An over-checking detection method, electronic equipment and storage medium

Abstract

This application discloses a method, electronic device, and storage medium for detecting over-examination, relating to the field of knowledge graphs. The method includes: extracting patients' electronic medical record data; obtaining a set of source nodes based on the electronic medical record data, mapping the items to be examined to target nodes; using the set of source nodes as an initial semantic scalar, performing iterative computation using a graph convolutional network until the semantic distribution reaches a steady state, and extracting the convergent semantic features of the target nodes in the steady state, thereby extracting the main transmission links and generating a locally connected subgraph for isomorphic matching with a complication pathology structure database; if specific implicit clinical associations exist, expanding the locally connected subgraph into a text sequence to obtain a natural language inference chain; determining the medical rationality of the items to be examined under the current symptoms, and outputting a reasonable or excessive classification label. This application can effectively improve clinical diagnostic efficiency.

Description

Technical Field

[0001] This application relates to the field of knowledge graphs, and more particularly to an over-checking detection method, electronic device, and storage medium. Background Technology

[0002] With the deepening development of medical informatization, electronic medical record systems and clinical decision support systems have been widely used in medical institutions at all levels. To improve medical quality and control medical costs, regulatory authorities and medical institutions urgently need to review the rationality of examinations prescribed by doctors to identify and prevent over-medicalization. Excessive testing not only increases the economic burden and risk of physical harm to patients but also wastes medical resources. Therefore, establishing an effective mechanism to address excessive testing has significant clinical value and social implications.

[0003] In existing technologies, the detection of over-testing primarily relies on medical knowledge graphs built upon clinical guidelines and expert consensus. These knowledge graphs typically employ a "disease-symptom-examination" triple structure, recording strong correlations within standard treatment pathways. The detection system matches the patient's chief complaint or confirmed disease with nodes in the knowledge graph to determine whether the doctor's prescribed examinations fall within the standard pathway. If an examination is not directly related to the patient's condition in the graph, the system identifies it as potential over-testing and triggers an alert. This method demonstrates good accuracy and interpretability when handling simple cases with a single disease and typical symptoms.

[0004] However, in real-world clinical settings, many patients present with multiple coexisting complications, complex medical histories, and atypical symptoms. In such cases, while certain examinations may not have a direct first-order correlation with the patient's chief complaint or primary diagnosis, they may be clinically necessary due to the interaction between multiple diseases, the need to screen for rare complications, or interdisciplinary pathological mechanisms. These implicit diagnostic and treatment logics often require multi-hop reasoning and cross-disciplinary medical fields to establish a complete causal chain. Existing detection methods based on standard knowledge graphs, focusing only on direct relationships, fail to capture these long-tail, interdisciplinary implicit clinical connections, leading to the misjudgment of reasonable but atypical examinations as over-testing. This high false positive rate not only frequently interrupts doctors' diagnostic and treatment processes, requiring them to repeatedly fill out explanations for examinations, severely impacting diagnostic efficiency, but may also reduce doctors' trust in intelligent review systems.

[0005] In summary, the current technical shortcomings are that existing over-testing and detection methods cannot accurately identify implicit diagnostic and treatment pathways with interdisciplinary pathological connections in complex medical case scenarios, resulting in an excessively high misjudgment rate for reasonable but atypical examinations, which seriously affects the efficiency of clinical diagnosis and treatment. Summary of the Invention

[0006] This application provides a method, electronic device, and storage medium for detecting over-examination, which can accurately identify hidden diagnostic and treatment pathways with cross-disciplinary pathological associations in complex medical record scenarios, significantly reduce the misjudgment rate of reasonable but atypical examinations, and improve clinical diagnostic and treatment efficiency.

[0007] To achieve the above objectives, the embodiments of this application adopt the following technical solutions: Firstly, an over-checking detection method is provided, which includes: In response to receiving the doctor's order for examination items, the system retrieves the patient's electronic medical record data, which includes a set of entities for the chief complaint symptoms and a set of entities for previously diagnosed diseases. The entity set of chief complaint symptoms and the entity set of previously diagnosed diseases are aligned with nodes in a pre-built medical knowledge graph to obtain a source node set, and the items to be examined are mapped to target nodes. The source node set is used as the initial semantic scalar. The graph convolutional network is iteratively computed based on the initial semantic scalar until the semantic distribution reaches a steady state. The convergent semantic features of the target node under the steady state are then extracted. Based on the extraction of convergent semantic features, the main transmission links from the source node set to the target node are obtained, and a locally connected subgraph is generated based on the main transmission links; The local connected subgraphs are subjected to subgraph isomorphic matching with a pre-defined complication pathology structure library to determine whether there is a specific implicit clinical association between the target node and the source node set. If there is a specific implicit clinical association between the target node and the source node set, the local connected subgraph is expanded into a text sequence of node entities and edge relationships, and the text sequence is used as a limiting context to be concatenated into a preset prompt word template to obtain a natural language inference chain. The medical rationality of the item to be examined under the current symptoms is determined based on the text sequence, and the reasonable or excessive classification label and the corresponding natural language inference chain are output. If the category label is appropriate, allow the order to be issued for the item to be inspected. When the category label is "excessive", an interception pop-up is triggered and the natural language inference chain is displayed.

[0008] In one possible implementation of the first aspect, the set of entities representing chief complaints and the set of entities representing previously diagnosed diseases are aligned with nodes in a pre-constructed medical knowledge graph to obtain a source node set, including: The entity set of chief complaint symptoms and the entity set of previously diagnosed diseases are standardized to obtain a standardized entity set. Retrieve node identifiers that match the standardized entity set in the medical knowledge graph; Activate the corresponding graph node based on the node identifier, and mark the activated graph node as the source node set.

[0009] In another possible implementation of the first aspect, the source node set is used as an initial semantic scalar. Iterative computation of a graph convolutional network is performed based on this initial semantic scalar until the semantic distribution reaches a steady state. The convergent semantic features of the target nodes in the steady state are then extracted, including: Obtain the attribute information of edges in the medical knowledge graph, including the triggering relationship, the diagnostic method relationship, and the concurrency relationship; Based on the edge attribute information, initialize the semantic transmission damping coefficient between each adjacent node; Extract the clinical importance score of each source node in the source node set; The clinical importance score is mapped to the initial semantic scalar magnitude of each source node; Based on the semantic transmission damping coefficient, a semantic diffusion matrix of a graph structure is constructed. The initial semantic scalar magnitude is iteratively diffused along the semantic diffusion matrix, and the mutation rate of the global node scalar is calculated in two adjacent iterations. When the mutation rate of the global node scalar is lower than the preset stability convergence factor, the semantic distribution is determined to have reached a steady state, and the accumulated scalar value of the target node is extracted as the convergent semantic feature under the steady state.

[0010] In another possible implementation of the first aspect, the main propagation links from the source node set to the target node are extracted based on convergent semantic features, and a locally connected subgraph is generated based on the main propagation links, including: Tracing back the scalar source path of the convergent semantic features of the target node in steady state; Calculate the path decay integral for each scalar source path; All scalar source paths are sorted in ascending order based on the path attenuation integral, and the paths ranked in the preset first percentage interval are extracted as candidate transmission links. Cross-reference frequency statistics are performed on the nodes in the candidate propagation links, and isolated paths with zero cross-reference frequency are removed. The remaining paths are used as the main propagation links. Merge the node and edge relationships in all major transmission links to construct a locally connected subgraph.

[0011] In another possible implementation of the first aspect, a subgraph isomorphic matching is performed between the locally connected subgraph and a pre-defined complication pathology structure library to determine whether there is a specific implicit clinical association between the target node and the source node set, including: Obtain a set of baseline pathological topologies for confirmed complex complication cases from the complication pathology structure library; Graph feature extraction is performed on locally connected subgraphs to obtain topological feature vectors of the locally connected subgraphs; Graph feature extraction is performed on each benchmark pathological topology graph in the benchmark pathological topology graph set to obtain the benchmark topology feature vector; Calculate the structural similarity between the topological feature vectors of the locally connected subgraphs and each baseline topological feature vector; When the structural similarity is greater than the preset isomorphic matching threshold, it is determined that the local connected subgraph and the corresponding baseline pathological topology graph have an isomorphic relationship. Specific latent attributes marked in the baseline pathological topology map with isomorphic relationships are extracted, and specific latent clinical associations between the target node and the source node set are confirmed based on the specific latent attributes.

[0012] In another possible implementation of the first aspect, calculating the structural similarity between the topological feature vectors of the locally connected subgraphs and each reference topological feature vector includes: Calculate the minimum number of edit operations between the locally connected subgraph and the baseline pathological topology, where the edit operations include node insertion, node deletion, edge insertion, and edge deletion; Calculate the structural difference degree based on the minimum number of editing operations; The structural dissimilarity is converted into structural similarity, where structural similarity and structural dissimilarity are negatively correlated.

[0013] In another possible implementation of the first aspect, structural dissimilarity is converted into structural similarity, including: Obtain the penalty coefficient for graph complexity; Multiply the structural difference by the penalty coefficient to obtain the corrected difference term; The initial similarity score is obtained by subtracting the correction difference term from the preset theoretical maximum similarity constant. The initial similarity scores are subjected to exponential smoothing to obtain the structural similarity.

