A knowledge graph-based inspection data auditing method and system

Abstract

The application discloses a kind of based on knowledge graph's test data auditing method and system, it is related to medical information processing technical field, target test observation data set is substituted into double-layer heterogeneous knowledge graph logic expression operation, if it does not satisfy preset logic condition and does not trigger preset exemption condition, calculate deviation degree and generate conflict grading state identifier;Mapping node to second layer, constraint path is extracted based on conflict grading state identifier pruning, calculate comprehensive intervention score and extract main associated intervention node and intervention node candidate set;Generation intervention parameter back propagation, match preset numerical value recalculation function logic and generate alternative observation data set;Re-substitute into operation, if it does not satisfy condition, then based on bias proportion parameter attenuation intervention parameter and call intervention node candidate set iteration calculation, reach preset iteration number threshold value and output artificial intervention early warning mark.The application realizes the automatic positioning of test data conflict, intervention and resolution, improves the auditing efficiency and reliability.

Description

Technical Field

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

[0002] In recent years, with the widespread adoption of automated production lines and high-throughput testing equipment in hospital laboratories, clinical laboratory systems generate massive amounts of laboratory observation data daily. The accuracy and reliability of this data directly impact clinicians' diagnostic and treatment decisions. Therefore, efficient and accurate automated review of laboratory data, intercepting erroneous data caused by logical conflicts or equipment malfunctions, has become an indispensable core component of Laboratory Information Management Systems (LIS). Traditional automated review often relies on manually set single thresholds or simple linear rule engines, which struggle to handle complex multi-item constraints. Therefore, the introduction of artificial intelligence and knowledge graph technologies for intelligent review of laboratory data has become a significant current development trend.

[0003] Currently, Chinese invention patent application CN202410676222.9 discloses a method and system for reviewing clinical test results based on artificial intelligence and big data. This method involves constructing a first knowledge graph to acquire physiological disease characteristics and quality control information, historical test data, and medication information, generating a heterogeneous dataset, performing intelligent fusion, extracting and expanding semantic features to obtain high-quality sample data; acquiring real-time message data from testing equipment, performing security authentication, comparing adjacent sample data to generate sample association verification results, constructing an initial automatic review model and outputting automatic review results; comparing the automatic review results with manual approval results to obtain a high-precision automatic review model; inputting sample data into the high-precision automatic review model, performing semantic reasoning, judging abnormal risks and outputting intelligent review reports, calculating confidence levels and comparing them with confidence thresholds, and outputting intelligent review reports with confidence levels higher than the confidence threshold as credible review results. However, there are still obvious limitations in the relevant technologies: First, most of these methods only stay at the surface reasoning of single-layer data association, lacking layered modeling of the internal constraints of the test data and the external generation environment. This leads to unclear conflict localization and tracing paths when facing multiple concurrent underlying physical anomalies (such as reagent decay and equipment baseline drift occurring simultaneously), making it easy to fall into computing power explosion and unable to accurately locate the real fault node. Second, the correction of abnormal data seriously lacks rigorous logical deduction support, simply outputting "high risk" or "low confidence" alarms without proposing automated state intervention for the source node of the anomaly and a reverse transmission mechanism for intervention parameters. As a result, the final data error correction still relies heavily on manual investigation, reducing the degree of automation. Summary of the Invention

[0004] The technical problem addressed by this invention is that existing technologies struggle to accurately trace the underlying physical or operational nodes (such as testing equipment, quality control materials, and reagent batches) that cause structural conflicts in test data under complex multi-project joint constraints. Furthermore, it is difficult to automatically establish a logical deduction mechanism from intervention at the source node status to reverse transmission of observed data and numerical recalculation after data anomalies are detected. This results in existing automated review systems often only providing surface-level anomaly alerts and risk warnings, failing to automatically generate alternative observation data with clear medical logic support and complete conflict resolution verification. Ultimately, data correction still heavily relies on manual review, exhibiting technical deficiencies such as low automation, lack of a closed-loop conflict resolution mechanism, and poor interpretability of the intervention process.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a method for verifying inspection data based on knowledge graphs, comprising the following steps:

[0006] Step S1: Map the target inspection observation dataset to the local related substructure of the two-layer heterogeneous knowledge graph, substitute it into the logical expression for calculation. If the preset logical conditions are not met and the preset exemption conditions are not triggered, calculate the deviation degree to generate a conflict level status identifier and determine that the local related substructure is in a structural conflict state.

[0007] Step S2: Map the observation node involved in the structural conflict state from the first layer to the second layer, extract constraint paths based on the conflict level state identifier and calculate the comprehensive intervention score, and extract the main associated intervention node and intervention node candidate set;

[0008] Step S3: Generate intervention parameters for the main associated intervention node and pass them back to the first layer. Based on the conflict level status identifier, match the preset numerical recalculation function logic, and use the target adjustment weight to perform numerical recalculation processing to generate an alternative observation dataset.

[0009] Step S4: Substitute the alternative observation dataset into the logical expression for calculation. When the calculation result meets the preset logical conditions, output the data processing instruction and classification label. When the calculation result does not meet the preset logical conditions, attenuate the intervention parameter based on the deviation ratio parameter and call the intervention node candidate set for iterative calculation until the conflict is resolved or the preset iteration number threshold is reached, output the manual intervention warning mark.

[0010] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S1 specifically includes:

[0011] Step S11: Obtain the target inspection observation dataset, parse the observation item identifier field in the target inspection observation dataset, and map each observation item identifier field to the corresponding observation item node in the first layer to obtain the mapped observation item node;

[0012] The target test observation dataset includes observation item identifier field, numerical field, time field, source field, and clinical background feature field;

[0013] Step S12: Construct a two-layer heterogeneous knowledge graph, which includes a first layer and a second layer;

[0014] The first layer represents the constraint connection relationship between observation item nodes, the second layer represents the associated nodes that generate observation item data, and a cross-layer connection relationship is established between the first layer and the second layer.

[0015] Step S13: Using the mapped observation node as the starting node, extract the local related substructure in the two-layer heterogeneous knowledge graph, extract the constraint connection relationship connecting two or more observation nodes from the local related substructure, and obtain the logical expression corresponding to the constraint connection relationship.

[0016] Step S14: Substitute the numerical fields in the target test observation dataset into the logical expression for calculation. If the calculation result does not meet the preset logical conditions, substitute the time field and the clinical background feature field for secondary verification calculation. When the state attribute of the associated node in the local associated substructure cannot meet the logical expression and the secondary verification calculation result does not trigger the preset exemption condition, output the judgment result of the structural conflict state.

[0017] The determination result of the structural conflict state includes a conflict level identifier, a set of observation node nodes that cause the operation result to not meet the preset logical conditions, and constraint connection relationships and corresponding logical expressions that do not meet the preset logical conditions.

[0018] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S14 specifically includes:

[0019] Extract the data values ​​of the numerical fields corresponding to the mapped observation node in the target test observation dataset, substitute the data values ​​into the variable parameters in the logical expression for joint operation, and obtain the first logical operation value;

[0020] The logical expression includes a numerical joint operation function and a comparison judgment relationship, and the first logical operation value is the result of the numerical joint operation function.

[0021] When the first logical operation value does not meet the preset logical conditions, the time field is extracted and the historical value set corresponding to the observation item node is obtained from the preset historical database. The time change rate between the current data value and the historical value set is calculated, and the clinical background feature field is extracted and text matching is performed in the preset pathological feature dictionary.

[0022] The formula for calculating the rate of change over time is:

[0023] ;

[0024] in, The rate of change over time, For the current observation value, This is the most recent historical observation value. It should be a very small constant to avoid the denominator being zero;

[0025] When the rate of change over time is within the preset safe range of historical fluctuations or the text matching result of the clinical background feature field falls within the legal range of the preset pathological feature dictionary, a conflict exemption flag is generated and the current review and judgment operation is terminated.

[0026] When no conflict exemption flag is generated, traverse the associated nodes in the local associated substructure that are connected to the constraint connection relationship that does not meet the preset logical conditions, and obtain the state attribute value of the associated node.

[0027] The state attribute value is used as a supplementary condition to participate in the secondary judgment of the corresponding logical expression. When the state attribute values ​​of all associated nodes in the local associated substructure cannot make the result of the secondary operation judgment meet the preset logical condition, the absolute value of the difference between the first logical operation value and the standard value corresponding to the preset logical condition is calculated as the deviation degree. The deviation degree is compared with the preset classification threshold, and a conflict classification label is generated based on the comparison result.

[0028] Extract the mapped observation nodes corresponding to the variable parameters whose first logical operation value does not meet the preset logical conditions, combine them to obtain the observation node set, package and encapsulate the conflict level identifier, the observation node set, the constraint connection relationship and the corresponding logical expression, and output the judgment result of the structural conflict state.

[0029] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S2 specifically includes:

[0030] Step S21: Based on the set of observation item nodes and the cross-layer connection relationship, map the set of observation item nodes from the first layer to the second layer to obtain the starting node set of the second layer;

[0031] Step S22: Starting from each node in the second-layer initial node set, extract the corresponding connection type identifier rules based on the conflict level identifier in the second layer and perform path matching and pruning operations to obtain a candidate path set. The processing logic includes:

[0032] Extract the connection type identifier rules corresponding to the conflict level identifier, and construct the initial search path starting from each node in the second-level starting node set.

[0033] In the second-layer network topology, traverse the neighboring nodes connected to the current node and verify whether the node classification attributes of the neighboring nodes and their association connection relationship with the current node conform to the legal connection type defined by the connection type identification rules.

[0034] If the verification passes, the influence weight attribute value of the adjacent node is extracted. When the influence weight attribute value is less than the preset clinical significance threshold, the adjacent node is pruned and removed.

[0035] When the value of the influence weight attribute is greater than or equal to the preset clinical significance threshold, the adjacent node is added to the current search path, and the adjacent node is used as the current node to continue the traversal and verification operation until the node classification attribute arrangement of the current search path is completely consistent with the preset node type sequence. The completely consistent current search path is extracted and saved to the candidate path set.

[0036] If the verification fails, the adjacent node is determined to be an invalid node. The adjacent node is skipped and the downward expansion of the current search branch is blocked. The current node is returned to continue to traverse the other adjacent nodes connected to it for verification operations until the current search path can no longer be expanded and a candidate path is formed. The candidate path is then extracted and saved to the candidate path set.

[0037] Step S23: Perform connectivity determination on each candidate path in the candidate path set, and select candidate paths that simultaneously connect all nodes in the second-level starting node set as the constraint path set.

[0038] Step S24: Calculate the path intersection degree, historical failure rate attribute value, and influence weight attribute value for each node in the constraint path set.

[0039] Preset weight coefficients are assigned to the path convergence degree, the reciprocal of the historical failure rate attribute value, and the influence weight attribute value, and then a weighted sum is performed to obtain the comprehensive intervention score for each node. The calculation formula is as follows:

[0040] ;

[0041] in, For comprehensive intervention scoring, The first preset weighting coefficient, This is the second preset weighting coefficient. The third preset weighting coefficient, For path intersection degree, This refers to the historical failure rate attribute value. To ensure the denominator is a very small constant, avoid setting the denominator to zero. To influence the weighted attribute values;

[0042] The node with the highest comprehensive intervention score is identified as the primary intervention node, and the remaining nodes with comprehensive intervention scores greater than a preset secondary threshold are arranged in descending order of score to form a candidate set of intervention nodes.

[0043] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S23 specifically includes:

[0044] The candidate paths in the candidate path set are aggregated according to their end nodes. The end node identifiers of each candidate path in the candidate path set are extracted. The candidate paths in the candidate path set are classified according to their end node identifiers. Candidate paths with the same end node identifiers are grouped into the same group to obtain a path group set. Each group in the path group set is numbered and labeled, and the candidate paths in each group are determined as the corresponding path aggregation set.

[0045] The end node is the last node in the candidate path according to the search order;

[0046] For each path aggregation set, extract the starting node corresponding to each path in the path aggregation set to form a subset of covering nodes;

[0047] Perform a coverage determination between the covered subset of nodes and the second-level starting node set;

[0048] When the subset of covered nodes covers all nodes in the set of starting nodes in the second layer, it is determined that the path aggregation set satisfies the preset cross-node connectivity condition.

[0049] Perform topology fusion processing on each path in the path aggregation set that meets the preset cross-node connectivity conditions to generate a connected path structure, and add the connected path structure to the constrained path set.

[0050] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S3 specifically includes:

[0051] Step S31: Extract the status attribute value of the main associated intervention node, perform status intervention operation on the status attribute value of the main associated intervention node, and set the status attribute value of the main associated intervention node to a preset baseline value.

[0052] The preset baseline value is the normal state value pre-stored in the state attribute field of the main associated intervention node;

[0053] In the second-layer network topology, the incoming connection relationship with the main associated intervention node as the endpoint node is extracted, the incoming connection relationship is blocked, and intervention parameters are generated.

[0054] Intervention parameters include the state change amount blocking marker and the initial intervention intensity;

[0055] Step S32: Using the intervention parameters as initial input, determine the main associated intervention node as the propagation starting node, and propagate the intervention parameters hop by hop to the first layer along the constraint path. At each hop, adjust the intervention parameters according to the current node's state attribute value and the connection relationship type between the current node and the previous hop node to obtain the updated intervention amount, and propagate it to the previous hop node. Repeat this operation until the intervention amount is propagated to the observation node corresponding to the first layer, and output the intervention propagation amount for the first layer node state update. The processing logic is as follows:

[0056] Mark the main associated intervention node as the propagation start node, and encapsulate the intervention parameters as the initial propagation amount;

[0057] Set the propagation start node to the current node, and assign the initial propagation amount to the current propagation amount;

[0058] Proceed hop-by-hop along the constraint path to the first layer. In each hop, extract the previous hop node that is directly connected to the current node.

