Knowledge graph establishment method and device and computer device

Abstract

The application relates to the technical field of medical data processing, and discloses a knowledge graph establishment method and device and computer equipment, which comprises the following steps: obtaining literature mining data and clinical diagnosis and treatment data of a target disease; performing entity recognition and relationship extraction based on the literature mining data to establish an initial knowledge graph; wherein the initial knowledge graph comprises knowledge triples capable of representing semantic association relationships between medical entities, and each knowledge triple comprises an entity node and a relationship edge; performing statistical analysis based on the clinical diagnosis and treatment data, and performing granularity update processing on the initial knowledge graph by using obtained actual feature data to generate a target knowledge graph. The beneficial effect is that the dynamic optimization of the knowledge graph is driven by real-world clinical data, the problem of insufficient update granularity of a traditional knowledge graph construction method is effectively solved, and the generated target knowledge graph can provide more accurate knowledge support for clinical auxiliary decision-making.

Description

Technical Field

[0001] This application relates to the field of medical data processing technology, and in particular to a method, apparatus and computer device for establishing a knowledge graph. Background Technology

[0002] Knowledge graphs, as a core technology for structured knowledge representation and reasoning, have been widely applied in the field of medical clinical decision support. In the diagnosis and treatment of rare diseases, due to the low incidence rate, scattered cases, and limited clinical knowledge, constructing high-quality knowledge graphs is of great significance for improving diagnostic accuracy and reducing misdiagnosis and missed diagnosis.

[0003] However, most knowledge graph construction methods in related technologies are based on one-time knowledge extraction and static updates from medical literature. They are difficult to fully utilize detailed information such as symptom associations, examination thresholds, gene phenotype probabilities, and treatment effects formed in real-world clinical diagnosis and treatment. As a result, there are problems such as lagging knowledge graph updates, insufficient clinical applicability, and low update accuracy.

[0004] Therefore, there is a need for a method that can be based on real clinical data and flexibly update knowledge graphs to improve their dynamism and granularity. Summary of the Invention

[0005] This application aims to at least partially solve one of the technical problems in related technologies. To this end, this application proposes a method, apparatus, and computer device for building a knowledge graph. The main technical solutions adopted in this application include: Firstly, this application provides a method for establishing a knowledge graph, which includes: acquiring literature mining data and clinical diagnosis and treatment data of a target disease; performing entity recognition and relation extraction based on the literature mining data to establish an initial knowledge graph; wherein the initial knowledge graph includes knowledge triples that can represent semantic relationships between medical entities, and each knowledge triple contains entity nodes and relation edges; performing statistical analysis based on the clinical diagnosis and treatment data to obtain actual feature data; and performing granular update processing on the initial knowledge graph using the actual feature data to generate a target knowledge graph; wherein the granular update processing includes node updates for entity nodes and weight updates for relation edges.

[0006] Constructing an initial knowledge graph using literature mining data provides a basic framework for disease knowledge, ensuring broad knowledge coverage. Statistical analysis of clinical diagnosis and treatment data yields multi-dimensional real-world feature data, extracting fine-grained quantitative features of disease manifestations in the real world. Granularity update processing feeds these clinical statistical features back to the entity nodes and relation edges of the knowledge graph, elevating it from qualitative description to quantitative characterization. Ultimately, real-world clinical data drives the dynamic optimization of the knowledge graph, effectively addressing the problem of insufficient update granularity in traditional knowledge graph construction methods. The resulting target knowledge graph provides more accurate knowledge support for clinical decision support.

[0007] Optionally, the actual feature data includes symptom association data, examination distribution data, phenotypic probability data, and severity distribution data. Statistical analysis based on clinical diagnosis and treatment data yields the actual feature data, including: association rule mining based on symptom and sign data in the clinical diagnosis and treatment data to obtain symptom association data; statistical analysis based on laboratory data of confirmed patients in the clinical diagnosis and treatment data to obtain examination distribution data; probability calculation based on variation data and phenotypic data in the clinical diagnosis and treatment data to obtain phenotypic probability data; and property clustering based on symptom and severity data in the clinical diagnosis and treatment data to obtain severity distribution data.

[0008] By mining association rules to determine the true strength of associations between symptoms, the qualitative nature of association descriptions in literature is compensated for. Statistical analysis of test data yields disease-specific examination thresholds, improving the clinical applicability of diagnostic indicators. Probability calculations of variants and phenotypes clarify the correspondence between genotypes and clinical manifestations, providing a basis for precise diagnosis. Property clustering of symptom modification patterns enables fine-grained symptom characterization, enriching the detailed dimensions of knowledge. These four types of data quantify the real-world characteristics of diseases from different dimensions, providing a solid data foundation for subsequent granular updates to the knowledge graph.

[0009] Optionally, the weight update of the relation edge is performed as follows: For any knowledge triple, the corresponding target feature data is determined in the actual feature data based on the entity node and relation edge of the knowledge triple; the ratio is calculated based on the relation edge and the target feature data to determine the target balance coefficient; wherein, the relation edge has an initial weight; the target feature data has a statistical weight; an exponentially weighted moving average is calculated based on the target balance coefficient, the initial weight, and the statistical weight to obtain the target edge weight of the relation edge; the initial weight of the relation edge is replaced based on the target edge weight to complete the weight update of the relation edge.

[0010] By introducing a balance coefficient, a weighted fusion of historical knowledge and new clinical evidence is achieved, avoiding the knowledge mutations that may result from direct replacement and preventing the neglect of sample size differences by simple averaging. Furthermore, dynamically adjusting the balance coefficient based on sample size allows weight updates to adaptively reflect the strength of evidence; the larger the sample size, the greater the influence of new data on the weights, thus achieving a smooth and scientific evolution of knowledge.

[0011] Optionally, the relation edges also have historical sample sizes; the target feature data also has real sample sizes; the target balance coefficient is determined by calculating the ratio based on the relation edges and the target feature data, including: calculating the ratio based on the historical sample sizes and the real sample sizes to determine the target balance coefficient; wherein, the target balance coefficient is the ratio of the historical sample sizes to the total sample sizes; the total sample sizes are obtained by adding the real sample sizes and the historical sample sizes.

[0012] By dynamically determining the balance coefficient using the ratio of historical sample size to the current sample size, the weighting of historical knowledge and clinical data can be automatically adjusted based on the amount of new evidence: when the clinical sample size is large, the balance coefficient decreases accordingly, giving new evidence a higher weight in the update process; when the clinical sample size is small, the balance coefficient increases accordingly, retaining more weight for historical knowledge. This dynamic adjustment mechanism makes weight updates more scientific and reasonable, avoiding the lack of adaptability caused by fixed weight ratios.

[0013] Optionally, entity recognition and relation extraction are performed based on document mining data to establish an initial knowledge graph, including: preprocessing the document mining data to obtain standard text data; performing entity recognition and combination based on the standard text data to obtain candidate entity pairs; extracting relations based on the candidate entity pairs to obtain the target relation type of the candidate entity pairs; generating relation edges of the candidate entity pairs based on the target relation type to construct knowledge triples, and integrating all knowledge triples to form an initial knowledge graph.

[0014] By preprocessing the literature mining data to obtain standard text data, interference from irrelevant content is effectively eliminated. Then, entity recognition and combination are used to obtain candidate entity pairs, and relation extraction is performed on them, ultimately integrating them to form a complete initial knowledge graph. The entire process automates the transformation from unstructured medical literature to a structured knowledge base, providing a structured knowledge foundation for subsequent knowledge fusion and graph updates.

[0015] Optionally, entity recognition and combination are performed based on standard text data to obtain candidate entity pairs, including: entity recognition based on standard text data to obtain entity nodes; wherein, entity nodes have entity locations; and entity nodes are combined using entity locations to obtain candidate entity pairs.

[0016] By using entity location to restrict and orderly combine entities within the same sentence, the grammatical basis for constructing semantic relations is ensured for candidate entity pairs. Simultaneously, entity location narrows the candidate range for relation extraction, avoiding invalid pairings of entities across sentences, effectively improving the efficiency and accuracy of subsequent relation extraction.

[0017] Optionally, before performing granularity update processing on the initial knowledge graph using actual feature data, the method further includes: performing conflict detection based on actual feature data and the initial knowledge graph to obtain conflict identification results; if the conflict identification results indicate that the conflict originates from granularity differences, then performing granularity update processing based on actual feature data and the initial knowledge graph to generate the target knowledge graph.

[0018] By adding a conflict detection step before granularity updates, two different types of conflicts—granularity differences and data differences—can be effectively distinguished, enabling differentiated update strategies. Granularity differences are directly supplemented with finer-grained data to enrich the knowledge hierarchy; data differences undergo credibility assessment before weighted fusion to ensure knowledge accuracy. This conflict detection and triage mechanism improves the granularity and reliability of knowledge graph updates, providing more rigorous knowledge support for clinical decision support.

[0019] Optionally, the method further includes: if the conflict identification result indicates that the conflict originates from data differences, then performing a confidence calculation based on the actual feature data and the initial knowledge graph to obtain a confidence result; and performing granular update processing based on the confidence result, the actual feature data, and the initial knowledge graph to generate a target knowledge graph.

