Method for constructing cardiovascular and cerebrovascular disease mapping knowledge domain and related equipment

Abstract

The invention provides a cardiovascular and cerebrovascular disease knowledge graph construction method and related equipment. The method comprises the following steps: analyzing multi-modal scientific research data, and extracting candidate entities and relation information; based on the first ontology and the second ontology, constructing a dual-drive unified knowledge ontology structure through semantic association; performing cross-modal semantic alignment and fusion to form entity nodes and relation edges; calculating prior credibility based on the evidence source attribute, and fusing the prior credibility and the likelihood information by using a probability graph model to obtain posterior credibility; sorting and rule induction are carried out on paths in the knowledge graph based on posterior credibility, a high-confidence rule is used as structured prior information to be fed back to the calculation process of a probability graph model according to a stability evaluation mechanism, and the weight is dynamically adjusted according to time information. The accuracy and robustness of the cardiovascular and cerebrovascular disease mapping knowledge domain can be improved.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing and knowledge engineering technology, specifically relating to a method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases and related equipment. Background Technology

[0002] With the deepening of research on cardiovascular and cerebrovascular diseases, massive amounts of multimodal research data have been generated, including clinical texts, medical images (such as CT and MRI), physiological waveforms (such as ECG and hemodynamic signals), and omics data (such as genomics and proteomics). This data contains rich knowledge ranging from disease mechanisms and experimental findings to clinical evidence. However, existing technologies face significant challenges in processing this data: First, most medical knowledge graphs focus on clinical diagnostic relationships such as "disease-symptom-drug," lacking deep modeling of research semantics such as "experimental sample-observation indicator-molecular mechanism," making it difficult to effectively organize and utilize research data. Second, research data is highly multi-source, heterogeneous, and rapidly evolving, with new terms, relationships, and evidence constantly emerging, making it difficult for traditional static graphs to achieve incremental updates and consistency maintenance at the hourly or even minute level. Third, research conclusions from different sources may contradict each other, and the lack of an automated, quantitative conflict assessment and resolution mechanism affects the credibility and practicality of knowledge graphs.

[0003] Therefore, there is an urgent need for a method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases that can integrate clinical and scientific research perspectives, support incremental absorption, automatically resolve conflicts, and have continuous self-optimization capabilities. Summary of the Invention

[0004] This invention aims to provide a method and related equipment for constructing a knowledge graph of cardiovascular and cerebrovascular diseases, in order to solve the problems in the existing technology that knowledge graphs are difficult to effectively integrate clinical and scientific research semantics, cannot efficiently process rapidly evolving multi-source data, and lack automated conflict resolution and continuous optimization capabilities.

[0005] In a first aspect, the present invention provides a method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases, the method comprising the following steps: This involves analyzing multimodal research data in the field of cardiovascular and cerebrovascular diseases to extract candidate entities and relationship information. The multimodal research data includes text, images, waveforms, and omics data. Based on a first ontology containing clinical diagnosis and treatment concepts and a second ontology containing scientific research semantic concepts, a unified knowledge ontology structure driven by semantic association is constructed. Based on the graph neural network model, cross-modal semantic alignment and fusion are performed on candidate entities extracted from different modal data that point to the same real-world object, and the candidate entities are mapped to a unified knowledge ontology structure to form entity nodes and relation edges of the knowledge graph. For relational fact edges in a knowledge graph, prior confidence is calculated based on the evidence source attributes of the associated evidence. Then, a probabilistic graphical model is used to fuse the prior confidence with likelihood information based on model scores and repeated verification to obtain posterior confidence, so as to evaluate and resolve conflicts between different evidence sources. Based on posterior confidence, the paths in the knowledge graph are sorted and rules are summarized. According to the preset stability assessment mechanism, the high confidence rules summarized are fed back as structured prior information to the calculation process of the probabilistic graphical model. The weights are dynamically adjusted according to time information to form a constrained closed-loop self-optimization process, so that the knowledge graph continues to evolve with the addition of new data.

[0006] Optionally, the first ontology includes concepts and relationships related to cardiovascular and cerebrovascular diseases, clinical symptoms, therapeutic drugs, medical examinations, and surgical interventions; the second ontology includes concepts and relationships related to experimental samples, experimental procedures, physiological or biochemical observation indicators, molecular biological pathways, and scientific literature evidence; the first ontology and the second ontology are connected through cross-layer semantic relationships.