[0059] Based on the connection relationship type between the current node and the previous hop node, the current transmission amount is adjusted and calculated to obtain the updated intervention amount;

[0060] The connection types include associated connection relationships and cross-layer connection relationships;

[0061] The updated intervention amount is passed to the previous hop node, the previous hop node is reset to the current node, and the updated intervention amount is reset to the current amount passed.

[0062] The operation of extracting the previous hop node and resetting the current node is executed repeatedly until the current node reaches the corresponding observation node in the first layer through the cross-layer connection relationship. Then the operation is stopped and the final current transit quantity is output as the intervention transit quantity.

[0063] Step S33: In response to the intervention transmission amount being transmitted to the first layer via cross-layer connection, the set of observation node items that trigger the structural conflict state is located in the first layer.

[0064] Step S34: Map the intervention transmission amount to each node in the set of located observation nodes, and convert the intervention transmission amount into target adjustment weights acting on each observation node according to the preset conversion rules.

[0065] Step S35: Extract the conflict level identifier and observation item identifier fields from the determination result of the structural conflict state, perform index matching in the preset function mapping table, and extract the corresponding preset numerical recalculation function logic.

[0066] Based on the target adjustment weight, the numerical fields corresponding to the observation item node set in the target inspection observation dataset are substituted into the extracted preset numerical recalculation function logic to perform numerical recalculation processing, obtain the updated observation values, and replace the corresponding original values ​​with the updated observation values ​​to generate the alternative observation dataset.

[0067] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S35 specifically includes the following operations:

[0068] Parse the set of observation node nodes to obtain the set of node identifiers corresponding to each node;

[0069] Based on the node identifier set, identifier matching is performed in the target inspection observation dataset to extract the numerical field data corresponding to each node identifier, thus obtaining the original numerical set.

[0070] The target adjustment weights are mapped and parsed to establish the correspondence between node identifiers and weight adjustment parameters;

[0071] Extract successful historical conflict resolution records from the preset historical database that are of the same type as the current conflict classification identifier. Extract the average historical weights used in the successful historical conflict resolution records. Calculate the difference between the average historical weights and the current weight adjustment parameters as the self-learning compensation coefficient.

[0072] The self-learning compensation coefficient is added to the current weight adjustment parameter to obtain the optimized weight adjustment parameter.

[0073] The original set of numerical values ​​and the optimized weight adjustment parameters are used as input variables and substituted into the preset numerical recalculation function for calculation to obtain the updated set of observed numerical values. The updated set of observed numerical values ​​is then written into the corresponding numerical field position in the target test observation dataset to replace the original values.

[0074] While keeping the observation identifier field, time field, and source field unchanged, the target inspection observation dataset after replacement is restructured to obtain the alternative observation dataset.

[0075] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S4 specifically includes:

[0076] Step S41 involves parsing the field mapping of the alternative observation dataset, extracting the corresponding numerical field data based on the node identifiers in the observation item node set, and generating an updated observation value set. The processing logic is as follows:

[0077] Traverse the set of observation nodes, extract the node identifier of each observation node in the set of observation nodes one by one, and construct and update the set of observation values;

[0078] Using the node identifier as the index key, the observation item identifier field is traversed and mapped in the substitute observation dataset for matching. When the content of the observation item identifier field is completely consistent with the node identifier, the field mapping match is determined to be successful, and the corresponding data record row in the substitute observation dataset is locked.

[0079] Extract the updated observation values ​​stored in the numeric fields of the data record rows;

[0080] The updated observation values ​​are mapped to their corresponding node identifiers to form key-value pairs, and these key-value pairs are added to the updated observation value set until all node identifiers in the observation item node set have been traversed.

[0081] Step S42: Perform structural analysis on the determination result of the structural conflict state, and extract the constraint connection relationship and corresponding logical expression that do not meet the preset logical conditions;

[0082] Step S43: The updated set of observed values ​​is mapped to the variable parameters in the logical expression according to the variable correspondence and then subjected to joint operation to obtain the second logical operation value;

[0083] Step S44: Compare the second logical operation value with the preset logical conditions. When the logical comparison result is that the second logical operation value meets the preset logical conditions, generate a conflict resolution status marker and determine that the structural conflict state of the local associated substructure is resolved based on the conflict resolution status marker.

[0084] Step S45: When the logic comparison result is that the second logic operation value does not meet the preset logic condition, calculate the absolute difference between the updated observed value and the corresponding original value, and divide the absolute difference by the original value to obtain the deviation ratio parameter.

[0085] The deviation ratio parameter is passed back to the second layer, and the intervention parameter of the current main associated intervention node is updated by multiplying it by the preset decay constant, while triggering the preset iteration counter to increment by one.

[0086] Step S46: Determine whether the value of the preset iteration counter is less than the preset iteration number threshold;

[0087] If the value of the preset iteration counter is less than the preset iteration number threshold, then when the intervention parameter of the current main associated intervention node is less than the preset lower limit, the candidate node at the top of the intervention node candidate set is extracted and replaced as the main associated intervention node, and the updated intervention parameter is returned to the execution step S3 to generate a new alternative observation dataset for cyclic calculation.

[0088] If the value of the preset iteration counter is greater than or equal to the preset iteration number threshold, a manual intervention warning mark is generated, and a data processing instruction is generated based on the conflict level identifier in the structural conflict status determination result. The data processing instruction and classification identifier are then output.

[0089] As a preferred embodiment of the knowledge graph-based verification data review method described in this invention, step S46 specifically includes:

[0090] Analyze the results of the structural conflict status determination, and extract conflict classification indicators and manual intervention warning markers;

[0091] Extract the source and time fields from the target inspection observation dataset;

[0092] Extract the state operation code corresponding to the conflict level identifier from the preset state mapping table;

[0093] When a manual intervention warning marker is generated, extract the status operation code that is jointly corresponding to the conflict classification identifier and the manual intervention warning marker.

[0094] The source field, time field, and status operation code are concatenated to generate a data processing instruction;

[0095] Convert data processing instructions, category identifiers, and node identifiers into key-value pairs in a preset format;

[0096] The key-value pairs are combined to generate a dictionary data structure, and the dictionary data structure is written into a preset output variable field.

[0097] Secondly, a knowledge graph-based verification data review system includes a conflict determination module, an intervention positioning module, a data reconstruction module, and an instruction output module.

[0098] The conflict determination module is used to map the target inspection observation dataset to the local related substructure of the two-layer heterogeneous knowledge graph, substitute it into the logical expression for calculation, and if the preset logical conditions are not met and the preset exemption conditions are not triggered, calculate the deviation degree to generate a conflict level status identifier, and determine that the local related substructure is in a structural conflict state.

[0099] The intervention positioning module is used to map the observation item nodes involved in the structural conflict state from the first layer to the second layer, extract constraint paths based on the conflict level state identifier, calculate the comprehensive intervention score, and extract the main associated intervention node and intervention node candidate set.

[0100] The data reconstruction module is used to generate intervention parameters for the main associated intervention node and back-transmit them to the first layer. Based on the conflict level status identifier, it matches the preset numerical recalculation function logic, uses the target adjustment weight to perform numerical recalculation processing, and generates an alternative observation dataset.

[0101] The instruction output module is used to substitute the alternative observation dataset into the logical expression for calculation. When the calculation result meets the preset logical conditions, it outputs the data processing instruction and classification label. When the calculation result does not meet the preset logical conditions, it attenuates the intervention parameter based on the deviation ratio parameter and calls the intervention node candidate set for iterative calculation until the conflict is resolved or the preset iteration number threshold is reached, and outputs the manual intervention warning mark.

[0102] The beneficial effects of this invention are as follows: This invention introduces a two-layer heterogeneous knowledge graph, which hierarchically models the constraint connection relationships between observation items and the associated nodes that generate observation item data, and realizes the association calculation between information at different levels through cross-layer connection relationships. When a structural conflict occurs, path matching and connectivity determination are performed in the second layer to extract the associated intervention nodes involved in the conflict, and the values ​​of the relevant nodes are reversed and corrected based on state intervention operations and graph propagation mechanisms. On this basis, the corrected observation data is resubmitted into the logical expression corresponding to the constraint connection relationship for secondary calculation, so that the original operation relationship that does not meet the preset logical conditions is reconstructed in a consistent manner. Based on the full-process data association expression of structured records, not only is an automated closed loop from conflict determination, path extraction, state intervention to result output realized, but it also brings significant performance improvement in practical applications: relying on the intelligent exemption mechanism based on time sequence and clinical background, the false positive data interception rate caused by individual physiological differences and special medical history of patients is significantly reduced; at the same time, relying on multi-dimensional scoring to locate intervention nodes and automatically match correction logic, the automatic correction success rate of test data anomalies is greatly improved, and the overall automatic review efficiency is significantly improved. Attached Figure Description

[0103] Figure 1 A flowchart illustrating the steps of a knowledge graph-based verification data review method according to an embodiment of the present invention;

[0104] Figure 2 This is a basic flowchart of a knowledge graph-based verification data review system provided in one embodiment of the present invention. Detailed Implementation

[0105] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0106] Example 1, referring to Figure 1 This paper presents a knowledge graph-based method for verifying inspection data, which includes the following steps:

[0107] Step S1: Map the target inspection observation dataset to the local related substructure of the two-layer heterogeneous knowledge graph, substitute it into the logical expression for calculation. If the preset logical conditions are not met and the preset exemption conditions are not triggered, calculate the deviation degree to generate a conflict level status identifier and determine that the local related substructure is in a structural conflict state.

[0108] Step S2: Map the observation node involved in the structural conflict state from the first layer to the second layer, extract constraint paths based on the conflict level state identifier and calculate the comprehensive intervention score, and extract the main associated intervention node and intervention node candidate set;

[0109] Step S3: Generate intervention parameters for the main associated intervention node and pass them back to the first layer. Based on the conflict level status identifier, match the preset numerical recalculation function logic, and use the target adjustment weight to perform numerical recalculation processing to generate an alternative observation dataset.

[0110] Step S4: Substitute the alternative observation dataset into the logical expression for calculation. When the calculation result meets the preset logical conditions, output the data processing instruction and classification label. When the calculation result does not meet the preset logical conditions, attenuate the intervention parameter based on the deviation ratio parameter and call the intervention node candidate set for iterative calculation until the conflict is resolved or the preset iteration number threshold is reached, output the manual intervention warning mark.

[0111] In specific implementation, step S1 includes:

[0112] Step S11: Obtain the target inspection observation dataset, parse the observation item identifier field in the target inspection observation dataset, and map each observation item identifier field to the corresponding observation item node in the first layer to obtain the mapped observation item node;

[0113] The target test observation dataset includes observation item identifier field, numerical field, time field, source field, and clinical background feature field;

[0114] Step S12: Construct a two-layer heterogeneous knowledge graph, which includes a first layer and a second layer.

[0115] The first layer represents the constraint connection relationship between observation item nodes, the second layer represents the associated nodes that generate observation item data, and a cross-layer connection relationship is established between the first layer and the second layer.

[0116] Step S13: Using the mapped observation node as the starting node, extract the local related substructure in the two-layer heterogeneous knowledge graph, extract the constraint connection relationship connecting two or more observation nodes from the local related substructure, and obtain the logical expression corresponding to the constraint connection relationship.

[0117] Step S14: Substitute the numerical fields in the target test observation dataset into the logical expression for calculation. If the calculation result does not meet the preset logical conditions, substitute the time field and the clinical background feature field for secondary verification calculation. When the state attribute of the associated node in the local associated substructure cannot meet the logical expression and the secondary verification calculation result does not trigger the preset exemption condition, output the judgment result of the structural conflict state.

[0118] The determination result of structural conflict status includes conflict level identifier, set of observation node that causes the operation result to not meet the preset logical conditions, and constraint connection relationship and corresponding logical expression that do not meet the preset logical conditions.

[0119] Specifically, the target test observation dataset originates from clinical testing systems, laboratory information systems, or hospital information systems, and is obtained through data interface calls, database reading, or real-time message parsing. Each data entry includes at least an observation item identifier field, a numerical field, a time field, a source field, and a clinical background feature field. The clinical background feature field includes age information, gender information, and medical history label information. Subsequently, the observation item identifier field is parsed and converted into a standardized encoding form based on preset observation item encoding rules. The standardized encoding form includes LOINC encoding or hospital-wide unified encoding. Then, in the first layer of a pre-constructed two-layer heterogeneous knowledge graph, the standardized encoding is mapped to the corresponding observation item node through identifier matching. The two-layer heterogeneous knowledge graph consists of a first layer, a second layer, and cross-layer connections connecting the first and second layers. The first layer includes multiple observation nodes and constraint connections between them, each with a corresponding logical expression. The second layer includes associated nodes related to the observation data generation process, such as testing equipment nodes, quality control nodes, reagent nodes, or operator nodes. Cross-layer connections are used to establish generative, dependency, or validation relationships between observation nodes and associated nodes. When extracting local associated substructures, the local associated substructure is extracted from the two-layer heterogeneous knowledge graph, starting with the mapped observation node. This substructure includes observation nodes, associated nodes, and corresponding constraint and cross-layer connections, and contains constraint connections connecting two or more observation nodes and their corresponding logical expressions.

