Dynamic health record automatic updating and management method based on cloud platform

Abstract

The invention relates to a dynamic health record automatic updating and management method based on a cloud platform. The method comprises the following steps: acquiring multi-source heterogeneous data and performing event packaging to generate a standard health event stream; performing cross-modal time sequence alignment processing on the standard health event stream to form a time sequence event sequence; an event-state double-layer time sequence knowledge graph is constructed, and real-time updating of health state vectors is achieved by means of the dynamic reasoning ability of a graph neural network in combination with a time sequence event sequence; and dynamic updating of the archive is completed through vectorization archiving. By the adoption of the method, the defects of traditional health archives in the aspects of data fusion depth, updating timeliness and state characterization precision can be effectively overcome, an accurate and real-time data basis is provided for chronic disease management and other long-term health intervention scenes, and the initiative and individuation level of health management is remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of record updating and management technology, and in particular relates to a method for automatic updating and management of dynamic health records based on a cloud platform. Background Technology

[0002] With the deep integration of cloud computing, the Internet of Things, and artificial intelligence technologies in the healthcare field, health management services are gradually developing towards digitalization, intelligence, and personalization. Against this backdrop, cloud platforms, as the carrier supporting the aggregation and processing of large-scale health data, provide the technological foundation for collaborative health management among medical institutions, health check centers, insurance companies, and enterprises. Among these, dynamic health records, as the core carrier recording an individual's health status throughout their entire life cycle, have become crucial for improving the quality and efficiency of health services due to their automatic updates and precise management capabilities. Currently, this field mainly relies on traditional health record management methods based on rule-based data integration and periodic archiving.

[0003] In traditional technologies, health record updates often employ passive triggering or timed batch processing mechanisms. Specifically, participating parties (such as hospital information systems, insurance claims platforms, and wearable device manufacturers) typically upload health data to a cloud platform through predefined data interfaces. The platform then standardizes the data format and stores it in a centralized database. The update logic often relies on simple triggering conditions, such as pushing new records after the completion of a medical event or the generation of a physical examination report, or batch integrating historical data according to fixed time windows. At the data fusion level, traditional methods typically employ timestamp-based linear alignment or keyword-based semantic mapping to achieve preliminary association of multi-source information.

[0004] However, current health record management methods still have significant limitations. Due to the significant differences in collection frequency, structure, and semantic description of multi-source data (medical, insurance, behavioral, etc.), traditional methods struggle to achieve deep integration of cross-modal data at both temporal and causal levels, resulting in a fragmentation of the continuity and integrity of health status evolution. Update mechanisms lack the ability to perceive dynamic changes in the health context and cannot adaptively adjust record content based on real-time data streams, causing record updates to lag behind actual changes in health status. Furthermore, existing methods utilize heterogeneous information such as unstructured text and continuous sensor data at a superficial level, failing to effectively mine the implicit clinical semantics and health risk signals, thus limiting the supporting value of records in long-term health intervention scenarios such as chronic disease management. Summary of the Invention

[0005] Therefore, it is necessary to provide a health record management method that can achieve deep fusion of multi-source heterogeneous data, support real-time tracking of health status evolution, and have context-aware dynamic update capabilities to address the above-mentioned technical problems.

[0006] Firstly, this application provides a method for the automatic updating and management of dynamic health records based on a cloud platform, including:

[0007] Acquire multi-source heterogeneous health data; and encapsulate the multi-source heterogeneous health data into events to generate standardized health event streams;

[0008] Perform cross-modal temporal alignment processing on standardized health event streams to generate temporal health event sequences;

[0009] A temporal knowledge graph is constructed; and based on the temporal knowledge graph and the temporalized health event sequence, the high-dimensional health state vector is updated through a graph neural network model to obtain the updated high-dimensional health state vector; the high-dimensional health state vector is used to represent the overall health status of an individual.

[0010] The dynamic health profile is updated based on the updated high-dimensional health state vector, resulting in the updated dynamic health profile.

[0011] Furthermore, the method also includes:

[0012] Extract high-dimensional health state vectors within a preset time range from the updated dynamic health records to form a health state evolution sequence;

[0013] The health status evolution sequence is input into a trained long short-term memory network model, and the risk prediction result is output.

[0014] Based on risk prediction results and the current high-dimensional health status vector, health intervention suggestions are generated using a large language model.

[0015] Furthermore, multi-source heterogeneous health data is encapsulated into events to generate standardized health event streams, including:

[0016] Numerical sensor data in multi-source heterogeneous health data are smoothed and denoised to obtain cleaned regular integer value data.

[0017] Quantitative feature vectors of medical image data from multi-source heterogeneous health data are extracted using a pre-trained convolutional neural network model; the quantitative feature vectors include electrocardiogram waveform features and medical image features.

[0018] For unstructured text data in multi-source heterogeneous health data, structured medical fact triples are generated through a large language model;

[0019] The cleaned integer value data, quantitative feature vectors, and structured medical fact triples are encapsulated according to preset event object rules to generate health event objects; the health event object includes event type, occurrence time, data source, and event payload;

[0020] All health event objects are combined to generate a standardized health event stream.

[0021] Furthermore, the standardized health event stream undergoes cross-modal temporal alignment processing to generate a temporally sequenced health event sequence, including:

[0022] Starting with an individual's birth time, a high-granularity virtual timeline is constructed; this high-granularity virtual timeline serves as a reference benchmark for time sequence alignment.

[0023] The occurrence timestamp of each health event object in the standardized health event stream is mapped to a virtual timeline to obtain a primary mapped event set;

[0024] From the primary mapping event set, select the health event sequence pairs that need to be aligned according to the preset association rules; the health event sequence pairs that need to be aligned include the source event sequence and the target event sequence to be aligned;

[0025] For health event sequence pairs that need to be aligned, the curved path with the minimum cumulative distance between the two sequences is determined by an improved dynamic time warping algorithm.

[0026] Based on the curved path, each event point in the source event sequence to be aligned is aligned to the corresponding time point in the target event sequence to generate a preliminary aligned event set;

[0027] The initial aligned event set is sorted according to the virtual timeline to obtain a time-series health event sequence.

[0028] Furthermore, a temporal knowledge graph is constructed; and based on the temporal knowledge graph and the temporally sequenced health event sequence, the high-dimensional health state vector is updated using a graph neural network model to obtain the updated high-dimensional health state vector, including:

[0029] Constructing a temporal knowledge graph; the temporal knowledge graph includes an event layer and a state layer;

[0030] The initial health event in the time-series health event sequence is identified as the initial health event node in the event layer of the time-series knowledge graph;

[0031] Based on the baseline information of the initial health events, a central health state node representing an individual is created in the state layer of the temporal knowledge graph; and a current high-dimensional health state vector is generated based on the central health state node.

[0032] When a new health event is received, the current high-dimensional health state vector is updated using a graph neural network model based on predefined clinical pathways and a causal reasoning rule base, resulting in an updated high-dimensional health state vector. New health event objects include subsequent events within a time-series of health events and newly added emergencies in real time.

[0033]

[0034] in, This is the updated high-dimensional health state vector. Let W be the high-dimensional health state vector before the update, W be the learnable weight matrix, || be the vector concatenation operation, and AGGREGATE be the aggregation function. The attention weight of the central healthy state node i at time t. Let σ be the feature representation of neighbor node i, and let σ be the activation function.

[0035] Furthermore, based on the risk prediction results and the current high-dimensional health state vector, a large-scale language model is used to generate health intervention recommendations, including:

[0036] Construct a structured prompt word template; the structured prompt word template includes a health status summary, a description of the risk prediction results, the target user role, and the application scenario;

[0037] The prompt word template is filled in based on the clinical significance interpretation of the current high-dimensional health status vector, the specific content of the risk prediction results, the target user identity, and the suggested application scenario information to obtain the filled prompt words;

[0038] The filled-in prompts are input into a pre-trained large-scale language model to generate a first draft of natural language text; the first draft of natural language text conforms to grammar and medical professionalism.

