339-0 contains: Parent concept: LOINC:2345-7 (glucose^fasting).

[0069] Relevant testing method: SNOMED:123456 (blood glucose test strip method).

[0070] Related disease: SNOMED:44054006 (Type 2 diabetes).

[0071] Normal range: LOINC: 2342-4 (normal range for fasting blood glucose).

[0072] A graph isomorphic transformer (GIT) is used to fuse the semantic vector (768-dimensional), type encoding (one-hot), and node degree of a node into an initial representation, which is then processed through three layers of GIT encoding to obtain the final representation.

[0073] (III) Comparative Learning and Verification Phase After Git sorting, the Top-3 candidates are: LOINC:2339-0 (Match score 0.94) LOINC:2341-6 (Match score 0.78) LOINC:14771-0 (Match score 0.62) The comparative learning validator performed secondary scoring on the Top-3 results, finding that candidate 1 had the smallest semantic distance to the query and was consistent with the context (data source: blood glucose meter, sampling time: early morning). Therefore, LOINC:2339-0 was ultimately confirmed as the mapping result, based on the overall confidence level. .

[0074] Suppose the components of a certain mapping decision are as follows: .

[0075] (Fields are from blood glucose meter devices and mapped to blood glucose-related concepts, with consistent context).

[0076] The data source mapped the same field to the same concept 10 times previously, with a consistency rate of 100%.

[0077] (The target concept LOINC:2339-0 has a rich neighborhood in the graph) Weighting coefficients: .

[0078] Overall confidence level: .

[0079] Automatic confirmation.

[0080] for Mapping decisions (within the range of 0.50 ≤ S < 0.75) are pushed to the manual review queue by the system. Reviewers find that the field should actually map to another concept and make corrections. The corrected mapping pairs are added to the training set as correction samples to adjust the model's decision boundaries.

[0081] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0082] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0083] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0084] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0086] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for collecting vital sign monitoring data on an online health platform, characterized in that, Includes the following steps: Integrate international standards, clinical guidelines, medical literature, and expert knowledge to construct a health domain knowledge graph that includes entities, attributes, and relationships; For each data source accessed in the online health platform, a data source semantic description model is constructed based on the data source metadata, data field semantic descriptions, and the known mapping relationship between the data source's unique coding system and the standard coding system. The data field to be mapped is input into the knowledge graph as a query node. A graph neural network is used for multi-hop reasoning to find the standard concept node with the best semantic match in the knowledge graph, forming a mapping model. The mapping result is then output using the mapping model. A semantic mapping engine is deployed at the data collection entry point to perform real-time semantic parsing and standardized mapping on the collected raw data fields, and to convert heterogeneous data into a standard format before storing it in the online health platform. A confidence score is output for each mapping, and the semantic description model of the data source is continuously optimized based on the confidence score.

2. The method for collecting vital sign monitoring data on an online health platform according to claim 1, characterized in that, By integrating international standards, clinical guidelines, medical literature, and expert knowledge, a health domain knowledge graph is constructed that includes entities, attributes, and relationships, specifically: Integrate international standards, clinical guidelines, medical literature, and expert knowledge to define core entity types and relationship types; For different expressions of the same concept in international standards, clinical guidelines, medical literature and expert knowledge, UMLSCUI can achieve cross-terminology system normalization and complete terminology level alignment. Referring to the BiolinkModel standardized predicate system, relation types from different sources are mapped to a unified set of relation predicates to complete the structural layer alignment. Entity disambiguation is performed by combining the SapBERT model with the edit distance algorithm to complete the instance layer alignment. After alignment at the terminology layer, structure layer, and instance layer, a health domain knowledge graph containing entities, attributes, and relationships is constructed. At the same time, the semantic embedding of entities is generated using the SapBERT model, and synonymous entities are identified through vector similarity. Merge and simplify semantically equivalent or approximate relational paths, identify and retain key paths in knowledge propagation based on centrality indicators in graph theory, and remove edge noise.

3. The method for collecting vital sign monitoring data on an online health platform according to claim 1, characterized in that, For each data source accessed in the online health platform, a data source semantic description model is constructed based on the data source metadata, data field semantic descriptions, and the known mapping relationship between the data source's unique encoding system and the standard encoding system. Specifically, this includes: For each data source connected to the online health platform, a metadata layer is constructed based on the data source name, the type of preset device, the version number, and the access time. Based on each data field in the data source, record the field name, data type, semantic label, value range, unit, and sampling frequency to construct a field semantic layer, and record the known partial mapping relationship between the data source's unique local encoding system and the standard encoding system to form an encoding mapping layer; Record the scene information of data collection and the relevant context information of the population characteristics to form a context layer, and establish a standardized semantic description model based on the meta-information layer, field semantic layer, encoding mapping layer and context layer; When a new data source is connected to the platform, the data source submits a DSSP through a standardized interface, inserts the field description information in the DSSP as a new node into the knowledge graph, establishes the data source entity, and associates it with the data source node through relevant relation types, and automatically finds the best standard concept mapping for each field of the current data source. After the mapping results are evaluated for confidence, if the confidence level meets the standard, the predetermined relationship type is confirmed and established; otherwise, it is transferred to the manual review queue.