[0120] The construction process of the two-layer heterogeneous knowledge graph specifically includes: First, parsing the observation item identifier field in the target test observation dataset and standardizing the observation items based on preset observation item coding rules to establish a set of observation item nodes; based on the test rules or medical constraint relationships between observation items, constructing constraint connection relationships between observation item nodes, and configuring corresponding logical expressions for each constraint connection relationship, thus forming the first layer; Second, acquiring auxiliary information data related to the observation item data generation process, constructing a set of associated nodes based on the entity types in the auxiliary information data, including test equipment nodes, quality control nodes, reagent nodes, or operator nodes; based on the auxiliary information data... To aid information data, the association records between entities are used to establish connections between related nodes, and connection type identifiers are configured for these connections, thus forming the second layer. Then, based on the correspondence between observation items and their generation sources, cross-layer connections are established between observation item nodes in the first layer and related nodes in the second layer. These cross-layer connections include generation associations between observation items and testing equipment, dependency associations between observation items and reagents, or verification associations between observation items and quality control nodes. Finally, the first layer, the second layer, and the cross-layer connections are integrated to form a two-layer heterogeneous knowledge graph structure containing multiple types of nodes and connections. Pre-defined observation item coding rules are used to standardize the identification of observation items in the target testing observation dataset. This set of mapping rules includes the correspondence between observation item names and standard codes. During parsing, the observation item identification is matched and parsed based on the text content in the observation item identification field, and the matching results are converted into a unified standard coding form. The standard coding form is a coded representation that uniformly identifies the same observation item under different data sources, different naming methods, or different expression formats. Locally related substructures are subgraph structures extracted from the knowledge graph based on the mapped observation item nodes.

[0121] Compared to traditional data review methods based on rule engines or single knowledge graphs, this application introduces a two-layer heterogeneous knowledge graph, hierarchically modeling the constraint relationships between observations and the associated nodes that generate the data. This allows conflict detection to move beyond threshold judgments for individual data items, enabling joint calculations of the state attributes of multiple observations and associated nodes within local associated substructures. This significantly improves the accuracy and interpretability of conflict determination. More importantly, this application innovatively introduces a time-dimensional volatility check and a patient's individual clinical background adaptation layer into conflict determination. When logical conflicts exist on the surface of the data, the system can automatically perform secondary verification and exemption based on the patient's actual condition, thereby greatly reducing false positives caused by special physiological characteristics (such as abnormal indicators in diabetic patients). Simultaneously, its structured conflict determination output, including conflict level identifiers, conflict node sets, constraint relationships, and logical expressions, provides a complete information foundation for subsequent path search, state intervention, and automatic correction. This enables a dynamic intervention mechanism based on severity classification, overcoming the technical shortcomings of traditional methods such as inaccurate conflict localization, inability to automatically associate intervention nodes, and susceptibility to batch false interception.

[0122] In specific implementation, step S14 includes:

[0123] Extract the data values ​​of the numerical fields corresponding to the mapped observation node in the target test observation dataset, substitute the data values ​​into the variable parameters in the logical expression for joint operation, and obtain the first logical operation value;

[0124] When the first logical operation value does not meet the preset logical conditions, the time field is extracted and the historical value set corresponding to the observation item node is obtained from the preset historical database. The time change rate between the current data value and the historical value set is calculated, and the clinical background feature field is extracted and text matching is performed in the preset pathological feature dictionary.

[0125] When the rate of change over time is within the preset safe range of historical fluctuations or the text matching result of the clinical background feature field falls within the legal range of the preset pathological feature dictionary, a conflict exemption flag is generated and the current review and judgment operation is terminated.

[0126] When no conflict exemption flag is generated, traverse the associated nodes in the local associated substructure that are connected to the constraint connection relationship that does not meet the preset logical conditions, and obtain the state attribute value of the associated node.

[0127] The state attribute value is used as a supplementary condition to participate in the secondary judgment of the corresponding logical expression. When the state attribute values ​​of all associated nodes in the local associated substructure cannot make the result of the secondary operation judgment meet the preset logical condition, the absolute value of the difference between the first logical operation value and the standard value corresponding to the preset logical condition is calculated as the deviation degree. The deviation degree is compared with the preset classification threshold, and a conflict classification label is generated based on the comparison result.

[0128] Extract the mapped observation nodes corresponding to the variable parameters whose first logical operation value does not meet the preset logical conditions, combine them to obtain the observation node set, package and encapsulate the conflict level identifier, the observation node set, the constraint connection relationship and the corresponding logical expression, and output the judgment result of the structural conflict state.

[0129] Specifically, substituting data values ​​into logical expressions for joint operations is actually accomplished through variable binding and expression evaluation. The logical expression contains several variable parameters, each mapped to an observation node. The system extracts the numerical field corresponding to that observation node from the target test observation dataset, assigns the value to the corresponding variable in the expression, and then performs mathematical or logical operations, such as addition, subtraction, multiplication, division, comparison operations, and logical AND / OR, to obtain the first logical operation value, whose calculation expression is:

[0130] ;

[0131] in, The first logical operation value, For the union operation function corresponding to the logical expression, To examine the numerical field data of the corresponding observation item node in the target observation dataset, The number of observations participating in the joint operation;

[0132] The logical expression is used to compare and judge the value of the first logical operation, and its judgment form is:

[0133] or ;

[0134] in, For logical expressions, For comparison relational operators, The standard reference threshold, These are the standard interval boundary values; when the above relationship is satisfied, it is determined that the preset logical condition is met.

[0135] Joint operation functions include ratio operation functions, difference operation functions, interval judgment functions, and multivariable logic constraint functions;

[0136] Preset logical conditions refer to the expected judgment criteria of a logical expression under normal circumstances. They are usually expressed as a Boolean condition, such as the result of the operation should equal a certain value, fall within a certain range, or satisfy a certain relationship. This transforms medical constraints into computable judgment rules. When the result of the operation does not meet the condition, it indicates that there are anomalies in the observed data. Variable parameters are replaceable placeholders in the logical expression, representing symbols for different observed values. For example, blood glucose and glycated hemoglobin are variable parameters in the expression "|blood glucose - 0.9 * glycated hemoglobin| < 0.5".

[0137] A pre-defined historical database refers to a structured storage system that stores patients' past test observation data and equipment historical calibration / quality control data. It is typically deployed as a standalone time-series database or directly synchronized with the hospital's Laboratory Information System (LIS). Its construction method is as follows: using the patient's unique identifier (such as a patient card number) or the equipment's unique identifier as the primary key, and the timestamp as the index, it stores the corresponding test items and values ​​in chronological order. Human physiological indicators or equipment operating status have temporal continuity; isolated judgments detached from historical baselines are highly prone to misinterpretation. This database provides the system with a data benchmark for calculating the rate of change over time. The formula for calculating the rate of change over time is:

[0138] ;

[0139] in, The rate of change over time, For the current observation value, This is the most recent historical observation value. It is a very small constant used to avoid the denominator being zero;

[0140] A pre-defined pathological feature dictionary is a rule base that associates and stores clinical disease diagnostic labels with the allowable abnormal value ranges of corresponding laboratory tests. Its construction process includes: standardizing disease description information in medical record texts based on clinical medical testing guidelines and expert consensus to generate corresponding diagnostic labels, including labels for severe diabetes and uremia; subsequently, establishing a mapping relationship between diagnostic labels and the corresponding legal abnormal ranges of laboratory tests, and storing this mapping using a key-value pair. The mapping content includes diagnostic labels, laboratory tests, and corresponding legal exemption ranges. For example, the legal exemption range for fasting blood glucose can be set to 10.0 mmol / L to 25.0 mmol / L.

[0141] Preset logical conditions are typically set based on normal physiological indicator ranges; for example, the normal blood glucose range can be set to 3.9 mmol / L to 6.1 mmol / L. When the target subject has a disease state, some test indicators may remain within the abnormal range for an extended period. If a uniform physiological benchmark is still used for judgment, misjudgments are likely to occur. By introducing a preset pathological feature dictionary, clinical disease states are associated with the legitimate abnormal ranges of corresponding test items. This allows the system to incorporate clinical background feature fields to make appropriate judgments on abnormal results during logical verification, thereby reducing the probability of misjudgments due to individual pathological differences.

[0142] The preset historical fluctuation safety range refers to the reasonable range of numerical variation allowed for different test items within a preset time period. This historical fluctuation safety range is set based on the biological variability of the corresponding test item. For routine biochemical items, the historical fluctuation safety range can be set within ±15% of the historical average, or within ±2 standard deviations of the historical data average. When calculating the rate of change over time, the system calculates the magnitude of change between the current data value and historical data based on the historical value set of the corresponding observation item in the target test observation dataset, and compares the magnitude of change with the preset historical fluctuation safety range. When the magnitude of change is within the historical fluctuation safety range, the current data value is considered to be consistent with the historical trend; when the magnitude of change exceeds the historical fluctuation safety range, the current data value is considered to have abnormal fluctuations. Since the test indicators of some target objects may be in a stable abnormal state for a long time, the introduction of the historical fluctuation safety range allows the system to combine historical trends to make continuous judgments on the current data, thereby distinguishing between normal pathological fluctuations and abnormal mutations. Abnormal mutations include abnormal numerical changes caused by abnormal testing equipment, reagent failure, or abnormal data acquisition. When the rate of change of the current data value over time is within the safe range of historical fluctuations, the system generates a corresponding conflict exemption flag, thereby reducing unnecessary anomaly interception.

[0143] When the first logical operation value does not meet the preset logical conditions, the system does not directly determine that the target test observation data is abnormal. Instead, it further performs secondary verification processing based on time series features and clinical background features. Specifically, the system extracts historical test records of the corresponding target object within a preset time range based on the time field, and calculates the time change rate corresponding to the current data value based on the historical test records. At the same time, it parses the clinical background feature field to extract the diagnostic label information contained therein. When the time change rate does not exceed the historical fluctuation safety range of the corresponding observation item, the current data value is determined to conform to the historical change trend; or, when the clinical background feature field contains a preset disease diagnosis label, and the disease diagnosis label has an allowable abnormal value range in the preset pathological feature dictionary, the current data value is determined to match the corresponding pathological state. When the time change rate meets the historical fluctuation safety range or the current data value conforms to the corresponding pathological state, the system generates a conflict exemption flag and terminates the current abnormal interception process. By combining historical time series change features and clinical pathological background information to jointly verify abnormal results, the system can make adaptability judgments on abnormal test results generated under special pathological states, thereby reducing misjudgments.

[0144] When no conflict exemption flag is generated, the system locates the associated nodes in the local associated substructure that correspond to the constraint connections that do not meet the preset logical conditions. These associated nodes include quality control nodes, equipment nodes, and reagent nodes. Subsequently, the system extracts the state attribute values ​​corresponding to each associated node, including whether quality control has passed, whether the equipment is within its calibration cycle, and reagent status information. Then, the state attribute values ​​are substituted as supplementary conditions into the corresponding logical expressions for secondary computation, allowing the state attribute values ​​to participate in the logical judgment process as conditional constraints. When the state attribute values ​​of all associated nodes in the local associated substructure fail to make the logical expression satisfy the preset logical conditions, the current structural conflict state is determined to be an irresolvable conflict state.

[0145] After determining that the current structural conflict state is an irresolvable conflict state, the system calculates the absolute value of the difference between the first logical operation value and the standard reference value corresponding to the preset logical condition, and determines the absolute value of the difference as the deviation degree, the calculation formula of which is:

[0146] ;

[0147] in, For deviation, The threshold value is the standard reference value corresponding to the preset logical condition;

[0148] Subsequently, the deviation is compared with a preset classification threshold, and a corresponding conflict classification identifier is generated based on the comparison result. The preset classification threshold includes a first threshold and a second threshold;

[0149] when When a minor conflict occurs, a conflict level identifier, Level_1_Minor, is generated.

[0150] when At that time, a conflict classification identifier Level_2_Critical corresponding to the critical conflict is generated;

[0151] when At that time, a conflict level identifier Level_3_Fatal corresponding to the fatal conflict is generated;

[0152] in, The first threshold, The second threshold is set as follows: the first threshold can be set to 10% of the corresponding standard reference value threshold, and the second threshold can be set to 30% of the corresponding standard reference value threshold.

[0153] The system extracts the observation nodes corresponding to the variable parameters that cause the first logical operation value to fail to meet the conditions, forming an observation node set. Finally, the conflict level identifier, the observation node set, the constraint connection relationship, and the corresponding logical expression are packaged and encapsulated to form a structured conflict determination result. This encapsulation typically uses a dictionary structure, JSON, or key-value pairs to organize different types of information into a single data packet, so as to serve as a unified input for path search and status intervention, ensuring the complete transmission and traceability of conflict information throughout the review process. Status attribute values ​​are used as additional constraint conditions in the logical expression judgment process and represent attribute parameters of the associated node's operating status or business status, including equipment status, quality control status, or reagent status; observation nodes are nodes mapped from the observation identifier field.

[0154] The order rules for conflict exemption judgment and conflict classification identifier generation are as follows: When the logical comparison result corresponding to the first logical operation value is "not satisfied with the preset logical condition", the system performs secondary verification and classification processing in the following order:

[0155] Based on the time field, a set of historical values ​​is obtained and the time change rate is calculated. When the time change rate falls into the historical fluctuation safety range corresponding to the observation item, a conflict exemption flag is generated and the current structural conflict determination is cancelled.