[0014] In another possible implementation of the first aspect, if a specific implicit clinical association exists between the target node and the source node set, the locally connected subgraph is expanded into a text sequence of node entities and edge relationships. This text sequence is then used as a limiting context and concatenated into a preset prompt word template to obtain a natural language inference chain, including: If there is a specific implicit clinical association between the target node and the source node set, perform a depth-first search traversal on the locally connected subgraph, and extract the medical standard names of the node entities passed through during the traversal. Extract the attribute description fields of the edges connecting each node entity; Following the traversal order, the medical specification name and attribute description fields are alternately concatenated to generate a linear text sequence; Load a preset prompt word template containing blank context slots, and concatenate a linear text sequence as a limiting context into the prompt word template to obtain a natural language inference chain. The prompt word template includes task instructions, inference constraints, and output format requirements.

[0015] Secondly, this application provides an electronic device, comprising: The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the over-checking detection method described above.

[0016] Thirdly, this application provides a machine-readable storage medium storing instructions that cause a machine to perform the aforementioned over-checking detection method.

[0017] The above technical solution establishes a graph structure representation from complex medical records to examination items by aligning the patient's chief complaint symptom entity set and the previously diagnosed disease entity set in a medical knowledge graph to form a source node set, and mapping the examination items to be examined to target nodes. Iterative computation using a graph convolutional network until the semantic distribution reaches a steady state allows the semantic information of the source nodes to propagate and aggregate along the edge relationships of the knowledge graph in multiple hops. This enables the target node to gather semantic features from multiple source nodes through multi-layered propagation, thus overcoming the limitation of traditional methods that only focus on first-order direct associations and achieving automatic discovery of long-distance, interdisciplinary implicit association paths. By extracting the main propagation links from the source node set to the target node and generating locally connected subgraphs, these locally connected subgraphs actually represent the atypical diagnostic and treatment logic that may exist between the patient's complex condition and the examination items. A subgraph isomorphic matching mechanism with a pre-defined complication pathology structure library is introduced. This mechanism can verify whether automatically discovered transmission links correspond to real clinical pathology mechanisms, especially in complex scenarios involving multiple complication interactions and rare complication screening. By matching with known pathology structures, it can effectively distinguish between genuine implicit clinical associations and noisy paths in the graph, thus significantly reducing the false positive rate. When a specific implicit clinical association is confirmed, the locally connected subgraph is converted into a text sequence and combined with prompt word templates to generate a natural language inference chain. This provides interpretability for the judgment results and facilitates doctors' quick understanding of complex multi-hop inference logic, avoiding the trust issues caused by the lack of explanation in traditional systems. A differentiated processing strategy is adopted based on the medical rationality judgment results. Reasonable examinations are directly approved, while inference chains are displayed for doctors' reference for excessive examinations. This ensures effective interception of truly excessive examinations and avoids frequent false alarms by accurately identifying implicit associations in complex cases. It fundamentally solves the problem of high false positive rates and impact on diagnostic and treatment efficiency in existing technologies when dealing with patients with multiple complications and complex medical histories.

[0018] Other features and advantages of the embodiments of this application will be described in detail in the following detailed description section. Attached Figure Description

[0019] Figure 1 A schematic flowchart illustrating an over-checking detection method provided in an embodiment of this application; Figure 2 A schematic diagram of a smart review and processing mechanism for medical examination prescriptions provided in this application embodiment; Figure 3 This is a flowchart illustrating the generation of a locally connected subgraph, as provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for illustration and explanation of the embodiments of this application and are not intended to limit the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0021] It should be noted that if the embodiments of this application involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0022] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0023] Figure 1 The illustration schematically shows a flowchart of an over-checking detection method according to an embodiment of this application. Figure 1 As shown in the embodiment of this application, an over-checking detection method is provided, which may include the following steps.

[0024] S110. In response to receiving the doctor's order for examination items, extract the patient's electronic medical record data, wherein the electronic medical record data includes a set of chief complaint symptoms and a set of previously diagnosed diseases. S120. Align the entity set of chief complaint symptoms and the entity set of previously diagnosed diseases in a pre-constructed medical knowledge graph to obtain the source node set, and map the items to be examined to the target nodes. S130. Take the source node set as the initial semantic scalar, perform iterative computation of the graph convolutional network based on the initial semantic scalar until the semantic distribution reaches a steady state, and extract the convergent semantic features of the target node in the steady state. S140. Based on the convergence semantic features, the main transmission links from the source node set to the target node are extracted, and a local connected subgraph is generated based on the main transmission links; S150. Perform subgraph isomorphic matching between the locally connected subgraph and the pre-defined complication pathology structure library to determine whether there is a specific implicit clinical association between the target node and the source node set. S160. If there is a specific implicit clinical association between the target node and the source node set, the local connected subgraph is expanded into a text sequence of node entities and edge relationships, and the text sequence is used as a limiting context to be concatenated into a preset prompt word template to obtain a natural language inference chain. S170. Determine the medical rationality of the item to be examined under the current symptoms based on the text sequence, and output the reasonable or excessive classification label and the corresponding natural language inference chain. S180. When the category label is reasonable, allow the order to be issued for the item to be inspected. S190. When the category label is excessive, trigger an interception pop-up and display the natural language inference chain.

[0025] Among these, determining the medical rationale for the items to be examined under the current symptoms based on text sequences includes: Determine the deduction path and conclusion markers based on the text sequence; Verify whether the pathological transmission logic in the deduction path conforms to the unidirectional causal constraints of the medical knowledge graph; If the unidirectional causal constraint is met and the conclusion is marked as reasonable, then the medical rationality is approved. If the unidirectional causal constraint is not met or the conclusion is marked as excessive, then the determination of medical rationality is not passed.

[0026] In this embodiment, the following steps are also included: If the pop-up blocking window is triggered, monitor whether the doctor performs a forced prescription action; In response to doctors' actions to force prescriptions, a medical reason input interface is provided in the blocking pop-up window; Receive the appeal text entered by the doctor through the medical reason input interface; The appeal text is associated and stored with the corresponding items to be examined, the patient's electronic medical record data, and the natural language inference chain; The associated stored data is submitted to the back-end management system for manual review by medical quality management personnel. The medical knowledge graph and complication pathology structure database are updated based on the results of manual review. Updates include adding new related edges or adjusting the transmission coefficients of existing edges.

[0027] Specifically, the medical knowledge graph is updated based on the results of manual review, including: Identify cases of forced invoicing that are deemed reasonable in the results of manual review; Extract patient symptom nodes, examination item nodes, and intermediate pathological nodes described in doctor's appeal texts from reasonable cases of mandatory prescriptions; Search the medical knowledge graph to see if there is a connection path via an intermediate pathology node between patient symptom nodes and examination item nodes. In the absence of a connection path, create new edges in the medical knowledge graph to establish a connection path, and assign initial conduction coefficients to the newly created edges; In the presence of a connection path, increase the conduction coefficient of the relevant edges on that path. The extent of the increase in the conduction coefficient is determined based on the clinical validation strength of the case.

[0028] In one embodiment of this invention, the method further includes combining the patient's time-series medication history to calibrate convergent semantic features, as follows: Extract the drug components from the patient's recent prescription records; Locate the antagonistic nodes corresponding to the entities of each drug component in the medical knowledge graph; Determine whether a locally connected subgraph contains nodes with antagonistic interactions; If antagonistic nodes are included, calculate the drug half-life decay rate at the antagonistic nodes; The aggregation semantic features of the target node are reversed and corrected by using the drug half-life decay ratio. The locally connected subgraph is regenerated based on the corrected convergence semantic features.

[0029] To eliminate the semantic amplification effect caused by cyclic citations in the graph structure, the following steps are included before extracting the convergent semantic features of the target node in steady state: During the iterative computation of the graph convolutional network, an in-degree history access stack is established for each node. When any node receives a semantic scalar, it checks the source link of the semantic scalar. Compare the source link with the in-degree historical access stack of the node; If the source link is found to contain the node's own identifier, a circular citation loop is determined to exist. Cut off the last directed edge in the loop reference closure and force the semantic transmission damping coefficient of that directed edge to zero; After eliminating all cyclic citation loops, continue with the semantic scalar iterative computation.

[0030] When generating the text sequence, the method also includes: Obtain a dictionary of instrument invasiveness ratings for the items to be inspected; Query the corresponding rating scale of the target node in the instrument invasiveness rating dictionary; When the rating scale indicates non-surface non-invasive examination, obtain the path depth of the local connected subgraph; The product of the degree of invasiveness of the medical device and the path depth is used as the weight for the necessity of secondary diagnosis; Convert the necessity weight of secondary diagnosis into a descriptive text format; The descriptive text format is forcibly inserted as a constraint into the inference constraint module of the preset prompt word template.

[0031] When the interception pop-up is triggered, the method also includes the following steps: Obtain real-time environmental heart rate monitoring data from the department terminal where the doctor is currently issuing the prescription; Calculate the volatility index of heart rate monitoring device data; Obtain the time difference sequence of queuing for registration in the current department and calculate the time urgency parameter; The intensity of fluctuation index and the time urgency parameter are normalized and fused to obtain the emergency status assessment value; When the category label is "excessive" and the emergency status assessment value indicates an emergency status, cancel the screen-top blocking attribute of the pop-up blocker. Convert the blocked pop-up to an asynchronous sidebar notification mode and generate a post-review mark in the system background.

[0032] When updating the complication pathology structure database based on manual review results, the method also includes the following steps: Obtain multiple doctor appeal texts submitted by different hospitals that involve the same patient's symptom nodes and examination item nodes; Extract intermediate pathological nodes described in each doctor's appeal text; Statistical histogram of frequency distribution of appeal cases with the same intermediate pathological node; Gaussian smoothing was applied to the frequency distribution histogram, and intermediate pathological nodes corresponding to local peaks were extracted as consensus pathological nodes. A new baseline pathology topology is constructed using only locally connected subgraphs containing consensus pathology nodes; Add the new baseline pathology topology to the complication pathology structure library.