[0039] Standardized post-processing of the initial natural language text yields health intervention recommendations.

[0040] Furthermore, the method also includes:

[0041] Collect medical expense data; and standardize the medical expense data to generate standardized expense event records; the standardized expense event records include expense item name, expense amount, payment status, payer type, service provider, and expense occurrence timestamp;

[0042] By associating and matching standardized expense event records with time-series health event sequences, a network of relationships between health events and expense records is established.

[0043] Based on the correlation network between health events and expense records, a multi-dimensional expense data graph is constructed. The multi-dimensional expense data graph uses time as the vertical axis and payer, service provider, and expense category as dimensions.

[0044] Calculate the total cost and cost percentage for each dimension in the multi-dimensional cost data graph to obtain the calculation results; and generate a Sankey diagram data source for cost flow based on the calculation results.

[0045] The Sankey diagram data source is visualized to generate a Sankey diagram of medical expense flow. The Sankey diagram of medical expense flow is used to show the path and flow ratio of funds from different payers to various service providers, broken down to specific expense items. It supports filtering the expense flow within a specific time period by time slider.

[0046] Secondly, this application also provides a cloud-based device for the automatic updating and management of dynamic health records, including:

[0047] The data acquisition module is used to acquire multi-source heterogeneous health data and encapsulate the multi-source heterogeneous health data into events to generate standardized health event streams.

[0048] The sequence generation module is used to perform cross-modal temporal alignment processing on standardized health event streams to generate temporalized health event sequences.

[0049] The vector update module is used to construct a temporal knowledge graph; and based on the temporal knowledge graph and the temporalized health event sequence, it updates the high-dimensional health state vector through a graph neural network model to obtain the updated high-dimensional health state vector; the high-dimensional health state vector is used to represent the overall health status of an individual.

[0050] The health record update module is used to update the dynamic health record based on the updated high-dimensional health state vector, and obtain the updated dynamic health record.

[0051] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the at least one instruction, the at least one program, the code set or instruction set is loaded and executed by the processor to implement the cloud platform-based dynamic health record automatic update and management method described in any of the embodiments of this application.

[0052] Fourthly, this application also provides a computer-readable storage medium storing at least one piece of program code, which is loaded and executed by a processor to implement the cloud-based dynamic health record automatic update and management method described in any of the embodiments of this application.

[0053] The aforementioned cloud-based method for the automatic updating and management of dynamic health records generates a standard health event stream by encapsulating multi-source heterogeneous data into events; performs cross-modal temporal alignment processing on the standard health event stream to form a temporal event sequence; constructs a two-layer event-state temporal knowledge graph; and leverages the dynamic reasoning capabilities of graph neural networks to achieve real-time updates of health state vectors; and completes dynamic updates of records through vectorized archiving. This method effectively overcomes the shortcomings of traditional health records in terms of data fusion depth, update timeliness, and accuracy of state representation, providing a precise and real-time data foundation for long-term health intervention scenarios such as chronic disease management, and significantly improving the initiative and personalization of health management. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating a cloud-based method for automatically updating and managing dynamic health records in one embodiment.

[0056] Figure 2 This is a flowchart illustrating the steps of encapsulating multi-source heterogeneous health data into events and generating a standardized health event stream in one embodiment.

[0057] Figure 3 This is a schematic diagram of a cloud-based automatic update and management device for dynamic health records in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, a method for automatic updating and management of dynamic health records based on a cloud platform is provided. This embodiment illustrates the application of this method to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through the interaction between the terminal and the server. Figure 1 As shown, in this embodiment, the method includes the following steps:

[0060] Step S101: Obtain multi-source heterogeneous health data; and encapsulate the multi-source heterogeneous health data into events to generate a standardized health event stream.

[0061] Among them, event-based encapsulation refers to encapsulating "action / state changes" into independent "events", which is a standardized encapsulation of the callback mechanism.

[0062] For example, health data is asynchronously collected from multiple sources, including hospital information systems and other medical institution information systems, physical examination center platforms, insurance claims databases, and personal wearable devices. This data exhibits significant heterogeneity, encompassing structured test values ​​(such as blood glucose levels), semi-structured examination reports, unstructured clinical medical records, continuous physiological sensor signals (such as heart rate waveforms), and insurance claim codes. This multi-source, heterogeneous health data is then encapsulated into standardized health event streams.

[0063] Step S102: Perform cross-modal temporal alignment processing on the standardized health event stream to generate a temporalized health event sequence.

[0064] Cross-modal time alignment refers to aligning multimodal data in time to avoid misunderstandings caused by time misalignment and ensure that multi-source information corresponds to the same event or state in the timeline.

[0065] For example, cross-modal temporal alignment processing is performed on a standardized health event stream to obtain a temporally sequenced health event sequence in which all health events in the health event stream are arranged in chronological order on a virtual timeline and are semantically consistent.

[0066] Step S103: Construct a temporal knowledge graph; and based on the temporal knowledge graph and the temporalized health event sequence, update the high-dimensional health state vector through a graph neural network model to obtain the updated high-dimensional health state vector; the high-dimensional health state vector is used to represent the overall health status of an individual.

[0067] Among them, graph neural network models are deep learning models used to process graph structure data. They can model the topology of nodes, edges and the entire graph, and learn the representation of nodes or graph by aggregating information from neighboring nodes. Common graph neural network models include graph convolutional networks, graph attention networks and graph autoencoders.

[0068] For example, an event-state two-layer temporal knowledge graph is initialized. The event layer nodes consist of events from a temporally sequenced health event sequence, while the state layer has a central node representing an individual's overall health status. The initial health event in the sequence (e.g., birth record) is used as the starting node of the event layer, and an initial high-dimensional health state vector is generated based on the baseline information of this event (e.g., birth weight). The vector dimension can be preset to 128 dimensions, covering multiple dimensions such as physiological, behavioral, and clinical features. When a new health event (e.g., abnormal physical examination event) arrives, a predefined clinical pathway and causal reasoning rule base are loaded, triggering an update of the central state node. The update process is completed by a graph neural network model based on an attention mechanism: this network model takes the current health state vector, the new event node, and its neighboring nodes in the graph (e.g., related past events) as input, calculates the weights between nodes through an attention mechanism, aggregates relevant information, and generates an updated state vector. Among them, the initial health event (such as the birth record) is the earliest occurrence in an individual's health information sequence and has fundamental identifying significance. In terms of content, it is not simply a record of the birth fact, but includes key initial data on health dimensions, such as the newborn's birth weight, gestational age, presence of congenital abnormalities, Apgar score at birth (an important indicator for assessing the degree of neonatal asphyxia), and key health information of the mother during pregnancy. The predefined clinical pathway is a standardized and structured medical plan developed in advance in the medical field to standardize the diagnosis and treatment process of specific diseases / surgeries. By unifying the pathway, it balances the quality, efficiency and cost of diagnosis and treatment and reduces the arbitrariness of clinical decision-making. The causal reasoning rule base is the core knowledge set that supports the causal reasoning process. The rules can adopt the IF-THEN (if-then) form, for example: IF event type = blood glucose test AND value > 7.0 mmol / L THEN update diabetes risk level.

[0069] Step S104: Update the dynamic health profile based on the updated high-dimensional health state vector to obtain the updated dynamic health profile.

[0070] For example, the updated high-dimensional health status vector is associated with the list of event IDs (Identifications) that triggered the update, and written as a new version into the dynamic health profile. The profile can adopt a versioned management model, generating a timestamped version record for each update and retaining historical versions for retrospection. By comparing the differences in status vectors of consecutive versions, a health status change log is automatically generated, such as the diabetes risk level rising from medium risk to high risk. In addition, the latest status vector is synchronized to a specified field in the cloud platform database. Here, the versioned management model refers to a management method that achieves traceable changes and controllable risks by recording and tracking the iteration history of files, code, data, or project deliverables, assigning a unique version identifier (e.g., v1.0) to deliverables at different stages, and standardizing the change process; the cloud platform database is a database system that relies on cloud computing technology, is deployed on cloud servers, and provides data storage and management services through the network.