[0156] If no exemption flag is generated, perform text matching on the clinical background feature field. When a pathology label is matched and the legal abnormal interval of the observation item corresponding to the pathology label covers the current observation value, generate a conflict exemption flag and cancel the current structural conflict determination.

[0157] If no exemption flag is generated, the deviation degree is calculated and compared with the preset grading threshold to generate a conflict grading flag. At the same time, the conflict grading flag, the set of observation node nodes that cause the default logic conditions to be not met, the constraint connection relationship flag and the corresponding logic expression flag are encapsulated and output as input for subsequent tracing and intervention steps.

[0158] In specific implementation, step S2 includes:

[0159] Step S21: Based on the set of observation item nodes and the cross-layer connection relationship, map the set of observation item nodes from the first layer to the second layer to obtain the corresponding set of starting nodes in the second layer;

[0160] Step S22: Starting from each node in the second-layer initial node set, extract the corresponding connection type identifier rules based on the conflict level identifier in the second layer and perform path matching and pruning operations to obtain a candidate path set. The processing logic includes:

[0161] Extract the connection type identifier rules corresponding to the conflict level identifier, and construct the initial search path starting from each node in the second-level starting node set.

[0162] In the second-layer network topology, traverse the neighboring nodes connected to the current node and verify whether the node classification attributes of the neighboring nodes and their association connection relationship with the current node conform to the legal connection type defined by the connection type identification rules.

[0163] If the verification passes, the influence weight attribute value of the adjacent node is extracted. When the influence weight attribute value is less than the preset clinical significance threshold, the adjacent node is pruned and removed.

[0164] When the value of the influence weight attribute is greater than or equal to the preset clinical significance threshold, the adjacent node is added to the current search path, and the adjacent node is used as the current node to continue the traversal and verification operation until the node classification attribute arrangement of the current search path is completely consistent with the preset node type sequence. The completely consistent current search path is extracted and saved to the candidate path set.

[0165] If the verification fails, the adjacent node is determined to be an invalid node, the adjacent node is skipped and the downward expansion of the current search branch is blocked. The current node is returned to continue traversing the remaining adjacent nodes connected to it for verification.

[0166] Step S23: Perform connectivity determination on each candidate path in the candidate path set, and select candidate paths that simultaneously connect all nodes in the second-level starting node set as the constraint path set.

[0167] Step S24: Calculate the path intersection degree, historical failure rate attribute value, and influence weight attribute value for each node in the constraint path set.

[0168] Preset weight coefficients are assigned to the path intersection degree, the reciprocal of the historical failure rate attribute value, and the influence weight attribute value, and a weighted sum is calculated to obtain the comprehensive intervention score for each node.

[0169] The node with the highest comprehensive intervention score is identified as the primary intervention node, and the remaining nodes with comprehensive intervention scores greater than a preset secondary threshold are arranged in descending order of score to form a candidate set of intervention nodes.

[0170] Specifically, mapping the set of observation nodes from the first layer to the second layer is accomplished through predefined cross-layer connection relationships. Each observation node in the first layer establishes a corresponding cross-layer connection relationship with its associated nodes in the second layer. These cross-layer connections characterize the data generation or influence relationships between the observation nodes and their associated nodes. Associated nodes include testing equipment nodes, quality control nodes, reagent nodes, or operator nodes. Based on the set of observation nodes obtained in step S3, associated nodes in the second layer are retrieved along the corresponding cross-layer connection relationships to obtain the associated nodes connected to the observation nodes. These obtained associated nodes are then used as the starting node set for the second layer. Through this mapping process, observation nodes involved in structural conflict are associated with their corresponding data generation source nodes, thereby achieving the tracing and association of the source path of the target testing observation data.

[0171] This application eliminates the traditional fixed node type search sequence and innovatively introduces connection type identification rules dynamically generated based on conflict severity identifiers. This is because conflicts of different severity often point to different physical sources. In the second layer, associated nodes have different classification attributes, such as equipment nodes, quality control nodes, reagent nodes, and operator nodes. The system parses the conflict severity identifiers generated in step S1 and dynamically extracts the legal connection types that match them. For example, for fatal conflicts, the connection type identification rules will force the path search to prioritize the direction of strong physical associations such as equipment-quality control products; while for minor conflicts, it may prioritize the direction of reagent batch-operator. This dynamic guidance based on semantic rules avoids meaninglessly traversing all paths in the huge second-layer network topology, thereby significantly improving search efficiency and the clinical relevance of the results.

[0172] Connection type identification rules are a pre-established set of association type mapping rules used to establish the correspondence between different conflict level identifiers and the types of legal association connections in the two-layer heterogeneous knowledge graph. Connection type identification rules are used to limit the path traversal direction and the types of association connections allowed in the second-layer network topology. In the medical testing scenario, various association connections are established between related nodes in the second layer, with different association connections corresponding to different connection type identifiers. Connection type identifiers include detection execution associations, calibration dependency associations, batch associations, and operation associations. Specifically, detection execution associations represent the execution association between observation data and testing equipment; calibration dependency associations represent the calibration association between observation data and quality control nodes; batch associations represent the batch association between observation data and reagent nodes; and operation associations represent the operation association between observation data and operator nodes. The system extracts the corresponding legal connection types from the connection type identification rules based on the conflict level identifier corresponding to the current structural conflict state, and filters the association connections in the second layer based on the legal connection types, thereby limiting the range of related nodes allowed during path traversal.

[0173] The system dynamically extracts a list of legal connection types based on the conflict level identifier corresponding to the current structural conflict state. When the conflict level identifier corresponds to a fatal conflict, the extracted legal connection types include detection execution associations and calibration dependency associations, prioritizing the search for associated paths to device nodes and quality control nodes. When the conflict level identifier corresponds to a minor conflict, the extracted legal connection types include operation associations, searching for associated paths to operator nodes. During path traversal, the system reads the connection type identifiers corresponding to the association connections between the current node and its neighboring nodes and matches and verifies these identifiers against the current list of legal connection types. If the connection type identifier exists in the list of legal connection types, the current association connection is deemed to meet the traversal conditions, and path extension to the corresponding neighboring node is allowed. If the connection type identifier does not exist in the list of legal connection types, the current association connection is deemed not to meet the traversal conditions, and further extension of the current path branch is terminated. By dynamically adjusting the range of legal connection types based on the conflict level identifier, the system can perform targeted searches for corresponding associated nodes according to different structural conflict states, thereby reducing the invalid traversal range in the second-layer network topology and improving the efficiency of associated path search.

[0174] The preset clinical significance threshold is a lower limit used to measure the degree of influence of a change in the state of an auxiliary node on the final test value. In specific embodiments of this application, this threshold cannot be an absolutely rigid fixed value, but rather a range set according to the biological variation patterns of the test items and the sensitivity of the reagents. For example, it can be set to 5% of the coefficient of variation or a normalized influence weight of 0.1. In medical testing scenarios, the second-layer association network in a two-layer heterogeneous knowledge graph usually has high complexity. Although some edge nodes are related to the current test items, their influence on the final test results is extremely low. For example, if a batch of cleaning solution has a slight concentration deviation or the auxiliary sample loading robot arm has a slight displacement, even if the node is further traced or its state is corrected, it is difficult to substantially improve the current conflicting results. Therefore, when traversing the second-layer association nodes, the system calculates the degree of influence corresponding to each candidate node and compares it with the preset clinical significance threshold. When the impact level is below a preset clinical significance threshold, the system determines that the node does not have effective conflict resolution value and terminates further expansion of the path corresponding to that node. When the impact level reaches or exceeds the preset clinical significance threshold, the node is retained as a valid candidate node for subsequent analysis. Through this method, the system can eliminate a large number of low-value associated paths in advance during the search phase, allowing subsequent conflict localization and recalculation processes to focus on highly relevant core nodes, thereby reducing graph traversal complexity and improving anomaly localization efficiency.

[0175] In the second layer of the two-layer heterogeneous knowledge graph, each associated node is assigned different node classification attributes based on its physical entity or business role. These attributes include equipment nodes, quality control nodes, reagent nodes, and operator nodes. Traditional knowledge graph path search typically uses a fixed, preset sequence of node types for constraint, requiring the search path to expand according to a pre-defined node type order. For example, the path must pass through equipment nodes, quality control nodes, and reagent nodes in sequence. While this method can limit the search scope, due to the complex origins, dynamic changes in the propagation chain, and concurrent coupling of multiple factors in medical testing anomalies, a fixed sequence of node types is difficult to adapt to the actual traceability needs under different conflict scenarios, easily leading to rigid search paths and generating a large number of invalid traversals unrelated to the current conflict. Therefore, this application eliminates this fixed, preset sequence of node types and innovatively upgrades it to dynamically calling the corresponding connection type identification rules based on conflict level identifiers. Connection type identification rules are used to define the types of legal associated connections that are allowed to continue traversing under the current conflict level. The associated connection type is a semantic label corresponding to the associated edge, including at least "detection performed on", "calibration depends on", "batch associated", and "operation associated with". During graph traversal, the system does not require specifying the type of node to be traversed. Instead, it reads the connection type field corresponding to the edge between the current node and its neighbors and matches this field with the set of legal connection types corresponding to the current conflict level. If a match is successful, the system allows further expansion to the corresponding neighboring node; otherwise, the search for that path is terminated. For example, when the current conflict level corresponds to a critical conflict, the system prioritizes traversing connections of the "detection performed on" and "calibration dependent on" types, quickly focusing the search on equipment and quality control nodes. When the current conflict level corresponds to a minor conflict, it further allows traversing connections of the "operation associated with" type, enabling traceability analysis of operator nodes. By using dynamic semantic connection constraints instead of fixed node type sequence constraints, the search path can adaptively adjust according to the current conflict level. This not only improves the semantic rationality of graph traversal but also reduces invalid expansion in complex second-layer topologies, improving the efficiency of anomaly source localization and the accuracy of clinical analysis.

[0176] In the second layer, the nodes are connected through association relationships, which are directed edges with connection type identifiers, representing the semantic association structure between the nodes. During traversal, the system starts with the current node, obtains its directly connected neighboring nodes, and reads the connection type identifiers corresponding to the association relationships between the current node and its neighboring nodes. Then, the connection type identifiers are matched against the valid connection types defined in the connection type identifier rules. If the connection type identifier meets the valid connection type conditions, further expansion to the corresponding neighboring nodes is allowed; if the connection type identifier does not meet the valid connection type conditions, the traversal of the current path is terminated. In the above verification process, the system further introduces a path pruning mechanism to reduce the ineffective search overhead in the complex association network of the second layer. Specifically, the system extracts the influence weight attribute value corresponding to the neighboring node. The influence weight attribute value is used to characterize the degree of influence of the neighboring node's state change on the final detection value. This influence weight attribute value can be set based on historical statistical results, historical anomaly backtracking results, or expert experience rules. Subsequently, the system compares the influence weight attribute value with a preset clinical significance threshold. When the influence weight attribute value is less than the preset clinical significance threshold, it indicates that although the adjacent node is related to the current conflict, its actual impact on the final test result is low and it does not have effective clinical intervention value. For example, if a reagent batch has a slight concentration deviation, but this deviation is not enough to cause a significant abnormality in the current test result, the system determines the corresponding adjacent node as an invalid node and skips it, while simultaneously blocking the downward expansion of the current search branch. When the influence weight attribute value reaches or exceeds the preset clinical significance threshold, the corresponding adjacent node is retained as a valid candidate node, and subsequent graph traversal operations continue. Through the above path pruning process, the system can preemptively eliminate weakly related nodes and low-value search paths in the second-layer association network, avoiding large-scale invalid expansion of the graph database during traversal, thereby concentrating computational resources on highly relevant core nodes, improving the efficiency of anomaly source localization and overall graph search efficiency.

[0177] Candidate paths are node sequences formed by hop-by-hop expansion in the second layer, starting from the second layer's initial node and based on connection type identification rules. The constraint path set is a set of paths in the candidate path set that can simultaneously cover all nodes in the second layer's initial node set after connectivity determination, or a connected path structure obtained by topological fusion of a set of paths that meet the coverage conditions.

[0178] The terminal node is the last node in the candidate path in the search order. When the current node does not have any expandable adjacent nodes that meet the connection type identification rules and whose influence weight is not lower than the clinical significance threshold, it is determined to be the terminal node.

[0179] Path convergence degree measures the hub status of a node within a set of constrained paths; its calculation is based on the node's position. The proportion of paths appearing in the constraint path set to the total number of constraint paths is calculated using the following formula:

[0180] ;

[0181] in, For path intersection degree, For containing nodes The number of constraint paths, To constrain the total number of paths;

[0182] Historical failure rate is used to characterize the frequency of anomalies of the entity corresponding to a node within a preset time window. It can be calculated by dividing the number of failures by the number of runs or directly using the historical failure rate recorded in the node attribute field.

[0183] The influence weight is used to characterize the contribution of node state changes to the deviation of observed values. Its value can be configured by historical anomaly backtracking statistics or expert rules and stored as a node attribute field. Pruning is performed when the influence weight is preset to a clinical significance threshold.

