Medical trauma data authenticity monitoring method and system based on knowledge graph

Abstract

The application provides a medical trauma data authenticity monitoring method and system based on a knowledge graph, and relates to the technical field of medical data processing, comprising the application. The application can effectively identify strong dependence fields by constructing a medical knowledge graph and extracting field combination rules, construct a field dependence graph, and filter effective dependence chains, so as to locate a linked abnormal field cluster. For the abnormal field cluster, the associated path integrity score is calculated by searching for reference cases and weighted matching, and the optimal correction suggestion is output to monitor the authenticity of medical data, accurately locate and correct abnormal data, and improve the data quality and reliability.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method and system for monitoring the authenticity of medical trauma data based on knowledge graphs. Background Technology

[0002] In the field of healthcare informatics, ensuring the authenticity and consistency of medical trauma data, such as electronic medical records, is crucial for improving medical quality and supporting clinical decision-making. Current technologies primarily rely on rule-based systems or statistical anomaly detection models to monitor the authenticity of medical data. Conventional practices typically involve verifying data fields one by one using predefined business rules or logical constraints, such as checking numerical ranges, format compliance, or simple logical inconsistencies. Another common approach is to utilize the statistical characteristics of historical data, such as the distribution patterns of field values, and identify anomalous data points that deviate from normal patterns by setting thresholds. These methods can, to some extent, detect obvious data entry errors or non-compliant data.

[0003] However, the aforementioned conventional monitoring methods have significant drawbacks. Methods based on predefined rules lack flexibility and adaptability. Medical knowledge is complex and constantly evolving, and fixed rules struggle to cover all potential, especially deep-seated logical contradictions and semantic inconsistencies across fields. Rule bases are costly to maintain and cannot effectively handle novel, undefined data anomaly patterns. Traditional statistical anomaly detection models often view individual fields in isolation or perform superficial correlation analysis, ignoring the complex, multi-hop semantic relationship networks between entities in medical data. This makes it difficult for models to detect chain anomalies formed by multiple fields interconnected through hidden paths, and also prevents them from providing semantically reasonable data correction suggestions based on rich domain knowledge networks after anomaly detection, thus limiting the intelligence and practicality of the monitoring system. Summary of the Invention

[0004] This invention provides a knowledge graph-based method and system for monitoring the authenticity of medical trauma data, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for monitoring the authenticity of medical trauma data based on knowledge graphs, comprising:

[0006] Acquire patient medical record data, extract medical entities and identify the relationships between entities, and construct a medical knowledge graph;

[0007] The values ​​of each field in the patient's medical record data are used as starting nodes to perform multi-hop diffusion in the medical knowledge graph. Path identifiers are generated for the diffusion path. Based on the path identifiers, weight coefficients are assigned to entity nodes and cumulative association weights are calculated to form field association domains. Field combination patterns are extracted.

[0008] The mutual information strength of each field pair is calculated based on the field combination rules. After semantic modulation, strong dependent field pairs are identified. A field dependency graph is constructed based on the strong dependent field pairs, and effective dependency chains are screened based on the semantic diversity of the path. When a field anomaly is detected, the anomaly propagation probability is calculated along the effective dependency chain, and the chained anomaly fields are located to form an anomaly field cluster.

[0009] Extract the field association domain structure of the abnormal field clusters, retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract the candidate value set and combine and arrange it, calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strong dependent field pairs, and select the combination with the highest association path integrity score as the correction suggestion output.

[0010] In one optional embodiment, the field values ​​in the patient's medical record data are used as starting nodes for multi-hop diffusion in the medical knowledge graph. Path identifiers are generated for the diffusion paths. Based on the path identifiers, weight coefficients are assigned to entity nodes, and cumulative association weights are calculated to form field association domains. The field combination patterns are extracted, including:

[0011] The values ​​of each field in the patient's medical record data are used as starting nodes and mapped to the medical knowledge graph. Multi-hop diffusion is carried out from the starting node along the directed edges. During the diffusion process, a path identifier is generated for each diffusion path, including the starting node, the sequence of edge types traversed, and the sequence of entity nodes reached. When the number of diffusion hops reaches the preset upper limit, the diffusion is terminated and the path identifier is saved to obtain the diffusion path identifier set of the starting node.

[0012] Extract the entity node sequence from the diffusion path identifier, mark the first entity node in the sequence as a first-order associated entity and assign it a first-order weight coefficient, mark the subsequent entity nodes as multi-order associated entities and assign them decreasing multi-order weight coefficients according to their positions in the sequence, count the number of times the entity node appears in the diffusion path identifier set, multiply the number of occurrences by the corresponding weight coefficient to obtain the cumulative association weight of the entity node, and filter the entity nodes whose cumulative association weight exceeds the preset weight threshold and aggregate them with the starting node to form a field association domain.

[0013] Traverse historical medical record data, extract field value combinations and corresponding field association domains from the medical records, identify the intersection entity nodes between different field association domains, count the frequency of field value combinations containing the same intersection entity nodes in historical medical record data, and record field value combinations and intersection entity nodes whose frequency exceeds a preset frequency threshold as field combination patterns.

[0014] In an optional embodiment, marking subsequent entity nodes as multi-level associated entities and assigning decreasing multi-level weight coefficients based on their position in the sequence includes:

[0015] Extract the last edge type traversed to reach subsequent entity nodes from the diffusion path identifier, query the preset edge type importance scoring table in the medical knowledge graph, and obtain the corresponding importance score;

[0016] The initial multi-order weight coefficients are calculated based on the position of the subsequent entity nodes in the entity node sequence, and the modulated multi-order weight coefficients are calculated by combining the initial multi-order weight coefficients with the edge type importance score.

[0017] In one optional embodiment, identifying the intersection entity nodes between different field association domains and counting the frequency of occurrence of field value combinations containing the same intersection entity nodes in historical medical record data includes:

[0018] The intersection operation identifies entity nodes that simultaneously exist in at least two related fields as intersection entity nodes.

[0019] For each intersection entity node, query the cumulative association weight of the intersection entity node in each field association domain, calculate the arithmetic mean of all cumulative association weights, and determine the intersection association strength of the intersection entity nodes.