[0071] In this embodiment, standard health event streams are generated by encapsulating multi-source heterogeneous data into events; cross-modal temporal alignment of these standard health event streams is performed to form temporal event sequences; a two-layer temporal knowledge graph is constructed, and the dynamic reasoning capabilities of graph neural networks are used to achieve real-time updates of health status vectors; and vectorized archiving is used to complete the dynamic updates of health records. This effectively overcomes the shortcomings of traditional health records in terms of data fusion depth, update timeliness, and accuracy of status representation, providing a precise and real-time data foundation for long-term health intervention scenarios such as chronic disease management, and significantly improving the initiative and personalization of health management.

[0072] In one exemplary embodiment, the method further includes:

[0073] Step S201: Extract high-dimensional health state vectors within a preset time range from the updated dynamic health records to form a health state evolution sequence.

[0074] The preset time range is a time interval pre-configured based on the needs of health management scenarios. It can be stored in a configuration file or determined by user-defined parameters, covering common scenarios such as short-term follow-up, chronic disease management, and annual health assessment, ensuring that the extracted vectors can reflect changes in health status at a specific stage.

[0075] For example, after updating the dynamic health record, a sliding search is performed within a preset time range. The search criteria use the current timestamp as a baseline, tracing back to a specified time point to filter out high-dimensional health status vectors corresponding to all version records within that range. Each vector is marked with a timestamp accurate to milliseconds. These vectors are sorted according to their timestamps. During the sorting process, potential vector missing issues are automatically detected and handled. When a vector at a certain time point is found to be missing, linear interpolation is used to fill it appropriately based on the vector values ​​of adjacent time points, thus ensuring the integrity of the sequence. These high-dimensional health status vectors arranged in chronological order are constructed into a multi-dimensional array structure, namely the health status evolution sequence. The row dimension of this sequence represents the time step, and the column dimension corresponds to the various feature components of the health status vector, such as blood pressure trends and blood sugar fluctuations, forming a complete time series data matrix. Linear interpolation is a basic mathematical method for estimating unknown values ​​based on the linear relationship between two points. The logic is that two points determine a straight line, and the value of any point on the line can be calculated from the coordinates of the two points.

[0076] Step S202: Input the health status evolution sequence into the trained long short-term memory network model and output the risk prediction result.

[0077] Among them, the well-trained Long Short-Term Memory (LSTM) network model is a finished model that is based on the LSTM network structure and trained on a large-scale labeled health state evolution sequence dataset. It has the ability to process specific sequence tasks. The number of neurons in its input layer is perfectly matched with the dimension of the health state vector, and the hidden layer contains multiple LSTM units. Each unit learns long-term dependencies in the sequence through an ingenious gating mechanism. The gating mechanism includes an input gate, a forget gate, and an output gate.

[0078] For example, the health state evolution sequence is input into an LSTM network step by step. The network dynamically updates cell states based on the current input and the hidden states of the previous time step, thus capturing the evolutionary patterns of health states. After processing by a fully connected layer, a multi-dimensional risk probability vector is output. Each dimension of the vector corresponds to the probability of a specific health risk, such as a cardiovascular event or diabetic complications, occurring within a specific future time window. A configurable probability threshold can be set; when the probability of a risk exceeds the threshold, the risk is considered a positive prediction, and its confidence level is recorded. All these risk prediction results, along with their prediction time range, confidence level, and other related metadata, are encapsulated into a structured risk prediction report object.

[0079] Step S203: Based on the risk prediction results and the current high-dimensional health status vector, generate health intervention recommendations using a large language model.

[0080] Among them, the Large Language Model (LLM) is an artificial intelligence model based on deep learning technology, used to understand and generate human language, and has the ability to process various language-related tasks such as text, dialogue, and creation.

[0081] For example, based on risk prediction results and the current high-dimensional health state vector, a large-scale language model, leveraging its powerful language understanding and generation capabilities and incorporating internalized medical knowledge, outputs a grammatically correct and professionally designed initial draft of natural language text. This initial draft undergoes post-processing to generate personalized, actionable health intervention recommendations. These recommendations clearly identify risk factors, recommend action plans, and outline expected goals. Post-processing refers to the process of optimizing the text quality through a series of manual or procedural operations after the initial text is automatically generated by artificial intelligence and algorithms, ensuring the text better meets the readability, accuracy, and compliance requirements of real-world usage scenarios.

[0082] In this embodiment, a time-seriesd state vector sequence is extracted from dynamic health records. An LSTM network is used to deeply mine the evolutionary patterns of health states and accurately predict future risks. Combined with the intelligent text generation capabilities of LLM, complex analysis results are transformed into personalized action guidelines. This improves the accuracy and foresight of health risk assessments, promotes the implementation of health interventions by generating easily understandable personalized suggestions, and provides efficient and accurate health management decision support for medical institutions, insurance companies, and individual users.

[0083] In one embodiment, such as Figure 2 As shown, multi-source heterogeneous health data is encapsulated into events to generate a standardized health event stream, including:

[0084] Step S301: Perform smoothing and denoising processing on the numerical sensor data in the multi-source heterogeneous health data to obtain cleaned regular integer value data.

[0085] For example, the raw numerical health data acquired from wearable devices, medical sensors, etc., is quality-assessed to identify and mark potentially anomalous data points, including abrupt changes caused by transient sensor interference, null values ​​due to signal loss, and data significantly exceeding reasonable physiological ranges due to transmission errors. Subsequently, a moving average filtering algorithm based on a sliding window is used to smooth the marked data. For missing values, linear interpolation or predictive imputation based on the characteristics and duration of the missing data can be used to fill in the gaps, depending on the data point's characteristics and the duration of the gap. After smoothing and imputation, all data from different devices, potentially using different units of measurement, are converted to a preset international standard unit system. Finally, a batch of time-series continuous, numerically stable, and dimensionlessly consistent integer data is output. Among them, the moving average filtering algorithm is a basic time-domain signal smoothing algorithm. Within a fixed-length time window, it replaces the value of the center point of the window with the arithmetic mean of all data points within the window, thereby effectively suppressing random high-frequency noise while preserving the overall trend of data change. The prediction imputation method based on the trend of data at adjacent time points is a method to solve the problem of missing values ​​in data by utilizing the continuity and trend correlation of time series data. By determining the time point of missing data, and then selecting the effective data adjacent to the missing point, the change pattern is calculated, such as a linear trend. For example, if the data at the previous 3 time points are 10, 12, and 14 respectively, showing a linear increase of +2 each time, the missing value is then derived using the analyzed trend. For example, if the data before the missing point is 14 and the trend is +2 each time, then the missing value is imputed as 16.

[0086] Step S302: Extract quantitative feature vectors from medical image data in multi-source heterogeneous health data using a pre-trained convolutional neural network model; the quantitative feature vectors include electrocardiogram waveform features and medical image features.

[0087] Among them, the pre-trained convolutional neural network (CNN) model refers to a model that has been trained on a large-scale medical image dataset. Through its multiple convolutional and pooling layers, the model can automatically and hierarchically learn the features in the image: the shallow network captures basic edges and textures, while the deep network identifies more complex morphological structures and pathological patterns.

[0088] Exemplarily, for the received medical image data, such as electrocardiogram waveforms, X-ray films, computed tomography images, etc., first adjust the image size to the fixed resolution required by the pre-trained convolutional neural network model, perform grayscale normalization on the adjusted image to eliminate brightness differences, and perform contrast enhancement when necessary to highlight key regions. Input the pre-processed image into the pre-trained convolutional neural network model, and extract the activation values of the layer before the last fully connected layer of the model as the abstract feature representation of the image. These activation values form a high-dimensional and dense numerical vector, that is, a quantitative feature vector, which encodes the key clinical information in the image in a machine-readable form.