[0007] Optionally, the graph neural network model is a graph attention network model; In the process of information propagation, the graph attention network model introduces gating weights based on the posterior credibility of edges, which attenuate edges with credibility below a preset threshold.

[0008] Optionally, prior confidence can be calculated based on the evidence source attributes of its associated evidence, including: The authority level of the source of the evidence, the size of the study sample, the quality score of the experimental methodology, and the freshness of the evidence; Each attribute value is converted into an intensity score using a monotonically increasing calibration function; The intensity scores are combined into a comprehensive log-odds value using a nonlinear aggregation function in the log-odds domain. The combined logarithmic odds value is converted into prior confidence using the Sigmoid function.

[0009] Optionally, the expression for the log-odds domain nonlinear aggregation function is:

[0010] in, This represents the combined logarithmic probability value. Indicates the bias term. Indicates the first One source of evidence attribute, , Indicates the level of authority of the source. Indicates the number of research samples. Indicates the quality score of the experimental methodology. Indicates the age or novelty of the evidence. Indicates corresponding to The monotonically increasing calibration function, Indicates the corresponding to the first Individual evidence source attributes The non-negative weight coefficients are obtained by training on a historical conflict evidence dataset with the goal of minimizing prediction error and imposing sparsity constraints.

[0011] Optionally, a probabilistic graphical model can be used to fuse prior confidence with likelihood information based on model scoring and repeated validation, including: Prior confidence Mapped to Prior parameters of the probability distribution; where the mapping relationship is: , , The equivalent sample size to characterize the strength of prior beliefs Represents prior parameters, ; Using likelihood information extracted from evidential data that reflects the model's internal consistency and external repeated validation results, the model is analyzed. The probability distribution is updated using conjugate methods to obtain the posterior distribution parameters. and ; where the posterior confidence level is determined by the posterior mean. Provided.

[0012] Optional, preset stability assessment mechanisms include: Define the system state potential energy function System state potential energy function It is used to measure the difference between the posterior probability distribution of the knowledge graph in the current round and the previous round, as well as the degree to which the current rule violates the preset consistency constraint; If and only if the system state change caused by rule feedback satisfies At that time, high-confidence rules are fed back as structured prior information to the probabilistic graphical model; among them, It is a positive number. This represents the change in parameters of a probabilistic graphical model.

[0013] Optionally, the constrained closed-loop self-optimization process also includes shadow update and drift circuit breaking steps; The shadow update and drift circuit breaker steps include: Before applying rule feedback to the main knowledge graph, the update effect is simulated and verified in an offline shadow environment. If the verification indicators do not improve or significant concept drift in the data distribution is detected, the feedback loop is interrupted, the application of structured prior information is paused, and the system is stabilized until it recovers.

[0014] Optionally, the weights can be dynamically adjusted based on time information, including: By assigning decay coefficients associated with timestamps to nodes and edges in the knowledge graph, the knowledge graph can present a dynamically evolving subgraph structure according to different time windows following the occurrence of clinical events of cardiovascular and cerebrovascular diseases.

[0015] Secondly, the present invention provides a device for constructing a knowledge graph of cardiovascular and cerebrovascular diseases, comprising: The data preprocessing module is used to analyze multimodal research data in the field of cardiovascular and cerebrovascular diseases and extract candidate entity and relationship information; multimodal research data includes text, images, waveforms and omics data; The ontology construction module is used to construct a unified knowledge ontology structure driven by semantic association based on a first ontology containing clinical diagnosis and treatment concepts and a second ontology containing scientific research semantic concepts. The semantic alignment module is used to perform cross-modal semantic alignment and fusion of candidate entities that point to the same real-world object extracted from different modal data based on a graph neural network model, and to map the candidate entities to a unified knowledge ontology structure to form entity nodes and relation edges of the knowledge graph. The conflict resolution module is used to calculate the prior confidence of relational fact edges in the knowledge graph based on the evidence source attributes of their associated evidence, and use a probabilistic graphical model to fuse the prior confidence with the likelihood information based on model scores and repeated verification to obtain the posterior confidence, so as to evaluate and resolve conflicts between different evidence sources. The knowledge graph optimization module is used to sort paths and summarize rules in the knowledge graph based on posterior confidence. According to the preset stability evaluation mechanism, the summarized high-confidence rules are fed back as structured prior information to the calculation process of the probabilistic graphical model, and the weights are dynamically adjusted according to time information to form a constrained closed-loop self-optimization process, so that the knowledge graph continues to evolve with the addition of new data.