[0020] Filter the intersection entity nodes whose intersection association strength exceeds the preset association threshold, and mark the filtered intersection entity nodes as core intersection entity nodes;

[0021] Traverse historical medical record data, extract field value combinations and corresponding field association fields, determine whether the field association fields contain core intersection entity nodes, and only retain field value combinations whose field association fields contain core intersection entity nodes;

[0022] The frequency of occurrence of retained field value combinations in historical medical record data.

[0023] In one optional embodiment, the mutual information strength of each field pair is calculated based on the field combination rules. After semantic modulation, strongly dependent field pairs are identified. A field dependency graph is constructed using the strongly dependent field pairs, and valid dependency chains are filtered based on path semantic diversity. When a field anomaly is detected, the anomaly propagation probability is calculated along the valid dependency chain, and the linked anomaly fields that form an anomaly field cluster are located, including:

[0024] Extract the joint occurrence count and marginal occurrence count of each field pair from the field combination rules, and calculate the basic mutual information strength of each field pair;

[0025] Extract the intersection entity nodes associated with each field, query the intersection association strength of the intersection entity nodes as a semantic modulation factor, and perform weighted modulation on the basic mutual information strength to obtain the modulated mutual information strength.

[0026] Field pairs whose modulation mutual information strength exceeds a preset strength threshold are selected as strong dependency field pairs;

[0027] A weighted directed field dependency graph is constructed using strong dependency fields, a transitive closure operation is performed to identify reachable paths, the path dependency strength is adjusted based on the path semantic diversity index, and effective dependency chains are selected.

[0028] When an anomaly is detected in any field, the valid dependency chain starting from the anomaly field is retrieved, the field nodes downstream of the anomaly field in the valid dependency chain are extracted, the anomaly propagation probability is calculated based on the path dependency strength, and the field nodes whose anomaly propagation probability exceeds the preset propagation threshold are selected as chained anomaly fields.

[0029] Exception fields and chained exception fields are grouped into exception field clusters.

[0030] In one optional embodiment, a weighted directed field dependency graph is constructed using strong dependency fields, a transitive closure operation is performed to identify reachable paths, the path dependency strength is adjusted based on the path semantic diversity index, and valid dependency chains are screened, including:

[0031] A weighted directed field dependency graph is constructed using strongly dependent field pairs as directed edges, fields as nodes, and modulation mutual information strength as edge weights.

[0032] Perform transitive closure operation on a weighted directed field dependency graph to identify all reachable paths in the graph;

[0033] For each reachable path, extract the strong dependency field pairs corresponding to each directed edge in the path and their intersection entity nodes, and count the number of different intersection entity nodes in the path as a path semantic diversity index.

[0034] When the path semantic diversity index is less than or equal to the preset lower limit of diversity, a decay factor of less than 1 is applied to the cumulative value of path edge weights. When the path semantic diversity index is greater than the preset upper limit of diversity, an enhancement factor of greater than 1 is applied to the cumulative value of path edge weights, thus obtaining the path dependency strength adjusted by semantic diversity.

[0035] Paths whose path dependency strength exceeds a preset path strength threshold are selected as valid dependency chains.

[0036] In one optional embodiment, the field association domain structure of the abnormal field cluster is extracted, reference medical record cases with the same field association domain structure are retrieved from the medical knowledge graph, a candidate value set is extracted and combined, and a weighted matching is performed to calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strongly dependent field pairs. The combination with the highest association path integrity score is selected as the correction suggestion output, including:

[0037] Extract the field association fields corresponding to each field in the abnormal field cluster to determine the field association field structure of the abnormal field cluster;

[0038] Retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract the field values ​​corresponding to each field in the abnormal field cluster from the reference medical case cases, and form a set of candidate values ​​for each field.

[0039] The candidate value sets of each field are combined and arranged by Cartesian product to generate candidate field value combinations;

[0040] For each candidate field value combination, extract the entity node corresponding to each field value in the medical knowledge graph, query the cumulative association weight of each entity node in the corresponding field association domain, and calculate the cumulative weight of the association path between entity nodes corresponding to different field values.

[0041] The cumulative weight of the associated path and the modulation mutual information strength of the corresponding strong dependency field pair in the abnormal field cluster are weighted and matched to calculate the associated path integrity score of the candidate field value combination.

[0042] The candidate field value combination with the highest association path integrity score is selected as the correction suggestion output.

[0043] A second aspect of this invention provides a knowledge graph-based system for monitoring the authenticity of medical trauma data, comprising:

[0044] The knowledge graph unit is used to acquire patient medical record data, extract medical entities and identify the relationships between entities, and construct a medical knowledge graph.

[0045] The association domain extraction unit is used to take the field values ​​in the patient's medical record data as the starting node for multi-hop diffusion in the medical knowledge graph, generate path identifiers for the diffusion path, assign weight coefficients to entity nodes based on the path identifiers and calculate the cumulative association weights to form field association domains, and extract field combination patterns.

[0046] An anomaly propagation unit is used to calculate the mutual information strength of each field pair according to the field combination rules, identify strongly dependent field pairs after semantic modulation, construct a field dependency graph with strongly dependent field pairs, and filter effective dependency chains based on path semantic diversity. When a field anomaly is detected, the anomaly propagation probability is calculated along the effective dependency chain, and the chained anomaly fields are located to form an anomaly field cluster.

[0047] The correction suggestion unit is used to extract the field association domain structure of abnormal field clusters, retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract candidate value sets and arrange them in combination, calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strong dependent field pairs, and select the combination with the highest association path integrity score as the correction suggestion output.

[0048] A third aspect of the present invention provides an electronic device, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] In this embodiment of the invention, by constructing a medical knowledge graph and utilizing multi-hop diffusion technology, the implicit field combination patterns and deep semantic relationships in medical record data are extracted. This effectively overcomes the shortcomings of traditional rule-based verification methods, such as narrow coverage and difficulty in discovering complex logical contradictions, providing a comprehensive and interpretable structured knowledge foundation for data authenticity assessment. By screening effective dependency chains with high semantic diversity, this method establishes a robust anomaly propagation model. When a single field exhibits an anomaly, the method automatically calculates and locates the chained anomaly fields affected by it along the dependency chain, forming anomaly field clusters. This significantly improves the ability to discover and locate complex data anomalies, especially correlational anomalies caused by hidden logical conflicts. By using cumulative association weights and mutual information strength for weighted matching, the method calculates the association path integrity score of different candidate correction value combinations, ensuring that the correction suggestions are not only reasonable at the numerical level but also maintain a high degree of consistency and integrity at the level of medical logic and knowledge association. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the knowledge graph-based method for monitoring the authenticity of medical trauma data according to an embodiment of the present invention.