[0089] Step S303: For the unstructured text data in the multi-source heterogeneous health data, generate structured medical fact triples through a large language model.

[0090] Exemplarily, perform text preprocessing on unstructured text data such as clinical notes written by doctors, written descriptions of physical examination reports, chief complaint records, etc., including word segmentation, removing stop words (such as meaningless words like "of"), and standardizing medical terms. For example, unify myocardial infarction and myocardial infarction into myocardial infarction. Perform named entity recognition (NER) and relation extraction (RE) processing on unstructured text data through a large language model. The NER task is responsible for identifying medical entities mentioned in the text, such as diseases (such as diabetes), symptoms (such as polydipsia), drugs (such as metformin), examination items (such as fasting blood glucose), etc., and annotating the type of each entity. The RE task further analyzes the semantic relationships between these entities. For example, there is a relationship of suffering from between the patient and diabetes, and there is a relationship of being used for treatment between metformin and diabetes. The LLM combines the identified entities and relationships into structured medical fact triples in the form of "subject-relation-object", such as <patient, symptom, polydipsia>.

[0091] Step S304: Package the cleaned and regularized numerical data, quantitative feature vectors, and structured medical fact triples according to the preset event object rules to generate a health event object; the health event object includes event type, occurrence time, data source, and event payload.

[0092] Among them, the preset event object rules refer to the standardized logic set in advance for regulating the generation, circulation, and response of event objects; this event object rule usually exists in the form of JSON Schema; JSON Schema is a specification for describing and validating JSON data structures, and it is a special JSON document used to define "what a qualified JSON data should look like", which solves the problems of inconsistent JSON data formats and cumbersome verification.

[0093] For example, based on pre-defined event object rules, the three types of standardized data obtained in the preceding steps are encapsulated into a unified health event object. These rules clearly define the essential attributes of the event object and the data type of its values. The encapsulation process is as follows: assign a globally unique identifier to each event; record the exact timestamp of the event occurrence, accurate to milliseconds; indicate the source of the original data for the event (e.g., a hospital information system); determine and assign the event type based on the data content (e.g., vital sign measurement). The cleaned integer value data, quantitative feature vectors extracted from images, or medical fact triples parsed from text are used as the specific content of the event to fill the event payload attributes, generating a structurally complete and information-rich health event object. Here, the event payload is the content carrier of the event; for example, in a "patient visit event," the "event payload" is the specific information of that event.

[0094] Step S305: Combine all health event objects to generate a standardized health event flow.

[0095] For example, the individual health event objects continuously generated in step S304 are strictly ordered according to the event occurrence timestamps recorded internally. After sorting, these event objects are organized into an ordered, continuous data sequence. This sequence is output as a standardized health event stream. Each event in this health event stream has a unified structure and semantic specification.

[0096] In this embodiment, numerical sensor data is sequentially smoothed, denoised, and normalized; deep feature extraction is performed on medical images; and semantic parsing and triple extraction are performed on unstructured text. All these heterogeneous data are then uniformly encapsulated into structured health event objects and aggregated into a standard event stream. This effectively solves the inconsistencies in format, semantics, and granularity of multi-source health data, achieving efficient conversion from raw, mixed data to standardized events that can be uniformly understood and processed by machines.

[0097] In one embodiment, a cross-modal temporal alignment process is performed on a standardized health event stream to generate a temporally sequenced health event sequence, including:

[0098] Step S401: Construct a high-granularity virtual timeline starting from the individual's birth time; the high-granularity virtual timeline serves as a reference benchmark for time sequence alignment.

[0099] For example, a reference frame for time alignment is first determined, using the individual's birth time as the absolute origin. This time point is typically obtained from an identity database or the earliest medical records and converted to Coordinated Universal Time (UTC) format to ensure global consistency. A high-granularity virtual timeline is constructed, which can employ a linearly continuous time representation method. Its smallest time unit can be set to the millisecond or microsecond level according to data precision requirements. The construction of the virtual timeline essentially establishes a unified time scale independent of the local time of any specific data source. This timeline is typically implemented in memory as an ordered linked list or an array data structure with timestamp indexes, supporting efficient range and point lookup operations. The linearly continuous time representation method abstracts time as a straight line without beginning or end, without interruption, and flowing uniformly. Each time point on this line is uniquely marked by a numerical value, and it can continuously describe any moment between two time points, aiming to map the flow of time in the physical world using a continuous real number axis model in mathematics.

[0100] Step S402: Map the occurrence timestamp of each health event object in the standardized health event stream to the virtual timeline to obtain the primary mapped event set.

[0101] For example, a standardized health event stream is traversed, reading the raw timestamp of each health event object. These timestamps may come from different clocks and have different formats. These raw timestamps are first uniformly converted to a high-precision absolute time representation based on UTC (Coordinated Universal Time). The precise time difference, i.e., the time offset, between the absolute timestamp of each event and the origin of the virtual timeline is calculated, and the event is marked at the corresponding position on the virtual timeline. For duration-based events, such as a week-long medication course, the mapping positions of its start and end timestamps on the virtual timeline are recorded to determine a time interval. After mapping all events, a primary mapped event set is obtained. This set contains the positional information of all events on the unified virtual timeline, but at this point, the events are only roughly arranged chronologically.

[0102] Step S403: From the primary mapping event set, select the health event sequence pairs that need to be aligned according to the preset association rules; the health event sequence pairs that need to be aligned include the source event sequence and the target event sequence to be aligned.

[0103] Among them, the preset association rules refer to the rules predefined based on medical knowledge or business logic. For example, the rules may specify that the "medication event sequence" needs to be aligned with the "blood glucose monitoring event sequence" to analyze the efficacy of the drug, or the "exercise event sequence" needs to be aligned with the "heart rate event sequence" to assess the exercise intensity.

[0104] For example, based on preset association rules, source event sequences and target event sequence pairs requiring fine-grained time-series alignment are intelligently identified from the primary mapping event set. The rule engine analyzes metadata such as event type and data source, as well as their proximity relationships on the virtual timeline, automatically pairing sequences with potential causal or correlational relationships. Source event sequences are typically event streams with low sampling frequencies or discrete time points, such as daily medication records; target event sequences are typically high-frequency sampled continuous data streams, such as heart rate monitoring data every minute. The rule engine is a tool that automatically makes decisions based on preset logic, extracting business rules (such as "if condition A is met, execute operation B") from the code, allowing non-technical personnel to configure and adjust rules without modifying the underlying program.

[0105] Step S404: For the health event sequence pairs that need to be aligned, the curved path with the minimum cumulative distance between the two sequences is determined by the improved dynamic time warping algorithm.

[0106] Dynamic Time Warping (DTW) is an algorithm that uses dynamic programming to find a curved path that minimizes the cumulative distance between two sequences, allowing the sequences to be non-linearly stretched or compressed in the time dimension to achieve optimal alignment. The improved DTW algorithm is enhanced by using a composite metric, d(i,j), which is not a simple Euclidean distance but incorporates event semantic similarity. This metric considers whether the event types are the same, the differences in key values ​​within the event payloads, and the strength of semantic associations between event types based on a medical knowledge graph. A medical knowledge graph is a database that stores key information about diseases, symptoms, drugs, examinations, and treatments in the medical field in a structured "entity-relationship" format, such as "pneumonia → typical symptoms → fever".