[0020] By introducing a credibility calculation mechanism for data discrepancies, a scientific basis for integration can be provided when new and old knowledge contradict each other, avoiding knowledge bias caused by simply accepting one side. At the same time, by comprehensively considering multiple dimensions of factors such as the authority, timeliness, and confidence of the source, the updated knowledge weights are made more objective and reliable, ensuring the overall quality of the target knowledge graph.

[0021] Secondly, this application provides a knowledge graph creation apparatus, the apparatus comprising: The raw data acquisition module is used to acquire literature mining data and clinical diagnosis and treatment data for the target disease; The initial knowledge graph construction module is used to perform entity recognition and relation extraction based on literature mining data to build an initial knowledge graph. The initial knowledge graph includes knowledge triples that can represent the semantic association between medical entities, and each knowledge triple contains an entity node and a relation edge. The real data acquisition module is used to perform statistical analysis based on clinical diagnosis and treatment data to obtain actual characteristic data. The granular update execution module is used to perform granular update processing on the initial knowledge graph using actual feature data to generate the target knowledge graph; the granular update processing includes node updates for entity nodes and weight updates for relation edges.

[0022] Thirdly, this application also provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 A flowchart of a method for building a knowledge graph according to an embodiment of this application; Figure 2 This is a flowchart of a method for constructing an initial knowledge graph according to an embodiment of this application; Figure 3 This is a flowchart of a method for constructing a target knowledge graph according to an embodiment of this application; Figure 4 This is a structural block diagram of a knowledge graph building apparatus according to an embodiment of this application; Figure 5 This is an internal structural diagram of a computer device provided according to an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] It should be noted that in the field of rare diseases, the construction of knowledge graphs for rare diseases mainly relies on two technical approaches.

[0027] One approach is a static construction method based on literature mining, which uses natural language processing techniques to extract entities and relationships from medical literature, clinical guidelines, and gene databases to form an initial knowledge graph. The other approach is a feedback optimization method based on clinical data, such as using real-world clinical data to optimize the knowledge base or decision-making model.

[0028] However, the above methods have obvious limitations: First, the sources of knowledge are limited, making it difficult to balance breadth and depth.

[0029] Among related technologies, while literature mining-based methods can acquire macro-level knowledge such as new diseases and genes, literature knowledge is generalized and lacks fine-grained characterization of the diversity of disease manifestations in the real world. Methods based on clinical data, while reflecting actual clinical experience, are limited by the sparsity of cases, making it difficult to discover new types and associations of rare diseases. Furthermore, these technologies lack a dual-source driving mechanism that effectively integrates literature knowledge and clinical experience, resulting in a difficulty in balancing the breadth and depth of knowledge graph coverage.

[0030] Second, knowledge updates lag behind, and there is a lack of a closed-loop evolution mechanism.

[0031] Traditional knowledge graphs are mostly static structures, relying on manual organization or periodic batch reconstruction for updates, and cannot continuously evolve with clinical practice. When new medical discoveries or new diagnostic and treatment experiences emerge, the knowledge graph struggles to respond in a timely manner, causing the knowledge support of clinical decision support systems to lag behind medical advancements.

[0032] Third, clinical feedback is not fully utilized, and there is a lack of fine-grained update methods.

[0033] Traditional methods involve clinical data feedback, but are only used to screen high-quality cases or optimize model parameters. They do not feed back the detailed feature information contained in the clinical data (such as the specific manifestation pattern of symptoms, the dynamic change pattern of examination indicators, and individual differences in treatment response) to the knowledge graph itself, and cannot achieve fine-grained updates of knowledge graph nodes, edge weights, and attribute dimensions.

[0034] Based on this, according to the embodiments of this application, an embodiment of a method for establishing a knowledge graph is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] This embodiment provides a method for establishing a knowledge graph, such as... Figure 1 As shown, the method includes the following steps: S110. Obtain literature mining data and clinical diagnosis and treatment data for the target disease.

[0036] The target disease can refer to the specific disease type that the knowledge graph construction is aimed at. For example, it can be a rare disease with characteristics such as low incidence, scattered cases, and limited clinical knowledge.

[0037] Literature mining data can refer to text data containing knowledge related to a target disease.

[0038] Specifically, for literature mining data, the latest literature data from multi-source medical literature databases can be captured periodically, and the captured raw literature data can be parsed for format and extracted for content. Then, standardized preprocessing operations can be performed on the extracted text to obtain literature mining data.

[0039] For example, an external medical literature database can be connected via an API interface, and then a search can be performed at a preset time (e.g., once a week) using the target disease as a keyword. This retrieves the latest literature whose titles or abstracts contain the target disease and uses it as literature mining data related to the target disease. Furthermore, it can also respond to various supplementary materials related to the target disease manually uploaded by users to supplement the literature mining data.

[0040] Clinical diagnosis and treatment data refers to real clinical data that comprehensively reflects the full information of diagnosed patients. Specifically, clinical diagnosis and treatment data can be obtained from the hospital's electronic medical record system.

[0041] For example, a search is first performed using the target disease as a keyword to obtain the raw clinical data of patients suffering from the target disease. This raw clinical data may include the patient's basic information (age, gender, ethnicity, family history), chief complaint and medical history (symptoms described in unstructured text, onset time, and disease course characteristics), physical examination records (description of signs and abnormalities), laboratory test results (laboratory indicator values, imaging descriptions, and pathology reports), diagnostic records (preliminary diagnosis, final diagnosis, and differential diagnosis), and treatment plans and efficacy (medication records, surgical records, follow-up results, and efficacy evaluation). The collected raw clinical data is then preprocessed. For example, duplicate records can be removed, missing values ​​can be processed, and the data format can be standardized. Then, named entity recognition and relation extraction can be performed on the unstructured text to transform it into a structured representation.

[0042] It is important to emphasize that information involving patient privacy (name, ID number, contact information, and specific date of visit) also needs to be anonymized to ensure compliance with medical data security and privacy protection regulations. Ultimately, this results in clinical diagnosis and treatment data that has been cleaned, structured, and anonymized.

[0043] S120. Entity identification and relation extraction are performed based on literature mining data to establish an initial knowledge graph.

[0044] The initial knowledge graph can refer to a knowledge network structure that contains medical entities related to the target disease and their semantic relationships, which may include knowledge triples that can represent the semantic relationships between medical entities.

[0045] The knowledge triple can be a basic knowledge unit that can describe the semantic relationship between two entities, and each knowledge triple contains an entity node and a relation edge.

[0046] For example, a knowledge triple can be formatted as (head entity, relation, tail entity), where the head entity and tail entity are entity nodes, and the relation is a relation edge. For instance, (GLA gene, pathogenic relation, Fabry disease) can be a knowledge triple, where "GLA gene" and "Fabry disease" are entity nodes, and "pathogenic relation" is a relation edge.

[0047] Specifically, an initial knowledge graph can be constructed using the following method: First, text preprocessing is performed on the literature mining data to obtain standard text data. Then, a pre-trained sequence labeling model is used to perform entity recognition on the standard text data to identify entity nodes such as diseases, symptoms, signs, genes, mutation types, drugs, examination indicators, and treatment plans, and the entity position of each entity is recorded. Next, for multiple entities identified within the same sentence, all possible entity pairs are enumerated as candidate entity pairs. Then, a pre-trained relation classification model is used to extract relations from the candidate entity pairs to determine whether a semantic relationship exists between each candidate entity pair and the target relation type (which may include pathogenic relations, clinical manifestation relations, diagnostic relations, treatment relations, prognostic relations, etc.). Finally, relation edges for candidate entity pairs are generated based on the target relation type, that is, entity pairs with semantic relationships and their relation types are combined into (head entity, relation, tail entity) forms to construct knowledge triples, and all knowledge triples are integrated to form the initial knowledge graph.

[0048] S130. Statistical analysis is performed based on clinical diagnosis and treatment data to obtain actual characteristic data.

[0049] Among them, actual characteristic data can refer to quantitative data that can reflect the actual manifestation characteristics of the disease, such as statistical characteristics such as distribution patterns and probability values ​​derived from real cases.

[0050] Specifically, actual characteristic data may include symptom association data, examination distribution data, phenotypic probability data, and severity distribution data.

[0051] Among them, symptom association data can refer to statistical association indicators calculated based on the co-occurrence frequency and temporal relationship of symptoms and signs in clinical diagnosis and treatment data.

[0052] For example, taking Fabry disease as the target disease, if the co-occurrence probability of "acromial sensory abnormalities" and "angiokeratoma" is 78% in the Fabry disease patient population, this probability value can be used as a symptom association data.

[0053] The test distribution data can refer to the threshold range or diagnostic cutoff value obtained after statistical analysis of the test index values ​​of confirmed patients.

[0054] For example, continuing from the previous example, assuming that the 90% reference range for Lyso-GB3 concentration in Fabry disease patients is calculated to be 1.2 ng / mL to 8.5 ng / mL using the percentile method; or that the optimal diagnostic cutoff value for Lyso-GB3 concentration is determined to be 3.5 ng / mL using ROC curve analysis, the above reference range and cutoff value can be used as a test distribution data.

[0055] Phenotypic probability data can refer to the frequency or conditional probability of occurrence of various clinical manifestations based on statistical analysis of patient groups with different gene mutation types in clinical diagnosis and treatment data.

[0056] For example, assuming that after statistical analysis of a patient population carrying the GLA gene c.427G>A mutation, the probability of kidney involvement is 43% and the probability of heart involvement is 28%, then the fact that the "GLA gene c.427G>A mutation" has a 43% probability of causing "kidney involvement" and a 28% probability of causing "heart involvement" can be used as a phenotypic probability data.