[0033] In response to receiving the doctor's order for examinations, the system first extracts the patient's complete medical record information from the hospital's electronic medical record database. The extraction of electronic medical record data employs structured parsing, and multiple data interface adapters are pre-configured to accommodate the format differences of electronic medical record systems used by different hospitals. For electronic medical record systems using the HL7 standard, basic patient information is obtained by parsing the PID segment in the HL7 message segment, and observation results and symptom descriptions are obtained through the OBX segment. During the extraction process, the acquisition of the chief complaint symptom entity set focuses on sections such as "chief complaint," "present illness," and "physical examination" in the medical record, using a pre-trained medical named entity recognition model to extract entities from these texts. This model is based on the BERT-BiLSTM-CRF architecture and has been fine-tuned on a large-scale Chinese medical corpus, enabling it to identify entity types such as symptoms, signs, and locations. For example, from the text "The patient complains of recurrent chest tightness and shortness of breath for 3 days, accompanied by paroxysmal nocturnal dyspnea," symptom entities such as "chest tightness," "shortness of breath," and "paroxysmal nocturnal dyspnea" can be extracted.

[0034] The set of previously diagnosed disease entities is extracted from the "Past Medical History" and "Diagnosis" sections of the medical record, with a focus on identifying the disease names corresponding to the ICD-10 codes. For patients with multiple historical diagnoses, all confirmed disease records are retained, along with diagnosis timestamp information.

[0035] After extraction, the entities representing chief complaint symptoms and previously diagnosed diseases are organized into two separate sets to prepare for subsequent knowledge graph alignment. This step ensures the accurate extraction of key medical entities from unstructured or semi-structured medical record texts, laying a data foundation for subsequent reasoning and analysis.

[0036] When aligning the sets of entity classes representing chief complaints and entity classes representing previously diagnosed diseases within a pre-constructed medical knowledge graph, the extracted entities first need to be standardized. Medical terminology contains numerous synonyms and abbreviations; for example, "myocardial infarction," "heart attack," and "MI" actually refer to the same disease. Standardization employs medical terminology mapping technology to uniformly map various expressions to standardized names in standard medical dictionaries such as SNOMED CT or ICD-10.

[0037] The mapping process uses a hybrid matching algorithm based on edit distance and semantic similarity. First, the character-level distance between the entity to be matched and each candidate in the dictionary is calculated, and a candidate set with a distance less than a threshold is selected. Then, a pre-trained medical word vector model is used to calculate semantic similarity, and the standard term with the highest similarity is selected as the mapping result.

[0038] After obtaining the standardized entity set, the system retrieves node identifiers that match these standardized entities within the medical knowledge graph. The medical knowledge graph is stored using the Neo4j graph database, and each node contains unique attributes such as a node ID, standardized name, and entity type. The retrieval process uses the Cypher query language, employing the MATCH statement to precisely match graph nodes based on their standardized names.

[0039] For each retrieved node, its status is marked as "activated," and the IDs of all activated nodes are collected into the source node set. Simultaneously, the items to be examined are mapped to corresponding examination item nodes in the knowledge graph; for example, "coronary CT angiography" is mapped to the CTA node in the graph, and this node is marked as the target node. Through this node alignment mechanism, the patient's clinical presentation is transformed into a structured representation in the knowledge graph, enabling subsequent reasoning and computation using the graph structure.

[0040] When using the source node set as the initial semantic scalar for iterative computation of a graph convolutional network, the semantic scalar values ​​of each node must first be initialized. For each node in the source node set, a different initial scalar amplitude is assigned based on its importance in clinical diagnosis. The clinical importance score is derived by analyzing the correlation between the symptom or disease and the final diagnosis in large-scale electronic medical record data, with a score range of 0 to 1. For example, for a patient with acute myocardial infarction, the clinical importance score for the chief symptom "chest pain" might be 0.9, while the score for the accompanying symptom "mild dizziness" might only be 0.3. These scores are directly used as the initial semantic scalar amplitudes of the source nodes. For non-source nodes, the initial scalar value is set to 0.

[0041] Next, a semantic diffusion matrix is ​​constructed, which describes how semantic information propagates along the edges in the knowledge graph. The edges in the medical knowledge graph have different attribute types, such as "triggering relationship" indicating that a certain symptom may trigger a certain disease, "diagnostic means relationship" indicating that a certain test can diagnose a certain disease, and "concurrency relationship" indicating that two diseases may occur concurrently.

[0042] Based on the attribute type of the edge, a semantic transmission damping coefficient is assigned to each edge. The damping coefficient for causal relationships is typically set to 0.7 to 0.9, indicating strong causal transmission; the damping coefficient for diagnostic means relationships is set to 0.8 to 0.95, indicating a strong association between the examination and the disease; and the damping coefficient for concurrent relationships is set to 0.5 to 0.7, indicating a relatively weak indirect association.

[0043] Semantic diffusion matrix elements Indicates from node To the node The semantic transitivity coefficient, if an edge exists between the two nodes, then It equals the damping coefficient of that side; otherwise, it is 0.

[0044] In the In the round of iteration, nodes semantic scalar Updated to After each iteration, the mutation rate of the global node scalar is calculated, defined as... ,in This represents the total number of nodes in the graph. When... When the convergence factor is less than a preset stability factor (e.g., 0.001), the semantic distribution is considered to have reached a steady state. In the steady state, the accumulated scalar value of the target node is the convergent semantic feature, reflecting the semantic strength of the convergence from all source nodes through multi-hop propagation to the target node. This graph convolution iterative mechanism captures long-distance implicit associations between source and target nodes, overcoming the limitation of traditional methods that only consider direct neighbors.

[0045] When extracting the main transmission links based on convergent semantic features, it is necessary to backtrack the scalar source of the target node. During the graph convolution iteration, the source node and the amount of semantic scalar received by each node in each iteration are recorded. The backtracking process starts from the target node, traces back all its predecessor nodes that received the scalar in the last iteration, and then recursively traces the predecessors of these predecessor nodes until it backtracks to the set of source nodes. This yields multiple paths from the source node to the target node, each path consisting of a series of nodes and edges.

[0046] To assess the importance of each path, the path attenuation integral is calculated, defined as the product of the damping coefficients of all edges along the path, i.e., for each path... Its attenuation integral is ,in For the edge The damping coefficient. The larger the attenuation integral, the stronger the semantic transmission capability of the path.

[0047] All paths are sorted in descending order of decay integral, and the paths in the top 20% (e.g., the first 20%) are extracted as candidate conduction pathways. However, some paths may be noise paths in the graph, irrelevant to actual clinical logic. To filter out this noise, the cross-reference frequency of nodes in the candidate conduction pathways is counted, i.e., how many candidate paths each node appears in. If a path contains nodes that never appear in other paths (cross-reference frequency is zero), this path is likely an isolated noise path and is discarded. The retained paths are the main conduction pathways, representing the main reasoning path from patient symptoms and disease to the items to be examined.

[0048] By merging all nodes and edges involved in the main transmission links, a locally connected subgraph is constructed. This subgraph is a subset of the original medical knowledge graph, specifically describing the association structure between the current patient's condition and the items to be examined. In this way, the complex global knowledge graph is simplified into a local reasoning structure specific to a particular patient, facilitating subsequent pathological structure matching and interpretability analysis.

[0049] When performing subgraph isomorphic matching between locally connected subgraphs and a complication pathology structure database, the first step is to load a set of baseline pathology topologies from confirmed complex complication cases into the database. These baseline topologies are typical pathological association patterns extracted through analysis of a large number of real complex cases, with each topology representing a specific combination of complications and its diagnostic and treatment pathway. For example, a baseline topology might describe the typical pathway requiring echocardiography and renal function tests when "diabetes mellitus-coronary artery disease-renal insufficiency" occurs concurrently.

[0050] Graph feature extraction is performed on locally connected subgraphs using a Graph Isomorphic Network (GIN) architecture from graph neural networks. GIN encodes the subgraph's topology into a fixed-dimensional vector representation through multiple layers of message passing and aggregation operations. Specifically, for each node in the subgraph, its attributes (such as entity type and standard name) are first encoded into an initial feature vector, and then aggregated using GIN's aggregation function. The update node indicates that, For nodes The neighborhood group, The parameters are learnable; MLP stands for Multilayer Perceptron.

[0051] go through After layer iteration, the representations of all nodes are summed and pooled to obtain the topological feature vector of the entire subgraph. The same method is applied to each graph in the baseline pathological topological graph set to obtain the corresponding baseline topological feature vector. The structural similarity between the topological feature vector of the locally connected subgraph and each baseline topological feature vector is calculated, using cosine similarity or graph edit distance-based similarity measures. The graph edit distance method is more accurate, measuring structural differences by calculating the minimum number of edit operations required to transform one graph into another. Edit operations include node insertion, node deletion, edge insertion, and edge deletion, each assigned a different cost weight. A... The search algorithm finds the optimal edit sequence and obtains the minimum number of edit operations. .

[0052] Structural dissimilarity is defined as ,in Let be the number of nodes in the two graphs. The number of edges is used, and normalization is applied to ensure that the dissimilarity is between 0 and 1.

[0053] Considering the impact of graph complexity on matching difficulty, a penalty coefficient is introduced. The difference item was corrected as follows: Structural similarity is calculated as follows: And perform exponential smoothing. ,in For smoothing parameters. When When the value exceeds a preset isomorphic matching threshold (e.g., 0.75), the locally connected subgraph is determined to have an isomorphic relationship with the corresponding baseline pathological topology. Specific latent attributes, such as labels like "rare complication screening" and "cross-departmental pathological association," are extracted from the successfully matched baseline topology to confirm a specific latent clinical association between the target node and the source node set. This matching mechanism effectively distinguishes between real, complex pathological associations and random noise paths in the graph, significantly improving the accuracy of the judgment.