[0107] For example, for each selected pair of health event sequences that needs alignment, an improved dynamic time warping algorithm is applied. This algorithm iteratively calculates the cumulative distance D(i,j), and its recursive formula is as follows:

[0108]

[0109] in, The minimum cumulative distance from the start of the sequence to the i-th point of the source sequence and the j-th point of the target sequence. Let be the local distance between the i-th event point in the source sequence and the j-th event point in the target sequence, where i is the i-th event point in the source sequence and j is the j-th event point in the target sequence. The choice is based on the minimum cumulative distance to the previous state. The source sequence moves forward one step, while the target sequence stops (corresponding to stretching the source sequence). As the target sequence moves forward one step, the source sequence stops (corresponding to the compressed source sequence). To synchronize the two sequences (align them), for example, the process eventually backtracks to find the optimal curved path that minimizes the total cumulative distance. This optimal curved path precisely describes how to temporally and curvedly map each event point in the source sequence to one or more event points in the target sequence.

[0110] Step S405: Based on the curved path, align each event point in the source event sequence to be aligned to the corresponding time point in the target event sequence to generate a preliminary aligned event set.

[0111] For example, after obtaining the optimal curved path, each event point in the source event sequence is repositioned to a new time position on the virtual time axis that matches the corresponding point in the target sequence, based on the mapping relationship provided by the curved path. For instance, the curved path might indicate that the i-th event in the source sequence (such as a medication administration) should be associated with the j-th to k-th events in the target sequence (such as heart rate changes over a period of time after medication administration). In this case, the timestamp of the medication administration event is adjusted to the time point of the j-th event in the target sequence, or a new time point is created, and the medication administration event is associated with the heart rate events from j to k. After aligning all sequence pairs, all aligned events are merged with events that did not require complex alignment or events already included in the target sequence to form a preliminary aligned event set.

[0112] Step S406: Sort the preliminary aligned event set according to the virtual timeline to obtain a time-series health event sequence.

[0113] For example, all events in the initially aligned event set are globally and strictly sorted in ascending order according to their new positions on the virtual timeline (timestamps adjusted by the improved DTW alignment). This ensures that all events are arranged in a uniform, high-precision chronological order. The final output is a time-seriesd sequence of health events.

[0114] In this embodiment, a unified virtual timeline benchmark is established to initially map and locate standardized events. Key sequence pairs are selected based on medical association rules, and an improved DTW algorithm is used for nonlinear bending alignment to generate a globally ordered temporal event sequence. This effectively solves the complex temporal asynchrony problem caused by differences in collection frequency, time asynchrony, and physiological response delays in multi-source health data.

[0115] In one embodiment, a temporal knowledge graph is constructed; and based on the temporal knowledge graph and a temporally sequenced health event sequence, a high-dimensional health state vector is updated using a graph neural network model to obtain an updated high-dimensional health state vector, including:

[0116] Step S501: Construct a temporal knowledge graph; the temporal knowledge graph includes an event layer and a state layer.

[0117] The event layer is specifically used to store and connect all health event nodes extracted from the time-series health event sequence. Each node is an entity with rich attributes, including at least a unique event identifier, a timestamp of the event occurrence, an event type classification code, a data source system identifier, and specific event payload content. The state layer is used to maintain one or more core nodes that represent the overall health status of an individual. These nodes do not directly correspond to a specific event, but are an abstraction and summary of the health status over a period of time.

[0118] For example, a temporal knowledge graph is constructed, which uses an attribute graph model as its basic data structure and is explicitly divided into two logical layers: an event layer and a state layer. The two layers are connected by directed edges, which not only represent temporal order relationships but may also carry semantic labels describing how events affect health status, such as causing, mitigating, and associating.

[0119] Step S502: The initial health event in the time-series health event sequence is determined as the initial health event node in the event layer of the time-series knowledge graph.

[0120] For example, the generated time-series health event sequence is scanned, and the initial health event is identified according to preset rules, such as selecting the event with the earliest timestamp or the first event with complete baseline information, such as a birth record. A node instance corresponding to this initial event is created in the event layer of the knowledge graph. The creation process includes assigning a unique in-graph node ID to the node and filling the node's attribute set with all event attributes, such as event ID, timestamp, type, and payload, in key-value pairs. The newly created initial health event node serves as the temporal starting point for the entire knowledge graph at the event layer. Information about the successful creation of this node is recorded and used as the entry point for subsequent state layer construction and graph traversal operations.

[0121] Step S503: Based on the baseline information of the initial health event, create a central health state node representing the individual in the state layer of the temporal knowledge graph; and generate the current high-dimensional health state vector based on the central health state node.

[0122] For example, the payload information carried in the initial health event node is parsed to extract baseline data that can be used to describe an individual's initial health state. A central health state node is created in the state layer of the knowledge graph, and this node also has its own unique node ID and attribute set. Based on the extracted baseline information, a high-dimensional numerical vector, i.e., the initial high-dimensional health state vector, is generated through predefined mapping rules. Each dimension of this vector represents a specific aspect of the health state, such as cardiovascular function, metabolic level, etc., and the value of the dimension is derived from the baseline information. This initial vector is stored as a key attribute of the central health state node. At the same time, a directed edge is established from the initial health event node to the central health state node, and the attribute of the edge records that this state vector was initialized by the initial event. Here, the predefined mapping rules refer to the pre-agreed input-output correspondence, for example: input: raw health-related information, which may be discrete, such as whether or not one smokes; output: the initial high-dimensional health state vector, such as [blood pressure value, blood sugar value, smoking identifier, allergy type 1,...], where each element is a numerical value, and the dimension depends on the number of features.

[0123] Step S504: When a new health event is received, the current high-dimensional health state vector is updated using a graph neural network model based on a predefined clinical pathway and causal reasoning rule base, resulting in an updated high-dimensional health state vector. The new health event object includes subsequent events within a time-series health event sequence and newly added emergencies in real time.

[0124]

[0125] in, This is the updated high-dimensional health state vector. Let W be the high-dimensional health state vector before the update, W be the learnable weight matrix, || be the vector concatenation operation, and AGGREGATE be the aggregation function. The attention weight of the central healthy state node i at time t. Let σ be the feature representation of neighbor node i, and let σ be the activation function.

[0126] Among them, AGGREGATE is an operation used to summarize the information of neighboring nodes. Common methods include summation, averaging, or taking the maximum value. In graph neural networks, it is responsible for aggregating the feature vectors of multiple neighboring nodes into a single, fixed-length vector, representing the collective information from the neighbors. This can be calculated through a trainable attention mechanism, for example: based on the correlation between the features of node i and the features of the center node, the higher the weight, the greater the influence of the neighboring node.

[0127] For example, new health events are continuously monitored and received, including subsequent events in a time-series event sequence and real-time pushed emergencies. Whenever a new event arrives, a corresponding new event node is created in the graph event layer and connected to existing nodes in chronological order. The rule engine then searches for a match in a predefined clinical pathway and causal reasoning rule base. Successfully matched rules are activated, collectively defining the logic for state updates in the current context. A pre-trained graph neural network model is invoked. This model, based on traditional graph neural networks for processing graph-structured data, incorporates an attention mechanism to achieve differentiated information aggregation. The model has undergone parameter optimization through data training and can be directly used for inference tasks. The vector representation of the current central health state node, the feature vector of the new event node, and information about their neighboring nodes (such as preceding event nodes) in the current graph are input into the graph neural network model. By utilizing attention weights to distinguish the importance of information from different neighboring nodes, the updated high-dimensional health state vector is calculated according to a formula. Finally, the updated high-dimensional health state vector is used to overwrite the current state vector attribute of the central health state node, completing the update. An influence edge is established from the new event node to the state node, and metadata such as the update timestamp is recorded.

[0128] In this embodiment, by constructing and dynamically maintaining a two-layer event-state temporal knowledge graph, discrete health events are organized at the event layer, while a centralized, vectorized health state representation is maintained at the state layer. This ensures that each new event triggers a computable and interpretable state update process based on clinical rules and graph neural networks. This approach captures and utilizes the complex spatiotemporal relationships and causal semantics between health events, providing a powerful cognitive engine for deeply understanding the dynamic evolution of individual health in scenarios such as precision medicine and chronic disease management. This significantly improves the intelligence and accuracy of health state assessment and prediction.