[0184] After extracting the set of constrained paths, the system abandons the traditional method of making decisions based solely on the intersection degree of a single path, and instead constructs a more scientific multi-dimensional comprehensive intervention scoring model. Specifically, the system extracts three key dimension parameters: First, path intersection degree, which is the proportion of times a node is traversed in all constrained paths, representing its pivotal position in the topology; second, historical failure rate attribute value, representing the frequency of failures of the node within a historical period, with its reciprocal used to characterize the health and stability of the equipment or as a penalty coefficient; and third, influence weight attribute value, representing the absolute influence of the node on data conflicts. The system assigns preset weight coefficients to these three dimensions and performs a linear weighted summation calculation to obtain the final comprehensive intervention score, the formula of which is:

[0185] ;

[0186] in, For comprehensive intervention scoring, The first preset weighting coefficient, This is the second preset weighting coefficient. The third preset weighting coefficient, For path intersection degree, This refers to the historical failure rate attribute value. To ensure the denominator is a very small constant, avoid setting the denominator to zero. To influence the weighted attribute values;

[0187] A preset secondary threshold is a baseline score used to screen alternative intervention targets. In a specific embodiment of this application, this threshold can be set as a specific percentage of the highest overall intervention score, such as 70% to 85% of the highest score, or as an absolute lower limit of the normalized score, such as greater than 0.6. In real-world testing scenarios, the root causes of data conflicts are often hidden. If only the highest-scoring primary node is used for intervention, and this primary node is not the real source of failure—for example, the device has the highest score but is not actually damaged, and the reagent with the second-highest score has completely failed—simply relying on the highest-scoring node for recalculation will result in the conflict never being resolved, and the system will be stuck in a dead end. By setting this secondary threshold, the system filters out low-scoring, low-success-rate garbage nodes, accurately retaining the candidate set of secondary core nodes with high intervention potential. When the primary node intervention fails, the system can quickly and reliably switch to secondary targets for a new round of testing, greatly improving the success rate and robustness of the system's fully automated conflict resolution.

[0188] After scoring, the system sorts the nodes in descending order of scores and precisely identifies the node with the highest comprehensive intervention score as the primary intervention node, serving as the preferred intervention entry point. More importantly, the system retains other high-scoring nodes with scores above a preset secondary threshold, forming a candidate set of intervention nodes. This candidate set design provides ample backup targets for subsequent responses to single intervention failures, forming the core foundation for achieving multi-round iterations and automatic closed-loop resolution.

[0189] Compared with traditional methods that rely solely on a single constraint or intervene directly based on anomalies in observed items, this application combines conflict detection with node tracing in a graph structure. By dynamically extracting connection type identification rules and employing pruning mechanisms, it achieves highly efficient semantic-guided search, ensuring that the searched constraint paths have medical intervention significance while maintaining extremely low system latency. Simultaneously, through multi-dimensional weighted summation and candidate set mechanisms, it completely resolves the shortcomings of past practices such as missing key weakly correlated nodes or selecting multiple invalid high-intersection nodes. This achieves a higher-dimensional approach, moving from blind attempts to multi-dimensional precise targeting and rotation of backup targets, greatly improving the accuracy, controllability, and robustness of data review and conflict resolution.

[0190] In specific implementation, step S23 includes:

[0191] The candidate paths in the candidate path set are aggregated according to their end nodes. The end node identifiers of each candidate path in the candidate path set are extracted. The candidate paths in the candidate path set are classified according to their end node identifiers. Candidate paths with the same end node identifiers are grouped into the same group to obtain a path group set. Each group in the path group set is numbered and labeled, and the candidate paths in each group are determined as the corresponding path aggregation set.

[0192] The end node is the last node in the candidate path according to the search order;

[0193] For each path aggregation set, extract the starting node corresponding to each path in the path aggregation set to form a subset of covering nodes;

[0194] Perform a coverage determination between the covered subset of nodes and the second-level starting node set;

[0195] When the subset of covered nodes covers all nodes in the set of starting nodes in the second layer, it is determined that the path aggregation set satisfies the preset cross-node connectivity condition.

[0196] Perform topology fusion processing on each path in the path aggregation set that meets the preset cross-node connectivity conditions to generate a connected path structure, and add the connected path structure to the constrained path set.

[0197] Specifically, the search order refers to the sequence in which the path matching operation in step S22 expands outward hop by hop from the second-layer starting node, based on dynamically extracted connection type identifier rules. During each path expansion, the system filters the neighboring nodes of the current node according to the legal connection type and the pruning conditions corresponding to the influencing weight attribute values, adding the neighboring nodes that meet the conditions to the current search path. As the path continues to expand, when there are no next-hop nodes that meet the legal connection type conditions or the preset clinical significance threshold conditions, the system determines that the current path has reached the search boundary of the second-layer association network. At this point, the last node in the current path is determined as the end node. The search order is used to limit the directionality and hierarchy of the path expansion process, ensuring that each candidate path expands in an orderly manner from the second-layer starting node to the bottom-layer associated source nodes, thereby forming a hierarchical tracing path structure for the current data conflict.

[0198] The preset cross-node connectivity condition is used to determine whether a group of paths can jointly connect all nodes in the second-level starting node set to a common ending node. The specific operation is as follows: First, candidate paths are grouped according to their ending node identifiers; all paths corresponding to the same ending node form a path aggregation set. Then, the starting node of each path in this aggregation set is extracted to form a covering node subset. Next, this covering node subset is compared with the second-level starting node set. If the covering node subset contains all nodes in the second-level starting node set, it means that a connectivity structure originating from all starting nodes and ultimately converging to this ending node can be established through this ending node.

[0199] For a set of paths that meet the preset cross-node connectivity conditions, topology fusion processing refers to merging multiple paths with the same end node into a unified connected path structure. Since these paths share the same end node, but may have branches or overlaps between the starting and ending nodes, the topology fusion operation integrates multiple paths into a directed graph structure through node merging and edge deduplication. Specifically, the system traverses each path in the set, merging nodes and edges into a new graph structure according to their sequential order. Nodes and edges shared by multiple paths are retained only once, while branching points retain their branch structures, ultimately generating a connected subgraph that starts from multiple starting nodes, passes through intermediate nodes, and finally converges to the same end node. This connected path structure retains complete path information from all starting and ending nodes while eliminating duplicate nodes and edges, providing a clear graph propagation path for subsequent steps to propagate intervention parameters upstream from the end node.

[0200] In specific implementation, step S3 includes:

[0201] Step S31: Extract the status attribute value of the main associated intervention node, perform status intervention operation on the status attribute value of the main associated intervention node, and set the status attribute value of the main associated intervention node to a preset baseline value.

[0202] In the second-layer network topology, the incoming connection relationship with the main associated intervention node as the endpoint node is extracted, the incoming connection relationship is blocked, and intervention parameters are generated.

[0203] Step S32: Using the intervention parameters as initial input, determine the main associated intervention node as the propagation starting node, and propagate the intervention parameters hop by hop to the first layer along the constraint path. At each hop, adjust the intervention parameters according to the current node's state attribute value and the connection relationship type between the current node and the previous hop node to obtain the updated intervention amount, and propagate it to the previous hop node. Repeat this operation until the intervention amount is propagated to the observation node corresponding to the first layer, and output the intervention propagation amount for the first layer node state update. The processing logic is as follows:

[0204] Mark the main associated intervention node as the propagation start node, and encapsulate the intervention parameters as the initial propagation amount;

[0205] Set the propagation start node to the current node, and assign the initial propagation amount to the current propagation amount;

[0206] Proceed hop-by-hop along the constraint path to the first layer. In each hop, extract the previous hop node that is directly connected to the current node.

[0207] Based on the connection relationship type between the current node and the previous hop node, the current transmission amount is adjusted and calculated to obtain the updated intervention amount;

[0208] Connection types include associative connections and cross-level connections;

[0209] The updated intervention amount is passed to the previous hop node, the previous hop node is reset to the current node, and the updated intervention amount is reset to the current amount passed.

[0210] The operation of extracting the previous hop node and resetting the current node is executed repeatedly until the current node reaches the corresponding observation node in the first layer through the cross-layer connection relationship. Then the operation is stopped and the final current transit quantity is output as the intervention transit quantity.

[0211] Step S33: In response to the intervention transmission amount being transmitted to the first layer via cross-layer connection, the set of observation node items that trigger the structural conflict state is located in the first layer.

[0212] Step S34: Map the intervention transmission amount to each node in the set of located observation nodes, and convert the intervention transmission amount into target adjustment weights acting on each observation node according to the preset conversion rules.

[0213] Step S35: Extract the conflict level identifier and observation item identifier fields from the determination result of the structural conflict state, perform index matching in the preset function mapping table, and extract the corresponding preset numerical recalculation function logic.

[0214] Based on the target adjustment weight, the numerical fields corresponding to the observation item node set in the target inspection observation dataset are substituted into the extracted preset numerical recalculation function logic to perform numerical recalculation processing, obtain the updated observation values, and replace the corresponding original values ​​with the updated observation values ​​to generate the alternative observation dataset.

[0215] Specifically, state intervention refers to forcibly setting the state attribute value of the associated intervention node to a preset baseline value. The preset baseline value typically represents the value of the node under normal or ideal conditions. For example, for a quality control node, the baseline value might be the status code corresponding to passing quality control; for a device node, the baseline value might be within the calibration period or during normal operation. In this embodiment, setting the preset baseline value adjusts the associated intervention node to a state that eliminates conflicts, thereby positively correcting the observed data through subsequent propagation. Blocking the connection relationships pointing to this associated intervention node temporarily shields the input influence of other nodes on the intervention node in the graph structure, ensuring that the intervention effect is not interfered with by other external factors, and keeping the propagation path of the intervention parameters singular and controllable.

[0216] The preset baseline value is the normal state value corresponding to the classification attribute of the second-level associated node. It is pre-configured and stored in the node status attribute field. The preset baseline value of the equipment node can indicate that the calibration is valid or the baseline is stable. The preset baseline value of the quality control node can indicate that the quality control has passed. The preset baseline value of the reagent node can indicate that the batch is valid or within the expiration period. The preset baseline value of the operator node can indicate that the authorization is valid or the operation is compliant.

[0217] The intervention parameters are the propagation inputs formed after performing state intervention on the primary associated intervention node and setting blocking flags on its inbound connections. The intervention parameters include the state change blocking flag and the initial intervention strength. When propagating hop-by-hop back along the constraint path, the intervention quantity update can use proportional propagation or type-dependent propagation rules.

[0218] (1) Proportional propagation: ;

[0219] (2) Type-related propagation: ;

[0220] in, The influence coefficient of the connection relationship. For connection relationship type, For the first The updated intervention amount after the jump. For the first Jump to the corresponding current intervention level. This is a propagation adjustment function related to connection type; when the propagation reaches the first layer through cross-layer connection relationships, the intervention quantity is converted into the weight adjustment parameter on the observation side according to the node type mapping rule, which serves as the source of the target adjustment weight.

[0221] During the hop-by-hop transmission of intervention parameters, the intervention parameters are adjusted and calculated based on the connection type between the current node and the previous hop node. When the connection is an associated connection, the intervention parameters are proportionally adjusted according to a preset association weight. When the connection is a cross-layer connection, the intervention parameters are transformed between layers according to preset mapping rules. The preset mapping rules are a set of rules used to transform intervention parameters between nodes at different levels, describing the transformation method of intervention parameters under cross-layer connections. The preset mapping rules include mapping rules based on node type or transformation rules based on parameter structure. When intervention parameters are transmitted via cross-layer connections, the intervention parameters are mapped and transformed based on the difference in node types between the current node and the target node. The preset association weight is a weight coefficient set within a preset value range, with values ​​in the range of real numbers greater than 0. The preset association weight is set based on the connection type of the associated connection, the association strength between nodes, or historical statistical information. A propagation blocking marker is added to incoming connections with the associated intervention node as the endpoint. The purpose of the propagation blocking marker is to prevent the superposition of influence terms from other incoming connections during the intervention parameter propagation calculation. This involves shielding other incoming influence terms during propagation calculations without deleting graph edges.

[0222] The process of encapsulating intervention parameters into initial propagation quantities involves converting the information generated by the intervention operation into a data packet that can be propagated hop-by-hop in the graph. This information includes state changes, weighting coefficients, or adjustment magnitudes. This data packet typically contains the direction and intensity of the intervention, as well as the identifier of the starting node for propagation. During propagation along the constraint path to the first layer, the current propagation quantity is adjusted based on the connection type between the current node and the previous hop node. For example, if the connection is an influence coefficient, the current propagation quantity is multiplied by the influence coefficient to obtain the updated intervention quantity; if the connection is an additive offset, an additive adjustment is performed. This adjustment ensures that the intervention quantity accurately reflects the physical or logical semantics of the relationships between nodes during propagation. Furthermore, to ensure robustness in the propagation process under complex topologies, this application introduces a fault-tolerant propagation mechanism in hop-by-hop propagation. The system monitors the network topology in real time. If, during a hop, the expected node data is missing or the connection is interrupted, the system automatically triggers intervention parameter completion or constraint path reselection, thereby ensuring that the intervention parameters are delivered to the first layer stably and reliably.

[0223] When the intervention propagation data reaches the first layer via cross-layer connections, the system maps and matches the intervention propagation data with the located set of observation nodes. This mapping and matching is accomplished through node identification: the intervention propagation data always carries the identification information of the target observation node during propagation, or the corresponding observation node is determined based on the endpoint of the propagation path upon reaching the first layer. Subsequently, a one-to-one correspondence is established between the weight information in the propagation data and these observation nodes to form the target adjustment weights. Preset conversion rules are used to convert intervention propagation data from one or more associated intervention nodes into target adjustment weights. These conversion rules include one or more of the following: weighted summation, multiplicative adjustment, priority selection, or threshold filtering.