[0057] Severity distribution data can refer to the fine-grained manifestation patterns of symptoms and their distribution frequencies obtained by extracting and clustering symptom modifiers from unstructured disease records in clinical diagnosis and treatment data.

[0058] For example, assuming that for the symptom node of "pain", by mining the medical records, information such as "paroxysmal upper abdominal colic" accounts for 35% and "persistent dull pain in the lower back" accounts for 28% can be obtained. These modifiers and specific distribution information can be used as severity distribution data.

[0059] Specifically, statistical analysis based on clinical diagnosis and treatment data to obtain actual characteristic data can include: First, for symptom-related data, association rules can be mined based on symptom and sign data in clinical diagnosis and treatment data to obtain symptom-related data.

[0060] Symptom data can refer to textual descriptions of a patient's subjective feelings of discomfort. For example, it may include abnormal sensations in the extremities, pain, and weakness.

[0061] Vital signs data can refer to descriptive information about abnormal findings obtained through objective examination. For example, this may include elevated blood pressure, heart murmurs, and hepatosplenomegaly.

[0062] Specifically, the first step is to screen out patients diagnosed with the target disease from clinical diagnosis and treatment data, and extract all symptoms and signs that appear in their medical records and physical examination records.

[0063] A filtering threshold can then be set to filter out symptom combinations that occur infrequently in the data, and symptom combinations that occur more frequently than the filtering threshold can be designated as frequent itemsets.

[0064] Next, association rule mining is performed. Assuming that X and Y are two different symptoms or signs, association rules in the form of "X→Y" can be extracted from the frequent itemset, and the confidence (i.e., the conditional probability that a transaction containing X also contains Y) and lift (i.e., the ratio of the probability of Y occurring when X occurs to the probability of Y occurring independently) of each rule are calculated.

[0065] Subsequently, confidence and lift are used to determine the validity of a rule. For example, a minimum confidence threshold (e.g., 0.5) and a minimum lift threshold (e.g., 1) can be set. Only association rules exceeding these thresholds can be considered valid rules. After determining the validity of a rule, the confidence and lift of the valid rule can be used as the symptom association data for that rule.

[0066] For example, consider Fabry disease. First, select 100 patients diagnosed with Fabry disease from clinical data. For each patient, extract all symptoms and signs from their medical records. Suppose that statistically, 78 patients presented with "acromial paresthesia" and 65 patients presented with "angiokeratoma," with 55 patients presenting with both. Then, set the selection threshold to 0.1. Since both 78 / 100 = 0.78 and 65 / 100 = 0.65 are greater than 0.1, both of these presentations can be included in the frequent itemset.

[0067] Next, the confidence and lift of the association rule are calculated. Specifically, assuming the minimum confidence threshold is 0.5 and the minimum lift threshold is 1, the confidence of "acromial paresthesia → hemangiokeratoma" is approximately 0.705 (55 / 78), indicating that about 70.5% of patients with acromial paresthesia also have hemangiokeratoma. Its lift is approximately 1.085 (0.705 / (65 / 100)). After threshold testing, this association rule is considered valid, and therefore, the confidence of "acromial paresthesia → hemangiokeratoma" (confidence 0.705, lift 1.085) can be considered a symptom association data point.

[0068] Secondly, regarding the distribution data of examinations, statistical analysis can be performed based on the test data of confirmed patients in clinical diagnosis and treatment data to obtain the distribution data of examinations.

[0069] The test data can refer to the quantitative indicators generated by laboratory tests or auxiliary examinations, that is, quantifiable physiological or pathological parameters obtained through medical testing methods. For example, it may include Lyso-GB3 concentration value, creatine kinase activity value, white blood cell count, and echocardiographic parameters.

[0070] When calculating the distribution data of the examination, it is also necessary to first screen out patients diagnosed with the target disease from the clinical diagnosis and treatment data, and extract the numerical results of the corresponding test indicators.

[0071] Subsequently, the analysis method can be selected based on whether a control group is available, such as the percentile method or the ROC curve analysis method.

[0072] Specifically, if no control group data is available, the percentile method can be used to describe the distribution of the disease population itself. This involves sorting the indicator values, calculating the specified percentile values, and using the resulting reference interval as the examination distribution data. If control group data is available, the ROC curve method can be used. This involves iterating through each candidate cutoff value and calculating the sensitivity and specificity separately, selecting the cutoff value with the largest Youden index as the optimal threshold. Finally, the calculated optimal threshold is used as the examination distribution data.

[0073] For example, taking Lyso-GB3, a biomarker for Fabry disease, as an example, we can first select 100 confirmed patients (range 1.5–15.0 ng / mL) from clinical data as the case group and 150 healthy individuals (range 0.8–2.2 ng / mL) as the control group. Then, after removing obviously erroneous values, the concentration values ​​are sorted from smallest to largest.

[0074] For the percentile method, a specified percentile (such as the 5th percentile and the 95th percentile) can be calculated to obtain a 90% reference interval, which can be used as the reference range for the test indicator in the context of the target disease.

[0075] For the ROC curve method, each value can be used sequentially as a candidate cutoff value. For each candidate cutoff value, the sensitivity (the proportion of cases with a value greater than or equal to that value) and specificity (the proportion of controls with a value less than that value) can be calculated. Then, an ROC curve is plotted with 1-specificity on the x-axis and sensitivity on the y-axis. The Youden index is then calculated, and the cutoff value that maximizes the Youden index is selected as the optimal threshold, thus serving as the examination distribution data.

[0076] Subsequently, for phenotypic probability data, probability calculations can be performed based on the variation data and phenotypic data in clinical diagnosis and treatment data to obtain phenotypic probability data.

[0077] The variation data can refer to gene mutation information recorded in the patient's gene testing results. For example, this could include specific mutation types such as GLA c.427G>A and GLA c.679C>T. The phenotypic data can refer to the patient's clinical manifestations, for example, including kidney involvement, cardiac involvement, and nervous system involvement.

[0078] Specifically, the process begins by identifying a patient group carrying a specific gene mutation from clinical data and counting the total number of this group. Next, the number of patients exhibiting a specific clinical manifestation within this group is counted. Finally, the frequency of this specific clinical manifestation is calculated by dividing the number of patients exhibiting the manifestation by the total number of patients carrying the mutation, yielding the conditional probability that the gene mutation leads to this phenotype. This conditional probability is then used as the phenotypic probability data.

[0079] For example, consider the GLA gene c.427G>A mutation. Suppose 100 Fabry disease patients carrying this mutation are identified from clinical data, and 43 of them develop kidney involvement. Calculating the conditional probability, the probability of kidney involvement is 43 / 100 = 0.43. Then, by summarizing the mapping relationship between the mutation type and its corresponding phenotypic probability, we obtain the phenotypic probability data: [GLA gene c.427G>A mutation has a 0.43 probability of kidney involvement].

[0080] Finally, for the severity distribution data, we can perform property clustering based on the symptom data and severity data in the clinical diagnosis and treatment data to obtain the severity distribution data.

[0081] Among them, severity data can refer to descriptive words that can describe the characteristics of symptoms. These can include descriptive language that describes the nature, location, temporal pattern, aggravating and relieving factors, and severity of symptoms, such as nature / type modifiers (e.g., "paroxysmal", "persistent", "dull pain" and "sharp pain"), severity / intensity modifiers (e.g., "mild", "severe" and "intense"), triggering / aggravating / relieving factors (e.g., "worsened after eating" and "relieved after activity"), time / frequency modifiers (e.g., "nighttime" and "lasts for 2 hours"), and location / radiation modifiers (e.g., "upper abdomen" and "radiating to the back").

[0082] Specifically, symptom entities can first be identified from clinical diagnosis and treatment data. Then, dependency parsing combined with rule matching is used to extract modifiers corresponding to the symptom entities. Subsequently, multi-dimensional clustering is performed using modifiers of the same symptom, grouping similar combinations of "nature, location, and time" into the same subtype. Next, the frequency of occurrence of each subtype is counted, and the frequency of occurrence of each subtype is divided by the total number of samples to obtain the distribution frequency of each subtype, which is ultimately used as the severity distribution data.

[0083] For example, consider the symptom "pain". From 1000 medical records containing descriptions related to "pain", the symptom entity "pain" is identified. Dependency parsing is used to extract modifiers such as the nature, location, and temporal pattern of pain from each record. These modifiers are then combined and clustered to obtain different subtypes, and their distribution frequencies are statistically analyzed. For instance, the combination "paroxysmal + upper abdomen + triggered by eating" can be grouped into one subtype, appearing 350 times (35% frequency). The combination "persistent + lower back + at rest" can be grouped into another subtype, appearing 280 times (28% frequency). The combination "paroxysmal + extremities + nocturnal" can be grouped into a third subtype, appearing 220 times (22% frequency). The combination "burning pain + sole of the foot + after activity" can be grouped into a fourth subtype, appearing 150 times (15% frequency). These subtypes and their frequencies can then be used as severity distribution data.

[0084] By mining association rules to determine the true strength of associations between symptoms, the qualitative nature of association descriptions in literature is compensated for. Statistical analysis of test data yields disease-specific examination thresholds, improving the clinical applicability of diagnostic indicators. Probability calculations of variants and phenotypes clarify the correspondence between genotypes and clinical manifestations, providing a basis for precise diagnosis. Property clustering of symptom modification patterns enables fine-grained symptom characterization, enriching the detailed dimensions of knowledge. These four types of data quantify the real-world characteristics of diseases from different dimensions, providing a solid data foundation for subsequent granular updates to the knowledge graph.