[0054] If a specific implicit clinical association is confirmed, the locally connected subgraph is converted into a text sequence that can be understood by a large language model. The conversion process uses a depth-first search to traverse the subgraph, starting from a node in the source node set and visiting adjacent nodes along the edges, prioritizing unvisited nodes, until all nodes have been traversed. During the traversal, the medical terminology of each node is recorded, such as "type 2 diabetes" or "coronary atherosclerotic heart disease." Simultaneously, attribute description fields of the edges connecting each node are extracted, such as "may cause," "requires diagnosis through…," and "common complications are…."

[0055] Following the traversal order, node names and edge descriptions are alternately concatenated to generate a linear text sequence. For example, a path might be transformed into "The patient has a history of type 2 diabetes; diabetes may lead to coronary atherosclerotic heart disease; coronary heart disease needs to be diagnosed through coronary CT angiography." For subgraphs containing multiple paths, the text sequences of each path are connected using delimiters.

[0056] Next, a preset prompt template is loaded. This template contains three parts: task instructions, inference constraints, and output format requirements. The task instructions describe the goal of the inference task, such as "Based on the following pathological association path, determine the medical rationality of the examination items." The inference constraints specify the boundary conditions for inference, such as "Only consider the provided pathological path, without introducing external knowledge" and "Must follow unidirectional causal logic." The output format requirements specify the structure of the model output, such as "Output format: Inference process: Conclusion: Reasonable / Excessive." The generated linear text sequence is then inserted into the blank context slots of the prompt template to form the complete prompt text.

[0057] The prompt text is input into a pre-trained large language model (such as GPT-4 or a medically fine-tuned LLaMA model). The model infers based on the provided pathological path, generating a natural language reasoning chain. This reasoning chain details the logical connection between the patient's symptoms and disease and the examination to be performed. For example, "The patient has a history of diabetes, which is a high-risk factor for coronary heart disease. When a patient with coronary heart disease experiences chest tightness, a coronary CTA is needed to screen for coronary artery stenosis; therefore, this examination is medically justified." This natural language form of reasoning chain has good interpretability, allowing doctors to quickly understand the system's judgment basis.

[0058] When determining the medical rationality of an item to be examined based on the generated text sequence, the reasoning path and conclusion label are first extracted from the output of the large language model. The reasoning path describes the reasoning process from the source node to the target node, and the conclusion label is "reasonable" or "excessive".

[0059] To ensure the correctness of the reasoning, it is necessary to verify whether the pathological transmission logic in the deduction path conforms to the unidirectional causal constraints of the medical knowledge graph. Medical causal relationships are usually directional; for example, "disease causes symptoms" is a reasonable causal direction, while "symptoms cause disease" violates causal logic.

[0060] The verification process maps each step of the reasoning path back to an edge in the knowledge graph and checks whether the direction of the edge is consistent with the reasoning direction. If a reverse causal relationship appears in the reasoning path (such as deriving a disease from an item being examined, and then deriving symptoms), it is determined that it does not comply with the unidirectional causal constraint.

[0061] If the deduction path passes the validation and the conclusion is marked as "reasonable," then the medical rationality is deemed successful, and the "reasonable" category label and corresponding reasoning chain are output. If the deduction path does not meet the constraints or the conclusion is marked as "excessive," then the medical rationality is deemed unsuccessful, and the "excessive" category label and reasoning chain are output. The category label and reasoning chain are returned to the clinical decision support system for subsequent processing decisions. This reasoning chain-based judgment method not only provides the judgment result but also offers a detailed reasoning process, enhancing the system's transparency and credibility.

[0062] Reference Figure 2 When the classification label is "reasonable," the order for the examination is directly approved without any intervention in the doctor's treatment process. The approval is achieved by sending a confirmation signal to the examination request module of the hospital information system, allowing the examination order to flow normally to the medical technology department. Simultaneously, the review record (including patient information, examination items, review results, and inference chain) is stored in the review log database for subsequent quality traceability and statistical analysis. The rapid approval of reasonable examinations avoids unnecessary process interruptions, ensuring the continuity and efficiency of clinical diagnosis and treatment.

[0063] When the category label is "excessive," an interception pop-up is triggered, displaying a warning message as a modal dialog box on the doctor's workstation interface. The interception pop-up consists of three main parts: first, a prominent warning title, such as "Examination Reasonableness Warning"; second, a display area showing the natural language reasoning chain, presenting the reasoning process in a structured way, such as displaying "Patient Condition Analysis," "Examination Item Correlation Analysis," and "Basis for Reasonableness Judgment" in segments; and finally, an operation button area, providing two options: "Cancel Order" and "Force Order."

[0064] After reviewing the reasoning chain, if the doctor agrees with the assessment, they can choose "Cancel Order" to withdraw the examination request. If the doctor believes the assessment is incorrect or there are special clinical considerations, they can choose "Force Order." In this case, a medical reason input interface will pop up, requiring the doctor to fill in an appeal text explaining the medical basis for forcing the order. This design prevents excessive testing while preserving the doctor's final decision-making power, balancing intelligent supervision with clinical autonomy.

[0065] Upon triggering the intercept pop-up, the system continuously monitors whether the doctor performs the forced prescription action. This monitoring is achieved through an event listener mechanism. When the doctor clicks the "Force Prescription" button, the system captures and responds to this event. During the response, the intercept pop-up interface switches to a medical reason input mode, displaying a multi-line text input box prompting the doctor to explain the medical reasons for the forced prescription in detail. A reference template is provided below the input box, such as "Please explain: 1. The patient's specific medical history or physical signs; 2. Specific clinical considerations for this examination; 3. Which alternative examination options have been ruled out." After the doctor completes the appeal text input and clicks the "Submit" button, the system receives and stores the text accordingly.

[0066] The associated storage packages the appeal text along with the corresponding items to be examined, the patient's electronic medical record data (including chief symptoms and past illnesses), the generated natural language inference chain, and the locally connected subgraph structure into a single review case record. This record is then submitted to the pending review queue of the backend management system for manual review by medical quality management personnel.

[0067] During the manual review process, quality management personnel comprehensively evaluate the doctor's reasons for appeal, the patient's actual condition, and the reasoning logic to determine whether the forced prescription is reasonable. Review results are categorized into three types: "Confirmed Reasonable," "Confirmed Over-Confirmed," and "Questionable and Needs Discussion." For cases deemed "Confirmed Reasonable," it indicates a misjudgment, requiring an update to the knowledge base to prevent similar future misjudgments. Updates include adding new related edges to the medical knowledge graph or adjusting the transmission coefficients of existing edges to enhance the semantic transmission capability of specific paths. For cases deemed "Over-Confirmed," the doctor's inappropriate behavior is recorded and included in performance evaluations or training plans. For cases "Questionable and Needs Discussion," they are submitted to a multidisciplinary expert committee for in-depth discussion, and the knowledge base is updated only after consensus is reached. This closed-loop mechanism of human-machine collaboration allows the system to continuously learn from clinical practice and optimize its judgment accuracy.

[0068] When updating the medical knowledge graph based on manual review results, the first step is to identify mandatory prescription cases marked as "confirmed reasonable" in the review results. These cases represent reasonable treatment pathways not currently covered by the knowledge base. Key information is extracted from these cases, including patient symptom nodes (such as "chest tightness" and "shortness of breath"), examination item nodes (such as "coronary CTA"), and intermediate pathological nodes described by doctors in their appeal texts (such as "coronary microcirculation disorder" and "Syndrome X").

[0069] The extraction of intermediate pathological nodes employs medical named entity recognition technology to parse the appeal text and identify entities such as diseases, pathological mechanisms, and complications. After extraction, the medical knowledge graph is searched to determine if there are any connecting paths between patient symptom nodes and examination item nodes via these intermediate pathological nodes. The search uses graph traversal algorithms, such as breadth-first search, starting from the symptom nodes and exploring whether it is possible to reach the examination item nodes through intermediate nodes. If no such connecting path exists, it indicates that the diagnostic logic is missing from the knowledge graph, and new edges need to be created.

[0070] The creation process involves adding directed edges between symptom nodes and intermediate pathology nodes, and between intermediate pathology nodes and examination item nodes. The edge type is determined based on the association characteristics (e.g., "may indicate", "needs investigation"), etc. Initial transmission coefficients are assigned to newly created edges, typically set to low values ​​(e.g., 0.5), as these new paths have not yet undergone large-scale clinical validation. If a connection path already exists, it indicates that although the path exists in the knowledge graph, its transmission coefficient is too low, resulting in insufficient semantic convergence during graph convolution calculations. In this case, the transmission coefficient of the relevant edges on that path is increased, with the increase determined by the clinical validation strength of the case. Clinical validation strength is assessed through the evidence level of the case, such as whether there are multiple similar complaints from independent physicians, whether there is supporting literature, etc. Cases with high validation strength receive a larger increase in transmission coefficient (e.g., an increase of 0.2), while cases with low validation strength receive a smaller increase (e.g., an increase of 0.05). Through this incremental update mechanism, the medical knowledge graph can gradually absorb new knowledge from clinical practice, covering more long-tail diagnostic and treatment scenarios and reducing false positive rates.