[0224] After obtaining the target adjustment weights, the system does not use a rigid, single calculation formula. Instead, it innovatively introduces an adaptive matching and historical data self-learning mechanism. First, the system extracts the conflict level identifier from the structural conflict state determination results to characterize the severity and type of the conflict. It then performs joint index matching in a preset function mapping table to adaptively extract the preset numerical recalculation function logic that best fits the conflict scenario. For example, for minor conflicts caused by equipment baseline drift, linear weighting logic or additive offset logic might be matched; for high-level conflicts caused by severe reagent attenuation, proportional adjustment logic or more complex deviation compensation logic might be matched.

[0225] More importantly, before substituting the data into the function for final calculation, the system uses historical intervention results to self-learn and optimize the target adjustment weights. Specifically, the system extracts successful conflict resolution records of the same type as the current conflict classification from a preset historical database, and calculates the average historical weight that actually resolved the conflict in these successful records. Then, it calculates the difference between this average historical weight and the currently extracted weight adjustment parameter, and uses this difference as a self-learning compensation coefficient added to the current weight adjustment parameter. The self-learned optimized weight adjustment parameter and the original values ​​are then substituted into the matching function logic for mathematical operations, ultimately obtaining the updated observation values. While keeping the observation item identification field, time field, and source field unchanged, the replaced target inspection observation dataset is structurally reorganized to obtain the alternative observation dataset.

[0226] The preset function mapping table is shown in Table 1;

[0227] Table 1

[0228]

[0229] The specific mathematical expression of the preset numerical recalculation function logic: After the system adaptively matches the specific function logic through the above table, it will extract the original numerical values ​​from the target test observation dataset. ) and the weight adjustment parameters optimized through graph propagation and self-learning ( Substitute these values ​​into the specific mathematical model below to obtain the updated observed values. ):

[0230] Logic A: Linear weighted / additive offset logic:

[0231] ;

[0232] When the conflict manifests as a constant system deviation, the system directly adds the optimized weight adjustment parameter as an additive constant to the original value. For example, if the fixed mechanical error setting of the sampling needle is -0.2μL, the system uses the optimized weight adjustment parameter as a fixed compensation amount to participate in the numerical correction.

[0233] Logic B: Proportional Adjustment / Multiplicative Correction Logic

[0234] ;

[0235] When the conflict manifests as a proportional system deviation, the system multiplies the original value with the proportional adjustment factor. For example, if a batch of biochemical reagents experiences a 5% decrease in enzyme activity during continuous use, the corresponding target adjustment weight will be calculated. =0.05. Based on this target, the system adjusts the weights to generate a corresponding proportional adjustment factor and performs proportional correction on the original value. Since high-concentration samples usually produce larger numerical deviations under reagent activity decay conditions, the multiplicative correction method is more in line with the numerical change characteristics under proportional error scenarios.

[0236] Logic C: Deviation Compensation Synthesis Logic:

[0237] ;

[0238] For complex conflicts affected by multiple factors, the graph propagation algorithm generates a value that includes a slope adjustment factor. ) and intercept adjustment factor ( The system uses a parameter matrix to simultaneously perform proportional correction on the original data through a linear combination equation and offset correction on the original data using an intercept adjustment factor, thereby jointly adjusting data deviations under complex error scenarios.

[0239] Instead of applying a uniform, fixed correction method to all conflicting data, the system adjusts the target observation data based on the characteristics of different conflict types, calling corresponding data correction models. Specifically, when the conflict exhibits constant systematic bias, an additive correction method is used; when the conflict exhibits proportional systematic bias, a multiplicative correction method is used; and when the conflict is influenced by multiple factors, a linear combination correction method is used. By employing corresponding correction methods for different conflict characteristics, the consistency between the surrogate observation dataset and the actual testing mechanism is improved. The generated surrogate observation dataset not only satisfies the constraint logic but also conforms to the data variation patterns of the corresponding test items, thereby enhancing the clinical reliability of the AI-corrected results.

[0240] The numerical recalculation in this application is a computational process for interpretable correction of the numerical values ​​of observations under the assumption of source intervention, and its substitution is based on at least the following constraints:

[0241] (1) Hypotheses on the sources of deviation of the main related intervention nodes and their state interventions obtained from the second-level source tracing and localization;

[0242] (2) The intervention parameters are backpropagated along the constraint path and mapped to obtain the target adjustment weights;

[0243] (3) Preset numerical recalculation function logic that matches the conflict level identifier.

[0244] After the alternative observation dataset is generated, it needs to be recalculated and verified by substituting the logical expression corresponding to the original constraint connection relationship. Only when the recalculation result meets the preset logical conditions will the automatic processing instruction and classification label be output. When the conflict level label is critical or the preset logical conditions cannot be met within the preset iteration number threshold, the automatic substitution result will not be output. Instead, a rejection or manual intervention warning mark will be output to limit the applicable boundary of the alternative observation dataset and ensure medical rationality.

[0245] Traditional methods often only identify data conflicts but fail to automatically provide correction solutions, or they directly modify observed data without a basis for correction. This application, however, applies intervention operations to the associated intervention nodes in the second layer and utilizes a graph propagation mechanism to transmit the intervention amount backward along the constraint path. This ensures that the final correction of the observed data truly originates from the adjustment of the root cause node of the conflict, and the entire intervention process is supported by a clear propagation path and adjustment logic. This not only improves the interpretability and credibility of data correction but also allows the intervention operations to be reused in similar conflict scenarios, avoiding repetitive manual analysis and processing, and significantly improving the automation level and consistency of verification data review.

[0246] In practice, step S35 includes the following specific operations:

[0247] Parse the set of observation node nodes to obtain the set of node identifiers corresponding to each node;

[0248] Based on the node identifier set, identifier matching is performed in the target inspection observation dataset to extract the numerical field data corresponding to each node identifier, thus obtaining the original numerical set.

[0249] The target adjustment weights are mapped and parsed to establish the correspondence between node identifiers and weight adjustment parameters;

[0250] Extract successful historical conflict resolution records from the preset historical database that are of the same type as the current conflict classification identifier. Extract the average historical weights used in the successful historical conflict resolution records. Calculate the difference between the average historical weights and the current weight adjustment parameters as the self-learning compensation coefficient.

[0251] The self-learning compensation coefficient is added to the current weight adjustment parameter to obtain the optimized weight adjustment parameter.

[0252] The original set of numerical values ​​and the optimized weight adjustment parameters are used as input variables and substituted into the preset numerical recalculation function for calculation to obtain the updated set of observed numerical values. The updated set of observed numerical values ​​is then written into the corresponding numerical field position in the target test observation dataset to replace the original values.

[0253] While keeping the observation identifier field, time field, and source field unchanged, the target inspection observation dataset after replacement is restructured to obtain the alternative observation dataset.

[0254] Specifically, parsing the observation node set refers to extracting the unique identifier of each observation node from the observation node set located in step S3. Since the observation node set has already been generated in a structured form in step S14, containing the identifier information of each observation node involved in the conflict, the parsing process essentially involves traversing this set, extracting the identifier field of each node one by one, forming a node identifier set, which serves as the index basis for subsequent data location and weight matching.

[0255] In the target verification observation dataset, identifier matching is performed by using each identifier in the node identifier set as a matching key to search within the observation item identifier field of the target verification observation dataset. Since the target verification observation dataset is typically in the form of a table or key-value pair list, each record corresponds to an observation item, containing its identifier, value, time, and source fields. The system precisely compares the node identifier with the observation item identifier field. When they match, the corresponding data record row is located, and the value in the value field of that record is then extracted to form the original value set corresponding to the node identifier.

[0256] Mapping and parsing the target adjustment weights and establishing the correspondence between node identifiers and weight adjustment parameters means splitting the intervention transmission quantity obtained in step S34 according to the observation item nodes, extracting the weight part related to each observation item node from the transmission quantity, and forming a mapping table from node identifiers to weight adjustment parameters. This ensures that the value of each observation item can be assigned the correct correction coefficient during subsequent numerical recalculation, avoiding weight mismatch.

[0257] The pre-defined numerical recalculation function is a mathematical function used to generate updated values ​​based on the original values ​​and weight adjustment parameters. It can be predefined according to the type of constraint and application scenario, commonly including linear weighted correction, proportional adjustment, and deviation compensation. For example, for a constraint like "A and B should satisfy a linear relationship," the recalculation function can use the weight adjustment parameter as the adjustment range, calculating it in the form of "updated value = original value × (1 + weight adjustment parameter)" or "updated value = original value + weight adjustment parameter." The purpose of setting this function is to transform abstract intervention weights into actual corrections to specific values, while ensuring that the correction process is repeatable, adjustable, and flexibly configurable according to different constraint types.

[0258] Structured reorganization, while maintaining the observation identifier, time, and source fields unchanged, refers to the system reorganizing the structure of the target inspection observation dataset after writing the updated observation values ​​back to their corresponding numerical field positions. This ensures that each record still contains all four fields: observation identifier, value, time, and source, and that, except for the numerical field, the content of the remaining fields is completely consistent with the original data. The reorganization process typically involves reorganizing the updated records into a table structure, key-value pair list, or dictionary array with the same format as the original data, thereby generating a substitute observation dataset with the same structure as the original dataset, except that the numerical parts have been corrected. This design guarantees that the substitute observation dataset can seamlessly replace the original data for subsequent review and judgment without causing downstream processing logic anomalies due to structural changes.

[0259] In specific implementation, step S4 includes:

[0260] Step S41 involves parsing the field mapping of the alternative observation dataset, extracting the corresponding numerical field data based on the node identifiers in the observation item node set, and generating an updated observation value set. The processing logic is as follows:

[0261] Traverse the set of observation nodes, extract the node identifier of each observation node in the set of observation nodes one by one, and construct and update the set of observation values;

[0262] Using the node identifier as the index key, the observation item identifier field is traversed and mapped in the substitute observation dataset for matching. When the content of the observation item identifier field is completely consistent with the node identifier, the field mapping match is determined to be successful, and the corresponding data record row in the substitute observation dataset is locked.

[0263] Extract the updated observation values ​​stored in the numeric fields of the data record rows;

[0264] The updated observation values ​​are mapped to their corresponding node identifiers to form key-value pairs, and these key-value pairs are added to the updated observation value set until all node identifiers in the observation item node set have been traversed.

[0265] Step S42: Perform structural analysis on the determination result of the structural conflict state, and extract the constraint connection relationship and corresponding logical expression that do not meet the preset logical conditions;

[0266] Step S43: The updated observation value set is mapped to the variable parameters in the logical expression according to the variable correspondence and then subjected to joint operation to obtain the second logical operation value, the calculation formula of which is:

[0267] ;

[0268] in, The value of the second logical operation. For logical evaluation functions, The logical expression corresponding to the constraint connection relationship. For the first in the logical expression One variable parameter, The number of variable parameters, To update the set of observed values ​​and variable parameters Updated observation values ​​with a mapping relationship;

[0269] Step S44: Compare the second logical operation value with the preset logical conditions. When the logical comparison result is that the second logical operation value meets the preset logical conditions, generate a conflict resolution status marker and determine that the structural conflict state of the local associated substructure is resolved based on the conflict resolution status marker.

[0270] Step S45: When the logic comparison result is that the second logic operation value does not meet the preset logic condition, calculate the absolute difference between the updated observed value and the corresponding original value, and divide the absolute difference by the original value to obtain the deviation ratio parameter.

[0271] The deviation ratio parameter is passed back to the second layer, and the intervention parameter of the current main associated intervention node is updated by multiplying it by the preset decay constant, while triggering the preset iteration counter to increment by one.

[0272] Step S46: Determine whether the value of the preset iteration counter is less than the preset iteration number threshold;

[0273] If the value of the preset iteration counter is less than the preset iteration number threshold, then when the intervention parameter of the current main associated intervention node is less than the preset lower limit, the candidate node at the top of the intervention node candidate set is extracted and replaced as the main associated intervention node, and the updated intervention parameter is returned to the execution step S3 to generate a new alternative observation dataset for cyclic calculation.

[0274] If the value of the preset iteration counter is greater than or equal to the preset iteration number threshold, a manual intervention warning mark is generated, and a data processing instruction is generated based on the conflict level identifier in the structural conflict status determination result. The data processing instruction and classification identifier are then output.

[0275] Specifically, constructing the updated observation value set involves extracting conflict-related values ​​from the alternative observation dataset and forming a structured set. The system first iterates through the observation node set obtained in step S14, extracting the node identifier for each observation node as an index key. Then, in the alternative observation dataset, it iterates and maps the observation identifier field, checking each record in the dataset and precisely comparing the record's observation identifier field content with the node identifier. When they match perfectly, the record is locked, and the updated observation value stored in its value field is extracted. The node identifier and the value are paired as a key-value pair and added sequentially to the updated observation value set until all node identifiers are matched. This set is essentially a mapping table from observation nodes to corrected values, providing accurate data input for subsequent logical expressions.

