An ai-based cardiovascular chronic disease data management method

Abstract

This invention relates to the field of medical data technology and discloses an AI-based method for managing cardiovascular chronic disease data. The method constructs a cardiovascular data pool containing patients' historical medical records and real-time physiological signals; establishes a dynamic ontology in the cardiovascular data pool based on a clinical knowledge graph, which defines the semantic relationships between different data entities; maps and generates data evolution trajectories on the dynamic ontology according to a preset treatment target path; activates an intelligent agent configured to traverse the dynamic ontology along the data evolution trajectory, resolving causal constraints between data entities based on the semantic relationships encountered during the traversal; the intelligent agent adaptively reorganizes new data flowing into the cardiovascular data pool in real time according to the resolved causal constraints, generating data slices with temporal correlation; inputs the data slices into a pre-trained deep learning network for feature fusion, outputting a dynamic profile of the patient's cardiovascular health status.

Description

Technical Field

[0001] This invention relates to the field of medical data technology, specifically to an AI-based method for managing cardiovascular chronic disease data. Background Technology

[0002] Current cardiovascular chronic disease data management primarily employs electronic medical record storage and simple statistical analysis. Existing technologies for patient data management are largely limited to structured storage, failing to effectively correlate historical treatment data with real-time physiological signals. Data organization is simplistic, lacking the definition and utilization of semantic relationships between medical concepts. Treatment process tracking methods are crude, unable to establish a mapping relationship between treatment goals and data changes. Data analysis processes are static, making it difficult to uncover hidden causal constraints within data entities. New data integration mechanisms are rigid, unable to dynamically reorganize real-time incoming data according to medical logic. Feature extraction methods are limited, and deep learning model training relies on fixed-format data inputs. Health status assessments are linear, resulting in output assessments lacking dynamism and interpretability. Existing methods need to address key technical challenges such as data semantic modeling, treatment path mapping, causal relationship mining, and dynamic feature fusion.

[0003] Traditional cardiovascular data management systems suffer from significant shortcomings in terms of intelligence and precision. Data pool construction is simplistic, resulting in low standardization of multi-source, heterogeneous data. Knowledge graph applications are superficial, with incomplete definitions of relationships between clinical concepts. Ontology construction methods are static and rigid, unable to adapt to updates in medical knowledge and individual patient differences. Treatment target pathways are set based on experience, lacking quantitative indicators and standardized definitions. Data evolution trajectory generation is linear, failing to consider the non-linear characteristics of disease development. Intelligent agent decision-making logic is simplistic, with limited semantic relationship parsing capabilities. Causal constraint mining methods are crude, confusing statistical correlation with medical causality. Data reorganization algorithms are inefficient, leading to high latency in large-scale real-time data processing. Feature fusion network structures are complex, prone to gradient vanishing or exploding during training. Dynamic profile updates are infrequent, failing to reflect changes in health status in real time. Existing technologies necessitate the establishment of a comprehensive intelligent management solution encompassing data organization and status assessment. Summary of the Invention

[0004] The purpose of this invention is to provide an AI-based method for managing cardiovascular chronic disease data, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an AI-based method for managing cardiovascular chronic disease data, the method comprising:

[0006] Construct a cardiovascular data pool that includes patients' historical medical records and real-time physiological signals;

[0007] A dynamic ontology is established in the cardiovascular data pool based on a clinical knowledge graph. The dynamic ontology is used to define the semantic relationships between different data entities.

[0008] Based on the preset treatment target path, a data evolution trajectory is generated and mapped on the dynamic ontology;

[0009] A smart agent is initiated, which is configured to traverse the dynamic ontology along the data evolution trajectory and resolve causal constraints between data entities based on the semantic relationships encountered during the traversal.

[0010] The intelligent agent adaptively reorganizes the new data flowing into the cardiovascular data pool in real time based on the causal constraints obtained from the analysis, generating a set of data slices with temporal correlation.

[0011] The data slices are input into a pre-trained deep learning network for feature fusion, outputting a dynamic profile of the patient's cardiovascular health status.