[0071] This embodiment establishes a data foundation for intelligent review from clinical practice by responding to doctors' orders for examinations and extracting patients' electronic medical record data, ensuring that the review process is based on complete patient information rather than partial data. Aligning chief symptoms and past illnesses with nodes in the medical knowledge graph achieves the transformation from unstructured medical record text to structured knowledge representation, laying the foundation for subsequent graph reasoning. Employing graph convolutional networks for multi-hop semantic propagation computation overcomes the limitations of traditional methods that only focus on direct associations, enabling the discovery of long-distance implicit association paths across disciplines and pathological mechanisms, effectively solving the reasoning challenges in complex complication scenarios. By extracting the main transmission links and generating locally connected subgraphs, the complex global knowledge graph is simplified into a local reasoning structure specific to a particular patient, improving the targeting and efficiency of reasoning. Introducing a subgraph isomorphic matching mechanism with the complication pathology structure library, through comparison and verification with real pathological patterns, effectively distinguishes reasonable implicit associations from graph noise, significantly reducing the misjudgment rate. Converting locally connected subgraphs into natural language reasoning chains provides interpretability for the judgment results, enhancing doctors' trust in the system and avoiding resistance caused by black-box decision-making. A differentiated processing strategy is adopted based on the medical rationality judgment results: reasonable examinations are directly approved, while the reasoning chain is displayed for doctors' reference for excessive examinations. This ensures both the effectiveness of supervision and avoids frequent false alarms interfering with the treatment process. A human-machine collaborative closed-loop learning mechanism is established. By collecting doctors' appeals against forced prescriptions and conducting manual review, the knowledge graph and pathology structure database are continuously updated, enabling the system to continuously optimize and gradually cover more long-tail scenarios. In summary, this effectively solves the problems of high misjudgment rates, lack of interpretability, and impact on clinical diagnosis and treatment efficiency in the treatment of patients with multiple complications and complex medical histories. It achieves accurate identification of excessive examinations and effective protection of reasonable examinations, promoting the improvement of medical quality and the rational use of medical resources.

[0072] In one embodiment of this example, the set of entity objects representing chief complaints and the set of entity objects representing previously diagnosed diseases are aligned with nodes in a pre-constructed medical knowledge graph to obtain a source node set, including the following steps: S210. Standardize the entity set of chief complaint symptoms and the entity set of previously diagnosed diseases to obtain a standardized entity set; S220. Retrieve node identifiers that match the standardized entity set in the medical knowledge graph; S230. Activate the corresponding graph node according to the node identifier, and mark the activated graph node as the source node set.

[0073] When standardizing the entity set of chief complaint symptoms and the entity set of previously diagnosed diseases, the first step is to address the inconsistency in medical terminology. In real-world clinical scenarios, the same symptom or disease may be expressed in multiple ways, including standard medical terminology, colloquial descriptions, local dialect expressions, abbreviations, and different versions of disease classification codes. For example, "acute myocardial infarction" may be expressed as "heart attack," "AMI," "acute myocardial infarction," or "acute phase of myocardial infarction," among other forms. This diversity in terminology directly leads to difficulties in aligning knowledge graph nodes. If different expressions cannot be mapped to unified standard terminology, it will result in the problem of relevant knowledge existing but being unable to be retrieved, thus affecting the completeness and accuracy of subsequent reasoning.

[0074] The standardization process employs a multi-level mapping strategy. First, each entity in the entity set undergoes text preprocessing, including punctuation removal, case neutralization, and simplified / traditional character conversion. Next, a standardized medical terminology dictionary is used for initial mapping. This dictionary integrates multiple authoritative medical terminology systems, including the ICD-10 International Classification of Diseases, the SNOMED CT systematic medical terminology, and Chinese disease diagnosis codes, establishing mapping relationships from various non-standard expressions to standard terms. For mapping relationships directly existing in the dictionary, the conversion is performed directly.

[0075] For expressions not included in the dictionary, a deep learning-based semantic matching model is used for fuzzy mapping. This model uses a pre-trained BERT model in the medical field as the encoder, and has been fine-tuned on a large-scale medical literature and electronic medical record corpus to understand the semantics of medical terms. The entities to be standardized and candidate standard terms are encoded into vector representations, cosine similarity is calculated, and the standard term with the highest similarity exceeding a threshold (e.g., 0.85) is selected as the mapping result.

[0076] For rare terms or neologisms that still cannot be mapped, mark them as "pending manual review" and temporarily retain the original expressions, while triggering a background notification to remind the knowledge base administrator to manually annotate them. After standardization is completed, a set of standardized entities is obtained, where each entity corresponds to a standardized name in the medical terminology system.

[0077] When retrieving node identifiers that match a standardized set of entities in a medical knowledge graph, it is necessary to quickly locate the target node within the large-scale graph data. Medical knowledge graphs typically contain millions of nodes and tens of millions of edges, covering various entity types and their relationships, including diseases, symptoms, examinations, drugs, and pathological mechanisms. The retrieval process employs an efficient index-based query mechanism, building B-tree indexes and full-text search indexes for the standard name fields of nodes in the graph database.

[0078] B-tree indexes support exact match queries, allowing retrieval in logarithmic time complexity when the standardized entity name exactly matches the standardized name of a graph node. Full-text search indexes, based on inverted index structures, support partial matching after node name segmentation, handling subtle differences that may exist during standardization.

[0079] During retrieval, the system first attempts to perform exact matching using a B-tree index. If a match is successful, the node identifier is returned directly. If an exact match fails, a full-text search index is used for fuzzy matching. Standardized entities are segmented into words, and candidate nodes containing these terms are retrieved. A matching score is calculated based on term coverage and TF-IDF weights, and the node with the highest score is selected as the matching result.

[0080] For entities with clearly defined types (e.g., explicitly labeled as diseases or symptoms), an entity type filter is added during retrieval, searching only within the corresponding subset of nodes, further improving retrieval efficiency and accuracy. After the retrieval is complete, a unique identifier for the matching node is returned. This retrieval process is performed on each entity in the standardized entity set, ultimately resulting in a set of node identifiers.

[0081] When activating the corresponding graph node based on the node identifier and marking it as the source node set, it is necessary to establish the initial state for reasoning computation in the knowledge graph. The essence of the activation operation is to set the initial conditions for subsequent iterative computation of the graph convolutional network, clarifying which nodes are the sources of semantic propagation.

[0082] In the graph database, based on the retrieved node identifiers, specific node objects are located using either the Cypher query language or the Gremlin graph traversal language. For each located node, an "Activity Status" flag is added or updated in its attribute field, setting the value of this field to "active". Simultaneously, an initial semantic scalar value is assigned to each active node, reflecting the node's clinical importance in the current patient's condition.

[0083] The initial semantic scalar is assigned a value considering multiple factors, including the position of the symptom or disease in the patient's chief complaint (chief complaint symptoms have higher weight than accompanying symptoms), the severity of the symptom (e.g., "severe chest pain" has higher weight than "mild discomfort"), and the time of diagnosis (recently diagnosed diseases have higher weight than old diseases). For nodes corresponding to chief complaint symptom entities, the initial semantic scalar is usually set to a high value (e.g., 0.8 to 1.0). For nodes corresponding to previously diagnosed disease entities, the initial semantic scalar is set to a medium to high value (e.g., 0.5 to 0.9) based on the disease's activity status and its relevance to the current chief complaint.

[0084] The identifiers of all activated nodes are collected into a set data structure, which is the source node set. This source node set serves as the starting point for semantic diffusion in subsequent graph convolution iterations; its initial semantic scalar will propagate along the edge relationships of the knowledge graph to adjacent nodes. By explicitly labeling the source nodes and assigning them initial semantic values, a mapping from the patient's specific condition to the knowledge graph structure is established, enabling abstract graph knowledge to be combined with specific clinical scenarios.

[0085] This embodiment effectively solves the key technical problem of difficulty in aligning knowledge graph nodes due to the diversity of medical terminology. Through a multi-level mapping strategy and deep learning semantic matching, various non-standard medical terms are uniformly converted into standardized expressions, directly eliminating knowledge retrieval obstacles caused by terminological inconsistencies—a prerequisite for accurate node alignment. By establishing B-tree indexes and full-text search indexes, fast and accurate node positioning is achieved in large-scale knowledge graphs, significantly reducing retrieval time and meeting the performance requirements of real-time clinical review, avoiding interruptions in the diagnosis and treatment process due to retrieval delays. By assigning initial semantic scalars reflecting clinical importance to source nodes, a precise mapping is established between individual patient conditions and the general knowledge graph. This allows subsequent graph convolutional inference to be personalized based on the patient's specific situation, avoiding inference bias caused by knowledge generalization in traditional methods. Ultimately, a complete conversion process from unstructured medical record text to structured graph representation is formed, providing an accurate and reliable data foundation for subsequent steps and fundamentally improving the system's inference accuracy and clinical applicability in complex complication scenarios.

[0086] In one embodiment of this example, the source node set is used as an initial semantic scalar, and a graph convolutional network is iteratively computed based on the initial semantic scalar until the semantic distribution reaches a steady state. The convergent semantic features of the target nodes in the steady state are then extracted, including the following steps: S310. Obtain the attribute information of the edges in the medical knowledge graph, wherein the attribute information of the edges includes the triggering relationship, the diagnostic method relationship, and the concurrency relationship; S320. Based on the edge attribute information, initialize the semantic transmission damping coefficient between each adjacent node; S330. Extract the clinical importance score of each source node in the source node set; S340. Map the clinical importance score to the initial semantic scalar magnitude of each source node; S350. Based on the semantic transmission damping coefficient, construct a semantic diffusion matrix for graph structures; S360. Iteratively diffuse the initial semantic scalar magnitude along the semantic diffusion matrix and calculate the mutation rate of the global node scalar in two adjacent iterations. S370. When the mutation rate of the global node scalar is lower than the preset stability convergence factor, the semantic distribution is determined to have reached a steady state, and the accumulated scalar value of the target node is extracted as the convergent semantic feature under the steady state.