[0016] The present invention has at least the following beneficial effects: By integrating clinical diagnosis ontology and scientific research semantic ontology, this approach overcomes the limitations of traditional knowledge graphs lacking scientific semantics. It provides a unified structured representation framework for the complete knowledge chain in the field of cardiovascular and cerebrovascular diseases, from basic research to clinical application, significantly improving the accuracy of the knowledge graph. Evidence source attributes are transformed into prior confidence through nonlinear aggregation, and a Bayesian conjugate update model is used to fuse prior and likelihood information to obtain posterior confidence. This provides an automated and quantifiable solution for handling contradictory conclusions in the field of cardiovascular and cerebrovascular diseases, greatly improving the consistency, confidence, and decision support value of the knowledge graph. By using inductively derived high-confidence rules as structured priors for conflict resolution, combined with stability assessment, the performance of the knowledge graph during continuous evolution is improved. This effectively avoids the amplification of erroneous rules and oscillations in the common view, ensuring the knowledge graph remains robust when receiving real and rapidly changing scientific research data streams. Attached Figure Description

[0017] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0018] Figure 1 This is a flowchart of a method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases in one embodiment of this application; Figure 2 This is a structural diagram of a device for constructing a knowledge graph of cardiovascular and cerebrovascular diseases in another embodiment of this application. Detailed Implementation

[0019] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0021] To address the shortcomings of existing knowledge graph technologies, such as the inability to effectively integrate clinical and research semantics, the inability to efficiently process rapidly evolving multi-source data, and the lack of automated conflict resolution and continuous optimization capabilities, this invention provides a method and related equipment for constructing a knowledge graph for cardiovascular and cerebrovascular diseases. By integrating clinical diagnostic and treatment ontology with research semantic ontology, it overcomes the limitations of traditional graphs lacking research semantics, providing a unified structured representation framework for the complete knowledge chain in the field of cardiovascular and cerebrovascular diseases, from basic research to clinical application. This significantly improves the accuracy of the knowledge graph. Furthermore, it transforms evidence source attributes into prior knowledge through nonlinear aggregation. The reliability of the knowledge graph is improved by fusing prior and likelihood information using a Bayesian conjugate update model to obtain posterior reliability. This provides an automated and quantifiable solution for handling contradictory conclusions in the field of cardiovascular and cerebrovascular diseases, significantly improving the consistency, reliability, and decision support value of the knowledge graph. By using the inductively derived high-confidence rules as structured priors to conflict resolution and combining them with stability assessment, the performance of the knowledge graph is improved during continuous evolution. This effectively avoids the amplification of erroneous rules and the oscillation of the common view, ensuring that the knowledge graph remains robust when receiving real and rapidly changing scientific research data streams.

[0022] Example 1 like Figure 1 As shown, the method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases provided by the present invention specifically includes steps 11 to 15.

[0023] Step 11: Analyze multimodal research data in the field of cardiovascular and cerebrovascular diseases and extract candidate entity and relationship information.

[0024] In this embodiment of the invention, multimodal scientific data includes text, images, waveforms, and omics data.

[0025] In one feasible implementation, the text data includes electronic medical records, experimental records, and abstracts and full texts of research papers. Entities (such as "myocardial infarction", "ApoE gene knockout mouse", "low-density lipoprotein cholesterol") and relations (such as "induction", "expression", "correlation") are extracted using natural language processing techniques, such as named entity recognition (NER) and relation extraction.

[0026] The imaging data includes coronary CTA images. Entities (such as "left anterior descending artery," "plaque," and "stenotic region") are identified and quantified using image segmentation and feature extraction algorithms, and representation vectors are calculated.

[0027] Waveform data includes electrocardiogram (ECG) signals and blood pressure waveforms. Features (such as ST segment elevation and heart rate variability) are extracted and feature vectors are generated using time-frequency analysis and other methods.

[0028] Omics data includes gene expression profiles. After standardization and annotation, it is linked with standard gene and pathway databases to form structured data.

[0029] It should be noted that all extracted candidate entities and relationships, along with their original data sources, timestamps, and other metadata, are sent to subsequent processes. This transforms unstructured and semi-structured raw data into structured information units with semantic annotations that can be used by knowledge graphs, laying the foundation for subsequent semantic fusion.

[0030] Step 12: Based on the first ontology containing clinical diagnosis and treatment concepts and the second ontology containing scientific research semantic concepts, construct a unified knowledge ontology structure driven by semantic association.