[0012] Preferably, the step of mapping and generating a data evolution trajectory on the dynamic ontology according to a preset treatment target path includes:

[0013] Extract key milestone events from the treatment target path; anchor each key milestone event to a corresponding data entity in the dynamic ontology; calculate the optimal path connecting the anchored data entities based on the semantic relationships defined in the dynamic ontology; and instantiate the optimal path as the data evolution trajectory.

[0014] Preferably, the calculation of the optimal path connecting the anchored data entities includes:

[0015] Different weight values ​​are assigned to different types of semantic relations in the dynamic ontology; a graph traversal algorithm is used to find a path with the minimum total weight, which passes through all anchored data entities in sequence; all data entities and semantic relation sequences passed through by the path with the minimum total weight are recorded.

[0016] Preferably, the intelligent agent adaptively reorganizes new data flowing into the cardiovascular data pool in real time based on the parsed causal constraints, including:

[0017] The intelligent agent monitors the cardiovascular data pool and captures new data; identifies the data entity type to which the new data belongs in the dynamic ontology; retrieves other data entities that have direct causal constraints with the data entity type; determines whether the new data meets the conditions for triggering data reorganization based on the retrieved causal constraint rules; when the conditions are met, extracts historical data related to the causal constraint rules from the cardiovascular data pool and splices and encapsulates it with the new data in chronological order.

[0018] Preferably, the method further includes:

[0019] A policy library is maintained for the intelligent agent, and the policy library contains various data reorganization patterns;

[0020] After determining whether the new data meets the conditions for triggering data reorganization, the method further includes: selecting a matching data reorganization pattern from the strategy library based on the characteristics of the new data and the triggered causal constraint rules; extracting historical data related to the causal constraint rules based on the selected data reorganization pattern, and splicing and encapsulating it with the new data in chronological order.

[0021] Preferably, the step of inputting the data slices into a pre-trained deep learning network for feature fusion includes:

[0022] The deep learning network contains multiple parallel feature extraction branches, each of which specializes in processing a type of data slice. The output of each feature extraction branch is passed to an attention mechanism layer, which dynamically calculates the importance scores of the output features from different branches. Based on the importance scores, the features from different branches are weighted and summed to generate a fused feature vector.

[0023] Preferably, the method further includes:

[0024] The fused feature vector is input into a state decoder, which maps the fused feature vector into a multi-dimensional health state vector, where each dimension of the health state vector corresponds to a quantitative assessment value of a cardiovascular health risk indicator.

[0025] Preferably, the method further includes:

[0026] An intervention knowledge base is established, which stores various interventions and their expected impact patterns on the health state vector; the health state vector is compared with an ideal state vector to generate a difference vector; and the interventions or combinations of interventions that minimize the difference vector are queried from the intervention knowledge base.

[0027] Preferably, the step of querying the intervention knowledge base for interventions or combinations of interventions that minimize the difference vector includes:

[0028] A content-based recommendation algorithm is used to calculate the similarity between the difference vector and the influence pattern of each intervention in the intervention knowledge base; several interventions with the highest similarity are selected; and the selected interventions are ranked according to the patient's individual preference information and clinical feasibility.

[0029] Preferably, the method further includes:

[0030] The system periodically receives feedback from clinicians, which is used to correct the output of the deep learning network or adjust the traversal strategy of the intelligent agent. The results of the correction and adjustment are used as new training samples to incrementally learn the process of constructing the dynamic ontology, thereby updating the mapping relationship of the clinical knowledge graph in the cardiovascular data pool.