[0087] When retrieving attribute information from edges in a medical knowledge graph, it is necessary to extract metadata describing the type and strength of relationships between nodes from the graph database. Edges in a medical knowledge graph not only represent associations between nodes, but more importantly, describe the specific nature and clinical significance of those associations.

[0088] The attribute information of the edges mainly includes three types of core relationships: the causal relationship describes the causal connection between the cause and the result, such as "hypertension causes stroke" and "diabetes causes retinopathy", which reflects the development and evolution of the disease; the diagnostic means relationship describes the diagnostic dependence between the disease and the examination items, such as "coronary heart disease needs to be diagnosed by coronary angiography" and "lung cancer needs to be diagnosed by pathological biopsy", which reflects the standard path of clinical diagnosis; and the comorbidity relationship describes the comorbidity pattern of multiple diseases coexisting, such as "diabetes is often complicated by hypertension" and "chronic kidney disease is often complicated by anemia", which reflects the interactive influence of complex diseases.

[0089] Different types of relationships have different transmission strengths and reliability in clinical reasoning. Causative relationships usually have strong causal certainty, diagnostic means relationships have clear diagnostic orientation, while concurrent relationships are relatively weak and vary from person to person.

[0090] In the graph database, edge attribute information is stored in JSON format or as relational attribute fields, including fields such as relation type label, relation strength coefficient, clinical evidence level, and number of supporting documents. A Cypher query is used to traverse all edges in the knowledge graph, extract the values ​​of these attribute fields, and categorize them according to relation type. For multiple edges of different types that may exist between the same pair of nodes, the attribute information of each edge is recorded separately.

[0091] When initializing the semantic transmission damping coefficient between adjacent nodes based on the edge attribute information, it is necessary to quantify the clinical characteristics of the medical relationship into numerical parameters. The semantic transmission damping coefficient determines the degree of attenuation of semantic information when it propagates from one node to its adjacent nodes; a larger coefficient indicates stronger transmission ability, and a smaller coefficient indicates weaker transmission ability.

[0092] For causal relationships, because they reflect a clear causal chain and have strong transmission capacity, the damping coefficient is usually set between 0.75 and 0.90. The specific value is adjusted according to the level of clinical evidence for the causal relationship. If the causal relationship is supported by large-scale epidemiological studies and the causal relationship is clear, the coefficient is set to 0.90; if it is only supported by small sample studies or case reports, the coefficient is set to 0.75.

[0093] Regarding the relationship between diagnostic methods, since it reflects the strong correlation between examination and disease and is the gold standard for clinical diagnosis, the damping coefficient is set between 0.80 and 0.95. For examinations with both high diagnostic sensitivity and specificity (such as pathological biopsy), the coefficient is set at 0.95; for auxiliary examinations (such as imaging screening), the coefficient is set at 0.80.

[0094] For comorbidities, since they reflect a statistical pattern of comorbidity rather than a necessary causal relationship, and there are significant individual differences, the damping coefficient is set between 0.50 and 0.70. The specific value is determined based on the epidemiological incidence of the comorbidities: if the comorbidity rate of the two diseases exceeds 50%, the coefficient is set at 0.70; if the comorbidity rate is between 20% and 50%, the coefficient is set at 0.60; and if the comorbidity rate is below 20%, the coefficient is set at 0.50.

[0095] The initialization process traverses each edge in the knowledge graph, selects or calculates the corresponding damping coefficient value from a preset coefficient range based on its relation type and clinical evidence level, and stores the value in the edge's attribute field.

[0096] When extracting the clinical importance score for each source node in the source node set, it is necessary to quantify the diagnostic value of different symptoms and diseases in the current patient's condition. The clinical importance score reflects the strength of a symptom or disease's indicative role in the final diagnosis; a higher score indicates a greater weight of the node in diagnostic reasoning.

[0097] The scoring system comprehensively considers multiple dimensions of clinical information. First, there's the symptom typicality dimension. Characteristic symptoms of a particular disease (such as typical chest pain in acute myocardial infarction) receive higher importance scores, while non-specific symptoms (such as fatigue and dizziness) receive lower scores. The typicality score is calculated by analyzing the conditional probabilities of symptoms and final diagnoses within large-scale electronic medical record data. Calculations show that the higher the conditional probability, the higher the typicality score.

[0098] Secondly, there is the dimension of symptom severity. Severity descriptive words extracted from medical record text (such as "intense," "persistent," and "progressively worsening") will increase the importance score, while mild symptoms (such as "occasional" and "mild") will have lower scores. Severity is identified by natural language processing technology using degree adverbs and modifiers in medical records and mapped to a quantitative score of 0 to 1.

[0099] Thirdly, there is the time dimension. For chief complaints, the importance score is higher than for past symptoms; for recently diagnosed diseases, the score is higher than for chronic diseases. The time factor is evaluated through an exponential decay function. Calculation, where It represents the number of days since the current time. The fourth dimension is the patient's individual characteristics, considering the impact of factors such as age, gender, and underlying diseases on the importance of symptoms.

[0100] Based on the above dimensions, the formula for calculating the clinical importance score is as follows: ,in For typicality scoring, Severity rating, As a time factor, Adjustment factor for individual characteristics The weight coefficients for each dimension are determined through regression fitting using training data annotated by clinical experts. For each node in the source node set, relevant information is extracted from the patient's medical records based on its corresponding symptom or disease entity, and then substituted into the above formula to calculate the clinical importance score. The score value is normalized to between 0 and 1.

[0101] By mapping clinical importance scores to the initial semantic scalar magnitude of each source node, a transformation mechanism from clinical evaluation to graph computation parameters is established. The initial semantic scalar magnitude serves as the starting input for graph convolution iterative computation, directly determining the initial intensity distribution of semantic propagation.

[0102] The mapping process employs a strategy combining linear transformation and nonlinear adjustment. First, the clinical importance score is... The basic mapping formula is used to transform the target amplitude range through linear mapping. ,in and These are the minimum and maximum values ​​of the initial semantic scalar magnitude, respectively, typically set to 0.3 and 1.0, to prevent any source node from having an initial magnitude that is too low and being ignored.

[0103] After linear mapping, for nodes with particularly high clinical importance scores (such as...) The magnitude is further enhanced by applying a nonlinear enhancement function, which takes the form of a power function. This causes high-scoring nodes to increase their amplitude faster, highlighting their dominant role in reasoning. For nodes with lower clinical importance scores (such as...), Applying the decay function Appropriate compression should be implemented to avoid secondary symptoms interfering too much with reasoning.

[0104] After mapping, the initial semantic scalar magnitudes of all source nodes are normalized so that the sum of the magnitudes of all source nodes equals the preset total semantic quantity (e.g., the number of source nodes), ensuring the conservation of the total quantity during semantic propagation. The normalization formula is as follows: ,in The number of source nodes. For nodes The amplitude before normalization This is the normalized amplitude.

[0105] The calculated initial semantic scalar magnitude is assigned to the corresponding source node, and the "semantic scalar" attribute field of the node is updated in the graph database. For non-source nodes, their initial semantic scalar magnitude is uniformly set to 0, indicating that these nodes do not carry any semantic information in the initial state and need to iteratively calculate the semantics received from the source node.

[0106] When constructing the semantic diffusion matrix of a graph structure based on the semantic transit damping coefficient, the graph's topological structure and edge attribute information are integrated into a unified mathematical representation. The semantic diffusion matrix is ​​the core data structure for iterative computation of graph convolutional networks, describing the semantic transitive relationships between each pair of nodes in the knowledge graph.

[0107] Assuming the knowledge graph contains Nodes, semantic diffusion matrix for A square matrix, matrix elements Indicates from node To the node The semantic transfer coefficients. The matrix construction process traverses all edges in the knowledge graph, for each node... Pointing to node For a directed edge, the semantic transmission damping coefficient of that edge. Assigning values ​​to matrix elements If node and nodes If there is no direct edge connecting them, then For undirected edges, it is necessary to set them simultaneously in the matrix. and Both values ​​are the damping coefficients of that side.

[0108] To prevent excessive semantic aggregation at highly connected nodes, the matrix is ​​normalized. The normalized matrix elements are: , upcoming node The sum of the propagation coefficients to all its neighbors is normalized to 1, making the node... The semantic scalar is proportionally distributed among neighboring nodes during propagation. Degree normalization ensures the numerical stability of the semantic propagation process, avoiding semantic explosion or disappearance problems caused by differences in node degree.

[0109] For isolated nodes (i.e., nodes without any edges connecting them), the corresponding matrix rows and columns are all 0, indicating that the node does not participate in semantic propagation. The constructed semantic diffusion matrix is ​​stored in sparse matrix format because medical knowledge graphs are usually sparse graphs, and most node pairs do not have direct connections. Sparse storage can significantly reduce memory usage and computational overhead.

[0110] The iterative diffusion calculation of the initial semantic scalar magnitude along the semantic diffusion matrix simulates the propagation process of semantic information in a knowledge graph. The iterative calculation is implemented using matrix multiplication, and at the... In each iteration, the semantic scalar vector of all nodes Through formula Update, in which This is the transpose of the semantic diffusion matrix. This is the semantic scalar vector from the previous iteration.

[0111] The matrix transpose operation allows semantics to propagate from predecessor nodes to successor nodes, consistent with the propagation direction of directed edges in a knowledge graph. Initially, Let be the initial semantic scalar magnitude vector of the source node, and the elements of non-source nodes are 0. After each iteration, the semantic scalar value of each node is equal to the sum of the semantic scalar values ​​of all its predecessor nodes multiplied by the transit coefficients of the corresponding edges, i.e. .