[0055] Figure 2 A schematic diagram of the logic flow for processing field dependency graphs. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0058] Figure 1 This is a flowchart illustrating the knowledge graph-based method for monitoring the authenticity of medical trauma data according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0059] Acquire patient medical record data, extract medical entities and identify the relationships between entities, and construct a medical knowledge graph;

[0060] The values ​​of each field in the patient's medical record data are used as starting nodes to perform multi-hop diffusion in the medical knowledge graph. Path identifiers are generated for the diffusion path. Based on the path identifiers, weight coefficients are assigned to entity nodes and cumulative association weights are calculated to form field association domains. Field combination patterns are extracted.

[0061] The mutual information strength of each field pair is calculated based on the field combination rules. After semantic modulation, strong dependent field pairs are identified. A field dependency graph is constructed based on the strong dependent field pairs, and effective dependency chains are screened based on the semantic diversity of the path. When a field anomaly is detected, the anomaly propagation probability is calculated along the effective dependency chain, and the chained anomaly fields are located to form an anomaly field cluster.

[0062] Extract the field association domain structure of the abnormal field clusters, retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract the candidate value set and combine and arrange it, calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strong dependent field pairs, and select the combination with the highest association path integrity score as the correction suggestion output.

[0063] In one optional embodiment, the field values ​​in the patient's medical record data are used as starting nodes for multi-hop diffusion in the medical knowledge graph. Path identifiers are generated for the diffusion paths. Based on the path identifiers, weight coefficients are assigned to entity nodes, and cumulative association weights are calculated to form field association domains. The field combination patterns are extracted, including:

[0064] The values ​​of each field in the patient's medical record data are used as starting nodes and mapped to the medical knowledge graph. Multi-hop diffusion is carried out from the starting node along the directed edges. During the diffusion process, a path identifier is generated for each diffusion path, including the starting node, the sequence of edge types traversed, and the sequence of entity nodes reached. When the number of diffusion hops reaches the preset upper limit, the diffusion is terminated and the path identifier is saved to obtain the diffusion path identifier set of the starting node.

[0065] Extract the entity node sequence from the diffusion path identifier, mark the first entity node in the sequence as a first-order associated entity and assign it a first-order weight coefficient, mark the subsequent entity nodes as multi-order associated entities and assign them decreasing multi-order weight coefficients according to their positions in the sequence, count the number of times the entity node appears in the diffusion path identifier set, multiply the number of occurrences by the corresponding weight coefficient to obtain the cumulative association weight of the entity node, and filter the entity nodes whose cumulative association weight exceeds the preset weight threshold and aggregate them with the starting node to form a field association domain.

[0066] Traverse historical medical record data, extract field value combinations and corresponding field association domains from the medical records, identify the intersection entity nodes between different field association domains, count the frequency of field value combinations containing the same intersection entity nodes in historical medical record data, and record field value combinations and intersection entity nodes whose frequency exceeds a preset frequency threshold as field combination patterns.

[0067] In one specific implementation, medical records are read from a medical trauma patient medical record database, and structured fields and their corresponding field values ​​are identified. Fields include medical entities such as trauma type, injury location, severity score, clinical manifestations, and treatment measures. Each field value is used as an independent starting node, and its corresponding entity node is located in a constructed medical knowledge graph. For the trauma type field value "open fracture," its corresponding disease entity node is found in the knowledge graph as the starting point for diffusion.

[0068] Starting from the initial node, the diffusion traversal proceeds along the directed edges in the knowledge graph. A preset maximum of 3 hops is set. In the first hop, the "open fracture" node is connected to symptom entity nodes such as "local swelling" and "severe pain" via the "accompanied by symptoms" edge, and to treatment entity nodes such as "debridement" and "internal fixation" via the "requires treatment" edge. A path identifier record is created for each diffusion path, in the format {initial node: open fracture, edge type sequence: [accompanied by symptoms], entity node sequence: [local swelling]}. The second hop continues, where the "local swelling" node may connect to pathological entity nodes such as "inflammatory response" and "tissue damage". The diffusion stops when 3 hops are reached, and all path identifiers are stored in the diffusion path identifier set of the initial node.

[0069] Parse the entity node sequence in the diffusion path identifier set. For the path {open fracture → accompanied by symptoms → local swelling → inflammatory response}, mark the first arriving entity node "local swelling" as a first-order associated entity and assign it a first-order weight coefficient of 0.8. Mark the second entity node "inflammatory response" as a second-order associated entity and assign it a second-order weight coefficient of 0.5. The weight coefficient decreases with the number of diffusion hops, and the third-order weight coefficient is set to 0.3. Traverse the diffusion path identifier set and count the occurrences of the "local swelling" entity node in 25 paths, of which 15 times are first-order associated entities and 10 times are second-order associated entities. Calculate the cumulative association weight as 15 multiplied by 0.8 plus 10 multiplied by 0.5, which equals 17. Set a preset weight threshold of 8, and filter entity nodes whose cumulative association weight exceeds the threshold, including "local swelling", "severe pain", "debridement", etc. Aggregate these entity nodes with the starting node "open fracture" to form the field association domain of this field value.

[0070] Load a historical trauma medical record dataset containing 5000 records. Extract field value combinations from each record, such as {Trauma Type: Open Fracture, Injury Location: Midshaft Tibia, Treatment: Debridement}. Simultaneously extract the corresponding field association domains for each field value. Compare different field association domains to identify intersecting entity nodes. The field association domain for "Open Fracture" contains entities such as "Debridement" and "Antibiotic Use," while the field association domain for "Midshaft Tibia" contains entities such as "Internal Fixation" and "Antibiotic Use." The intersecting entity node is identified as "Antibiotic Use." Count the frequency of field value combinations containing this intersecting entity node in historical medical records. The combination {Open Fracture, Midshaft Tibia, Antibiotic Use} was found to appear 420 times in historical data. Set a preset frequency threshold of 50 times. Record field value combinations that exceed the threshold and their intersecting entity node "Antibiotic Use" as field combination patterns, storing them in a pattern knowledge base for subsequent anomaly monitoring.