[0031] Compared with the prior art, the beneficial effects of the present invention are:

[0032] A dynamic ontology is established within a cardiovascular data pool based on a clinical knowledge graph. This dynamic ontology defines the semantic relationships between different data entities. The clinical knowledge graph includes medical concepts such as diseases, symptoms, drugs, and examination indicators, along with their relationships. The dynamic ontology is described using the OWL language, supporting logical reasoning and semantic querying. Ontology construction is achieved through entity linking and relation extraction techniques, associating instances in the data pool with concepts in the knowledge graph. Semantic relationships include hierarchical, correlational, and causal relationships. The ontology structure supports dynamic expansion, allowing new concepts and relationships to be added at any time. The ontology reasoning engine performs semantic consistency checks, identifying logical contradictions in the data. Data evolution trajectories are generated on the dynamic ontology based on predefined treatment target paths. These target paths are defined through clinical guidelines and expert experience, encompassing target indicators for different treatment stages. The mapping process employs a graph matching algorithm to find subgraph structures within the ontology network that conform to the target path. The data evolution trajectory describes the expected trends of various indicators under the ideal treatment path. Trajectory generation considers patient baseline characteristics and complication status, enabling personalized adjustments. Trajectory visualization helps physicians understand treatment expectations. The trajectory update mechanism is dynamically adjusted based on the treatment effect to maintain consistency with the actual situation.

[0033] The intelligent agent traverses the dynamic ontology along the data evolution trajectory, resolving causal constraints between data entities based on semantic relationships encountered during the traversal. The intelligent agent employs a reinforcement learning framework, using the data evolution trajectory as its navigation path. The traversal prioritizes visiting ontology nodes relevant to the current treatment stage. Semantic relationship resolution is achieved through a combination of rule-based reasoning and statistical learning. Causal constraint mining considers temporal order, dose-response relationships, and confounding factors. Constraint strength is quantified using confidence metrics to eliminate spurious correlations. A constraint network is constructed to demonstrate complex causal relationships between multiple data entities. The resolution results are stored in the form of logical rules, supporting causal inference. The intelligent agent adaptively reorganizes new incoming data based on the resolved causal constraints, generating data slices with temporal correlations. The adaptive reorganization algorithm considers the strength and time decay characteristics of causal constraints. Data slices are divided by time windows, with the window size adaptively adjusted according to data frequency. Data within a slice is rearranged according to causal relationships, highlighting changes in key indicators. Temporal correlations are verified using cross-correlation and Granger causality tests. Slice labels include timestamps, causal context, and importance scores. The reorganization process supports streaming processing to meet real-time requirements. Sliced ​​data is stored in a columnar manner to improve read efficiency.

[0034] Data slices are input into a pre-trained deep learning network for feature fusion, outputting a dynamic profile of a patient's cardiovascular health status. The deep learning network employs a Transformer architecture to capture long-distance temporal dependencies. Feature fusion is achieved through a multi-head attention mechanism, weighted and integrating information from different data slices. The network is pre-trained using a large-scale cardiovascular disease dataset to learn general feature representations. Fine-tuning is tailored to individual patient data, adapting to specific disease characteristics. The dynamic profile includes multi-dimensional information such as risk scores, trend predictions, and anomaly warnings. The profile update frequency is synchronized with the data slice generation, ensuring real-time performance. Profile visualization uses a dashboard format to intuitively display changes in health status. Through the synergistic effect of dynamic ontology construction, semantic relationship parsing, causal constraint mining, and feature fusion, intelligent management of cardiovascular chronic disease data is achieved. The ontology provides the semantic foundation, intelligent agents realize causal reasoning, data reorganization optimizes feature extraction, and deep learning generates dynamic assessments. This integrated approach significantly improves the accuracy and timeliness of cardiovascular disease management. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating the working principle of the AI-based cardiovascular chronic disease data management method described in this invention.

[0036] Figure 2 A flowchart for generating the data evolution trajectory;

[0037] Figure 3 Flowchart for adaptive reorganization of new data;

[0038] Figure 4 A heatmap of causal constraints for cardiovascular data entities;