[0085] In addition, actual characteristic data may also include quantified efficacy data, which refers to statistical indicators obtained after analyzing the efficacy of treatment in a group of patients using the same treatment regimen. This may include efficacy rate and confidence interval, median time to response, and incidence of adverse reactions, etc.

[0086] Specifically, for quantitative data on effectiveness, distribution statistics can be performed based on treatment data and efficacy evaluation data in clinical diagnosis and treatment data to obtain quantitative data on effectiveness.

[0087] Treatment data can refer to information about the treatment regimens received by the patient in the clinical medical records. For example, this may include enzyme replacement therapy, symptomatic supportive treatment, and drug names and dosages. Efficacy evaluation data can refer to information on the evaluation of treatment effects in the clinical medical records. For example, this may include the degree of symptom relief, improvement in indicators, occurrence of adverse reactions, and onset time.

[0088] Specifically, the process begins by identifying and counting a group of patients using the same treatment regimen from clinical data. Then, the number of patients in this group who meet the preset efficacy criteria and the number of patients experiencing various adverse reactions are counted. Simultaneously, the onset time for each patient (the time interval from the start of treatment to reaching the efficacy criteria) is extracted. Finally, the efficacy rate (number of effective patients divided by the total number of patients), the 95% confidence interval for the efficacy rate, the incidence of adverse reactions, and the median onset time (the median after sorting onset times) are calculated, collectively forming the quantitative efficacy data.

[0089] Ultimately, symptom correlation data, examination distribution data, phenotypic probability data, severity distribution data, and effect quantification data together constitute the actual characteristic data.

[0090] S140. Perform granular update processing on the initial knowledge graph using actual feature data to generate the target knowledge graph.

[0091] Among them, the target knowledge graph can refer to a knowledge network structure that integrates the breadth of literature knowledge and the depth of clinical experience.

[0092] It should be noted that granular update processing can include node updates for entity nodes and weight updates for relation edges.

[0093] Specifically, node updates for entity nodes can be performed in the following ways: For any entity involved in the actual feature data, the initial knowledge graph can be traversed to determine whether there is a corresponding entity node in the initial knowledge graph (for example, by calculating the similarity of vectors; if the similarity exceeds a preset threshold, it is considered that there is a corresponding entity node).

[0094] If a corresponding entity node is determined to exist, there is no need to create a new node. If it is determined not to exist, the entity can be treated as a new entity, and a corresponding entity node can be created in the initial knowledge graph, thus completing the operation of adding an entity node. Simultaneously, if an entity node in the initial knowledge graph no longer has any corresponding clinical records or statistical support in the actual feature data, it can be marked as obsolete and deleted or archived, thus completing the operation of deleting an entity node.

[0095] For example, when a new mutation type "GLA c.1234T>C" not included in the initial knowledge graph appears in clinical data, and entity alignment confirms that there is no corresponding node for this mutation in the graph, a new entity node for this gene mutation can be added to the graph. If clinical data verifies that a symptom entity is a redundant node with an incorrect name, it will be deleted.

[0096] Specifically, the weight update for relation edges can be performed as follows: For any knowledge triple, firstly, the corresponding target feature data can be determined in the actual feature data based on the entity nodes and relation edges of the knowledge triple. Then, a ratio calculation is performed based on the relation edges and target feature data to determine the target balance coefficient. Next, an exponentially weighted moving average is calculated based on the target balance coefficient, initial weights, and statistical weights to obtain the target edge weight of the relation edge. Finally, the initial weight of the relation edge is replaced based on the target edge weight to complete the relation edge weight update.

[0097] Among them, target feature data can refer to specific statistical data items in the actual feature data that directly correspond to the entities and relations of the current knowledge triple, such as the confidence value of a certain symptom pair in the corresponding symptom association data, or a specific cutoff value in the inspection distribution data.

[0098] Specifically, the head entity name, relation type, and tail entity name of the knowledge triple can be used as matching criteria to search through each entry in the actual feature data. If, through similarity judgment, it is found that the disease / symptom / gene concept of a certain entry is consistent with the head entity, and the manifestation / indicator / treatment concept is consistent with the tail entity, then the statistical value corresponding to that entry can be used as the target feature data of the knowledge triple.

[0099] After determining the target feature data, the ratio of the relation edges and the target feature data can be calculated to determine the target balance coefficient.

[0100] The target balance coefficient can be a weighting coefficient that controls the distribution ratio between the weight of historical knowledge and the weight of new clinical evidence.

[0101] It should be noted that relation edges have a historical sample size, that is, the total number of case samples on which the relation edge was based during the initial construction of the knowledge graph or during historical updates. Target feature data, on the other hand, has a real sample size, that is, the number of real-world patient cases on which the target feature data was based during this statistical analysis.

[0102] Specifically, the ratio can be calculated by using the historical sample size and the actual sample size to determine the target balance coefficient.

[0103] The total sample size can be obtained by adding the actual sample size and the historical sample size, representing the total number of samples participating in this update calculation.

[0104] The target balance coefficient can be the ratio of the historical sample size to the total sample size, and can be calculated using the following formula: α = N old / N total =N old / (N clinical +N old ) In the formula, α represents the target balance coefficient; N old N represents the historical sample size. clinical N represents the actual sample size; total This represents the total sample size.

[0105] By dynamically determining the balance coefficient using the ratio of historical sample size to the current sample size, the weighting of historical knowledge and clinical data can be automatically adjusted based on the amount of new evidence: when the clinical sample size is large, the balance coefficient decreases accordingly, giving new evidence a higher weight in the update process; when the clinical sample size is small, the balance coefficient increases accordingly, retaining more weight for historical knowledge. This dynamic adjustment mechanism makes weight updates more scientific and reasonable, avoiding the lack of adaptability caused by fixed weight ratios.

[0106] Furthermore, the relation edges also have initial weights, and the target feature data also have statistical weights. After determining the target balance coefficient, an exponentially weighted moving average can be calculated based on the target balance coefficient, initial weights, and statistical weights to obtain the target edge weight of the relation edges.

[0107] The initial weight can refer to the original association strength value of the relation edge in the initial knowledge graph, while the statistical weight can refer to the corresponding statistical association strength value (such as confidence, effectiveness, and probability value) in the target feature data. The target edge weight can refer to the new association strength value determined after the corresponding relation edge has been updated in granularity.

[0108] For example, the target edge weight can be calculated using the following formula: W new =α×W old +(1-α)×W clinical In the formula, W new Represents the objective edge weight; α represents the objective balance coefficient; W old W represents the initial weights; clinical Represents statistical weight.

[0109] Finally, the initial weights of the relation edges can be replaced based on the target edge weights to complete the weight update of the relation edges. That is, the calculated target edge weight value is used to overwrite the original initial weight value of the relation edge, so that the association strength of the relation edge reflects the latest statistical distribution in the real world.

[0110] By introducing a balance coefficient, a weighted fusion of historical knowledge and new clinical evidence is achieved, avoiding the knowledge mutations that may result from direct replacement and preventing the neglect of sample size differences by simple averaging. Furthermore, dynamically adjusting the balance coefficient based on sample size allows weight updates to adaptively reflect the strength of evidence; the larger the sample size, the greater the influence of new data on the weights, thus achieving a smooth and scientific evolution of knowledge.

[0111] Optionally, each knowledge triple in the knowledge graph may also be accompanied by additional metadata, which may refer to auxiliary information features that can describe the source and quality attributes of the knowledge triple, and may include the initial weight of the knowledge triple, source identifier, and extraction timestamp, etc.

[0112] It should be noted that the relation classification model for relation extraction during the construction of knowledge triples can output a probability value between 0 and 1 for each candidate entity pair. This probability value, after being normalized by Softmax, becomes the model's confidence level (i.e., the initial weight) for the judgment result of the knowledge triple. At the same time, the system records the source document identifier (such as the PMID of PubMed) and extraction timestamp of the triple, which together constitute the additional metadata of the triple.

[0113] Correspondingly, in addition to updating the node of entity nodes and the weight of relation edges, granular updates can also include updating the attributes of additional metadata.

[0114] Specifically, attribute updates can be performed as follows: First, entity comparison is conducted using actual feature data and the initial knowledge graph to identify the target entity node in the initial knowledge graph whose attributes need updating. Then, detailed feature information matching the entity node can be extracted from the actual feature data (such as the common fever patterns of the symptom, the reference range of the test indicator in a specific disease subtype, or the effective probability of the treatment plan in different age groups). This detailed feature information is then organized into attribute key-value pairs according to a preset data structure format, and metadata such as confidence level, update time, and sample size are added to each attribute. Finally, the organized attribute key-value pairs are assigned to the attribute fields of the target entity node, completing the attribute update.

[0115] For example, consider the "fever" symptom node. Assume the initial knowledge graph for the "fever" node only contains the basic attribute "definition: elevated body temperature". The actual feature data's degree distribution includes a detailed statistical analysis of the "fever" symptom's manifestation patterns: remittent fever accounts for 65%, continuous fever for 20%, and intermittent fever for 15%. This statistic is based on real clinical data from 500 patients, with a confidence level of 0.94.