[0071] In an optional embodiment, marking subsequent entity nodes as multi-level associated entities and assigning decreasing multi-level weight coefficients based on their position in the sequence includes:

[0072] Extract the last edge type traversed to reach subsequent entity nodes from the diffusion path identifier, query the preset edge type importance scoring table in the medical knowledge graph, and obtain the corresponding importance score;

[0073] The initial multi-order weight coefficients are calculated based on the position of the subsequent entity nodes in the entity node sequence, and the modulated multi-order weight coefficients are calculated by combining the initial multi-order weight coefficients with the edge type importance score.

[0074] In one specific implementation, when assigning weight coefficients to subsequent entity nodes, it is necessary to comprehensively consider the node's position information in the diffusion path and the semantic importance of the associated edges traversed to reach that node. The diffusion path identifier is stored in a serialized encoding format of "starting entity-relationship type-intermediate entity-relationship type-target entity," which facilitates tracing the complete inference chain. When extracting the last hop edge type, the last relation type marker is located by parsing the path identifier string. This marker reflects the direct semantic connection from the predecessor node to the current node.

[0075] A pre-built edge type importance scoring table is established in the medical knowledge graph, based on medical diagnostic logic. Direct diagnostic relationships, such as "symptom-diagnosis-disease," are assigned high scores of 0.9 to 1.0; treatment relationships, such as "disease-treatment use-drug," are assigned medium-high scores of 0.7 to 0.85; indirect inference relationships, such as "drug-interaction-drug," are assigned medium scores of 0.5 to 0.7; and statistical relationships, such as "age-prevalence-disease," are assigned low scores of 0.3 to 0.5. During a query, the extracted edge type is used as the key to retrieve the corresponding importance score from the scoring table. .

[0076] When calculating the initial multi-order weight coefficients based on positional decay, the starting node is set as order zero, the node reached in the first hop is set as order one, and so on. For subsequent entity nodes at order n, the initial multi-order weight coefficients are calculated using an exponential decay function, expressed as: α is the decay factor, ranging from 0.6 to 0.8, and is typically set to 0.7. This design ensures that as the diffusion hop count increases, the correlation strength of entities far from the starting node gradually weakens, which aligns with the proximate cause priority principle in medical reasoning.

[0077] The modulated multi-order weight coefficients are calculated by fusing location information and edge type semantics. The specific calculation method is as follows: In this formula, 0.4 serves as a safety net weight to ensure all nodes maintain basic relevance, while 0.6 times the importance score acts as a modulation gain. When the edge type importance score reaches its maximum value of 1.0, the modulation coefficient reaches 1.0 for complete retention; when the score is 0.5, the modulation coefficient is 0.7 for moderate attenuation; when the score is below 0.3, the corresponding association path is considered semantically insufficient, and its weight coefficient is significantly compressed. This two-factor weighting mechanism reflects both the influence of diffusion distance and incorporates professional judgment in medical semantics, ensuring that distant nodes reached through strong semantic associations maintain high weights, while nearby nodes reached through weak associations have limited weights. This avoids introducing noisy data to interfere with subsequent field association domain construction and anomaly detection judgment processes.

[0078] In one optional embodiment, identifying the intersection entity nodes between different field association domains and counting the frequency of occurrence of field value combinations containing the same intersection entity nodes in historical medical record data includes:

[0079] The intersection operation identifies entity nodes that simultaneously exist in at least two related fields as intersection entity nodes.

[0080] For each intersection entity node, query the cumulative association weight of the intersection entity node in each field association domain, calculate the arithmetic mean of all cumulative association weights, and determine the intersection association strength of the intersection entity nodes.

[0081] Filter the intersection entity nodes whose intersection association strength exceeds the preset association threshold, and mark the filtered intersection entity nodes as core intersection entity nodes;

[0082] Traverse historical medical record data, extract field value combinations and corresponding field association fields, determine whether the field association fields contain core intersection entity nodes, and only retain field value combinations whose field association fields contain core intersection entity nodes;

[0083] The frequency of occurrence of retained field value combinations in historical medical record data.

[0084] In one specific implementation, when extracting patterns from field combinations, it is necessary to identify the intersection entity nodes between different field association domains. Specifically, for multiple constructed field association domains, a set intersection operation is used to compare the entity nodes contained in each field association domain one by one. If an entity node appears in at least two field association domains simultaneously, then that entity node is marked as an intersection entity node. For example, the field association domains of the patient admission time field include the entity nodes "emergency department," "trauma history," and "coagulation function," while the field association domains of the injury type field include the entity nodes "trauma history," "CT images," and "surgical record." Therefore, the set intersection operation can identify "trauma history" as an intersection entity node.

[0085] After identifying the intersection entity nodes, it is necessary to evaluate the association strength of these nodes across different field association domains. For each intersection entity node, the cumulative association weight value corresponding to that node in each field association domain containing it is queried. The cumulative association weight is obtained by accumulating the path weight coefficients during the multi-hop diffusion process, reflecting the tightness of the association between the node and the starting field value. The arithmetic mean of all the retrieved cumulative association weight values ​​is then calculated, and the average value is the intersection association strength of the intersection entity node. The higher the intersection association strength value, the higher the association status of the node in multiple field association domains.

[0086] To filter out truly meaningful intersectional entity nodes, a preset association threshold is set to filter the strength of intersectional associations. The association strength of each intersectional entity node is compared numerically with the preset threshold, and only nodes with an association strength exceeding the threshold are retained. These filtered nodes are marked as core intersectional entity nodes, representing key bridges connecting different fields. In practical applications, the preset association threshold can be set according to the characteristics of medical trauma data, typically ranging from 0.4 to 0.7.

[0087] Historical medical record data is analyzed systematically, with each record read from the database and field value combinations extracted. Each field value combination contains multiple fields and their corresponding values; for example, "admission time-injury type-surgical method" constitutes a field value combination. Simultaneously, the field association domain structure corresponding to this value combination is extracted, and it is determined whether these domains contain the previously identified core intersection entity nodes. Specifically, the set of core intersection entity nodes is matched with the set of entity nodes in the field association domains. If the field association domain contains at least one core intersection entity node, the field value combination is retained; otherwise, it is removed from the candidate set.