[0039] Figure 5 This is a comparison chart of the similarity of intervention measures for cardiovascular chronic diseases. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figure 1 This invention provides an AI-based method for managing cardiovascular chronic disease data. The method includes: constructing a cardiovascular data pool that integrates historical patient records from medical institutions and real-time physiological signals collected from wearable devices, forming a unified repository for structured and unstructured data; establishing a dynamic ontology within the cardiovascular data pool based on a clinical knowledge graph, where the ontology defines semantic relationships between data entities using entities and relationships within the graph, including symptom, diagnosis, treatment concepts, and their interactions; mapping a data evolution trajectory onto the dynamic ontology according to a pre-defined treatment goal path, reflecting the expected sequence of changes in patient data from the current state to the target state; activating an intelligent agent configured to traverse the dynamic ontology along the data evolution trajectory, resolving causal constraints between data entities during the traversal, these constraints stemming from logical dependencies in clinical knowledge; and adaptively reorganizing new data flowing into the cardiovascular data pool in real-time based on the resolved causal constraints, generating a set of time-series correlated data slices that capture the evolution of key data points over time. Data slices are input into a pre-trained deep learning network for feature fusion. The network extracts abstract features through multi-layer processing and outputs a dynamic profile of the patient's cardiovascular health status, which quantifies risk indicators and trends in the form of vectors or graphs.

[0042] Example 1: See Figure 2In practice, key milestone events are extracted from the treatment target pathway, which is defined based on clinical guidelines and includes events such as achieving blood pressure control targets and completing lipid management. Each key milestone event is anchored to a corresponding data entity in a dynamic ontology. The anchoring process is achieved by matching the event description with the semantic tags of the data entity in the dynamic ontology; for example, the blood pressure control target achievement event is associated with the blood pressure measurement value entity in the dynamic ontology. The optimal path connecting the anchored data entities is calculated based on the semantic relationships defined in the dynamic ontology. The calculation process involves analyzing the association types and path costs between data entities.

[0043] In some embodiments, different weight values ​​are assigned to different types of semantic relations in the dynamic ontology. These weight values ​​are allocated based on clinical importance or data source reliability; for example, diagnostic relations are assigned higher weights, while measurement relations are assigned lower weights. A graph traversal algorithm is used to find a path with the minimum total weight. This algorithm uses Dijkstra's algorithm to evaluate multi-hop paths from the starting data entity to the target data entity, sequentially traversing all anchored data entities. All data entities and semantic relation sequences traversed by the path with the minimum total weight are recorded, and this sequence is instantiated as a data evolution trajectory.

[0044] It is understandable that the calculation of the optimal path depends on a weighting strategy, which can be adjusted based on clinical expert input or historical data statistics. In practice, the formula for calculating the total path weight is:

[0045]

[0046] in: Indicates the total weight of the path. This represents the total number of semantic relations in the path. This represents the j-th semantic relation in the path. Representing semantic relations The weight value.

[0047] Optionally, the weight values ​​can be dynamically updated based on the semantic relationship type, such as adjusting the weight values ​​when new clinical evidence is added. In some embodiments, the graph traversal algorithm can be combined with the A* algorithm to optimize search efficiency, but the core objective is to find the path with the minimum total weight. It can be understood that after instantiating the data evolution trajectory, this trajectory is used for subsequent traversal and data reorganization processes by the intelligent agent. In a specific implementation, the data evolution trajectory is stored in the form of a directed graph, containing a sequence of nodes and edge relationships; nodes correspond to data entities, and edges correspond to semantic relationships.

[0048] Example 2: See Figure 3In practice, the intelligent agent monitors the cardiovascular data pool and captures new data, which can be real-time blood pressure readings transmitted from the patient's wearable device. It identifies the data entity type to which the new data belongs in the dynamic ontology, for example, identifying the blood pressure reading as the "ambulatory blood pressure monitoring" data entity type. It retrieves other data entities with direct causal constraints to the data entity type, querying the dynamic ontology for entities like "drug dosage adjustment" and "edema symptoms" that have a causal relationship with "ambulatory blood pressure monitoring." Based on the retrieved causal constraint rules, it determines whether the new data meets the conditions for triggering data recombination. The causal constraint rule is defined as triggering recombination when the systolic blood pressure exceeds 160 mmHg three times consecutively.