[0116] The aforementioned distribution information can then be organized into attribute key-value pairs in the format {attribute name: "common fever pattern distribution", attribute value: {relaxative fever: 0.65, sustained fever: 0.20, intermittent fever: 0.15}, confidence level: 0.94, update date: "2026-05-15", sample size: 500}, and assigned to the dynamic attribute field of the "fever" node. After the update is complete, the diagnostic model can obtain its fine-grained fever pattern distribution information when calling the "fever" node, thereby assisting doctors in achieving more accurate symptom matching and disease prediction.

[0117] Furthermore, quality verification can be performed after the granularity update is completed.

[0118] Specifically, quality verification can include consistency verification and confidence verification.

[0119] Consistency verification can refer to the process of checking whether there are logical contradictions between the updated entity nodes, relation edges, and additional metadata and the original knowledge of the knowledge graph. For example, it can be used to determine whether the same entity node is simultaneously assigned mutually exclusive attribute values ​​or relation labels.

[0120] Specifically, consistency verification can be performed in the following ways: First, a rule base for contradictions is predefined. This rule base can store judgment rules for various logical contradictions, such as mutually exclusive relationships including that the same drug should not be labeled as both "recommended drug" and "contraindicated drug" at the same time, and that the same disease should not be labeled as both "benign" and "malignant" at the same time.

[0121] Subsequently, after each granularity update operation, all knowledge triples and attribute information involved in this update are extracted, including entity type, relation type, attribute name, and attribute value. The extracted new knowledge is then retrieved and compared within the initial knowledge graph.

[0122] Specifically, for numerical attributes, interval algebra can be used to determine if there are irreconcilable contradictions in the value range; if so, it is considered contradictory knowledge. For relational knowledge, rules in the contradiction rule base are applied one by one for matching; if a contradictory rule is matched, it is considered contradictory knowledge. Finally, conflict markers are added to all contradictory knowledge for submission to manual review.

[0123] For example, consider the detection of conflicting drug relationships. Assume the initial knowledge graph already contains the knowledge "Amitriptyline → Recommended Drug → Fabry Disease Neuropathic Pain" (i.e., amitriptyline is a recommended drug for Fabry disease neuropathic pain). This update introduces a new knowledge "Amitriptyline → Contraindicated Drug → Fabry Disease Neuropathic Pain" (i.e., amitriptyline is a contraindicated drug for Fabry disease neuropathic pain). The system extracts the relationship type of the new knowledge as "Contraindicated Drug," and retrieves existing relationships in the initial knowledge graph where the head entity "Amitriptyline" and the tail entity "Fabry Disease Neuropathic Pain" have the same relationship type as "Recommended Drug."

[0124] The search results are matched with the contradiction rule base. The mutual exclusion relationship in the rule base stipulates that the same drug should not be labeled as both "recommended drug" and "contraindicated drug" at the same time. Therefore, the new knowledge "amitriptyline → contraindicated drug → Fabry disease neuralgia" is determined to be contradictory knowledge, a conflict mark is added to it, and it is pushed to the manual review queue for expert confirmation.

[0125] Confidence verification can refer to the verification process of evaluating and filtering the confidence of updates based on statistical sample size.

[0126] It should be noted that in the statistical analysis of actual feature data, each statistical value (such as the confidence level of symptom-related data, the cutoff value of examination distribution data, etc.) will have a corresponding statistical sample size. During updates, if the statistical sample size is higher than a preset threshold (e.g., 30 cases), the statistical value is considered to have passed the confidence check and can be updated normally. If the statistical sample size is lower than the preset threshold (e.g., 30 cases), the statistical value is considered to have low confidence and is considered a low-confidence update. Updates with low confidence can be temporarily suspended and await confirmation by human experts.

[0127] For example, consider the update of the association rule for "acromial sensory abnormalities → hemangiokeratoma". Through symptom association data mining of actual feature data, the confidence level of this association rule was found to be 0.78, corresponding to a statistical sample size of 25 cases. Assuming the preset minimum sample size threshold is 30 cases, since 25 < 30, this update entry is judged as a low-confidence update, a "low_confidence" status is added to it, and it is pushed to the manual review queue.

[0128] In the above implementation, constructing an initial knowledge graph using literature mining data provides a basic framework for disease knowledge, ensuring broad knowledge coverage. Statistical analysis of clinical diagnosis and treatment data yields multi-dimensional real-world feature data, extracting fine-grained quantitative features of disease manifestations in the real world. Granularity update processing feeds these clinical statistical features back to the entity nodes and relation edges of the knowledge graph, achieving an improvement from qualitative description to quantitative characterization. Ultimately, real-world clinical data drives the dynamic optimization of the knowledge graph, effectively solving the problem of insufficient update granularity in traditional knowledge graph construction methods. The generated target knowledge graph can provide more accurate knowledge support for clinical decision support in rare diseases.

[0129] In some implementation methods, please refer to the appendix. Figure 2 Entity identification and relation extraction are performed based on literature mining data to build an initial knowledge graph, including: S210. Perform text preprocessing on the literature mining data to obtain standard text data.

[0130] Standard text data can refer to standardized text data with a uniform format, that is, data obtained after processing the literature mining data by removing irrelevant content, segmenting paragraphs, splitting sentences, and tagging parts of speech.

[0131] Specifically, the first step is to clean the literature mining data, removing irrelevant text blocks and duplicate page information. Then, the cleaned text blocks are split into independent paragraphs according to paragraph boundaries, and each paragraph is further segmented into sentences, resulting in independent sentence units. Next, a pre-built medical dictionary is used to perform word segmentation and recognition on each sentence unit, and part-of-speech tagging is applied to each identified word. Finally, the sentence units with attached part-of-speech tags are used as standard text data.

[0132] For example, let's take a medical literature abstract related to Fabry disease as an example from the literature mining data. First, the format can be parsed to extract the main text. Then, regular expressions are used to match and remove irrelevant text blocks such as reference citations, funding statements, and conflict of interest statements. At the same time, the header and footer position information after PDF parsing is used to remove duplicate content.

[0133] Next, newline characters or double newline characters in the text can be used as paragraph boundaries to divide the cleaned text block into several natural paragraphs. Then, Chinese periods, exclamation marks, and question marks can be used as sentence boundaries to divide each paragraph into sentence units.

[0134] Then, a word segmenter loaded with a medical dictionary (which consists of about 5,000 terms covering medical terms related to rare diseases) can be used to perform word segmentation and recognition, thereby correctly identifying the medical terms contained in the sentence unit and marking them as a complete word unit.

[0135] Finally, the part-of-speech tagging function of natural language processing tools can be used to identify and tag all complete words in each sentence unit with part-of-speech tags such as disease nouns, symptom nouns, and modifiers, and finally obtain sentence units containing part-of-speech tags such as [disease noun: Fabry disease] and [symptom noun: acromegaly], which can be used as standard text data.

[0136] S220. Entity recognition and combination are performed based on standard text data to obtain candidate entity pairs.

[0137] Candidate entity pairs can refer to a combination of two entities identified from the same sentence that are to be judged.

[0138] Specifically, entity recognition and combination are performed based on standard text data to obtain candidate entity pairs, including: entity recognition based on standard text data to obtain entity nodes; and combination processing of entity nodes using entity locations to obtain candidate entity pairs.

[0139] Entity nodes can refer to medical concept units with independent semantics identified from text, that is, entity objects that exist as nodes in a knowledge graph.

[0140] Specifically, an entity node has an entity location, which can refer to positioning information used to determine the start and end boundaries of the entity object, such as the start character index and end character index of the entity object in standard text data.

[0141] Optionally, entity recognition based on standard text data to obtain entity nodes can be achieved by using a pre-trained sequence labeling model to perform entity recognition on all sentence units contained in the standard text data.

[0142] This sequence labeling model can use BioBERT as the underlying semantic encoder, and then stack a Conditional Random Field (CRF) layer on top of it. BioBERT is used to convert each character in the input into a vector representation that contains contextual semantics. The CRF layer is used to learn the transition constraints between labels and optimize the output label sequence.

[0143] Furthermore, the model can employ a BIO (Browser-Initialization) labeling scheme, which predicts a label for each input character, in the format {B / I / O}-{entity type}. Here, B represents the starting character of the entity, I represents a character within the entity, and O represents a non-entity character. After the model outputs the label sequence, all consecutive character segments labeled B-{entity type} and subsequent I-{entity type} are extracted as entity nodes. The entity position is then determined by the starting index of the character corresponding to the B label and the ending index of the character corresponding to the last I label. Finally, all identified entity nodes are aggregated, retaining their respective entity types and positions.

[0144] After obtaining the entity nodes, the entity nodes can be combined using their locations to obtain candidate entity pairs.

[0145] Specifically, sentence units can be used as the smallest combination unit. Only entity nodes identified within the same sentence unit are included in the combination scope. A full combination enumeration without replacement is performed on all entity nodes within each sentence to generate all possible unordered entity pairs. Then, according to the order of entity positions, the entity with the earlier starting position is determined as the head entity, and the entity with the later starting position is determined as the tail entity, forming the direction of the initial candidate pairs. Next, the number of characters separating the two entities in each initial candidate pair can be calculated. If the distance exceeds a preset character threshold, the pair is discarded, thus eliminating entity pairs that are too far apart (likely without direct relation). Finally, the remaining initial candidate pairs are combined to obtain the candidate entity pairs.