[0088] After the above filtering operations, a simplified set of field value combinations is obtained. These combinations all establish substantial relationships between fields through core intersection entity nodes. For each retained field value combination, its frequency of occurrence is counted in historical medical record data. Specifically, all historical medical record records are scanned, and when a field value in a record is exactly the same as the field value combination to be counted, the counter for that combination is incremented by one. Finally, the frequency value of each field value combination is obtained. This frequency reflects the occurrence pattern of specific field combinations in real medical scenarios, providing a statistical basis for subsequent mutual information strength calculations.

[0089] In one optional embodiment, the mutual information strength of each field pair is calculated based on the field combination rules. After semantic modulation, strongly dependent field pairs are identified. A field dependency graph is constructed using the strongly dependent field pairs, and valid dependency chains are filtered based on path semantic diversity. When a field anomaly is detected, the anomaly propagation probability is calculated along the valid dependency chain, and the linked anomaly fields that form an anomaly field cluster are located, including:

[0090] Extract the joint occurrence count and marginal occurrence count of each field pair from the field combination rules, and calculate the basic mutual information strength of each field pair;

[0091] Extract the intersection entity nodes associated with each field, query the intersection association strength of the intersection entity nodes as a semantic modulation factor, and perform weighted modulation on the basic mutual information strength to obtain the modulated mutual information strength.

[0092] Field pairs whose modulation mutual information strength exceeds a preset strength threshold are selected as strong dependency field pairs;

[0093] A weighted directed field dependency graph is constructed using strong dependency fields, a transitive closure operation is performed to identify reachable paths, the path dependency strength is adjusted based on the path semantic diversity index, and effective dependency chains are selected.

[0094] When an anomaly is detected in any field, the valid dependency chain starting from the anomaly field is retrieved, the field nodes downstream of the anomaly field in the valid dependency chain are extracted, the anomaly propagation probability is calculated based on the path dependency strength, and the field nodes whose anomaly propagation probability exceeds the preset propagation threshold are selected as chained anomaly fields.

[0095] Exception fields and chained exception fields are grouped into exception field clusters.

[0096] In one specific implementation, after obtaining the pattern of field combinations, it is necessary to conduct in-depth correlation analysis on each field pair in the medical record data. For any two fields F... i and F j The number of times a condition appears simultaneously in all medical record samples is counted as the joint frequency N(F). i F j Simultaneously, record the edge frequency N(F) of each field appearing individually. i ) and N(F j Based on these statistics, the basic mutual information strength is calculated. This value reflects the degree of co-occurrence association between two fields at the data level. The higher the value, the more likely the two fields are to appear in the medical record at the same time.

[0097] To introduce medical semantic constraints, it is necessary to extract semantic evidence supporting the association between the fields from the constructed medical knowledge graph. Specifically, the query field F... i and F j The shared entity nodes are obtained by calculating the intersection of the mapped set of entity nodes. For each intersection entity node, the association strength between that node and two fields is queried in the knowledge graph, and the average of the association strengths of all intersection nodes is taken as the semantic modulation factor. The role of this factor is to combine statistically based mutual information strength with semantic relationships in a medical knowledge base to calculate modulated mutual information strength. Set an intensity threshold. Filter to meet The field pairs are treated as strong dependency field pairs.

[0098] Construct a field dependency graph G based on the selected strong dependency field pairs. DIn this graph, fields are used as nodes, strong dependencies are used as directed edges, and edge weights represent the modulation mutual information strength. Transitive closure is performed on the graph to identify all reachable paths. For each path, the geometric mean of the edge weights is calculated as the initial path strength, while a path semantic diversity metric is introduced for adjustment. Path semantic diversity is measured by the number of entity types traversed by the path; a higher number of types indicates a broader coverage of medical knowledge. The final path dependency strength is obtained by multiplying the path strength by the semantic diversity factor, and the paths with the highest dependency strengths are retained as valid dependency chains.

[0099] During the anomaly detection phase, when a certain field F is detected... a When a value violates a preset rule, immediately search the field dependency graph for F. a This represents all valid dependency chains of the starting node. For each dependency chain located at F... a Each field node F downstream d Extract from F a To F d Path dependence strength value P ad This intensity value, after being normalized, is used as the anomaly propagation probability, representing the likelihood of an anomaly propagating from the originating field to downstream fields. (Setting...) A propagation threshold is set, and downstream fields with a propagation probability exceeding this threshold are marked as cascading anomalous fields. Finally, the initial anomalous field F... a These fields are merged with all identified chained anomaly fields to form an anomaly field cluster. This field cluster fully characterizes the propagation range of the anomaly within the field dependency network, providing a precise set of target fields for subsequent corrective actions.

[0100] In one optional embodiment, a weighted directed field dependency graph is constructed using strong dependency fields, a transitive closure operation is performed to identify reachable paths, the path dependency strength is adjusted based on the path semantic diversity index, and valid dependency chains are screened, including:

[0101] A weighted directed field dependency graph is constructed using strongly dependent field pairs as directed edges, fields as nodes, and modulation mutual information strength as edge weights.

[0102] Perform transitive closure operation on a weighted directed field dependency graph to identify all reachable paths in the graph;

[0103] For each reachable path, extract the strong dependency field pairs corresponding to each directed edge in the path and their intersection entity nodes, and count the number of different intersection entity nodes in the path as a path semantic diversity index.

[0104] When the path semantic diversity index is less than or equal to the preset lower limit of diversity, a decay factor of less than 1 is applied to the cumulative value of path edge weights. When the path semantic diversity index is greater than the preset upper limit of diversity, an enhancement factor of greater than 1 is applied to the cumulative value of path edge weights, thus obtaining the path dependency strength adjusted by semantic diversity.

[0105] Paths whose path dependency strength exceeds a preset path strength threshold are selected as valid dependency chains.

[0106] In the knowledge graph-based monitoring of the authenticity of medical trauma data, constructing a field dependency graph and identifying valid dependency chains are crucial steps. By analyzing the dependencies between fields, the system can determine whether medical trauma data conforms to the characteristic distribution of real data, thereby uncovering potential data fraud.

[0107] In one specific implementation, a medical trauma dataset is acquired, containing multiple fields such as patient basic information, trauma type, trauma severity, vital signs, and treatment process. Based on a large number of real medical trauma cases, the modulation mutual information value between any two fields is calculated to measure the dependence strength between them. The modulation mutual information value is obtained by calculating the logarithm ratio of the joint distribution of the two fields to their respective marginal distributions. When the calculated modulation mutual information value exceeds a preset threshold (e.g., 0.45), the two fields are determined to constitute a strongly dependent field pair. For example, in emergency trauma data, the modulation mutual information value between "injury severity" and "systolic blood pressure" is 0.52, exceeding the threshold of 0.45, thus constituting a strongly dependent field pair.