[0049] In some embodiments, a policy library is maintained for the intelligent agent. This library contains various data reconstruction patterns, including time window stitching, event-driven aggregation, and outlier association. A matching data reconstruction pattern is selected from the policy library based on the characteristics of the new data and the triggered causal constraint rules. For example, when the new data exhibits persistent hypertension characteristics, the time window stitching pattern is selected. Historical data related to the causal constraint rules is extracted from the cardiovascular data pool based on the selected data reconstruction pattern. This historical data includes recent medication records and weight change data, which are then stitched and encapsulated with the new data in chronological order.

[0050] It is understandable that the process of determining whether new data meets the triggering conditions for data reorganization can be formally expressed. In specific implementation, the triggering conditions are determined by the following logical function:

[0051]

[0052] in: Represents a new data instance The triggering state, This indicates the number of causal constraint rules. This represents the activation weight of the k-th rule. This represents the matching degree function of the new data instance under the k-th rule. Indicates the trigger threshold. This represents the sigmoid activation function.

[0053] Optionally, the data restructuring patterns in the strategy library can be expanded according to clinical practice needs, such as adding differentiated restructuring patterns based on patient grouping. In some embodiments, the operation of the time window splicing pattern includes tracing back a fixed period centered on the new data time point and extracting all historical data records within that period that conform to causal constraint rules. It is understood that the structure of the data slice during the encapsulation process includes a timestamp sequence, data type identifier, and numerical array, and the encapsulation format adopts a unified serialization standard. In specific implementations, the splicing operation ensures temporal continuity, and the encapsulation operation adds metadata descriptions to record the restructuring pattern and triggering rules. Optionally, the intelligent agent records a decision log during the restructuring process. The decision log includes details of trigger condition evaluation and the basis for strategy selection; the decision log is used for subsequent process auditing and algorithm optimization.

[0054] See Figure 4 This diagram is a core visualization of the intelligent agent's analysis of causal constraints between data entities. The rows and columns cover core cardiovascular data entities such as blood pressure, blood sugar, blood lipids, heart rate, and weight. Cell values ​​and color gradations on the right quantify the strength of causal relationships between these entities. This diagram visualizes the causal network structure between data entities, serving as a crucial basis for the intelligent agent to identify direct causal constraints and determine data recombination trigger conditions. It provides an intuitive quantitative reference for semantic relationships in the technical process of retrieving causal constraints and adaptively recombining data, helping to clarify the causal interaction patterns of different data entities in cardiovascular chronic disease data management.

[0055] Example 3: In specific implementation, the deep learning network includes multiple parallel feature extraction branches, each specializing in processing a type of data slice. For example, the first feature extraction branch processes vital sign data slices, the second branch processes laboratory test data slices, and the third branch processes medication record data slices. The output of each feature extraction branch is passed to an attention mechanism layer, which dynamically calculates the importance scores of the features output from different branches. The importance scores reflect the degree of contribution of each branch's features to the current health status assessment. Based on the importance scores, the features from different branches are weighted and summed to generate a fused feature vector, which integrates complementary information from multiple data sources.

[0056] In some embodiments, the attention mechanism layer calculates importance scores using a content-based attention mechanism, with the output feature vector of each feature extraction branch serving as the input to the attention mechanism. In a specific implementation, the formula for calculating the importance score is expressed as:

[0057]

[0058] in: This represents the importance score of the output feature of the p-th feature extraction branch. This represents the total number of feature extraction branches. It is an index variable used to traverse all feature extraction branches. This represents the output feature vector of the p-th feature extraction branch. The weight matrix represents the linear transformation. This represents the bias vector. The context vector representing the attention mechanism. This indicates the transpose operation. The transpose of the context vector representing the attention mechanism. This represents the hyperbolic tangent activation function.

[0059] It can be understood that the generation of the fused feature vector is achieved through weighted summation, which is calculated as the sum of the products of each branch feature vector and its corresponding importance score. The fused feature vector is input into a state decoder, which maps the fused feature vector into a multi-dimensional health state vector. Each dimension of the health state vector corresponds to a quantitative assessment value of a cardiovascular health risk indicator. For example, the first dimension corresponds to the heart failure risk index, the second dimension corresponds to the probability of arrhythmia, and the third dimension corresponds to the arteriosclerosis degree score.