[0112] This computational process allows the semantic information of the source node to gradually diffuse outwards along the edge relationships of the graph. After multi-hop propagation, nodes farther from the source node can also accumulate semantic contributions from the source node. To monitor the convergence state of the iterative process, the mutation rate of the global node scalar is calculated after each iteration, defined as... That is, the average change in semantic scalar of all nodes in two adjacent iterations.

[0113] The mutation rate reflects the rate of change in the semantic distribution. When the mutation rate gradually decreases and approaches 0, it indicates that the semantic distribution is stabilizing, and the semantic scalars of each node no longer change significantly. Iterative computations are set with a maximum number of iterations (e.g., 100 iterations) as a termination condition to prevent infinite iterations caused by abnormal graph structures. In actual computations, the mutation rate typically decreases below the convergence threshold after 10 to 30 iterations.

[0114] When the mutation rate of the global node scalar is lower than the preset stability convergence factor, the semantic distribution is determined to have reached a steady state, and the convergent semantic features of the target node in the steady state are extracted. The stability convergence factor is a threshold parameter for determining the termination of the iteration, usually set to 0.001 or 0.0001. This means that when the average change of the semantic scalar of all nodes is less than this threshold, the semantic distribution is considered to be sufficiently stable, and continuing the iteration will not bring significant semantic changes.

[0115] After each iteration, the calculated mutation rate will be... With stable convergence factor If a comparison is made, If a steady state is reached, the iterative computation terminates. (Steady state semantic scalar vector) It represents the final distribution state of semantic information after it has been fully propagated in the knowledge graph. The scalar value of each node reflects the semantic strength of that node from all source nodes.

[0116] For the target node (i.e., the node corresponding to the project to be inspected), extract its semantic scalar value in steady state. This value is the convergent semantic feature of the target node. The magnitude of the convergent semantic feature reflects the strength of the association between the target node and the source node set. The larger the value, the stronger the association, indicating that there is a strong reasoning path support between the item to be examined and the patient's symptoms and disease; the smaller the value, the weaker the association, indicating that the item to be examined is not sufficiently correlated with the patient's condition.

[0117] In addition to the scalar value itself, the semantic source distribution of the target node can also be extracted. That is, the proportion of the semantic scalar of the target node comes from each source node. By tracing back the propagation path in the iteration process, the proportion of each source node's semantic contribution to the target node can be calculated. This source distribution information provides a detailed basis for the subsequent generation of inference chains.

[0118] This embodiment effectively solves the core technical problem of how to achieve accurate multi-hop semantic propagation in medical knowledge graphs to discover implicit clinical associations through the above steps. The steps of obtaining edge attributes and initializing damping coefficients differentiate the different transmission characteristics of priming relationships, diagnostic means relationships, and concurrent relationships, and set differentiated damping coefficients for them. This ensures that graph convolution calculations accurately reflect the true clinical reasoning logic, avoiding reasoning distortion caused by all edge weights being the same. The steps of extracting clinical importance and mapping initial amplitude quantify the diagnostic value of symptoms and diseases from multiple dimensions and transform them into initial semantic scalars, ensuring that reasoning starts from the most important clinical clues and avoiding the blurring of reasoning focus caused by treating all source nodes equally. The step of constructing the diffusion matrix ensures the numerical stability of the semantic propagation process through degree normalization, avoiding semantic explosion or disappearance caused by differences in node degree. The iterative diffusion calculation step achieves multi-hop semantic propagation through multiple rounds of matrix multiplication, enabling target nodes to aggregate semantic contributions from multiple source nodes transmitted through different paths, thereby effectively discovering long-distance implicit associations. By determining the convergence state and extracting the steady-state semantic scalar of the target node, the complex multi-hop reasoning process is condensed into quantitative features, providing a reliable basis for subsequent path extraction and rationality judgment. The entire technical solution forms a complete computational process from graph structure to semantic propagation and feature extraction, realizing accurate multi-hop reasoning in complex medical knowledge graphs. It effectively solves the problem that traditional methods cannot discover cross-disciplinary and long-distance implicit clinical connections, providing core technical support for reducing the misjudgment rate of over-examination and improving the efficiency of clinical diagnosis and treatment.

[0119] In one embodiment of this example, refer to Figure 3 The main propagation links from the source node set to the target node are extracted based on convergent semantic features, and a locally connected subgraph is generated based on the main propagation links, including the following steps: S410, Tracing back the scalar source path of the convergent semantic features of the target node in steady state; S420. Calculate the path decay integral for each scalar source path; S430. Sort all scalar source paths in ascending order according to the path attenuation integral, and extract the paths arranged in the preset first percentage interval as candidate transmission links. S440. Count the cross-reference frequency of nodes in the candidate propagation links, remove isolated paths with zero cross-reference frequency, and use the remaining paths as the main propagation links. S450. Merge the node and edge relationships in all major transmission links to construct a locally connected subgraph.

[0120] When tracing the scalar source path of the converged semantic features of the target node in steady state, it is necessary to reverse-track the propagation trajectory of semantic information during the graph convolution iteration process. During the iterative diffusion calculation, detailed information of the semantic scalar received by each node in each iteration is recorded, including the identifier of the source node, the semantic quantity transmitted, and the identifier of the edges traversed during transmission.

[0121] The backtracking process begins with the target node. First, the semantic scalar value of the target node in the last iteration (steady state) is extracted. This value is formed by the sum of the semantic quantities passed down from all its predecessor nodes in the previous iteration. The identifiers of these predecessor nodes and their respective contributions are obtained by querying the iteration record table. For each predecessor node, the semantic source in the previous iteration is recursively traced back until a source node in the source node set is reached.

[0122] This forms a complete propagation path from the source node to the target node. The path consists of a series of nodes and directed edges connecting them, represented as follows: ,in As the source node, For the target node, The edges connecting the nodes. This represents the path length.

[0123] Due to the multi-hop propagation characteristics of graph convolutional networks, there may be a large number of different paths from the source node set to the target node, including source nodes with different starting points, different sequences of intermediate nodes, and different path lengths. The backtracking algorithm adopts a depth-first search strategy, systematically traversing all possible backtracking branches to ensure that no valid propagation path is missed.

[0124] To improve backtracking efficiency, a path length limit is set (e.g., a maximum of 10 hops). Paths exceeding this length will not continue backtracking, as excessively long paths typically indicate that semantic transmission has been excessively attenuated, contributing negligibly to the target node. The backtracking process also needs to handle potential loops in the graph structure. When a duplicate node is detected in a path, the backtracking of that branch is terminated to avoid getting stuck in an infinite loop.

[0125] After backtracking is complete, a set of all scalar source paths from the source node set to the target node is obtained. These paths represent various possible inference logics between patient symptoms and diseases and the items to be examined.

[0126] When calculating the path decay integral for each scalar source path, it is necessary to quantify the semantic propagation capability of the path. The path decay integral reflects the strength of semantic information retained as it propagates from the source node to the target node along the path. The larger the integral value, the stronger the propagation capability of the path, and the more significant the semantic contribution of the path to the target node.

[0127] The path attenuation integral is calculated based on the semantic transmission damping coefficients of all edges on the path, using a product approach, as shown in the formula: ,in For path Upper Damping coefficient of the strip edge, The path length (number of edges) is the damping coefficient. Since the damping coefficient ranges from 0 to 1, the product decreases rapidly as the path length increases. This is because the longer the path, the more transmission links the semantic information passes through, and the more severe the cumulative attenuation.

[0128] For example, for a path with three edges, the damping coefficients of 0.9, 0.8, and 0.85 are respectively. The attenuation integral of this path is... However, for a path with 5 edges, even if the damping coefficient of each edge is 0.8, its attenuation integral is only... It is significantly lower than the short path.

[0129] To more comprehensively assess the importance of a path, in addition to the decay integral, the initial semantic scalar magnitude of the path's source nodes can also be considered, and the combined contribution of the two can be obtained. ,in This represents the initial semantic scalar magnitude of the source node at the beginning of the path. The overall contribution considers both the path's propagation capability and the clinical importance of the source node, more accurately reflecting the path's actual contribution to the semantic convergence of the target node.

[0130] For all scalar source paths obtained through backtracking, calculate their path decay integral or comprehensive contribution for each one, and store the calculation results in association with the path identifier.

[0131] By sorting all scalar source paths in ascending order based on the path decay integral and extracting paths ranked within a preset percentage interval as candidate propagation links, a selection process from massive paths to critical paths is achieved. Due to the complexity of medical knowledge graphs and the multi-hop propagation characteristics of graph convolutional networks, there may be hundreds or even thousands of different paths from the source node set to the target node. These paths contain both clinically meaningful reasoning logic and a large number of noisy and weakly associated paths.

[0132] The sorting process arranges all paths in descending order according to their path decay integral (or overall contribution), with the path having the largest decay integral placed first, indicating its strongest transmission ability and greatest semantic contribution to the target node. The sorting algorithm uses quicksort or heapsort, with a time complexity of O(log n). ,in This represents the total number of paths.

[0133] After sorting, a filtering boundary is determined based on a preset percentage threshold (such as the top 20% or top 30%), and paths within this percentage range are extracted as candidate propagation links. For example, if there are 1000 paths in total and the percentage threshold is set to 20%, then the top 200 paths with the highest decay integrals are extracted.