[0108] A weighted directed field dependency graph is constructed based on the identified strongly dependent field pairs. In this graph, each field is treated as a node, and strongly dependent field pairs form directed edges, with the modulated mutual information value serving as the edge weight. For example, the directed edge weight from "injury severity" to "systolic blood pressure" is 0.52; the directed edge weight from "systolic blood pressure" to "pulse" is 0.61. The complete dependency graph contains all relevant field nodes and their dependencies in the medical trauma data.

[0109] Perform transitive closure operations on the constructed weighted directed field dependency graph to identify all reachable paths in the graph. The transitive closure operation is implemented using depth-first search or the Warshall algorithm to find all reachable paths between any two nodes in the graph. For example, reachable paths from "trauma type" to "recovery time" include multiple paths such as: "trauma type → injury severity → systolic blood pressure → recovery time" and "trauma type → treatment plan → recovery time".

[0110] For each identified reachable path, the strongly dependent field pairs corresponding to each directed edge in the path and their intersection entity nodes are extracted. Intersection entity nodes refer to entities in the medical knowledge graph that are simultaneously associated with two adjacent fields. The number of different intersection entity nodes in the path is counted as a path semantic diversity index. For example, the path "Trauma Type → Injury Severity → Systolic Blood Pressure → Recovery Time" contains three different intersection entity nodes: "Fracture," "Hemorrhagic Shock," and "Hypotension," resulting in a path semantic diversity index of 3.

[0111] The path dependency strength is adjusted based on the path semantic diversity index. A preset lower limit for diversity is set to 2, and an upper limit to 5. When the path semantic diversity index is less than or equal to the preset lower limit, a decay factor of 0.85 is applied to the cumulative value of the path edge weights. For example, the semantic diversity index for the path "blood pressure → pulse" is 1, which is less than the lower limit of 2; its original cumulative edge weight value of 0.61 needs to be multiplied by the decay factor of 0.85, resulting in an adjusted value of 0.52. When the path semantic diversity index is greater than the preset upper limit, an enhancement factor of 1.15 is applied to the cumulative value of the path edge weights. For example, the semantic diversity index for the path "trauma type → injury severity → systolic blood pressure → treatment plan → recovery time" is 7, which is greater than the upper limit of 5; its original cumulative edge weight value of 0.38 needs to be multiplied by the enhancement factor of 1.15, resulting in an adjusted value of 0.44.

[0112] The cumulative weight of each edge on a path is calculated using a multiplication method, which involves multiplying the weights of all edges along the path. To avoid excessively long paths leading to small cumulative weights, the logarithm of the multiplication result is taken and then normalized. After semantic diversity adjustment, reachable paths with a path dependency strength exceeding a preset path strength threshold (e.g., 0.40) are selected as valid dependency chains.

[0113] In practical application, a real medical trauma data case is presented: a patient suffered a car accident with multiple fractures, a severe injury, a systolic blood pressure of 85 mmHg, a pulse of 120 beats / min, an emergency surgical fixation treatment, and a recovery time of 60 days. The distribution characteristics of this data along the effective dependency chain are compared with the characteristic patterns of historical real data. Following the effective dependency chain "trauma type → injury severity → systolic blood pressure → pulse → treatment plan → recovery time" (dependency strength 0.46, exceeding the threshold of 0.40), the correlation between each node in the data is analyzed. The comparison reveals that, in cases of multiple fractures and severe injury, a historical distribution probability of systolic blood pressure between 80-90 mmHg, a pulse between 110-130 beats / min, an emergency surgical fixation treatment, and a recovery time between 45-70 days is 84%, indicating that this data conforms to the characteristic distribution of real medical trauma data.

[0114] If suspicious data shows the following: trauma type is "minor abrasion," injury severity is "mild," but systolic blood pressure is "85 mmHg," pulse is "120 beats / min," treatment plan is "emergency surgical fixation," and recovery time is "60 days," and effective dependency chain analysis reveals that this combination has a probability of only 0.2% in historical distribution, far below the normal threshold of 5%, the system determines that the data's authenticity is questionable and marks it as potentially false data.

[0115] The above implementation method establishes the association patterns among fields of medical trauma data by constructing a weighted directed field dependency graph and identifying valid dependency chains, providing a reliable basis for monitoring data authenticity. This method is particularly suitable for detecting fraudulent data where individual field values ​​are within reasonable ranges but the combination of multiple fields does not conform to real medical patterns, thus improving the accuracy and reliability of medical trauma data quality control.

[0116] like Figure 2 As shown, a schematic diagram of the field dependency graph processing logic is presented.

[0117] In one optional embodiment, the field association domain structure of the abnormal field cluster is extracted, reference medical record cases with the same field association domain structure are retrieved from the medical knowledge graph, a candidate value set is extracted and combined, and a weighted matching is performed to calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strongly dependent field pairs. The combination with the highest association path integrity score is selected as the correction suggestion output, including:

[0118] Extract the field association fields corresponding to each field in the abnormal field cluster to determine the field association field structure of the abnormal field cluster;

[0119] Retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract the field values ​​corresponding to each field in the abnormal field cluster from the reference medical case cases, and form a set of candidate values ​​for each field.

[0120] The candidate value sets of each field are combined and arranged by Cartesian product to generate candidate field value combinations;

[0121] For each candidate field value combination, extract the entity node corresponding to each field value in the medical knowledge graph, query the cumulative association weight of each entity node in the corresponding field association domain, and calculate the cumulative weight of the association path between entity nodes corresponding to different field values.

[0122] The cumulative weight of the associated path and the modulation mutual information strength of the corresponding strong dependency field pair in the abnormal field cluster are weighted and matched to calculate the associated path integrity score of the candidate field value combination.

[0123] The candidate field value combination with the highest association path integrity score is selected as the correction suggestion output.

[0124] In one specific implementation, after detecting an abnormal field cluster, it is necessary to provide reasonable correction suggestions by extracting the corresponding field association domain from each field in the abnormal field cluster. Assuming the abnormal field cluster contains three fields: "surgical method," "anesthesia type," and "postoperative medication," the association domain structure formed by these three fields during multi-hop diffusion is obtained, including the entity node types, semantic types of connecting edges, and path identification patterns involved in each association domain. This structural information is then combined to form the overall field association domain structure of the abnormal field cluster, which describes the graph topological features between the fields.