[0276] Structural parsing of the structural conflict determination result refers to decomposing the structured conflict determination result output in step S14 and subsequently retained into its constituent elements. In step S14, this determination result is packaged and encapsulated into a data structure containing a conflict level identifier, a set of observation node nodes, constraint connection relationships, and their corresponding logical expressions, typically existing in the form of a dictionary, a JSON object, or a list of key-value pairs. During parsing, the system extracts the constraint connection relationship field and the corresponding logical expression field according to the predefined data structure fields, for use in step S43.

[0277] When the second logical operation value still does not meet the preset logical conditions, it indicates that the data correction in the current round has failed to resolve the corresponding conflict. At this time, the system calculates the absolute difference between the updated observed value and the original value, and uses this absolute difference divided by the original value to obtain the deviation ratio parameter. The deviation ratio parameter is used to characterize the adjustment magnitude of the observed data in the current round. For example, when the original blood glucose value is 10 and the updated value is 12, the corresponding deviation ratio parameter is 20%. This parameter quantifies the "surgical intensity" of the system on the original data. The system transmits the deviation ratio parameter in reverse to the second layer of the two-layer heterogeneous knowledge graph to adjust the intervention parameter corresponding to the current main association intervention node. When the correction fails, it indicates that the current intervention intensity is too high or there is a deviation between the current main association intervention node and the actual conflict source. Therefore, the system uses a preset attenuation constant to attenuate the current intervention parameter. For example, the current intervention parameter is multiplied by 0.8 to reduce the intervention intensity in subsequent rounds, thereby reducing the occurrence of over-adjustment or numerical oscillation during the data correction process. At the same time, in order to prevent the system from falling into an infinite recalculation loop, this application designs a controlled multi-round iteration and backup target rotation mechanism. The system increments a preset iteration counter and checks if it has reached a preset iteration threshold. If the threshold is not reached, the system checks if the attenuated intervention parameters are below a preset lower limit, for example, if the weight has decayed to near zero. If below the lower limit, it indicates that the current main intervention node cannot resolve the conflict through adjustment. At this time, the system immediately selects the candidate node ranked second in the comprehensive intervention score from the intervention node candidate set pre-generated in step S24, promotes it to the new main associated intervention node, and seamlessly returns to step S3 with the initial intervention parameters to initiate a new round of intervention cycle. This mechanism can achieve multiple rounds of node switching and iterative correction based on the intervention node candidate set, thereby improving the system's automatic handling capability for complex conflict scenarios. When the preset iteration counter reaches or exceeds the preset iteration threshold, it indicates that the corresponding conflict has not been resolved under the current candidate node set. At this time, the system generates a manual intervention warning mark and stops the automatic data correction operation. At the same time, the system outputs the corresponding conflict tracing data and historical iteration records, generates an alarm command, and submits it to the manual review module for subsequent review and processing by a human physician.

[0278] The preset classification mapping table is a lookup table structure used to convert the identifiers and classification attributes of associated intervention nodes into business-recognizable classification identifiers;

[0279] The preset category mapping table is shown in Table 2;

[0280] Table 2

[0281]

[0282] The purpose of this table is to associate the technical nodes in the graph with the business classification system, so that the final output classification labels can be directly understood and used by downstream systems or human reviewers, thereby enhancing the interpretability and business adaptability of the review results.

[0283] The preset iteration counter is an integer variable used in the system to record the number of times the current conflict resolution loop is executed, and its initial value is set to 0. In step S4, when the second logical operation value still does not meet the preset logical condition and triggers the intervention parameter update, the preset iteration counter is automatically incremented by one. The preset iteration counter is used to limit the maximum number of iterations in the automatic intervention process, thereby avoiding the system from continuously performing invalid iterative calculations during repeated corrections.

[0284] In this embodiment, the preset iteration threshold is set to 4 times. Complex clinical test data conflicts may be caused by multiple concurrent factors. If artificial intelligence is given unlimited trial and error privileges, the system is highly likely to forcibly piece together a set of fabricated data that is mathematically perfectly self-consistent but biologically and physicochemically absurd by drastically distorting various weight parameters. Limiting this threshold to a very small number (such as 3 times) means that the system can only try the 2-3 most important intervention target points. Once the limit is exceeded, the system immediately determines that the conflict exceeds the safe control range of the automated algorithm, decisively triggers a manual intervention warning, immediately freezes the calculation, and generates an alarm work order.

[0285] The preset lower limit refers to the minimum weight threshold used to determine whether the current primary intervention node still has effective intervention capabilities. As the intervention parameter undergoes multiple rounds of attenuation processing, its value gradually decreases. In a specific embodiment of this application, the preset lower limit can be set as a floating-point constant close to zero, such as 0.01. When the intervention parameter is less than the preset lower limit, it indicates that the current primary intervention node has a weak regulatory effect on the current conflict, and continuing to adjust parameters based on this primary intervention node is unlikely to resolve the corresponding conflict. At this time, the system stops the parameter adjustment process based on the current primary intervention node and selects other candidate nodes with higher comprehensive intervention scores from the intervention node candidate set as new primary intervention nodes. Simultaneously, the corresponding intervention parameters are reset, and the intervention processing in step S3 is re-executed. By setting a preset lower limit, the system can dynamically switch between different intervention nodes during multiple iterations, thereby reducing invalid iteration calculations and improving the automatic conflict resolution capability in complex conflict scenarios.

[0286] Compared to traditional methods, conventional data review often only outputs anomaly alerts after a conflict is discovered, requiring manual intervention for judgment and correction. This application, however, automates the verification of conflict resolution results by substituting alternative observation datasets into logical expressions for secondary calculations and automatically determining whether the conflict has been resolved based on a logical evaluation function. Furthermore, by extracting classification identifiers from associated intervention nodes and combining them with conflict grading identifiers to generate standardized data processing instructions, the system can output instructions and classification results that can be directly executed by downstream automated processes. This transforms the entire review process from "conflict discovery" to an integrated closed loop of "automatic intervention, verification, and output," significantly improving the automation level and operability of data review results.

[0287] In specific implementation, step S46 includes:

[0288] Analyze the results of the structural conflict status determination, and extract conflict classification indicators and manual intervention warning markers;

[0289] Extract the source and time fields from the target inspection observation dataset;

[0290] Extract the state operation code corresponding to the conflict level identifier from the preset state mapping table;

[0291] When a manual intervention warning marker is generated, extract the status operation code that is jointly corresponding to the conflict classification identifier and the manual intervention warning marker.

[0292] The source field, time field, and status operation code are concatenated to generate a data processing instruction;

[0293] Convert data processing instructions, category identifiers, and node identifiers into key-value pairs in a preset format;

[0294] The key-value pairs are combined to generate a dictionary data structure, and the dictionary data structure is written into a preset output variable field.

[0295] Specifically, parsing the structural conflict status determination result refers to the field decomposition of the structured conflict determination data generated in step S14 and passed through subsequent processes. In step S14, this determination result is encapsulated into a data structure containing a conflict level identifier, a set of observation node values, constraint connection relationships, and corresponding logical expressions, typically existing in the form of a dictionary, JSON object, or key-value pair list. During parsing, the system directly extracts the value of the conflict level identifier field according to the predefined data structure field names, serving as the core basis for generating data processing instructions. Simultaneously, the system checks whether a manual intervention warning flag has been generated in the multi-round iteration mechanism of step S4, such as a boolean value FLAG_MANUAL_WARN=True.

[0296] The default state mapping table is a lookup table used to convert conflict classification identifiers into standardized state operation codes.

[0297] The preset state mapping table is shown in Table 3;

[0298] Table 3

[0299]

[0300] The purpose of this mapping table is to transform internal conflict identifiers into standardized operation instructions that downstream systems or execution modules can recognize, ensuring that data processing instructions have good universality and scalability. This application innovatively introduces a joint mapping logic, as shown in the last row of Table 3. When a minor or critical conflict that could have been resolved through recalculation generates a manual intervention warning flag due to multiple iterations timeouts caused by the complexity of its physical source, the system no longer outputs a normal success instruction. Instead, it outputs a unique timeout alarm operation code (OP_ITER_TIMEOUT) through joint key-value matching. This mechanism realizes the state distinction between the automatic conflict resolution process and the manual intervention process, and improves the system's information connection capability between the automatic processing stage and the manual review stage.

[0301] A preset format refers to a standardized data structure representation agreed upon within the system, commonly including JSON or key-value pair lists. When converting data processing instructions, category identifiers, and node identifiers into key-value pairs in a preset format, the system organizes these three types of information as different key-value entries. For example, in JSON format, the generated structure might be "{"data_instruction":"Laboratory A_202503271430_OP_ITER_TIMEOUT","classify_id":"REAG_GLU","node_id":"node_001","manual_warn":true}". This formatted key-value pair data allows subsequent modules to directly read data by field without further parsing, ensuring consistency and readability of data when transmitted between different modules within the system.

[0302] The pre-defined output variable field is a designated storage location reserved by the system in memory or data exchange structure to store the final output result. This field is typically defined as a structured variable, such as a dictionary object or a custom data object. After generating key-value pairs, the system assigns the entire key-value pair to this output variable field for use by upper-level calls or downstream processes. By standardizing the output variable field, the output interface of the entire auditing method is standardized, facilitating integration with other systems.

[0303] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. 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.

[0304] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention. All data acquisition actions in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located and with the authorization granted by the owner of the corresponding device.

Claims

1. A method for verifying inspection data based on knowledge graphs, characterized in that, Includes the following steps: Step S1: Map the target inspection observation dataset to the local related substructure of the two-layer heterogeneous knowledge graph, substitute it into the logical expression for calculation. If the preset logical conditions are not met and the preset exemption conditions are not triggered, calculate the deviation degree to generate a conflict level status identifier and determine that the local related substructure is in a structural conflict state. The two-layer heterogeneous knowledge graph includes a first layer and a second layer; The first layer represents the constraint connection relationship between observation item nodes, the second layer represents the associated nodes that generate observation item data, and a cross-layer connection relationship is established between the first layer and the second layer. Step S2: Map the observation node involved in the structural conflict state from the first layer to the second layer, extract constraint paths based on the conflict level state identifier and calculate the comprehensive intervention score, and extract the main associated intervention node and intervention node candidate set; Step S3: Generate intervention parameters for the main associated intervention node and pass them back to the first layer. Based on the conflict level status identifier, match the preset numerical recalculation function logic, and use the target adjustment weight to perform numerical recalculation processing to generate an alternative observation dataset. Step S4: Substitute the alternative observation dataset into the logical expression for calculation. When the calculation result meets the preset logical conditions, output the data processing instruction and classification label. When the calculation result does not meet the preset logical conditions, attenuate the intervention parameter based on the deviation ratio parameter and call the intervention node candidate set for iterative calculation until the conflict is resolved or the preset iteration number threshold is reached, output the manual intervention warning mark.

2. The knowledge graph-based verification data review method as described in claim 1, characterized in that, Step S1 specifically includes: Step S11: Obtain the target inspection observation dataset, parse the observation item identifier field in the target inspection observation dataset, and map each observation item identifier field to the corresponding observation item node in the first layer to obtain the mapped observation item node; The target test observation dataset includes observation item identifier field, numerical field, time field, source field, and clinical background feature field; Step S12: Construct a two-layer heterogeneous knowledge graph; Step S13: Using the mapped observation node as the starting node, extract the local related substructure in the two-layer heterogeneous knowledge graph, extract the constraint connection relationship connecting two or more observation nodes from the local related substructure, and obtain the logical expression corresponding to the constraint connection relationship. Step S14: Substitute the numerical fields from the target inspection observation dataset into the logical expression for calculation. When the calculation result does not meet the preset logical conditions, the time field and the clinical background feature field are substituted for secondary verification calculation. When the state attribute of the associated node in the local associated substructure cannot meet the logical expression and the secondary verification calculation result does not trigger the preset exemption condition, the judgment result of the structural conflict state is output. The determination result of the structural conflict state includes a conflict level identifier, a set of observation node nodes that cause the operation result to not meet the preset logical conditions, and constraint connection relationships and corresponding logical expressions that do not meet the preset logical conditions.

3. The knowledge graph-based verification data review method as described in claim 2, characterized in that, Step S14 specifically includes: Extract the data values ​​of the numerical fields corresponding to the mapped observation node in the target test observation dataset, substitute the data values ​​into the variable parameters in the logical expression for joint operation, and obtain the first logical operation value; The logical expression includes a numerical joint operation function and a comparison judgment relationship, and the first logical operation value is the operation result of the numerical joint operation function. When the first logical operation value does not meet the preset logical conditions, the time field is extracted and the historical value set corresponding to the observation item node is obtained from the preset historical database. The time change rate between the current data value and the historical value set is calculated, and the clinical background feature field is extracted and text matching is performed in the preset pathological feature dictionary. The formula for calculating the rate of change over time is: ； in, The rate of change over time, For the current observation value, This is the most recent historical observation value. It should be a very small constant to avoid the denominator being zero; When the rate of change over time is within the preset safe range of historical fluctuations or the text matching result of the clinical background feature field falls within the legal range of the preset pathological feature dictionary, a conflict exemption flag is generated and the current review and judgment operation is terminated. When no conflict exemption flag is generated, the associated nodes connected to the constraint connection relationship that does not meet the preset logical conditions are traversed in the local associated substructure, and the state attribute values ​​of the associated nodes are obtained. The state attribute value is used as a supplementary condition to participate in the secondary judgment of the corresponding logical expression. When the state attribute values ​​of all associated nodes in the local associated substructure cannot make the result of the secondary operation judgment meet the preset logical condition, the absolute value of the difference between the first logical operation value and the standard value corresponding to the preset logical condition is calculated as the deviation degree. The deviation degree is compared with the preset classification threshold, and a conflict classification label is generated based on the comparison result. The preset grading thresholds include a first threshold and a second threshold; when When a minor conflict occurs, a conflict level identifier, Level_1_Minor, is generated. when At that time, a conflict classification identifier Level_2_Critical corresponding to the critical conflict is generated; when At that time, a conflict level identifier Level_3_Fatal corresponding to the fatal conflict is generated; in, The first threshold, The second threshold, and ; Extract the mapped observation nodes corresponding to the variable parameters whose first logical operation value does not meet the preset logical conditions, combine them to obtain the observation node set, package and encapsulate the conflict level identifier, the observation node set, the constraint connection relationship and the corresponding logical expression, and output the judgment result of the structural conflict state.