[0060] Optionally, the state decoder can employ a multilayer perceptron structure to implement nonlinear mapping, where the multilayer perceptron contains multiple hidden layers and an output layer. In some embodiments, the feature extraction branch can be built based on a convolutional neural network to process waveform data, or based on a recurrent neural network to process time-series data. It is understood that the parameters of the attention mechanism layer are learned during model pre-training and are fixed during the inference phase. Specific implementations during model pre-training include training the network using a large-scale cardiovascular disease dataset. This dataset integrates diverse patient historical medical records and real-time physiological signals, ensuring data coverage of different disease stages and population characteristics. The pre-training phase employs a supervised learning framework, where the input data are preprocessed data slices, and the output labels are expert-annotated health status vectors. Network parameters are optimized by minimizing the loss function between the predicted output and the true label. After pre-training, the network parameters, including the weights and biases of the attention mechanism layer, are fixed and saved. During the inference phase, feature fusion for new data is directly loaded without further adjustment, thus ensuring model consistency and efficiency. The fine-tuning process involves a small number of iterative training iterations based on individual patient data to adapt to the specific patient's condition characteristics, but the parameters of the attention mechanism layer remain frozen. In practice, the number of dimensions in the health status vector is consistent with the types of risk indicators of clinical concern, and the numerical range of each dimension is normalized. Optionally, the output layer of the state decoder can use the sigmoid activation function to constrain the value of each dimension to between 0 and 1, where 0 represents the lowest risk state and 1 represents the highest risk state.

[0061] Example 4: In specific implementation, an intervention knowledge base is established. This knowledge base stores various interventions and their expected impact patterns on the health state vector. The knowledge base is implemented as a structured database, including, for example, medication adjustments, exercise prescriptions, and dietary recommendations. Each entry is associated with an impact pattern vector, which describes the expected direction and magnitude of change in each dimension of the health state vector caused by the intervention. The health state vector is compared with an ideal state vector to generate a difference vector. The ideal state vector is based on clinical guidelines or individualized treatment goals. The difference vector is calculated element-wise by subtraction, representing the deviation between the current health state and the target state. The intervention knowledge base is queried for interventions or combinations of interventions that minimize the difference vector. The query process is based on a similarity metric in the vector space, filtering out interventions whose impact pattern vectors and difference vectors are opposite in direction and have matching magnitudes.

[0062] In some embodiments, a content-based recommendation algorithm is used to calculate the similarity between the difference vector and the influence pattern of each intervention in the intervention knowledge base. The content-based recommendation algorithm uses cosine similarity to measure the alignment of the two vectors in direction. Several interventions with the highest similarity are selected, for example, the top five interventions with the highest similarity are chosen as a candidate set. The selected interventions are ranked according to the patient's individual preferences and clinical feasibility. Individual preferences include the patient's acceptance of the intervention or historical adherence, and clinical feasibility considers the availability of medical resources or contraindications.

[0063] In practice, the similarity calculation uses the following formula:

[0064]

[0065] in: This represents the similarity value between the difference vector and the influence pattern vector of the i-th intervention. Represents the difference vector. This represents the impact pattern vector of the i-th intervention measure. This represents the vector dot product operation. The Euclidean norm of the difference vector. Let represent the Euclidean norm of the pattern vector affected by the i-th intervention.

[0066] As can be understood (referring to Table 1), the construction of the intervention knowledge base relies on clinical evidence and expert knowledge, and the knowledge base entries are regularly updated to incorporate the latest research findings. Optionally, the influence pattern vector can be derived based on randomized controlled trial data or real-world evidence, with the vector dimension consistent with the health status vector. In some embodiments, the ranking process incorporates a multi-criteria decision-making approach, such as assigning weights to similarity, individual preferences, and clinical feasibility, and calculating a comprehensive score.