[0125] In the medical knowledge graph, a structural matching retrieval is performed to find historical medical records with the same field association domain structure. The matching process must ensure that the association domain of the corresponding field in the reference case is consistent with the association domain of each field in the abnormal field cluster in terms of path pattern. The actual values ​​of the corresponding fields are extracted from the retrieved reference medical records to form a candidate value set. For example, the "surgical method" field might yield the candidate value set {open surgery, laparoscopic surgery}, and the "anesthesia type" field might yield {general anesthesia, local anesthesia}.

[0126] Perform a Cartesian product operation on the candidate value set for each field to generate all possible combinations of field values. If the number of candidate values ​​for the three fields are 2, 2, and 3 respectively, then 12 candidate combinations are generated. For each candidate combination, locate the entity node corresponding to each field value in the medical knowledge graph and query the cumulative association weight of that node in its field association domain. Calculate the association path between entity nodes corresponding to different field values ​​within the combination, and sum the weight coefficients of each edge on the path to obtain the cumulative weight W of the association path. path .

[0127] Extract strongly dependent field pairs from the field dependency graph corresponding to the abnormal field clusters and obtain their modulation mutual information strength I. adj The cumulative weights of the associated paths are weighted and matched with the modulation mutual information strength, and the calculation formula is as follows: ,in and This is a balancing coefficient. The score comprehensively considers the semantic coherence of candidate combinations within the knowledge graph and the strength of statistical dependencies between fields. All candidate combinations are traversed to calculate their association path integrity score, and the combination with the highest score is selected as the final correction suggestion. This combination conforms to the entity association patterns in the medical knowledge graph and satisfies the strong dependency constraints between fields, ensuring that the corrected data is reasonable both semantically and statistically. The correction suggestion is output in a structured format, annotating the suggested values ​​for each field and their confidence scores.

[0128] The knowledge graph-based medical trauma data authenticity monitoring system of this invention includes:

[0129] The knowledge graph unit is used to acquire patient medical record data, extract medical entities and identify the relationships between entities, and construct a medical knowledge graph.

[0130] The association domain extraction unit is used to take the field values ​​in the patient's medical record data as the starting node for multi-hop diffusion in the medical knowledge graph, generate path identifiers for the diffusion path, assign weight coefficients to entity nodes based on the path identifiers and calculate the cumulative association weights to form field association domains, and extract field combination patterns.

[0131] An anomaly propagation unit is used to calculate the mutual information strength of each field pair according to the field combination rules, identify strongly dependent field pairs after semantic modulation, construct a field dependency graph with strongly dependent field pairs, and filter effective dependency chains based on path semantic diversity. When a field anomaly is detected, the anomaly propagation probability is calculated along the effective dependency chain, and the chained anomaly fields are located to form an anomaly field cluster.

[0132] The correction suggestion unit is used to extract the field association domain structure of abnormal field clusters, retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract candidate value sets and arrange them in combination, calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strong dependent field pairs, and select the combination with the highest association path integrity score as the correction suggestion output.

[0133] A third aspect of the present invention provides an electronic device, comprising:

[0134] processor;

[0135] Memory used to store processor-executable instructions;

[0136] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0137] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0138] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0139] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A medical trauma data authenticity monitoring method based on a knowledge graph, characterized in that, include: Acquire patient medical record data, extract medical entities and identify the relationships between entities, and construct a medical knowledge graph; The patient's medical record data is used as a starting node for multi-hop diffusion within a medical knowledge graph. Path identifiers are generated for the diffusion paths. Based on these path identifiers, entity nodes are assigned weight coefficients, and cumulative association weights are calculated to form field association domains. Field combination patterns are extracted. The field association domains formed by assigning weight coefficients to entity nodes based on path identifiers and calculating cumulative association weights include: Extract the entity node sequence from the diffusion path identifier, mark the first entity node in the sequence as a first-order associated entity and assign it a first-order weight coefficient, mark the subsequent entity nodes as multi-order associated entities and assign them decreasing multi-order weight coefficients according to their positions in the sequence, count the number of times the entity node appears in the diffusion path identifier set, multiply the number of occurrences by the corresponding weight coefficient to obtain the cumulative association weight of the entity node, and filter the entity nodes whose cumulative association weight exceeds the preset weight threshold and aggregate them with the starting node to form a field association domain. The mutual information strength of each field pair is calculated based on the field combination rules. After semantic modulation, strongly dependent field pairs are identified. A field dependency graph is constructed using these strongly dependent field pairs, and effective dependency chains are selected based on path semantic diversity. When a field anomaly is detected, the anomaly propagation probability is calculated along the effective dependency chain, and the linked anomaly fields are located to form an anomaly field cluster. A weighted directed field dependency graph is constructed using strongly dependent field pairs, and a transitive closure operation is performed to identify reachable paths. The path dependency strength is obtained by adjusting the path semantic diversity index. The selection of effective dependency chains also includes: A weighted directed field dependency graph is constructed using strongly dependent field pairs as directed edges, fields as nodes, and modulation mutual information strength as edge weights. Perform transitive closure operation on a weighted directed field dependency graph to identify all reachable paths in the graph; For each reachable path, extract the strong dependency field pairs corresponding to each directed edge in the path and their intersection entity nodes, and count the number of different intersection entity nodes in the path as a path semantic diversity index. When the path semantic diversity index is less than or equal to the preset lower limit of diversity, a decay factor of less than 1 is applied to the cumulative value of path edge weights. When the path semantic diversity index is greater than the preset upper limit of diversity, an enhancement factor of greater than 1 is applied to the cumulative value of path edge weights, thus obtaining the path dependency strength adjusted by semantic diversity. Filter reachable paths whose path dependency strength exceeds a preset path strength threshold as valid dependency chains; Extract the field association domain structure of the abnormal field clusters, retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract the candidate value set and combine and arrange it, calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strong dependent field pairs, and select the combination with the highest association path integrity score as the correction suggestion output.