[0146] Next, an interval filtering operation can be performed. For example, the number of characters separating the two entities in each initial candidate pair can be calculated. If the interval exceeds a preset character threshold, the pair is discarded, thus eliminating entity pairs that are too far apart and are unlikely to be directly related. Finally, the initial candidate pairs that pass the filtering are aggregated to obtain candidate entity pairs.

[0147] By using entity location to restrict and orderly combine entities within the same sentence, the grammatical basis for constructing semantic relations is ensured for candidate entity pairs. Simultaneously, entity location narrows the candidate range for relation extraction, avoiding invalid pairings of entities across sentences, effectively improving the efficiency and accuracy of subsequent relation extraction.

[0148] S230. Extract relations based on candidate entity pairs to obtain the target relation type of the candidate entity pairs.

[0149] The target relationship type can refer to the category label of the semantic association between two entities in a candidate entity pair.

[0150] It is understandable that different candidate entity pairs may have different target relationship types, such as pathogenic relationships (e.g., genes and diseases), clinical manifestation relationships (e.g., diseases and symptoms), diagnostic relationships (e.g., test indicators and diseases), treatment relationships (e.g., drugs and diseases), and prognostic relationships (e.g., diseases and outcomes).

[0151] Specifically, extracting the target relation type of candidate entity pairs based on candidate entity pairs can be achieved in the following ways: For each candidate entity pair, the sentence unit in which it belongs is first restored, and the position information of the sentence unit and the head and tail entities in the candidate entity pair are input into the pre-trained relation classification model.

[0152] It should be noted that this relation classification model can use RoBERTa-wwm-ext as the backbone encoder, with a bidirectional attention pooling layer and a fully connected classification layer superimposed on top. RoBERTa-wwm-ext can be used to convert each character or word in the input sentence into a vector representation carrying contextual semantic information. The bidirectional attention pooling layer can be used to calculate the attention weights between the head entity and each character in the sentence and sum them to obtain the head entity's context vector; similarly, the context vector of the tail entity can be calculated. The fully connected classification layer can be used to receive the concatenated comprehensive feature vector and output the classification result.

[0153] After inputting the sentence units and the positional information of the head and tail entities in the candidate entity pairs into the pre-trained relation classification model, the model first locates the start and end positions of the head and tail entities, and then performs average pooling on the character vectors within their respective position ranges to obtain the head entity vector and tail entity vector. A bidirectional attention pooling layer further calculates the head entity context vector and tail entity context vector. Subsequently, the head entity vector, tail entity vector, head entity context vector, tail entity context vector, and the difference between the head and tail entity vectors are concatenated to form a comprehensive feature vector.

[0154] Furthermore, the comprehensive feature vector can be input into a fully connected classification layer to output the original scores of each relation type. After Softmax normalization, the probability distribution (i.e., confidence level) of each relation type can be obtained. Finally, the relation type with the highest confidence level can be used as the target relation type of the candidate entity pair.

[0155] It should be noted that if the highest confidence level is still lower than the preset confidence threshold (e.g., 0.5), it can be determined that the candidate entity pair does not have any target relation type, and there is no need to generate the corresponding knowledge triple.

[0156] For example, for a candidate entity pair (Fabry disease, acrosensory abnormalities), assuming that the probabilities of each relation type output by the model are: pathogenic relation 0.03, clinical manifestation relation 0.92, diagnostic relation 0.02, treatment relation 0.01, prognostic relation 0.02, and the highest probability 0.92 is higher than the threshold 0.5, then the clinical manifestation relation can be used as the target relation type for the candidate entity pair.

[0157] S240. Generate relation edges for candidate entity pairs based on the target relation type to construct knowledge triples, and integrate all knowledge triples to form an initial knowledge graph.

[0158] Specifically, for candidate entity pairs whose target relation type has been determined in S230, the head entity and tail entity in the candidate entity pair are respectively used as the head entity and tail entity of the knowledge triple, and their target relation type is used as the relation edge connecting the head and tail entities, and combined into a structured unit in the form of (head entity, relation edge, tail entity), thereby completing the construction of a knowledge triple.

[0159] Then, all candidate entity pairs can be traversed, and the above operation can be repeated for each candidate entity pair with a determined target relation type to obtain several knowledge triples. Finally, all the above knowledge triples can be summarized and organized and imported according to the storage format of a graph database to form an initial knowledge graph.

[0160] Optionally, additional metadata can be added to the existing knowledge triples. As mentioned above, this additional metadata may include the source document identifier (such as the PMID of PubMed), the current timestamp, and the confidence level corresponding to the target relation type output by the relation classification model in S230.

[0161] For example, continuing from the previous example, it has been determined that the target relation type for the candidate entity pair (Fabry disease, acrosensory impairment) is "clinical manifestation relation" with a confidence level of 0.92. After generating relation edges and constructing knowledge triples, a knowledge triple of the form (Fabry disease, clinical manifestation relation, acrosensory impairment) can be obtained, where "Fabry disease" is the head entity node, "clinical manifestation relation" is the relation edge, and "acrosensory impairment" is the tail entity node. Additional metadata in the form of {Source Document PMID: "12345678", Extraction Timestamp: "2026-05-15T10:30:00Z", Confidence Level: 0.92} can also be added to this triple.

[0162] Similarly, assuming that the target relation type of the candidate entity pair (Fabry disease, GLA gene) is "pathogenic relation" after S230 judgment and the confidence level is 0.88, a knowledge triple (Fabry disease, pathogenic relation, GLA gene) can be generated and the corresponding additional metadata can be attached.

[0163] Finally, the generated knowledge triples (Fabry disease, clinical manifestation relationship, acrosensory abnormality) and (Fabry disease, pathogenic relationship, GLA gene) can be summarized, organized according to the node and edge format of the graph database, and then imported and stored to form an initial knowledge graph containing entity nodes such as "Fabry disease", "acromial sensory abnormality", and "GLA gene" and their associated edges.

[0164] In the above implementation, standard text data is obtained by preprocessing the literature mining data, effectively eliminating interference from irrelevant content. Then, candidate entity pairs are obtained through entity recognition and combination, and relationships are extracted, ultimately integrating them to form a complete initial knowledge graph. The entire process automates the transformation from unstructured medical literature to a structured knowledge base, providing a structured knowledge foundation for subsequent knowledge fusion and graph updates.

[0165] In some implementation methods, please refer to the appendix. Figure 3 Before performing granular update processing on the initial knowledge graph using actual feature data, the method also includes: S310. Conflict detection is performed based on actual feature data and the initial knowledge graph to obtain conflict identification results.

[0166] Among them, the conflict identification result can refer to the judgment result that characterizes whether the knowledge between the actual feature data and the initial knowledge graph is consistent, and can indicate whether a conflict exists and the type of conflict source.

[0167] For example, the conflict identification results can include three types of results: no conflict, conflict exists and the conflict originates from granularity differences (i.e., the new knowledge and the old knowledge are different in terms of coarseness at the conceptual level), and conflict exists and the conflict originates from data differences (i.e., the new knowledge and the old knowledge have substantial contradictions in terms of numerical values ​​or values).

[0168] Specifically, for any statistical feature item in the actual feature data, the entity node corresponding to that statistical feature item can be used as an index to search for whether a corresponding entity node and relation edge exist in the initial knowledge graph. If it does not exist, there is no need to perform conflict detection, and it is directly treated as new knowledge.

[0169] If corresponding entity nodes and relation edges exist, the large language model can be further utilized to make concept hierarchy judgments based on the entity and relation types corresponding to the actual feature data and the entity nodes and relation edges corresponding to the knowledge triples.

[0170] If the result of the concept level judgment indicates that the two are coarse-grained descriptions of the same concept (e.g., there is an ancestor-descendant relationship in a medical ontology), then the conflict can be determined to be due to the difference in granularity, and the corresponding conflict identification result can be generated.

[0171] If the conceptual level assessment indicates that the two are different linguistic representations of the same concept (i.e., the same concept in the medical ontology) and have identical relationship types, then the initial weights of the relation edges of the knowledge triple are further extracted, and the statistical values ​​in the actual feature data are compared with the initial weights in the initial knowledge graph. If the numerical comparison results show that they are inconsistent, then the conflict can be determined to originate from data differences, and a corresponding conflict identification result is generated.

[0172] For example, suppose the actual feature data includes a symptom association: the association strength between "fever" and "respiratory infection" is 0.5. The entity information for this entry is extracted as "fever" (symptom) and "respiratory infection" (disease), with a statistical value of 0.5. First, the entity nodes "fever" and "respiratory infection" can be used as indexes for searching the initial knowledge graph. The search reveals that the graph does not contain the entity node "fever," but it does contain the entity nodes "high fever" and "respiratory infection," and there is a "symptom present" relationship edge between "high fever" and "respiratory infection," with an initial weight of 0.6. Since there is no perfectly matching entity node "fever" in the graph, the system further queries a medical ontology (such as SNOMED CT), finding that "fever" and "high fever" have a parent-child relationship in the ontology (i.e., "high fever" belongs to "fever," and "high fever" is a subclass, while "fever" is a parent class). Therefore, it is determined that they belong to different granular descriptions of the same concept, and the conflict arises from the difference in granularity, generating corresponding conflict identification results.