4. The knowledge graph-based verification data review method as described in claim 3, characterized in that, Step S2 specifically includes: Step S21: Based on the set of observation item nodes and the cross-layer connection relationship, map the set of observation item nodes from the first layer to the second layer to obtain the starting node set of the second layer; Step S22: Starting from each node in the second-layer initial node set, extract the corresponding connection type identifier rules based on the conflict level identifier in the second layer and perform path matching and pruning operations to obtain a candidate path set. The processing logic includes: Extract the connection type identifier rules corresponding to the conflict level identifier, and construct the initial search path starting from each node in the second-level starting node set. In the second-layer network topology, traverse the neighboring nodes connected to the current node and verify whether the node classification attributes of the neighboring nodes and their association connection relationship with the current node conform to the legal connection type defined by the connection type identification rules. If the verification passes, the influence weight attribute value of the adjacent node is extracted. When the influence weight attribute value is less than the preset clinical significance threshold, the adjacent node is pruned and removed. When the value of the influence weight attribute is greater than or equal to the preset clinical significance threshold, the adjacent node is added to the current search path, and the adjacent node is used as the current node to continue the traversal and verification operation until the node classification attribute arrangement of the current search path is completely consistent with the preset node type sequence. The completely consistent current search path is extracted and saved to the candidate path set. If the verification fails, the adjacent node is determined to be an invalid node. The adjacent node is skipped and the downward expansion of the current search branch is blocked. The current node is returned to continue to traverse the other adjacent nodes connected to it for verification operations until the current search path can no longer be expanded and a candidate path is formed. The candidate path is then extracted and saved to the candidate path set. Step S23: Perform connectivity determination on each candidate path in the candidate path set, and select candidate paths that simultaneously connect all nodes in the second-level starting node set as the constraint path set. Step S24: Calculate the path intersection degree, historical failure rate attribute value, and influence weight attribute value for each node in the constraint path set. Preset weight coefficients are assigned to the path convergence degree, the reciprocal of the historical failure rate attribute value, and the influence weight attribute value, and then a weighted sum is performed to obtain the comprehensive intervention score for each node. The calculation formula is as follows: ； in, For comprehensive intervention scoring, The first preset weighting coefficient, This is the second preset weighting coefficient. The third preset weighting coefficient, For path intersection degree, This refers to the historical failure rate attribute value. To ensure the denominator is a very small constant, avoid setting the denominator to zero. To influence the weighted attribute values; The node with the highest comprehensive intervention score is identified as the primary intervention node, and the remaining nodes with comprehensive intervention scores greater than a preset secondary threshold are arranged in descending order of score to form a candidate set of intervention nodes.

5. The knowledge graph-based verification data review method as described in claim 4, characterized in that, Step S23 specifically includes: The candidate paths in the candidate path set are aggregated according to their end nodes. The end node identifiers of each candidate path in the candidate path set are extracted. The candidate paths in the candidate path set are classified according to their end node identifiers. Candidate paths with the same end node identifiers are grouped into the same group to obtain a path group set. Each group in the path group set is numbered and labeled, and the candidate paths in each group are determined as the corresponding path aggregation set. The end node is the last node in the candidate path according to the search order; For each path aggregation set, extract the starting node corresponding to each path in the path aggregation set to form a subset of covering nodes; Perform a coverage determination between the covered subset of nodes and the second-level starting node set; The coverage determination is that the subset of covered nodes contains all nodes in the set of starting nodes of the second layer; When the subset of covered nodes covers all nodes in the set of starting nodes in the second layer, it is determined that the path aggregation set satisfies the preset cross-node connectivity condition. Perform topology fusion processing on each path in the path aggregation set that meets the preset cross-node connectivity conditions to generate a connected path structure, and add the connected path structure to the constrained path set.

6. The knowledge graph-based verification data review method as described in claim 5, characterized in that, Step S3 specifically includes: Step S31: Extract the status attribute value of the main associated intervention node, perform status intervention operation on the status attribute value of the main associated intervention node, and set the status attribute value of the main associated intervention node to a preset baseline value. The preset baseline value is the normal state value pre-stored in the state attribute field of the main associated intervention node; In the second-layer network topology, the incoming connection relationship with the main associated intervention node as the endpoint node is extracted, the incoming connection relationship is blocked, and intervention parameters are generated. Intervention parameters include the state change amount blocking marker and the initial intervention intensity; Step S32: Using the intervention parameters as initial input, determine the main associated intervention node as the propagation starting node, and propagate the intervention parameters hop by hop to the first layer along the constraint path. At each hop, adjust the intervention parameters according to the current node's state attribute value and the connection relationship type between the current node and the previous hop node to obtain the updated intervention amount, and propagate it to the previous hop node. Repeat this operation until the intervention amount is propagated to the observation node corresponding to the first layer, and output the intervention propagation amount for the first layer node state update. The processing logic is as follows: Mark the main associated intervention node as the propagation start node, and encapsulate the intervention parameters as the initial propagation amount; Set the propagation start node to the current node, and assign the initial propagation amount to the current propagation amount; Proceed hop-by-hop along the constraint path to the first layer. In each hop, extract the previous hop node that is directly connected to the current node. Based on the connection relationship type between the current node and the previous hop node, the current transmission amount is adjusted and calculated to obtain the updated intervention amount; The connection types include associated connection relationships and cross-layer connection relationships; The updated intervention amount is passed to the previous hop node, the previous hop node is reset to the current node, and the updated intervention amount is reset to the current amount passed. The operation of extracting the previous hop node and resetting the current node is executed repeatedly until the current node reaches the corresponding observation node in the first layer through the cross-layer connection relationship. Then the operation is stopped and the final current transit quantity is output as the intervention transit quantity. Step S33: In response to the intervention transmission amount being transmitted to the first layer via cross-layer connection, the set of observation node items that trigger the structural conflict state is located in the first layer. Step S34: Map the intervention transmission amount to each node in the set of located observation nodes, and convert the intervention transmission amount into target adjustment weights acting on each observation node according to the preset conversion rules. Step S35: Extract the conflict level identifier and observation item identifier fields from the determination result of the structural conflict state, perform index matching in the preset function mapping table, and extract the corresponding preset numerical recalculation function logic. Based on the target adjustment weight, the numerical fields corresponding to the observation item node set in the target inspection observation dataset are substituted into the extracted preset numerical recalculation function logic to perform numerical recalculation processing, obtain the updated observation values, and replace the corresponding original values ​​with the updated observation values ​​to generate the alternative observation dataset.

7. The knowledge graph-based verification data review method as described in claim 6, characterized in that, Step S35 specifically includes the following operations: Parse the set of observation node nodes to obtain the set of node identifiers corresponding to each node; Based on the node identifier set, identifier matching is performed in the target inspection observation dataset to extract the numerical field data corresponding to each node identifier, thus obtaining the original numerical set. The target adjustment weights are mapped and parsed to establish the correspondence between node identifiers and weight adjustment parameters; Extract successful historical conflict resolution records from the preset historical database that are of the same type as the current conflict classification identifier. Extract the average historical weights used in the successful historical conflict resolution records. Calculate the difference between the average historical weights and the current weight adjustment parameters as the self-learning compensation coefficient. The self-learning compensation coefficient is added to the current weight adjustment parameter to obtain the optimized weight adjustment parameter. The original set of numerical values ​​and the optimized weight adjustment parameters are used as input variables and substituted into the preset numerical recalculation function for calculation to obtain the updated set of observed numerical values. The updated set of observed numerical values ​​is then written into the corresponding numerical field position in the target test observation dataset to replace the original values. While keeping the observation identifier field, time field, and source field unchanged, the target inspection observation dataset after replacement is restructured to obtain the alternative observation dataset.

8. The knowledge graph-based verification data review method as described in claim 7, characterized in that, Step S4 specifically includes: Step S41 involves parsing the field mapping of the alternative observation dataset, extracting the corresponding numerical field data based on the node identifiers in the observation item node set, and generating an updated observation value set. The processing logic is as follows: Traverse the set of observation nodes, extract the node identifier of each observation node in the set of observation nodes one by one, and construct and update the set of observation values; Using the node identifier as the index key, the observation item identifier field is traversed and mapped in the substitute observation dataset for matching. When the content of the observation item identifier field is completely consistent with the node identifier, the field mapping match is determined to be successful, and the corresponding data record row in the substitute observation dataset is locked. Extract the updated observation values ​​stored in the numeric fields of the data record rows; The updated observation values ​​are mapped to their corresponding node identifiers to form key-value pairs, and these key-value pairs are added to the updated observation value set until all node identifiers in the observation item node set have been traversed. Step S42: Perform structural analysis on the determination result of the structural conflict state, and extract the constraint connection relationship and corresponding logical expression that do not meet the preset logical conditions; Step S43: The updated set of observed values ​​is mapped to the variable parameters in the logical expression according to the variable correspondence and then subjected to joint operation to obtain the second logical operation value; Step S44: Compare the second logical operation value with the preset logical conditions. When the logical comparison result is that the second logical operation value meets the preset logical conditions, generate a conflict resolution status marker and determine that the structural conflict state of the local associated substructure is resolved based on the conflict resolution status marker. Step S45: When the logic comparison result is that the second logic operation value does not meet the preset logic condition, calculate the absolute difference between the updated observed value and the corresponding original value, and divide the absolute difference by the original value to obtain the deviation ratio parameter. The deviation ratio parameter is passed back to the second layer, and the intervention parameter of the current main associated intervention node is updated by multiplying it by the preset decay constant, while triggering the preset iteration counter to increment by one. Step S46: Determine whether the value of the preset iteration counter is less than the preset iteration number threshold; If the value of the preset iteration counter is less than the preset iteration number threshold, then when the intervention parameter of the current main associated intervention node is less than the preset lower limit, the candidate node at the top of the intervention node candidate set is extracted and replaced as the main associated intervention node, and the updated intervention parameter is returned to the execution step S3 to generate a new alternative observation dataset for cyclic calculation. If the value of the preset iteration counter is greater than or equal to the preset iteration number threshold, a manual intervention warning mark is generated, and a data processing instruction is generated based on the conflict level identifier in the structural conflict status determination result. The data processing instruction and classification identifier are then output.

9. The knowledge graph-based verification data review method as described in claim 8, characterized in that, Step S46 specifically includes: Analyze the results of the structural conflict status determination, and extract conflict classification indicators and manual intervention warning markers; Extract the source and time fields from the target inspection observation dataset; Extract the state operation code corresponding to the conflict level identifier from the preset state mapping table; When a manual intervention warning marker is generated, extract the status operation code that is jointly corresponding to the conflict level identifier and the manual intervention warning marker; The source field, time field, and status operation code are concatenated to generate a data processing instruction; Convert data processing instructions, category identifiers, and node identifiers into key-value pairs in a preset format; The key-value pairs are combined to generate a dictionary data structure, and the dictionary data structure is written into a preset output variable field.

10. A knowledge graph-based verification data review system, applied in a knowledge graph-based verification data review method as described in any one of claims 1-9, characterized in that, It includes a conflict determination module, an intervention and positioning module, a data reconstruction module, and an instruction output module; The conflict determination module is used to map the target inspection observation dataset to the local related substructure of the two-layer heterogeneous knowledge graph, substitute it into the logical expression for calculation, and if the preset logical conditions are not met and the preset exemption conditions are not triggered, calculate the deviation degree to generate a conflict level status identifier, and determine that the local related substructure is in a structural conflict state. The two-layer heterogeneous knowledge graph includes a first layer and a second layer; The first layer represents the constraint connection relationship between observation item nodes, the second layer represents the associated nodes that generate observation item data, and a cross-layer connection relationship is established between the first layer and the second layer. The intervention positioning module is used to map the observation item nodes involved in the structural conflict state from the first layer to the second layer, extract constraint paths based on the conflict level state identifier, calculate the comprehensive intervention score, and extract the main associated intervention node and intervention node candidate set. The data reconstruction module is used to generate intervention parameters for the main associated intervention node and back-transmit them to the first layer. Based on the conflict level status identifier, it matches the preset numerical recalculation function logic, uses the target adjustment weight to perform numerical recalculation processing, and generates an alternative observation dataset. The instruction output module is used to substitute the alternative observation dataset into the logical expression for calculation. When the calculation result meets the preset logical conditions, it outputs the data processing instruction and classification label. When the calculation result does not meet the preset logical conditions, it attenuates the intervention parameter based on the deviation ratio parameter and calls the intervention node candidate set for iterative calculation until the conflict is resolved or the preset iteration number threshold is reached, and outputs the manual intervention warning mark.

Images