[0031] In this embodiment of the invention, the first ontology includes concepts and relationships related to cardiovascular and cerebrovascular diseases, clinical symptoms, therapeutic drugs, medical examinations, and surgical interventions. For example, the concepts in the first ontology include diseases (e.g., angina pectoris, stroke), symptoms (e.g., chest pain, hemiplegia), examinations (e.g., coronary angiography, brain MRI), and treatments (aspirin, stent implantation). Relationships include the presence of symptoms and the treatment of diseases with medication.

[0032] The second ontology encompasses concepts and relationships related to experimental samples, experimental procedures, physiological or biochemical observation indicators, molecular biological pathways, and scientific literature evidence. For example, concepts in the second ontology include experimental subjects (e.g., mice, human umbilical vein endothelial cells), experimental paradigms (high-fat diet feeding, hypoxia-reoxygenation treatment), observation indicators (left ventricular ejection fraction, inflammatory factor levels), molecular mechanisms (signaling pathways, ACE genes), and evidence genres (randomized controlled trials, retrospective cohort studies). Relationships include experimental models, experimental measurement indicators, and gene regulatory pathways.

[0033] In this embodiment of the invention, the first ontology and the second ontology are connected through cross-layer semantic relations. For example, the clinical disease "atherosclerosis" can be associated with the research layer's "rat model" through the relation "existence of animal sample model"; the research indicator "serum total cholesterol" can be associated with the clinical disease "hyperlipidemia" through the relation "biomarkers used for...". This forms a dual-driven unified knowledge ontology structure of "clinical base + research upper layer".

[0034] Step 13: Based on the graph neural network model, perform cross-modal semantic alignment and fusion on candidate entities extracted from different modal data that point to the same real-world object, and map the candidate entities to a unified knowledge ontology structure to form entity nodes and relation edges of the knowledge graph.

[0035] In this embodiment of the invention, the graph neural network model is a graph attention network model.

[0036] In one feasible implementation, step 13 specifically includes steps 13.1 to 13.3.

[0037] Step 13.1, construct the initial alignment map.

[0038] Specifically, the concepts in the unified ontology are used as the baseline nodes. Each candidate entity extracted by S1 is used as a node to be aligned, and the initial similarity between them and each baseline node is calculated based on their text description, feature vector, etc., forming weighted edges and constituting an initial heterogeneous information network.

[0039] Step 13.2, Graph Neural Network Alignment.

[0040] Specifically, a Graph Attention Network (GAT) is used as the core model. This model uses multiple rounds of message passing to allow each node to aggregate information from its neighboring nodes.

[0041] In this process, node features include their semantic embedding vectors and source modality types. Specifically, for edges connecting candidate entities and ontology concepts, the calculation of their attention weights not only considers feature similarity but also introduces a gating mechanism based on the edge's posterior credibility (initially set to a default value or simply calculated based on the source). For edges with posterior credibility below a threshold (e.g., 0.3), their attention weights are significantly attenuated, thereby suppressing the influence of low-quality or conflicting evidence in information propagation. A slight penalty term is added to the model training loss. When two highly similar candidate entity nodes are aligned to positions in the ontology that are structurally inconsistent (e.g., not belonging to the same subtree or having no association), a small penalty is applied to encourage structurally reasonable alignment.

[0042] Step 13.3, Alignment Strategy and Blending.

[0043] Specifically, through information propagation and feature updates in the GAT model, each candidate entity node receives a final representation pointing to the most likely ontology concept. By using a threshold, high-confidence candidate entities are merged into their corresponding ontology concept nodes, with evidence and attributes from different modalities being integrated into that node. For example, the textual description of "hypertensive rat," the calculated "left ventricular hypertrophy index" from images, and the experimental record of "systolic blood pressure > 180 mmHg" can all be merged into the entity node "hypertension disease model" and its related attribute edges after alignment.

[0044] It is worth mentioning that this move solves the problem of "different descriptions of the same thing" in multi-source heterogeneous data, achieves accurate entity unification, and lays the foundation for building a high-quality, non-redundant knowledge graph.

[0045] In another feasible implementation, in response to new terms appearing in the new data, the present invention compares the new terms with existing concepts in the unified knowledge ontology structure through semantic matching, and merges or incorporates the new terms into the knowledge graph after review based on the comparison confidence, so as to achieve incremental evolution of the ontology. Specifically, when the new data is parsed (through step 11): firstly, entities that cannot be found to have a high confidence match in the existing unified ontology are identified and added to the "new term queue".