[0129] In one embodiment, based on risk prediction results and the current high-dimensional health state vector, a large language model is used to generate health intervention recommendations, including:

[0130] Step S601: Construct a structured prompt template; the structured prompt template includes a health status summary, a description of the risk prediction results, the target user role, and the application scenario.

[0131] For example, a structured cue word template is created and stored. This template is a text frame containing multiple predefined placeholders. These placeholders are carefully designed to guide a large language model to generate health intervention recommendations that meet specific needs. The key placeholders preset in the template include, but are not limited to: Used to insert a summary of key health indicators interpreted from a high-dimensional health state vector; Used to describe the specific risk prediction results output by the Long Short-Term Memory Network Model, including risk type, probability, and urgency; Used to specify the identity of the intended recipient of the advice, such as a clinician, insurance underwriter, health manager, or the patient themselves; This is used to clarify the application context of the recommendations, such as insurance product pricing, clinical follow-up plan development, or personal daily health management. By providing rich contextual information and clear role definitions, the generation direction of large language models is constrained.

[0132] Step S602: Fill the prompt word template with the clinical significance interpretation of the current high-dimensional health status vector, the specific content of the risk prediction results, the target user identity, and the suggested application scenario information to obtain the filled prompt words.

[0133] For example, a clinical knowledge parser can be used to interpret the current high-dimensional health status vector. This parser, based on a predefined medical knowledge graph or mapping rules, transforms the values ​​and trends of key dimensions in the vector into clinically meaningful natural language descriptions, such as: a fasting blood glucose level consistently above 7.0 mmol / L indicates poor blood glucose control. The resulting risk prediction is then converted into easily understandable text descriptions, such as: the risk level for developing diabetic microvascular complications within the next three months is moderate, with a probability of approximately 30%. Based on the context, the target user role (e.g., a health management specialist at an insurance company) and specific application scenario (e.g., designing a health incentive program for people with prediabetes) are determined. These processed text messages are then used to replace the corresponding placeholders in the prompt template, generating a filled prompt string.

[0134] Step S603: Input the filled-in prompt words into a pre-trained large-scale language model to generate a first draft of natural language text; the first draft of natural language text conforms to grammar and medical professionalism.

[0135] Among them, the pre-trained large-scale language model is based on the internal parameters and knowledge obtained from training on massive text data such as medical literature and clinical guidelines. It generates subsequent text token by token in an autoregressive manner. Autoregression is a statistical and machine learning approach that predicts the future based on the historical information of the data itself, using the past to predict the present / future.

[0136] For example, the filled-in prompts are sent as input text to a pre-trained large language model. Utilizing the medical knowledge learned by the model, the accuracy of the generated draft text in terms of medical facts is ensured, such as correctly mentioning relevant drug names, examination items, or lifestyle interventions. The generated draft text is a coherent natural language paragraph or list whose content directly responds to the requirements set in the prompts, forming a preliminary draft of professional health intervention recommendations.

[0137] Step S604: Standardize and post-process the initial draft of the natural language text to obtain health intervention recommendations.

[0138] For example, necessary post-processing of the original natural language text draft includes the following steps: secondary verification of medical facts, which utilizes a smaller rule base specializing in medical knowledge verification to quickly validate key medical statements mentioned in the text (such as drug dosage); text normalization, which eliminates potentially ambiguous expressions, colloquial language, or repetitive content to ensure that the recommendations are clear, concise, and unambiguous; formatting adjustment, which organizes the text into an easy-to-read structure according to predefined output specifications, such as bullet points, bold headings, or standard paragraphs; and sensitive information filtering, which checks for and removes any information that may involve privacy or is inappropriate to disclose in the current context. After these post-processing steps, a standardized health intervention recommendation document that can be directly delivered is output. Among them, rule bases specializing in medical knowledge verification provide verification standards for the accuracy, compliance, and clinical applicability of medical information. They typically integrate authoritative medical data sources and structure them into rules that can be directly invoked, such as evidence-based medicine evidence, drug / treatment guidelines, basic medical logic, and terminology standards. These rules are used to address the problems of high risk of information distortion in the medical field (such as incorrect medication recommendations), high professional threshold (requiring integration with clinical guidelines), and significant consequences (directly related to health and safety).

[0139] In this embodiment, health status, risk prediction, user roles, and scenario information are organized by designing prompt word templates. The powerful generation capabilities of a large language model are used to generate preliminary professional text. After standardization and post-processing, high-quality, actionable health intervention recommendations are output. This improves the personalization level, generation efficiency, and operability of health management recommendations, effectively addressing the bottleneck problems of insufficient standardization, high personalization costs, and heavy reliance on human experience in traditional health management intervention recommendations.

[0140] In one exemplary embodiment, the method further includes:

[0141] Step S701: Collect medical expense data; and standardize the medical expense data to generate standardized expense event records; the standardized expense event records include expense item name, expense amount, payment status, payer type, service provider, and expense occurrence timestamp.

[0142] For example, raw medical expense data is collected. This raw data typically contains different field naming conventions, diverse data formats, and non-standardized medical item names. The medical expense data undergoes data cleaning, including removing duplicate records, filling in missing values ​​in key fields, and correcting obvious formatting errors. Subsequently, a medical expense terminology standardization engine can be invoked. This engine, based on standard medical coding systems such as the International Classification of Diseases and the Classification and Code of Medical Services Operations, maps the raw expense item names to unified standard terms. Simultaneously, all currency amounts are converted to a base currency unit (e.g., RMB), and all date and timestamps are converted to Coordinated Universal Time (UTC) format. The processed data is encapsulated into standardized expense event records, whose structured fields include standardized expense item names, expense amounts in unified currency units, payment status (e.g., paid, unpaid, pending claim), payer type (e.g., out-of-pocket, basic medical insurance, commercial insurance), standard code of the service provider, and precise expense occurrence timestamp.

[0143] Step S702: The standardized expense event records are matched with the time-series health event sequences to establish a network of association between health events and expense records.

[0144] In this network of relationships, nodes represent health events or expense records, and edges represent the relationships between them. The attributes of the edges can describe the type of relationship (such as "caused") and the strength of the relationship. The strength of the relationship can be calculated based on temporal proximity and semantic matching.

[0145] For example, based on temporal proximity and semantic relevance of event content, standardized expense event records are associated and matched with previously constructed time-series health event sequences. First, for each expense record, all health events occurring within a time window are searched in the time-series health event sequence. The specific window size is configurable. Using predefined association rules and a medical knowledge graph, it is analyzed whether there is a logical causal or accompanying relationship between the content of these health events and the medical services described in the expense record. For example, an expense record for "coronary angiography" will be associated with health events in the sequence that are close in time, such as "chest pain consultation." This forms a semantically rich network of associations connecting health events and expense records. The predefined association rules are logical associations pre-set based on medical common sense and historical data. For example, prescribing "antibiotics" is usually associated with "bacterial infection diagnosis," using known patterns to anchor the direction of analysis.

[0146] Step S703: Based on the correlation network between health events and expense records, construct a multi-dimensional expense data graph; the multi-dimensional expense data graph uses time as the vertical axis and payers, service providers, and expense categories as dimensions.

[0147] For example, a cost data graph model for multidimensional analysis is constructed based on a network of relationships. This model is a three-dimensional data cube structure. Its vertical axis (row dimension) is the time axis, and the time granularity can be selected according to the analysis needs, such as aggregation by day, week, month, or year. Its dimensions include: payer dimension, service provider dimension, and cost category dimension. By traversing the cost event records in the network of relationships, and based on the cost occurrence timestamp, payer type, service provider, and the category to which the standardized cost item belongs, each record is merged into the corresponding multidimensional cell of the data cube.

[0148] Step S704: Calculate the total cost and the cost percentage of each dimension in the multi-dimensional cost data graph to obtain the calculation results; and generate the Sankey diagram data source for cost flow based on the calculation results.