[0173] Suppose another symptom association is: the association strength between "Fabry disease" and "acromial paresthesia" is 0.78. Then, by searching using the entity nodes "Fabry disease" and "acromial paresthesia" as indexes, it is found that there are already perfectly matching entity nodes "Fabry disease" and "acromial paresthesia" in the graph, and there is a "symptom present" relationship edge between the two, with an initial weight of 0.70.

[0174] Subsequently, a large language model was used to determine the conceptual hierarchy of the two entities.

[0175] If the large language model queries the medical ontology and returns two entities that correspond to the same concept in the UMLS semantic network (i.e., consistent CUI encoding), it means that there is no parent-child relationship or difference in granularity; they are different linguistic expressions of the same concept (i.e., synonymous relationships). Subsequently, since the relationship types are exactly the same (both are "symptom-related"), the initial weight of this relationship edge (0.70) can be extracted, and the statistical value of 0.78 in the actual feature data can be compared with the initial weight of 0.70. The comparison result shows that the two are inconsistent (difference of 0.08), therefore, it can be determined that the conflict originates from data differences, and the corresponding conflict identification result is generated.

[0176] S320. If the conflict identification result indicates that the conflict originates from granularity differences, then granularity update processing is performed based on the actual feature data and the initial knowledge graph to generate the target knowledge graph.

[0177] Specifically, if the conflict identification results indicate that the conflict originates from differences in granularity, meaning that the new and old knowledge are not substantially contradictory but rather differ in the coarseness of their conceptual levels, then the new and old knowledge need to be merged according to their ontology hierarchical relationships. Therefore, granularity update processing can be performed based on the actual feature data and the initial knowledge graph to generate the target knowledge graph.

[0178] It should be noted that the specific limitations of granularity update processing (including node updates of entity nodes, weight updates of relation edges, and attribute updates, etc.) can be found in the relevant descriptions in S140, and will not be repeated here.

[0179] Optionally, the method further includes: if the conflict identification result indicates that the conflict originates from data differences, then performing a confidence calculation based on the actual feature data and the initial knowledge graph to obtain a confidence result; and performing granular update processing based on the confidence result, the actual feature data, and the initial knowledge graph to generate a target knowledge graph.

[0180] It's important to note that, unlike granularity differences, data differences often manifest as substantial inconsistencies in numerical values ​​or interpretations between new and old knowledge at the same conceptual level; that is, there may be contradictions between the old and new knowledge. To ensure the accuracy and reliability of knowledge updates and prevent erroneous knowledge from contaminating the knowledge graph, credibility calculations are necessary to determine which side of the knowledge should be accepted.

[0181] Specifically, for conflicting entries identified as data discrepancies, the reliability information of old knowledge in the initial knowledge graph and the reliability information of new knowledge in the actual feature data are first obtained separately. Then, based on the reliability information of the old and new knowledge, the actual credibility score corresponding to the actual feature data and the initial credibility score corresponding to the initial knowledge graph are calculated according to preset scoring rules. Finally, the actual credibility score and the initial credibility score are compared to obtain the credibility result.

[0182] Reliability information refers to multi-attribute information used to assess the credibility of knowledge, specifically including the authority score, timeliness score, and confidence score of conflicting items. The authority score characterizes the authority of the knowledge source, the timeliness score characterizes the freshness of the knowledge, and the confidence score characterizes the certainty of the knowledge itself.

[0183] Specifically, for existing knowledge in the initial knowledge graph, the authority score can be determined by its source type. For example, it can be pre-defined that the authority score for knowledge from clinical guidelines is 1.0, for knowledge from randomized controlled trials is 0.9, for knowledge from literature reviews is 0.8, for knowledge from cohort studies is 0.7, and for knowledge from case reports is 0.5. For new knowledge in actual characteristic data, the authority score can be determined by its sample size. For example, a sample size less than 50 scores 0.4, a sample size between 50 and 100 scores 0.5, a sample size between 100 and 200 scores 0.6, a sample size between 200 and 300 scores 0.7, and a sample size greater than 300 scores 0.8, etc.

[0184] The timeliness score can be calculated based on the number of months since the publication or collection date, with higher scores for more recent publications. The confidence score is calculated by extracting the initial weights of the initial knowledge graph triples and the statistical weights of the actual feature data, which are then used as the confidence score.

[0185] Furthermore, after obtaining the reliability information, a weighted sum can be performed using the following formula to obtain the actual credibility score corresponding to the actual feature data and the initial credibility score corresponding to the initial knowledge graph.

[0186] Q1=q 1a ×β1+ q 1b ×β2+q 1c ×β3 Q2=q 2a ×β1+ q 2b ×β2+q 2c ×β3 In the formula, Q1 represents the actual credibility score; q 1aβ1 represents the authority score corresponding to the actual feature data; β1 represents the authority score weighting coefficient, which can be 0.5; q 1b β2 represents the timeliness score corresponding to the actual feature data; β2 represents the weighting coefficient of the timeliness score, which can be 0.2; q 1c β3 represents the confidence score corresponding to the actual feature data; β3 represents the confidence score weighting coefficient, which can be 0.3; Q2 represents the initial confidence score; q 2a q represents the authority score corresponding to the initial knowledge graph; 2b q represents the timeliness score corresponding to the initial knowledge graph; 2c This represents the confidence score corresponding to the initial knowledge graph.

[0187] Finally, the actual credibility score can be compared with the initial credibility score to obtain the credibility result. This credibility result can refer to a comparative conclusion that characterizes the degree of credibility of the two parties in the conflict.

[0188] For example, taking the data difference in the association strength between Fabry disease and "discomfort in the extremities" as an example, the confidence calculation process can be as follows: Regarding prior knowledge, assuming the initial knowledge graph originates from a literature review published 24 months ago, its authority score is 0.8; the timeliness score is calculated using exponential decay, i.e., the timeliness score is e. (-24 / 12) The initial confidence score is approximately 0.135, and the confidence score is based on an initial weight of 0.85. Therefore, after weighted calculation, the initial confidence score can be calculated as 0.5×0.8+0.2×0.135+0.3×0.85= 0.4+0.027+0.255 = 0.682.

[0189] Regarding the new knowledge, assuming the actual feature data comes from 275 real-world clinical cases collected this month, with an authority score of 0.7 (sample size between 200 and 300), a timeliness score of 1.0, and a confidence score with a statistical weight of 0.92, then after weighted calculation, the actual confidence score can be 0.5×0.7+0.2×1.0+0.3×0.92=0.35+0.2+0.276=0.826.

[0190] Then, by comparing the scores of both sides, since 0.826 > 0.682, the credibility result of the actual feature data is higher.

[0191] After obtaining the credibility result, granular update processing can be performed based on the credibility result, actual feature data, and initial knowledge graph to generate the target knowledge graph.

[0192] Specifically, for conflicting entries identified as data discrepancies, the decision of which side's knowledge to adopt can be made based on the credibility results. If the credibility results indicate that the old knowledge in the initial knowledge graph is more credible, the old knowledge is retained, and the new knowledge is marked as low priority or stored in the pending review queue. If the credibility results indicate that the new knowledge in the actual feature data is more credible, the update method in S140 can be used for granular update processing, which will not be elaborated here.

[0193] If the actual credibility score is similar to the initial credibility score (e.g., the difference is less than the preset threshold), the conflicting entry will be pushed to the manual review queue for expert confirmation.

[0194] Finally, after all conflicting items have been processed using this method, the target knowledge graph is generated.

[0195] By introducing a credibility calculation mechanism for data discrepancies, a scientific basis for integration can be provided when new and old knowledge contradict each other, avoiding knowledge bias caused by simply accepting one side. At the same time, by comprehensively considering multiple dimensions of factors such as the authority, timeliness, and confidence of the source, the updated knowledge weights are made more objective and reliable, ensuring the overall quality of the target knowledge graph.

[0196] In the above implementation, by adding a conflict detection step before granularity updates, it is possible to effectively distinguish between two different types of conflicts: granularity differences and data differences, thus enabling a differentiated update strategy. Granularity differences are directly supplemented with finer-grained data to enrich the knowledge hierarchy; data differences are first assessed for reliability and then weighted and fused to ensure knowledge accuracy. This conflict detection and triage mechanism improves the refinement and reliability of knowledge graph updates, providing more rigorous knowledge support for clinical decision support.

[0197] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple steps or stages, which are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0198] This specification also provides a knowledge graph creation apparatus 400, such as... Figure 4 As shown, it includes: a raw data acquisition module 410, an initial map construction module 420, a real data acquisition module 430, and a granular update execution module 440, wherein: The raw data acquisition module 410 is used to acquire literature mining data and clinical diagnosis and treatment data for the target disease.

[0199] The initial knowledge graph construction module 420 is used to perform entity recognition and relation extraction based on literature mining data to establish an initial knowledge graph. The initial knowledge graph includes knowledge triples that can represent the semantic association between medical entities, and each knowledge triple contains entity nodes and relation edges.

[0200] The real data acquisition module 430 is used to perform statistical analysis based on clinical diagnosis and treatment data to obtain actual characteristic data.

[0201] The granularity update execution module 440 is used to perform granularity update processing on the initial knowledge graph using actual feature data to generate the target knowledge graph; wherein, the granularity update processing includes node update for entity nodes and weight update for relation edges.