[0046] Subsequently, the cosine similarity between new terms and all existing concepts in the ontology is calculated periodically (e.g., every hour) using a pre-trained knowledge embedding model (such as BioBERT). Then, decisions and merging are made based on the cosine similarity and a preset similarity threshold. Specifically, if the highest similarity exceeds a high threshold (e.g., 0.9), the new term is considered an alias or variant of an existing concept, automatically merged, and the list of aliases for existing concepts is updated. If the highest similarity is between the high and low thresholds (e.g., 0.6-0.9), ambiguity exists, triggering a semi-automatic review. The new term, its contextual evidence, and several candidate concepts with the highest similarity are submitted to experts for review, who decide whether to merge, create, or discard it. The review results and context are recorded as learning samples. If the highest similarity is below a low threshold and is confirmed as a new concept after review, a new ontology concept is created, and its hierarchical or related relationships with existing ontologies are defined. Finally, the newly added terms and their relationships are used as new training samples and fed back into the fine-tuning process of the alignment model in step 13, enabling the model to better identify similar terms in the future.

[0047] It is worth mentioning that this move enables knowledge graphs to automatically absorb new concepts generated at the forefront of scientific research, maintaining their coverage and freshness.

[0048] Step 14: For the relational fact edges in the knowledge graph, calculate the prior confidence based on the evidence source attributes of the associated evidence, and use a probabilistic graphical model to fuse the prior confidence with the likelihood information based on model scores and repeated verification to obtain the posterior confidence, so as to evaluate and resolve conflicts between different evidence sources.

[0049] In one feasible implementation, the process of calculating prior confidence based on the evidence source attributes of its associated evidence includes steps 14.1.1 to 14.1.4.

[0050] Step 14.1.1: Obtain the authority level of the source of the evidence, the size of the research sample, the quality score of the experimental methodology, and the freshness of the evidence.

[0051] In one feasible implementation, four attributes are extracted for each source of evidence: the level of authority of the source (e.g., 5 for "XX Medical Journal" and 1 for preprints), the number of study samples, the quality score of experimental methodology (based on the Jadad scale, etc.), and the freshness of the evidence (years since publication).

[0052] Step 14.1.2: Convert each attribute value into an intensity score using a monotonically increasing calibration function.

[0053] Specifically, through a monotonically increasing calibration function Each attribute value Mapped to intensity score For example, sample size can be mapped using a grading function.

[0054] Step 14.1.3: Use the nonlinear aggregation function in the log-odds domain to synthesize the intensity scores into a comprehensive log-odds value.

[0055] In one feasible implementation, the expression for the log-odds domain nonlinear aggregation function is:

[0056] in, This represents the combined logarithmic probability value. Indicates the bias term. Indicates the first One source of evidence attribute, , Indicates the level of authority of the source. Indicates the number of research samples. Indicates the quality score of the experimental methodology. Indicates the age or novelty of the evidence. Indicates corresponding to The monotonically increasing calibration function, Indicates the corresponding to the first Individual evidence source attributes The non-negative weight coefficients are obtained by training on a historical conflict evidence dataset with the goal of minimizing prediction error and imposing sparsity constraints.

[0057] Step 14.1.4: Convert the combined logarithmic odds value into prior confidence using the Sigmoid function.

[0058] Through calculation formula To obtain the prior confidence level , This represents the Sigmoid function.

[0059] In this embodiment of the invention, the likelihood information comes from two aspects: first, the score given by the graph alignment model in step 13 for the existence of the relationship; and second, the number of times the relationship has been independently verified in different studies. This information is constructed into a likelihood function.

[0060] In one feasible implementation, the process of fusing prior confidence with likelihood information based on model scoring and repeated validation using a probabilistic graphical model includes steps 14.2.1 to 14.2.2.

[0061] Step 14.2.1, set the prior confidence level Mapped to Prior parameters of the probability distribution; where the mapping relationship is: , , The equivalent sample size to characterize the strength of prior beliefs Represents prior parameters, .

[0062] Step 14.2.2: Using the likelihood information extracted from the evidence data, which reflects the internal consistency of the model and the results of external repeated validation, to... The probability distribution is updated using conjugate methods to obtain the posterior distribution parameters. and ; where the posterior confidence level is determined by the posterior mean. The variance of the posterior distribution is given. Meanwhile, the variance of the posterior distribution reflects the uncertainty of this fact.