2. The method of claim 1, wherein, The values ​​of each field in the patient's medical record data are used as starting nodes for multi-hop diffusion within the medical knowledge graph. Path identifiers are generated for the diffusion paths, and weight coefficients are assigned to entity nodes based on these path identifiers. Cumulative association weights are calculated to form field association domains, and the patterns of field combinations are extracted, including: The values ​​of each field in the patient's medical record data are used as starting nodes and mapped to the medical knowledge graph. Multi-hop diffusion is carried out from the starting node along the directed edges. During the diffusion process, a path identifier is generated for each diffusion path, including the starting node, the sequence of edge types traversed, and the sequence of entity nodes reached. When the number of diffusion hops reaches the preset upper limit, the diffusion is terminated and the path identifier is saved to obtain the diffusion path identifier set of the starting node. Based on the path identifier, weight coefficients are assigned to entity nodes and cumulative association weights are calculated to form the field association domain; Traverse historical medical record data, extract field value combinations and corresponding field association domains from the medical records, identify the intersection entity nodes between different field association domains, count the frequency of field value combinations containing the same intersection entity nodes in historical medical record data, and record field value combinations and intersection entity nodes whose frequency exceeds a preset frequency threshold as field combination patterns.

3. The method of claim 2, wherein, Subsequent entity nodes are marked as multi-level associated entities and assigned decreasing multi-level weight coefficients based on their position in the sequence, including: Extract the last edge type traversed to reach subsequent entity nodes from the diffusion path identifier, query the preset edge type importance scoring table in the medical knowledge graph, and obtain the corresponding importance score; The initial multi-order weight coefficients are calculated based on the position of the subsequent entity nodes in the entity node sequence, and the modulated multi-order weight coefficients are calculated by combining the initial multi-order weight coefficients with the edge type importance score.

4. The method of claim 2, wherein, Identify the intersection entity nodes between different field association domains, and count the frequency of field value combinations containing the same intersection entity nodes in historical medical record data, including: The intersection operation identifies entity nodes that simultaneously exist in at least two related fields as intersection entity nodes. For each intersection entity node, query the cumulative association weight of the intersection entity node in each field association domain, calculate the arithmetic mean of all cumulative association weights, and determine the intersection association strength of the intersection entity nodes. Filter the intersection entity nodes whose intersection association strength exceeds the preset association threshold, and mark the filtered intersection entity nodes as core intersection entity nodes; Traverse historical medical record data, extract field value combinations and corresponding field association fields, determine whether the field association fields contain core intersection entity nodes, and only retain field value combinations whose field association fields contain core intersection entity nodes; The frequency of occurrence of retained field value combinations in historical medical record data.

5. The method of claim 1, wherein, The mutual information strength of each field pair is calculated based on the field combination rules. After semantic modulation, strongly dependent field pairs are identified. A field dependency graph is constructed using these strongly dependent field pairs, and valid dependency chains are filtered based on path semantic diversity. When a field anomaly is detected, the anomaly propagation probability is calculated along the valid dependency chain, and the linked anomaly fields that form the anomaly field cluster include: Extract the joint occurrence count and marginal occurrence count of each field pair from the field combination rules, and calculate the basic mutual information strength of each field pair; Extract the intersection entity nodes associated with each field, query the intersection association strength of the intersection entity nodes as a semantic modulation factor, and perform weighted modulation on the basic mutual information strength to obtain the modulated mutual information strength. Field pairs whose modulation mutual information strength exceeds a preset strength threshold are selected as strong dependency field pairs; A weighted directed field dependency graph is constructed using strong dependency fields, a transitive closure operation is performed to identify reachable paths, the path dependency strength is adjusted based on the path semantic diversity index, and effective dependency chains are selected. When an anomaly is detected in any field, the valid dependency chain starting from the anomaly field is retrieved, the field nodes downstream of the anomaly field in the valid dependency chain are extracted, the anomaly propagation probability is calculated based on the path dependency strength, and the field nodes whose anomaly propagation probability exceeds the preset propagation threshold are selected as chained anomaly fields. Exception fields and chained exception fields are grouped into exception field clusters.

6. The method of claim 1, wherein, Extract the field association domain structure of abnormal field clusters, retrieve reference medical record cases with the same field association domain structure from the medical knowledge graph, extract candidate value sets and arrange them in combination, calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strongly dependent field pairs, and select the combination with the highest association path integrity score as the correction suggestion output, including: Extract the field association fields corresponding to each field in the abnormal field cluster to determine the field association field structure of the abnormal field cluster; Retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract the field values ​​corresponding to each field in the abnormal field cluster from the reference medical case cases, and form a set of candidate values ​​for each field. The candidate value sets of each field are combined and arranged by Cartesian product to generate candidate field value combinations; For each candidate field value combination, extract the entity node corresponding to each field value in the medical knowledge graph, query the cumulative association weight of each entity node in the corresponding field association domain, and calculate the cumulative weight of the association path between entity nodes corresponding to different field values. The cumulative weight of the associated path and the modulation mutual information strength of the corresponding strong dependency field pair in the abnormal field cluster are weighted and matched to calculate the associated path integrity score of the candidate field value combination. The candidate field value combination with the highest association path integrity score is selected as the correction suggestion output.

7. A knowledge graph-based medical trauma data authenticity monitoring system, used to implement the method of any one of claims 1-6, characterized in that, include: The knowledge graph unit is used to acquire patient medical record data, extract medical entities and identify the relationships between entities, and construct a medical knowledge graph. The association domain extraction unit is used to take the field values ​​in the patient's medical record data as the starting node for multi-hop diffusion in the medical knowledge graph, generate path identifiers for the diffusion path, assign weight coefficients to entity nodes based on the path identifiers and calculate the cumulative association weights to form field association domains, and extract field combination patterns. An anomaly propagation unit is used to calculate the mutual information strength of each field pair according to the field combination rules, identify strongly dependent field pairs after semantic modulation, construct a field dependency graph with strongly dependent field pairs, and filter effective dependency chains based on path semantic diversity. When a field anomaly is detected, the anomaly propagation probability is calculated along the effective dependency chain, and the chained anomaly fields are located to form an anomaly field cluster. The correction suggestion unit is used to extract the field association domain structure of abnormal field clusters, retrieve reference medical case cases with the same field association domain structure in the medical knowledge graph, extract candidate value sets and arrange them in combination, calculate the association path integrity score based on the cumulative association weight in the field association domain and the mutual information strength of strong dependent field pairs, and select the combination with the highest association path integrity score as the correction suggestion output.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images