[0149] For example, the constructed multi-dimensional cost data graph is sliced, diced, and rotated according to different dimensional combinations. The cost amounts within the target cell are aggregated and calculated; for example, calculating the total cost from a specific payer to various service providers within a certain time period. Simultaneously, the percentage of each component is calculated; for example, calculating the percentage of a payer's total spending on a particular service provider. These calculation results (total amount and percentage) are organized into a specific data structure, which is the necessary data source for generating the Sankey diagram. The Sankey diagram data source typically contains two core types of information: a list of nodes, listing all involved payers, service providers, and further subdivided cost categories; and a list of links, detailing the flow of funds between nodes, with each link containing the source node, target node, and the amount transferred (i.e., flow rate).

[0150] Step S705: Visualize the Sankey diagram data source to generate a Sankey diagram of medical expense flow. The Sankey diagram of medical expense flow is used to show the path and flow ratio of funds from different payers to various service providers and broken down to specific expense items. It supports filtering the expense flow within a specific time period by time slider.

[0151] For example, a visualization rendering engine is invoked to draw strip-shaped flow lines of varying widths based on the proportion of link flow values, connecting nodes representing payers, service providers, and expense categories. The width of the flow lines visually reflects the size of the fund flow, and colors are typically used to distinguish different categories, such as using different colors for different payers. A time slider control can be integrated into the user interface, which is bound to the time axis of the multi-dimensional expense data graph. When the user drags the time slider to select different time periods, the aggregated expense data within that time period is dynamically recalculated, the Sankey diagram data source is updated, and the visualization engine is triggered to redraw, displaying the expense flow path and flow changes within the selected time period in real time. The generated Sankey diagram of medical expense flow provides a global perspective, revealing the source, destination, and allocation proportion of medical funds at each stage, and supports interactive exploration. The visualization rendering engine is a technical component that transforms digital scenes into images or videos that humans can intuitively perceive; it is a translation and drawing tool from data to visual presentation. The time slider control is a common graphical user interface interactive component used to allow users to intuitively select, locate, or control time-related content through sliding operations.

[0152] In this embodiment, raw cost data undergoes deep cleaning and standardization, and is then precisely spatiotemporally and semantically correlated with health event sequences. This constructs a data cube supporting multi-dimensional analysis, enabling the calculation of quantitative relationships in fund flows and generating a Sankey diagram data source. A dynamic, interactive visualization interface visually displays the complete flow path of medical expenses. This transforms previously scattered and isolated cost records into a clear global view revealing the logic of money flowing with illness, enhancing the understanding of medical consumption structures among medical institutions, insurance companies, businesses, and individuals. It provides powerful data insight support for cost control, resource optimization, insurance product design, and personal health financial management, effectively addressing the pain points of traditional health record systems: the disconnect between cost information and health events, and the lack of macro-level visualization analysis capabilities.

[0153] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0154] Based on the same inventive concept, this application also provides a cloud-based device for automatically updating and managing dynamic health records on a cloud platform, used to implement the aforementioned automatic updating and management of dynamic health records on a cloud platform. The solution provided by this device is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more embodiments of the cloud-based device for automatically updating and managing dynamic health records provided below can be found in the limitations of the cloud-based method for automatically updating and managing dynamic health records above, and will not be repeated here.

[0155] In one exemplary embodiment, such as Figure 3 As shown, a cloud-based device 300 for automatic updating and management of dynamic health records is provided, including:

[0156] The data acquisition module 301 is used to acquire multi-source heterogeneous health data and perform event-based encapsulation processing on the multi-source heterogeneous health data to generate a standardized health event stream.

[0157] The sequence generation module 302 is used to perform cross-modal temporal alignment processing on the standardized health event stream to generate a temporalized health event sequence.

[0158] The vector update module 303 is used to construct a temporal knowledge graph; and based on the temporal knowledge graph and the temporalized health event sequence, it updates the high-dimensional health state vector through a graph neural network model to obtain the updated high-dimensional health state vector; the high-dimensional health state vector is used to represent the overall health status of an individual.

[0159] The file update module 304 is used to update the dynamic health file based on the updated high-dimensional health state vector to obtain the updated dynamic health file.

[0160] In one exemplary embodiment, the apparatus further includes:

[0161] The sequence composition module is used to extract high-dimensional health state vectors within a preset time range from the updated dynamic health records to form a health state evolution sequence.

[0162] The risk prediction module is used to input the health status evolution sequence into the trained long short-term memory network model and output the risk prediction result.

[0163] The suggestion generation module is used to generate health intervention suggestions based on risk prediction results and the current high-dimensional health status vector, using a large language model.

[0164] In one embodiment, the data acquisition module 301 is further used for

[0165] Numerical sensor data in multi-source heterogeneous health data are smoothed and denoised to obtain cleaned regular integer value data.

[0166] Quantitative feature vectors of medical image data from multi-source heterogeneous health data are extracted using a pre-trained convolutional neural network model; the quantitative feature vectors include electrocardiogram waveform features and medical image features.

[0167] For unstructured text data in multi-source heterogeneous health data, structured medical fact triples are generated through a large language model;

[0168] The cleaned integer value data, quantitative feature vectors, and structured medical fact triples are encapsulated according to preset event object rules to generate health event objects; the health event object includes event type, occurrence time, data source, and event payload;

[0169] All health event objects are combined to generate a standardized health event stream.

[0170] In one embodiment, the sequence generation module 302 is further configured to:

[0171] Starting with an individual's birth time, a high-granularity virtual timeline is constructed; this high-granularity virtual timeline serves as a reference benchmark for time sequence alignment.

[0172] The occurrence timestamp of each health event object in the standardized health event stream is mapped to a virtual timeline to obtain a primary mapped event set;

[0173] From the primary mapping event set, select the health event sequence pairs that need to be aligned according to the preset association rules; the health event sequence pairs that need to be aligned include the source event sequence and the target event sequence to be aligned;

[0174] For health event sequence pairs that need to be aligned, the curved path with the minimum cumulative distance between the two sequences is determined by an improved dynamic time warping algorithm.

[0175] Based on the curved path, each event point in the source event sequence to be aligned is aligned to the corresponding time point in the target event sequence to generate a preliminary aligned event set;

[0176] The initial aligned event set is sorted according to the virtual timeline to obtain a time-series health event sequence.

[0177] In one embodiment, the vector update module 303 is further configured to:

[0178] Constructing a temporal knowledge graph; the temporal knowledge graph includes an event layer and a state layer;

[0179] The initial health event in the time-series health event sequence is identified as the initial health event node in the event layer of the time-series knowledge graph;

[0180] Based on the baseline information of the initial health events, a central health state node representing an individual is created in the state layer of the temporal knowledge graph; and a current high-dimensional health state vector is generated based on the central health state node.

[0181] When a new health event is received, the current high-dimensional health state vector is updated using a graph neural network model based on predefined clinical pathways and a causal reasoning rule base, resulting in an updated high-dimensional health state vector. New health event objects include subsequent events within a time-series of health events and newly added emergencies in real time.

[0182]

[0183] in, This is the updated high-dimensional health state vector. Let W be the high-dimensional health state vector before the update, W be the learnable weight matrix, || be the vector concatenation operation, and AGGREGATE be the aggregation function. The attention weight of the central healthy state node i at time t. Let σ be the feature representation of neighbor node i, and let σ be the activation function.

[0184] In one embodiment, it is suggested that the generation module is further configured to:

[0185] Construct a structured prompt word template; the structured prompt word template includes a health status summary, a description of the risk prediction results, the target user role, and the application scenario;

[0186] The prompt word template is filled in based on the clinical significance interpretation of the current high-dimensional health status vector, the specific content of the risk prediction results, the target user identity, and the suggested application scenario information to obtain the filled prompt words;

[0187] The filled-in prompts are input into a pre-trained large-scale language model to generate a first draft of natural language text; the first draft of natural language text conforms to grammar and medical professionalism.