[0067] Table 1: Knowledge Base Table of Intervention Measures

[0068] Intervention number Intervention name Expected impact pattern vector INT-001 Increase ACEI dosage [-0.2,-0.1,0.0] INT-002 Start aerobic exercise [-0.1,-0.05,-0.05] INT-003 Limit sodium intake [-0.15,0.0,0.0]

[0069] It is understandable that the generation of the difference vector depends on the definition of the ideal state vector, which can be dynamically adjusted according to patient risk stratification or treatment stage. In specific implementation, the goal of minimizing the difference vector is transformed into finding interventions that are negatively correlated with the influence pattern vector and the difference vector; a negative correlation indicates that the intervention can reduce bias. Optionally, the query for combinations of interventions can be achieved through linear programming or heuristic search to evaluate the combined effect of multiple interventions. In some embodiments, individual preference information is obtained through patient questionnaires or historical selection records, and clinical feasibility is determined based on hospital agreements or physician assessments.

[0070] See Figure 5 This diagram is the core visualization of the intervention recommendation mechanism. It shows the similarity scores between interventions such as combined antihypertensive medication, smoking cessation intervention, sodium restriction, initiation of aerobic exercise, and increased ACEI dosage, and the patient's health status difference vector. This diagram quantifies the potential match between different interventions and the improvement of the patient's cardiovascular health. It serves as a key basis for the content recommendation algorithm-based intervention selection process, combined with individual preferences and clinical feasibility ranking techniques. This provides data support for the precise selection of interventions or combinations in clinical practice, facilitating efficient recommendation of interventions in the management of cardiovascular chronic diseases.

[0071] Example 5: In specific implementation, feedback information from clinicians is received periodically. This feedback is input through a dedicated doctor interaction interface and includes correction information for the deep learning network output or suggestions for adjusting the intelligent agent traversal strategy. The results of the corrections and adjustments are used as new training samples. These new training samples contain the original input data, corrected output labels, and strategy adjustment records, and are stored in a standardized format in a training sample library. Incremental learning is performed on the process of constructing the dynamic ontology. This incremental learning process uses the new training samples to update the weights and confidence of the semantic relationship mapping, thereby updating the mapping relationship of the clinical knowledge graph in the cardiovascular data pool.

[0072] In some embodiments, the feedback information correcting the output of the deep learning network can be a numerical adjustment of a specific dimension in the health status vector, such as a doctor correcting the heart failure risk index from 0.7 to 0.6. Suggestions for adjusting the traversal strategy of the intelligent agent can be specifying the priority of specific data entities during the traversal process, such as increasing the access order of the blood pressure data entity in the data evolution trajectory. The incremental learning process employs an online learning algorithm, which gradually incorporates information from new training samples while retaining existing knowledge.

[0073] In practical implementation, the objective function of incremental learning is expressed as:

[0074]

[0075] in: Indicates in the training sample set The overall loss function on This represents the set of parameters for constructing a dynamic ontology model. This represents the complete training set containing the new training samples. Represents a single training sample loss function, The parameter is The model for input The predicted output, This indicates the corrected label. Represents the regularization coefficient. This represents the regularization term for the model parameters.

[0076] It is understandable that updating the mapping relationship of the clinical knowledge graph in the cardiovascular data pool involves adding, deleting, and modifying semantic relationships, such as strengthening the association between a certain symptom and a diagnosis based on feedback information. Optionally, the incremental learning process can use stochastic gradient descent to optimize model parameters, updating parameters each time with a small batch of new training samples. In some embodiments, the doctor's interface provides visualization tools to display the current structure of the dynamic ontology, allowing doctors to directly drag nodes or adjust relationship strength. It is understandable that the training sample library adopts a version management mechanism, retaining historical training samples for model rollback and effect comparison. In specific implementations, the incremental learning cycle of the dynamic ontology construction process is set to be executed once a week, or triggered when a certain number of new training samples are accumulated. Optionally, some old training samples are retained during the incremental learning process to prevent knowledge forgetting; old training samples are selected based on timestamps and importance scores. Feedback information processing includes consistency verification, which checks whether there are contradictions in the feedback from different doctors on similar cases, and special marks contradictory feedback for expert review.