[0063] In practice, values ​​above the threshold (e.g.) The fact edges are included in the public view of the graph, and their weights are set to... Evidence below the threshold but above a certain lower limit can be displayed only in the research view as disputed edges. Completely contradictory evidence with extremely low credibility is downgraded or temporarily stored.

[0064] It should be noted that traditional methods for handling conflicting evidence in knowledge graphs often rely on simple voting, fixed-weighting, or manual adjudication, making it difficult to perform fine-grained, automated quantitative evaluation of the quality of multi-source evidence. This application creatively proposes a conflict resolution framework based on a probabilistic graphical model. Its key features are: first, calculating prior credibility through nonlinear aggregation from four dimensions—evidence source, sample size, methodological quality, and timeliness—achieving a mathematical model transformation of medical evidence evaluation experience; then, fusing prior and data-driven likelihood information using Bayesian conjugate updates to obtain posterior credibility with uncertainty. This method transforms subjective, qualitative conflict judgment into objective, quantitative probability calculation, and its design requires a deep integration of clinical epidemiological criteria and machine learning probabilistic modeling thinking, and is not readily apparent. This mechanism significantly improves the automation, decision-making accuracy, and internal consistency of knowledge graphs when processing massive amounts of contradictory research data.

[0065] Step 15: Based on posterior confidence, the paths in the knowledge graph are sorted and rules are summarized. According to the preset stability assessment mechanism, the high confidence rules obtained by summarizing are fed back as structured prior information to the calculation process of the probabilistic graphical model. The weights are dynamically adjusted according to time information to form a constrained closed-loop self-optimization process, so that the knowledge graph continues to evolve with the addition of new data.

[0066] In one feasible implementation, the process of ranking paths in a knowledge graph based on posterior credibility includes: finding multi-hop paths connecting two entities (such as "a certain drug" and "a certain adverse reaction"), ranking them according to indicators such as the joint probability of posterior credibility of each edge on the path, path length, and novelty, and recommending potential, unstated implicit associations (causal or related clues).

[0067] In one feasible implementation, rule-based induction of paths in the knowledge graph based on posterior credibility includes: inducing logical rules from high-frequency, high-credibility path patterns. For example: "Drug X -> Inhibition -> Protein Y ∧ Protein Y -> Participation -> Pathway Z -> Drug X -> Regulation -> Pathway Z". Simultaneously, temporal rules are also induced, such as "Postoperative Day 3 -> Peak of inflammatory markers".

[0068] The inductively derived high-confidence rules are transformed into structured prior information. For example, a rule can be represented as a prior bias regarding the existence of a specific type of relationship. Before feeding these prior biases back to the Bayesian model in step 14, a pre-defined stability evaluation mechanism is used to safeguard stability. Specifically, in this embodiment of the invention, the pre-defined stability evaluation mechanism includes steps A and B: Step A, define the system state potential energy function. Among them, the system state potential energy function This is used to measure the difference between the posterior probability distribution of the knowledge graph in the current round and the previous round, as well as the degree to which the current rule violates the preset consistency constraint. In one feasible implementation, the system state potential function... The expression is ;in, Used to measure the posterior probability distribution of the entire map before and after this update. Divergence (degree of change) Used to measure the new rules against existing hard constraints The degree of violation of (e.g., "a drug cannot be both an agonist and an antagonist").

[0069] Step B, if and only if the system state change caused by rule feedback satisfies At that time, high-confidence rules are fed back as structured prior information to the probabilistic graphical model.

[0070] in, It is a positive number. This represents the change in parameters of a probabilistic graphical model.

[0071] In another embodiment of the invention, the constrained closed-loop self-optimization process further includes shadow update and drift circuit breaking steps.

[0072] Specifically, the shadow update and drift circuit breaker steps include: Before actually updating the main graph, this feedback is applied in a fully replicated shadow environment, and its effect is evaluated on a set of validation data streams (e.g., whether alignment consistency and conflict resolution accuracy have improved). Simultaneously, Page-Hinkley tests are used to check for concept drift in the monitoring data streams. If the validation metrics do not improve or significant drift is detected, the feedback loop is immediately shut down, the rule is ineffective, and only step 14 is used for regular passive evidence updates until the monitoring metrics stabilize. The prior biases of rules that pass the stability assessment are formally written back into the Bayesian network of step 14, affecting the credibility calculation of future new evidence (evidence consistent with the rule is strengthened, while conflicting evidence is suppressed).