[0188] Standardized post-processing of the initial natural language text yields health intervention recommendations.

[0189] In one exemplary embodiment, the apparatus further includes:

[0190] The record generation module is used to collect medical expense data and standardize the medical expense data to generate standardized expense event records. The standardized expense event records include the expense item name, expense amount, payment status, payer type, service provider, and expense occurrence timestamp.

[0191] The relationship network building module is used to associate and match standardized expense event records with time-series health event sequences to establish a relationship network between health events and expense records;

[0192] The data graph construction module is used to build a multi-dimensional cost data graph based on the relationship network between health events and cost records. The multi-dimensional cost data graph uses time as the vertical axis and payer, service provider, and cost category as dimensions.

[0193] The data source generation module is used to calculate the total cost and the cost percentage of each dimension in the multi-dimensional cost data graph, obtain the calculation results, and generate the Sankey diagram data source for cost flow based on the calculation results.

[0194] The data source visualization module is used to visualize the Sankey diagram data source and generate a Sankey diagram of medical expense flow. The Sankey diagram of medical expense flow is used to show the flow path and flow ratio of funds from different payers to various service providers and broken down to specific expense items. It supports filtering the expense flow within a specific time period by time slider.

[0195] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the cloud-based dynamic health record automatic update and management method described above.

[0196] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0197] For the device 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 components described as separate parts 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 disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0198] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for automatic updating and management of dynamic health records based on a cloud platform, characterized in that: The method includes: Acquire multi-source heterogeneous health data; and encapsulate the multi-source heterogeneous health data into events to generate a standardized health event stream; The standardized health event stream is subjected to cross-modal temporal alignment processing to generate a temporally sequenced health event sequence. A temporal knowledge graph is constructed; and based on the temporal knowledge graph and the temporalized health event sequence, a high-dimensional health state vector is updated using a graph neural network model to obtain an updated high-dimensional health state vector; the high-dimensional health state vector is used to represent the overall health status of an individual. The dynamic health profile is updated based on the updated high-dimensional health state vector to obtain the updated dynamic health profile.

2. The method according to claim 1, characterized in that, The method further includes: The high-dimensional health state vector within a preset time range is extracted from the updated dynamic health record to form a health state evolution sequence; The health status evolution sequence is input into a trained long short-term memory network model, and the risk prediction result is output. Based on the risk prediction results and the current high-dimensional health status vector, health intervention suggestions are generated using a large language model.

3. The method according to claim 1, characterized in that, The step of encapsulating the multi-source heterogeneous health data into an event-based format to generate a standardized health event stream includes: The numerical sensor data in the multi-source heterogeneous health data is smoothed and denoised to obtain cleaned regular integer value data. Quantitative feature vectors of medical image data from the multi-source heterogeneous health data are extracted using a pre-trained convolutional neural network model; the quantitative feature vectors include electrocardiogram waveform features and medical image features. For the unstructured text data in the multi-source heterogeneous health data, structured medical fact triples are generated using a large language model; The cleaned integer value data, the quantitative feature vector, and the structured medical fact triplet are encapsulated according to preset event object rules to generate a health event object; the health event object includes event type, occurrence time, data source, and event payload; All the health event objects are combined to generate a standardized health event stream.

4. The method according to claim 1, characterized in that, The process of performing cross-modal temporal alignment on the standardized health event stream to generate a temporally sequenced health event sequence includes: A high-granularity virtual timeline is constructed, starting from the individual's birth time; this high-granularity virtual timeline serves as a reference benchmark for time sequence alignment. The occurrence timestamp of each health event object in the standardized health event stream is mapped to the virtual timeline to obtain the primary mapped event set; From the primary mapping event set, health event sequence pairs that need to be aligned are selected according to preset association rules; the health event sequence pairs that need to be aligned include the source event sequence and the target event sequence to be aligned. For the health event sequence pairs that need to be aligned, the curved path with the minimum cumulative distance between the two sequences is determined by an improved dynamic time warping algorithm. Based on the curved path, each event point in the source event sequence to be aligned is aligned to the corresponding time point in the target event sequence to generate a preliminary aligned event set; The preliminary aligned event set is sorted according to the virtual timeline to obtain a time-series health event sequence.

5. The method according to claim 1, characterized in that, The process involves constructing a temporal knowledge graph; and based on the temporal knowledge graph and the temporally sequenced health event sequence, updating the high-dimensional health state vector using a graph neural network model to obtain the updated high-dimensional health state vector, including: Construct a temporal knowledge graph; the temporal knowledge graph includes an event layer and a state layer; The initial health event in the time-series health event sequence is determined as the initial health event node in the event layer of the time-series knowledge graph; Based on the baseline information of the initial health event, a central health state node representing the individual is created in the state layer of the temporal knowledge graph; and a current high-dimensional health state vector is generated based on the central health state node. When a new health event is received, the current high-dimensional health state vector is updated using a graph neural network model based on a predefined clinical pathway and causal reasoning rule base, resulting in an updated high-dimensional health state vector. The new health event object includes subsequent events within the time-series health event sequence and newly added emergencies in real time. in, This is the updated high-dimensional health state vector. Let W be the high-dimensional health state vector before the update, W be the learnable weight matrix, || be the vector concatenation operation, and AGGREGATE be the aggregation function. The attention weight of the central healthy node i at time t. Let σ be the feature representation of neighbor node i, and let σ be the activation function.

6. The method according to claim 2, characterized in that, The process of generating health intervention recommendations based on the risk prediction results and the current high-dimensional health state vector using a large language model includes: Construct a structured prompt word template; the structured prompt word template includes a health status summary, a description of risk prediction results, the target user role, and the application scenario; The prompt word template is filled in based on the clinical significance interpretation of the current high-dimensional health status vector, the specific content of the risk prediction results, the target user identity, and the suggested application scenario information to obtain the filled prompt words; The filled-in prompts are input into a pre-trained large-scale language model to generate a first draft of natural language text; the first draft of natural language text conforms to grammar and medical professionalism. The initial draft of the natural language text was standardized and post-processed to obtain health intervention recommendations.

7. The method according to claim 1, characterized in that, The method further includes: Collect medical expense data; and standardize the medical expense data to generate standardized expense event records; the standardized expense event records include expense item name, expense amount, payment status, payer type, service provider, and expense occurrence timestamp; The standardized expense event records are associated and matched with the time-series health event sequences to establish a network of association between health events and expense records; Based on the correlation network between the health events and expense records, a multi-dimensional expense data graph is constructed; the multi-dimensional expense data graph uses time as the vertical axis and payer, service provider, and expense category as dimensions. Calculate the total cost and the cost percentage under each dimension in the multi-dimensional cost data graph to obtain the calculation results; and generate a Sankey diagram data source for cost flow based on the calculation results. The Sankey diagram data source is visualized to generate a Sankey diagram of medical expense flow. The Sankey diagram of medical expense flow is used to show the flow path and flow ratio of funds from different payers to various service providers, broken down to specific expense items. It supports filtering the expense flow within a specific time period by a time slider.

8. A cloud-based device for automatic updating and management of dynamic health records, characterized in that: The device includes: The data acquisition module is used to acquire multi-source heterogeneous health data and perform event-based encapsulation processing on the multi-source heterogeneous health data to generate a standardized health event stream. The sequence generation module is used to perform cross-modal temporal alignment processing on the standardized health event stream to generate a temporally sequenced health event sequence. The vector update module is used to construct a temporal knowledge graph; and based on the temporal knowledge graph and the temporalized health event sequence, it updates the high-dimensional health state vector through a graph neural network model to obtain the updated high-dimensional health state vector; the high-dimensional health state vector is used to represent the overall health status of an individual. The file update module is used to update the dynamic health file based on the updated high-dimensional health state vector to obtain the updated dynamic health file.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images