[0077] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-based cardiovascular chronic disease data management method, characterized by, The method comprises: constructing a cardiovascular data pool containing patient historical medical records and real-time physiological signals; establishing a dynamic ontology in the cardiovascular data pool based on a clinical knowledge graph, the dynamic ontology being used to define semantic relationships between different data entities; generating a data evolution trajectory on the dynamic ontology according to a preset treatment target path; starting an intelligent agent configured to traverse the dynamic ontology along the data evolution trajectory and analyze causal constraints between data entities according to semantic relationships encountered during the traversal; the intelligent agent adaptively recombining new data flowing into the cardiovascular data pool in real time according to the analyzed causal constraints to generate a set of data slices with temporal correlation; inputting the data slices into a pre-trained deep learning network for feature fusion and outputting a dynamic portrait of the patient's cardiovascular health status; the generating of the data evolution trajectory on the dynamic ontology according to the preset treatment target path comprises: extracting key milestone events from the treatment target path; anchoring each key milestone event to a corresponding data entity in the dynamic ontology; calculating an optimal path connecting the anchored data entities according to the semantic relationships defined in the dynamic ontology; and instantiating the optimal path as the data evolution trajectory; the adaptive recombination of the new data flowing into the cardiovascular data pool in real time by the intelligent agent according to the analyzed causal constraints comprises: the intelligent agent monitors the cardiovascular data pool to capture new data; identifies the data entity type to which the new data belongs in the dynamic ontology; retrieves other data entities that have direct causal constraints with the data entity type; determines whether the new data meets the trigger data recombination condition according to the retrieved causal constraint rules; and when the condition is met, extracts historical data related to the causal constraint rules from the cardiovascular data pool and splices and encapsulates the new data in chronological order; the inputting of the data slices into a pre-trained deep learning network for feature fusion comprises: the deep learning network comprises multiple parallel feature extraction branches, each branch being dedicated to processing a type of data slice; the output of each feature extraction branch is passed to an attention mechanism layer that dynamically calculates the importance scores of different branch output features; and the features from different branches are weighted and summed based on the importance scores to generate a fused feature vector. 2.The AI-based cardiovascular chronic disease data management method of claim 1, wherein, the calculation of the optimal path connecting the anchored data entities comprises: assigning different weight values to different types of semantic relationships in the dynamic ontology; using a graph traversal algorithm to find a path with the minimum total weight, the path passing through all anchored data entities in sequence; and recording all data entities and semantic relationship sequences passed through by the path with the minimum total weight. 3.The AI-based cardiovascular chronic disease data management method of claim 2, wherein, the method further comprises: maintaining a strategy library for the intelligent agent, the strategy library containing multiple data recombination modes; The method further comprises: 4.The AI-based cardiovascular chronic disease data management method of claim 3, wherein, inputting the fusion feature vector into a state decoder, which maps the fusion feature vector into a multi-dimensional health state vector, each dimension of which corresponds to a quantitative evaluation value of a cardiovascular health risk indicator. The method further comprises: 5.The AI-based cardiovascular chronic disease data management method of claim 4, wherein, establishing an intervention knowledge base storing a plurality of interventions and their expected impact patterns on the health state vector; comparing the health state vector with an ideal state vector to generate a difference vector; and querying the intervention knowledge base for an intervention or a combination of interventions that can minimize the difference vector. The querying the intervention knowledge base for an intervention or a combination of interventions that can minimize the difference vector comprises: 6.The AI-based cardiovascular chronic disease data management method of claim 5, wherein, applying a content-based recommendation algorithm to calculate the similarity between the difference vector and the impact pattern of each intervention in the intervention knowledge base; screening a number of interventions with the highest similarity; and sorting the screened interventions according to the individual preference information of the patient and the clinical execution feasibility. The method further comprises: 7.The AI-based cardiovascular chronic disease data management method of claim 1, wherein, periodically receiving feedback information from a clinician, which is used to correct the output of the deep learning network or adjust the traversal strategy of the intelligent agent; and using the corrected and adjusted results as new training samples to perform incremental learning on the process of constructing the dynamic ontology, thereby updating the mapping relationship of the clinical knowledge graph in the cardiovascular data pool. ​

Images