4. The method for collecting vital sign monitoring data on an online health platform according to claim 1, characterized in that, The data field to be mapped is input into the knowledge graph as a query node. A graph neural network is used for multi-hop reasoning to find the standard concept node with the best semantic match in the knowledge graph, forming a mapping model. The mapping model is then used to output the mapping result, specifically: A mapping model is constructed, and the text describing the data source fields to be mapped is encoded into semantic vectors using the mapping model. The cosine similarity between semantic vectors is calculated, and the Top-K candidate standard concepts are recalled in the knowledge graph based on the cosine similarity of the semantic vectors. The first phase recall of a preset number of candidate concepts and their multi-hop neighborhood subgraphs in the knowledge graph are extracted, a candidate subgraph is constructed, a graph isomorphism transformer model is used to sort the candidate concepts, and the semantic vector, node type encoding, node degree and other features of the candidate nodes are fused into the initial node representation. The aggregation strategy of graph isomorphic network is adopted to aggregate the neighborhood information of each node in multiple rounds to capture the structural role information of the node. Self-attention calculation is performed on all candidate subgraph nodes in the global scope to capture the long-range dependencies between nodes. The semantic and structural features are adaptively weighted and fused using a gated aggregator to generate the final representation vector of each candidate concept. The matching score between the representation of the field to be mapped and the representation of each candidate concept is calculated. The candidate concepts with a preset ranking and their matching scores are output. The candidate concepts with a preset ranking and their matching scores are then scored a second time, and the final mapping result and confidence score are output. Finally, the mapping model is trained by comparing positive and negative samples.

5. The method for collecting vital sign monitoring data on an online health platform according to claim 1, characterized in that, A semantic mapping engine is deployed at the data acquisition entry point to perform real-time semantic parsing and standardized mapping on the collected raw data fields. This converts heterogeneous data into a standardized format before storing it in the online health platform. Specifically: A semantic mapping engine is deployed at the data acquisition entry point. The semantic mapping engine is used to parse the field names, data types and values ​​in the incoming data packets, and to query the local LRU cache. If the field already has a mapping record, it is returned directly. For fields that are not cached, the mapping model is invoked for real-time inference, and the original data fields are converted into standardized concept codes and standardized values ​​based on the mapping results. The standardized data is then encapsulated into standardized health events in a unified format. The system collects standardized health events in a unified format and distinguishes between high-frequency mapping pairs and full mapping pairs based on the historical number of each standardized health event. High-frequency mapping pairs are cached locally, while low-frequency mapping pairs are cached in a distributed manner. Meanwhile, the knowledge graph is used to store authoritative records of all mapping relationships, and the data is uniformly converted into a standard format and then stored in the online health platform.

6. The method for collecting vital sign monitoring data on an online health platform according to claim 1, characterized in that, A confidence score is output for each mapping, and the semantic description model of the data source is continuously optimized based on the confidence score, specifically as follows: For each mapping output, a comprehensive confidence score is calculated by considering factors such as model confidence, contextual consistency, historical consistency, and knowledge graph density, and hierarchical processing is performed based on the comprehensive confidence score. Manually confirmed mapping results are added to the training data as positive samples, and manually corrected mappings are added to the training dataset as hard negative samples or corrected samples to adjust the decision boundary of the model. Meanwhile, the mapping model is incrementally retrained regularly using the expanded training set, the semantic description model of the data source is continuously optimized, and the concept addition process of the knowledge graph is triggered when the case is marked as a new concept, the new concept and its position and relationship in the knowledge graph are defined, and the dynamic expansion of the knowledge graph is completed.

7. The method for collecting vital sign monitoring data on an online health platform according to claim 5, characterized in that, Also includes: All standardized health events are organized according to user and time dimensions to form a time-series event sequence for each user. In addition to the defined standard fields, each event is given derived time-series features such as event interval, trend label, and anomaly marker. For heterogeneous data from different data sources, we implement time granularity unification, missing value handling, time alignment window, and cross-modal time alignment for time-series relationship reasoning. We also map each standardized health event to a node in a knowledge graph and establish the association between events and standard concepts. The knowledge graph embedding method is used to embed event nodes and concept nodes into a unified low-dimensional vector space to construct a vector index, and the vector index is used to guide fast retrieval based on semantic similarity.

8. A vital sign monitoring data acquisition system for an online health platform, characterized in that, The device includes a memory and a processor. The memory includes a program for a vital sign monitoring data acquisition method for an online health platform. When the processor executes the program for the vital sign monitoring data acquisition method for an online health platform, it implements the steps of the vital sign monitoring data acquisition method for an online health platform as described in any one of claims 1-7.

Images