[0202] In some implementations, the actual feature data includes symptom association data, examination distribution data, phenotypic probability data, and severity distribution data. The real data acquisition module 430 is also used to perform association rule mining based on symptom data and sign data in clinical diagnosis and treatment data to obtain symptom association data; to perform statistical analysis based on the test data of confirmed patients in clinical diagnosis and treatment data to obtain examination distribution data; to perform probability calculation based on variation data and phenotypic data in clinical diagnosis and treatment data to obtain phenotypic probability data; and to perform property clustering based on symptom data and severity data in clinical diagnosis and treatment data to obtain severity distribution data.

[0203] In some implementations, the granularity update execution module 440 is further configured to, for any knowledge triple, determine the corresponding target feature data in the actual feature data based on the entity nodes and relation edges of the knowledge triple; perform a ratio calculation based on the relation edges and the target feature data to determine the target balance coefficient; wherein, the relation edges have initial weights; the target feature data have statistical weights; perform an exponentially weighted moving average calculation based on the target balance coefficient, the initial weights, and the statistical weights to obtain the target edge weight of the relation edge; and replace the initial weight of the relation edge based on the target edge weight to complete the weight update of the relation edge.

[0204] In some implementations, the relation edges also have historical sample sizes; the target feature data also has real sample sizes; the granularity update execution module 440 is further used to perform a ratio calculation based on the historical sample sizes and real sample sizes to determine the target balance coefficient; wherein, the target balance coefficient is the ratio of the historical sample sizes to the total sample sizes; the total sample sizes are obtained by adding the real sample sizes and the historical sample sizes.

[0205] In some implementations, the initial knowledge graph construction module 420 is also used to perform text preprocessing on the literature mining data to obtain standard text data; perform entity recognition and combination based on the standard text data to obtain candidate entity pairs; perform relation extraction based on the candidate entity pairs to obtain the target relation type of the candidate entity pairs; generate relation edges of the candidate entity pairs based on the target relation type to construct knowledge triples, and integrate all knowledge triples to form an initial knowledge graph.

[0206] In some implementations, the initial map construction module 420 is also used to perform entity recognition based on standard text data to obtain entity nodes; wherein, the entity nodes have entity locations; and the entity nodes are combined using the entity locations to obtain candidate entity pairs.

[0207] In some implementations, the knowledge graph building apparatus 400 further includes a verification module for performing conflict detection based on actual feature data and an initial knowledge graph to obtain conflict identification results; if the conflict identification results indicate that the conflict originates from granularity differences, then granularity update processing is performed based on actual feature data and an initial knowledge graph to generate a target knowledge graph.

[0208] In some implementations, the verification module is further configured to, if the conflict identification result indicates that the conflict originates from data differences, calculate the credibility based on the actual feature data and the initial knowledge graph to obtain the credibility result; and perform granular update processing based on the credibility result, the actual feature data, and the initial knowledge graph to generate the target knowledge graph.

[0209] For specific limitations regarding the knowledge graph creation apparatus, please refer to the limitations of the knowledge graph creation method described above, which will not be repeated here. Each module in the aforementioned knowledge graph creation apparatus can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independently of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the operations corresponding to each module.

[0210] In this embodiment, the knowledge graph building device is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0211] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application, such as... Figure 5As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processor can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on an external input / output device (such as a display device coupled to the interface). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations. Figure 5 Take a processor 10 as an example.

[0212] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0213] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0214] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0215] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0216] The computer device also includes an input device 30 and an output device 40. The processor 10, memory 20, input device 30, and output device 40 can be connected via a bus or other means. Figure 5 Taking the example of a connection between China and Israel via a bus.

[0217] Input device 30 can receive input numerical or character information, and generate key signal inputs related to user settings and function control of the computer device, such as a touchscreen, keypad, mouse, trackpad, touchpad, joystick, one or more mouse buttons, trackball, joystick, etc. Output device 40 may include display devices, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors). The aforementioned display devices include, but are not limited to, liquid crystal displays, light-emitting diodes, displays, and plasma displays. In some alternative embodiments, the display device may be a touchscreen.

[0218] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant regulations, and user authorization should be obtained.

[0219] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose whether to provide personal information to the software or hardware, such as the electronic device, application, server, or storage medium performing the operations of this disclosed technical solution, based on the prompt message.

[0220] As an optional but non-limiting implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0221] It is understood that the above notification and user authorization process is merely illustrative and does not constitute a limitation on the implementation of this disclosure. Other methods that meet the relevant regulations may also be applied to the implementation of this disclosure.

[0222] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the relevant regulations and related requirements.

[0223] It is understood that in the specific implementation of this application, data such as user information, location information, and navigation data are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with relevant regulations and standards.

[0224] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0225] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0226] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0227] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0228] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0229] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0230] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for establishing a knowledge graph, characterized in that, The method includes: Acquire literature mining data and clinical diagnosis and treatment data for the target disease; Entity recognition and relation extraction are performed based on the literature mining data to establish an initial knowledge graph; wherein, the initial knowledge graph includes knowledge triples that can represent the semantic association between medical entities, and each knowledge triple contains an entity node and a relation edge; Statistical analysis was performed on the aforementioned clinical diagnosis and treatment data to obtain actual characteristic data; The initial knowledge graph is processed using the actual feature data to generate a target knowledge graph; wherein the granular update process includes node updates for the entity nodes and weight updates for the relation edges.

2. The method according to claim 1, characterized in that, The actual characteristic data includes symptom association data, examination distribution data, phenotypic probability data, and severity distribution data; the statistical analysis based on the clinical diagnosis and treatment data to obtain the actual characteristic data includes: Association rule mining is performed on the symptom and sign data in the clinical diagnosis and treatment data to obtain the symptom association data; Statistical analysis was performed on the test data of confirmed patients in the clinical diagnosis and treatment data to obtain the test distribution data; Probability calculations are performed based on the variation data and phenotypic data in the clinical diagnosis and treatment data to obtain the phenotypic probability data; Based on the symptom and severity data in the clinical diagnosis and treatment data, property clustering is performed to obtain the severity distribution data.

3. The method according to claim 1, characterized in that, The weight update of the relation edge is performed in the following manner: For any knowledge triple, the corresponding target feature data is determined in the actual feature data based on the entity node and the relation edge of the knowledge triple; A ratio calculation is performed based on the relationship edges and the target feature data to determine the target balance coefficient; wherein, the relationship edges have initial weights; and the target feature data have statistical weights. An exponentially weighted moving average is calculated based on the target balance coefficient, the initial weight, and the statistical weight to obtain the target edge weight of the relation edge; The initial weights of the relation edges are replaced based on the target edge weights to complete the weight update of the relation edges.

4. The method according to claim 3, characterized in that, The relation edges also have historical sample sizes; the target feature data also has real sample sizes; the step of calculating the ratio based on the relation edges and the target feature data to determine the target balance coefficient includes: The target balance coefficient is determined by calculating the ratio between the historical sample size and the actual sample size; wherein the target balance coefficient is the ratio of the historical sample size to the total sample size; the total sample size is obtained by adding the actual sample size and the historical sample size.

5. The method according to claim 1, characterized in that, The step of performing entity recognition and relation extraction based on the document mining data to establish an initial knowledge graph includes: The document mining data is preprocessed to obtain standard text data; Entity recognition and combination are performed based on the standard text data to obtain candidate entity pairs; Relationship extraction is performed based on the candidate entity pairs to obtain the target relation type of the candidate entity pairs; Relation edges of the candidate entity pairs are generated based on the target relation type to construct the knowledge triples, and all the knowledge triples are integrated to form an initial knowledge graph.

6. The method according to claim 5, characterized in that, The step of performing entity recognition and combination based on the standard text data to obtain candidate entity pairs includes: Entity recognition is performed based on the standard text data to obtain entity nodes; wherein, the entity nodes have entity locations; The entity nodes are combined using the entity locations to obtain candidate entity pairs.

7. The method according to claim 1, characterized in that, Before performing granular update processing on the initial knowledge graph using the actual feature data, the method further includes: Conflict detection is performed based on the actual feature data and the initial knowledge graph to obtain conflict identification results; If the conflict identification result indicates that the conflict originates from granularity differences, then a granularity update process is performed based on the actual feature data and the initial knowledge graph to generate the target knowledge graph.

8. The method according to claim 7, characterized in that, The method further includes: If the conflict identification result indicates that the conflict originates from data differences, then a credibility calculation is performed based on the actual feature data and the initial knowledge graph to obtain a credibility result. Based on the credibility result, the actual feature data, and the initial knowledge graph, perform granular update processing to generate the target knowledge graph.

9. A knowledge graph construction device, characterized in that, The device includes: The raw data acquisition module is used to acquire literature mining data and clinical diagnosis and treatment data for the target disease; An initial knowledge graph construction module is used to perform entity recognition and relation extraction based on the literature mining data to establish an initial knowledge graph; wherein, the initial knowledge graph includes knowledge triples that can represent the semantic association between medical entities, and each knowledge triple contains an entity node and a relation edge; The real data acquisition module is used to perform statistical analysis based on the clinical diagnosis and treatment data to obtain actual characteristic data; The granularity update execution module is used to perform granularity update processing on the initial knowledge graph using the actual feature data to generate a target knowledge graph; wherein, the granularity update processing includes node updates for the entity nodes and weight updates for the relation edges.

10. A computer device, characterized in that, include: A memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, the processor executing the computer instructions to perform the method of any one of claims 1 to 8.

Images