[0073] It should be noted that traditional knowledge graph updates and optimizations are mostly passive incremental or based on static rules, lacking the ability to actively learn from data and safely provide feedback to guide system evolution. The innovation of this application lies in constructing a constrained closed loop of "conflict resolution - rule induction - feedback reinforcement." The system automatically induces logical and temporal rules from high-confidence graphs and feeds them back to the probabilistic model as structured priors to achieve knowledge reinforcement. Crucially, to ensure the stability of this self-optimization process, this application introduces a safeguard mechanism including potential function monitoring, shadow updates, and concept drift circuit breaking, ensuring that the system only accepts feedback when it can strictly guarantee monotonic performance improvement. This balanced design, which enables the system to actively learn while effectively preventing error amplification or performance oscillations, demonstrates a high level of engineering insight and creativity. Ultimately, this solution endows the knowledge graph with continuously evolving, self-improving, and highly robust intelligent characteristics.

[0074] Example 2 like Figure 2 As shown, this invention discloses a knowledge graph construction device for cardiovascular and cerebrovascular diseases. The knowledge graph construction device 200 for cardiovascular and cerebrovascular diseases includes: The data preprocessing module 201 is used to analyze multimodal research data in the field of cardiovascular and cerebrovascular diseases and extract candidate entity and relationship information; the multimodal research data includes text, images, waveforms and omics data; The ontology construction module 202 is used to construct a unified knowledge ontology structure driven by semantic association based on a first ontology containing clinical diagnosis and treatment concepts and a second ontology containing scientific research semantic concepts. The semantic alignment module 203 is used to perform cross-modal semantic alignment and fusion on candidate entities that point to the same real-world object extracted from different modal data based on a graph neural network model, and to map the candidate entities to a unified knowledge ontology structure to form entity nodes and relation edges of the knowledge graph. The conflict resolution module 204 is used to calculate the prior confidence of relational fact edges in the knowledge graph based on the evidence source attributes of their associated evidence, and use a probabilistic graphical model to fuse the prior confidence with the likelihood information based on model scores and repeated verification to obtain the posterior confidence, so as to evaluate and resolve conflicts between different evidence sources. The knowledge graph optimization module 205 is used to sort paths and summarize rules in the knowledge graph based on posterior confidence. According to the preset stability evaluation mechanism, the summarized high-confidence rules are fed back as structured prior information to the calculation process of the probabilistic graphical model, and the weights are dynamically adjusted according to time information to form a constrained closed-loop self-optimization process, so that the knowledge graph continues to evolve with the addition of new data.

[0075] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. Their specific functions and technical effects can be found in the method embodiments section, and will not be repeated here. Those skilled in the art will understand that, for ease of description and brevity, the above-mentioned division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0076] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0077] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases, characterized in that, include: This study analyzes multimodal research data in the field of cardiovascular and cerebrovascular diseases to extract candidate entities and relational information; the multimodal research data includes text, images, waveforms, and omics data. Based on a first ontology containing clinical diagnosis and treatment concepts and a second ontology containing scientific research semantic concepts, a unified knowledge ontology structure driven by semantic association is constructed. Based on the graph neural network model, cross-modal semantic alignment and fusion are performed on candidate entities that point to the same real-world object extracted from data of different modalities, and the candidate entities are mapped to the unified knowledge ontology structure to form entity nodes and relation edges of the knowledge graph. For the relational fact edges in the knowledge graph, the prior confidence is calculated based on the evidence source attributes of the associated evidence, and the prior confidence is fused with the likelihood information based on model scoring and repeated verification using a probabilistic graphical model to obtain the posterior confidence, so as to evaluate and resolve conflicts between different evidence sources. Based on the posterior confidence, the paths in the knowledge graph are sorted and rules are summarized. According to the preset stability evaluation mechanism, the summarized high-confidence rules are fed back as structured prior information to the calculation process of the probabilistic graphical model. The weights are dynamically adjusted according to time information to form a constrained closed-loop self-optimization process, so that the knowledge graph continues to evolve with the addition of new data.

2. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 1, characterized in that, The first ontology contains concepts and relationships related to cardiovascular and cerebrovascular diseases, clinical symptoms, therapeutic drugs, medical examinations, and surgical interventions; the second ontology contains concepts and relationships related to experimental samples, experimental procedures, physiological or biochemical observation indicators, molecular biological pathways, and scientific literature evidence; the first ontology and the second ontology are connected through cross-layer semantic relationships.

3. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 2, characterized in that, The graph neural network model is a graph attention network model; In the process of information propagation, the graph attention network model introduces gating weights based on the posterior credibility of edges to attenuate edges with credibility below a preset threshold.

4. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 3, characterized in that, The calculation of prior confidence based on the evidence source attributes of its associated evidence includes: Obtain the source authority level, study sample size, experimental methodological quality score, and the freshness of the evidence; Each attribute value is converted into an intensity score using a monotonically increasing calibration function; The intensity scores are combined into a comprehensive log-odds value using a nonlinear aggregation function in the log-odds domain. The combined logarithmic odds value is converted into the prior confidence level using the Sigmoid function.

5. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 4, characterized in that, The expression for the log-odds domain nonlinear aggregation function is: in, This represents the combined logarithmic probability value. Indicates the bias term. Indicates the first One source of evidence attribute, , Indicates the level of authority of the source. Indicates the number of research samples. Indicates the quality score of the experimental methodology. Indicates the age or novelty of the evidence. Indicates corresponding to The monotonically increasing calibration function, Indicates the corresponding to the first Individual evidence source attributes The non-negative weight coefficients are obtained by training on a historical conflict evidence dataset with the goal of minimizing prediction error and imposing sparsity constraints.

6. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 5, characterized in that, The method of fusing the prior confidence with likelihood information based on model scoring and repeated validation using a probabilistic graphical model includes: Prior confidence Mapped to Prior parameters of the probability distribution; where the mapping relationship is: , , The equivalent sample size to characterize the strength of prior beliefs Represents prior parameters, ; Using likelihood information extracted from evidence data that reflects the internal consistency of the model and the results of external repeated validation, the model is analyzed. The probability distribution is updated using conjugate methods to obtain the posterior distribution parameters. and Wherein, the posterior confidence level is determined by the posterior mean. Provided.

7. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 6, characterized in that, The preset stability assessment mechanism includes: Define the system state potential energy function System state potential energy function It is used to measure the difference between the posterior probability distribution of the knowledge graph in the current round and the previous round, as well as the degree to which the current rule violates the preset consistency constraint; If and only if the system state change caused by rule feedback satisfies At that time, the high-confidence rules are fed back as structured prior information to the probabilistic graphical model; wherein, It is a positive number. This represents the change in parameters of a probabilistic graphical model.

8. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 7, characterized in that, The constrained closed-loop self-optimization process also includes shadow update and drift circuit breaking steps; The shadow update and drift circuit breaker steps include: Before applying the rule feedback to the main knowledge graph, the update effect is simulated and verified in an offline shadow environment. If the verification metric does not improve or a significant concept drift in the data distribution is detected, the feedback loop is interrupted, and the application of the structured prior information is paused until the system returns to stability.

9. The method for constructing a knowledge graph of cardiovascular and cerebrovascular diseases according to claim 8, characterized in that, The dynamic adjustment of weights based on time information includes: The nodes and edges in the knowledge graph are assigned decay coefficients associated with timestamps, enabling the knowledge graph to present a dynamically evolving subgraph structure according to different time windows following the occurrence of clinical events of cardiovascular and cerebrovascular diseases.

10. A device for constructing a knowledge graph of cardiovascular and cerebrovascular diseases, characterized in that, include: The data preprocessing module is used to analyze multimodal research data in the field of cardiovascular and cerebrovascular diseases and extract candidate entity and relationship information; the multimodal research data includes text, images, waveforms and omics data; The ontology construction module is used to construct a unified knowledge ontology structure driven by semantic association based on a first ontology containing clinical diagnosis and treatment concepts and a second ontology containing scientific research semantic concepts. The semantic alignment module is used to perform cross-modal semantic alignment and fusion of candidate entities that point to the same real-world object extracted from different modal data based on a graph neural network model, and to map the candidate entities to the unified knowledge ontology structure to form entity nodes and relation edges of the knowledge graph. The conflict resolution module is used to calculate the prior confidence level of relational fact edges in the knowledge graph based on the evidence source attributes of their associated evidence, and to use a probabilistic graphical model to fuse the prior confidence level with the likelihood information based on model scoring and repeated verification to obtain the posterior confidence level, so as to evaluate and resolve conflicts between different evidence sources. The knowledge graph optimization module is used to sort and summarize the paths in the knowledge graph based on the posterior confidence. According to the preset stability evaluation mechanism, the high confidence rules summarized are fed back as structured prior information to the calculation process of the probabilistic graphical model. The weights are dynamically adjusted according to time information to form a constrained closed-loop self-optimization process, so that the knowledge graph continues to evolve with the addition of new data.

Images