[0134] Setting the pre-selection percentage threshold requires a balance between coverage and accuracy. A threshold that is too high will retain too many noisy paths, while a threshold that is too low may miss important suboptimal paths. In practical applications, the threshold can be dynamically adjusted based on the convergent semantic feature value of the target node. If the convergent semantic feature value is high (indicating strong association), the pre-selection percentage can be appropriately reduced, retaining only the optimal path; if the convergent semantic feature value is low (indicating weak association), the pre-selection percentage can be appropriately increased to retain more candidate paths and increase the chance of discovering implicit associations.

[0135] The set of candidate transmission links obtained by screening is significantly smaller than the original set of paths, but the paths that contribute the most to the semantics of the target node are retained. These paths are most likely to represent the real clinical reasoning logic.

[0136] By counting the cross-reference frequencies of nodes in the candidate conduction pathways and removing isolated paths with zero cross-reference frequencies, noisy paths were further filtered out, and consensus-based reasoning logic was extracted. Although the candidate conduction pathways are already high-quality paths selected based on conduction capacity, they may still contain some isolated paths that do not intersect with other paths. These isolated paths are often random connections or rare marginal associations in the graph and do not have general clinical significance.

[0137] Cross-reference frequency statistics analyze the overlap of nodes between different pathways to identify nodes that are traversed by multiple pathways. These high-frequency nodes typically represent key pathological mechanisms or diagnostic steps and have higher clinical reliability. The statistical process traverses the candidate pathway set, and for each node, calculates how many different paths pass through it, recording this count as the node's cross-reference frequency.

[0138] For example, if a node representing "coronary atherosclerosis" is traversed by 30 out of 50 candidate paths, its cross-reference frequency is 30. For each candidate path, the cross-reference frequency of all its intermediate nodes (excluding the starting source node and the ending target node) is checked. If the cross-reference frequency of all intermediate nodes in a path is 1 (i.e., these nodes only appear in this path and not in any other path), then the path is determined to be an isolated path.

[0139] Isolated paths are characterized by their reasoning logic not being supported by other paths. They may be noisy connections in the atlas or extremely rare clinical scenarios. Removing them can improve the reliability of the reasoning. After removing isolated paths, the remaining paths are the main conduction links. These paths not only have strong semantic conduction capabilities, but their reasoning logic is also cross-validated by multiple paths, resulting in higher clinical credibility. The number of main conduction links is typically between a few and dozens, representing an order-of-magnitude reduction compared to the original hundreds or thousands of paths.

[0140] When merging the node and edge relationships in all major transmission links to construct a locally connected subgraph, the discrete set of paths is transformed into a unified graph structure representation. Although the major transmission links already represent the key reasoning paths from the source node to the target node, these paths exist in the form of linear sequences, which is not convenient for structured analysis and pattern matching.

[0141] The locally connected subgraph integrates all major transmission links to form a subgraph structure containing all relevant nodes and edges. This subgraph is a local view of the original medical knowledge graph, specifically describing the network of associations between the current patient's condition and the items to be examined.

[0142] The construction process first initializes an empty graph structure, then traverses all major transmission links. For each path, it extracts all nodes and edges it contains. For a node, it checks whether the node already exists in the locally connected subgraph. If it does not exist, it is added to the node set of the subgraph, and the node's attribute information (such as node type, standard name, medical code, etc.) in the original knowledge graph is copied. If it already exists, it is skipped to avoid duplicate addition.

[0143] For edges, check if the edge already exists in the subgraph. If it does not exist, add it to the edge set of the subgraph and copy the edge's attribute information (such as relationship type, damping coefficient, evidence level, etc.). If it already exists, update the edge's weight information, for example, by accumulating the frequency of the edge's appearance in different paths as an indicator of the edge's importance.

[0144] After traversing all major transmission links, the locally connected subgraph contains all relevant nodes and edges, forming a connected graph structure. The topological properties of this subgraph reflect the complexity of the patient's condition. If the subgraph contains multiple source nodes and multiple intersecting paths, it indicates that the patient has a complex combination of multiple symptoms or diseases; if the subgraph presents a linear chain structure, it indicates that the reasoning logic is relatively simple and direct.

[0145] The construction of locally connected subgraphs also requires topology optimization, such as removing dangling nodes with a degree of 1 (these nodes are only connected to one neighbor and contribute little to the connectivity of the subgraph) or merging redundant nodes with the same set of neighbors. The optimized locally connected subgraph retains the key reasoning structure while maintaining good simplicity and readability. Locally connected subgraphs are stored in a graph database or adjacency list for easy subsequent subgraph isomorphic matching and visualization.

[0146] In this embodiment, the scalar source path tracing step transforms the black-box process of graph convolution computation into a traceable set of paths through a reverse tracing mechanism, providing a data foundation for the interpretability of the reasoning process. The path decay integral calculation step provides a scientific basis for ranking path importance by quantitatively evaluating the semantic transmission capability of paths, avoiding biases caused by simply filtering by path length or hop count. The candidate link sorting and filtering step simplifies the massive number of paths into high-quality candidate paths through descending order based on transmission capability and a pre-percentage truncation, significantly reducing the data scale while ensuring no key paths are missed. The cross-reference frequency filtering step identifies and eliminates isolated noisy paths through a consensus verification mechanism, preserving reasoning logic supported by multiple cross-paths, significantly improving the reliability and clinical credibility of the reasoning. By integrating discrete paths into a unified graph structure, the transformation from linear reasoning chains to a networked association pattern is achieved, providing a structured representation for subsequent pathological structure matching. The entire technical solution forms a complete conversion process from graph convolution calculation results to structured reasoning patterns, which not only ensures the representativeness and reliability of the reasoning path, but also optimizes the computational efficiency. It effectively solves the problem that traditional methods cannot extract key reasoning logic from complex calculation results, and lays a solid foundation for subsequent pathological structure matching and natural language reasoning chain generation.

[0147] In one embodiment of this invention, a subgraph isomorphic matching is performed between the locally connected subgraph and a pre-defined complication pathology structure library to determine whether there is a specific implicit clinical association between the target node and the source node set. This includes the following steps: S510. Obtain a set of baseline pathological topology maps of confirmed complex complication cases from the complication pathology structure library; S520. Extract graph features from the locally connected subgraph to obtain the topological feature vector of the locally connected subgraph; S530. Extract graph features from each benchmark pathological topology graph in the benchmark pathological topology graph set to obtain the benchmark topology feature vector; S540. Calculate the structural similarity between the topological feature vectors of the locally connected subgraphs and each reference topological feature vector; S550. When the structural similarity is greater than the preset isomorphic matching threshold, it is determined that the local connected subgraph and the corresponding baseline pathological topology graph have an isomorphic relationship. S560. Extract the specific latent attributes marked in the baseline pathological topology map that has isomorphic relationship, and confirm the specific latent clinical association between the target node and the source node set based on the specific latent attributes.

[0148] When retrieving a set of baseline pathological topologies for confirmed complex complication cases from the Complication Pathology Structure Database, it is necessary to load clinically validated typical pathological association patterns. The Complication Pathology Structure Database is a knowledge base built by analyzing a large number of real complex cases, specifically including typical diagnostic and treatment pathways for complex clinical scenarios involving multiple disease complications, screening for rare complications, and interdisciplinary pathological mechanisms.

[0149] These cases typically originate from hospital discussions of difficult cases, multidisciplinary consultation records, and complex cases reported in medical literature. They undergo review and annotation by a team of clinical experts to ensure their clinical authenticity and representativeness. Each baseline pathology topology diagram represents a specific complex pathological association pattern in a graph structure. Nodes in the diagram represent symptoms, diseases, pathological mechanisms, or examination items, while edges represent the clinical relationships between them.

[0150] For example, a baseline topology might describe a typical complication diagnosis and treatment pathway of "diabetes - microvascular disease - retinopathy - fundus fluorescein angiography", which includes nodes for diabetes, microvascular disease, retinopathy and fundus fluorescein angiography, as well as relational edges connecting them such as "cause", "manifests as", "requires diagnosis through...".

[0151] Each baseline topology map also includes tagging information, including metadata such as the type of disease involved, complication category, clinical department, level of evidence, and case source, as well as specific implicit attribute tags, such as "rare complication," "cross-departmental association," and "multi-system involvement." The complication pathology structure database is stored using a graph database, supporting efficient retrieval and matching based on the graph structure.

[0152] The acquisition process loads all baseline pathological topologies from the database via a query interface, or pre-filters based on the current patient's disease type, loading only baseline topologies for relevant disease areas to improve matching efficiency. The loaded set of baseline pathological topologies typically contains hundreds to thousands of different pathological patterns, covering common complex complication scenarios in clinical practice.

[0153] When extracting graph features from locally connected subgraphs to obtain their topological feature vectors, the structural information of the graph needs to be encoded into a fixed-dimensional vector representation. The goal of graph feature extraction is to capture the topological characteristics of the subgraph, including node connection patterns, degree distribution, path structure, clustering coefficients, and other graph-theoretic features, so that graphs with similar structures have similar feature vectors, and graphs with large structural differences have significantly different feature vectors.

[0154] Feature extraction employs a graph isomorphic network architecture from graph neural networks. This architecture is specifically designed to distinguish different graph structures and possesses strong expressive power. First, initial feature encoding is performed on each node in the locally connected subgraph. The initial features of each node include a one-hot encoding of the node type (e.g., symptom class, disease class, examination class, etc.), the node's degree, and the node's centrality index within the subgraph. These features are then concatenated to form the initial feature vector of the node. .

[0155] Then, the node features are updated through multi-layer graph convolution operations, at the... Layers, nodes The features are updated by aggregating the features of its neighboring nodes, and the update formula is: ,in For nodes The neighborhood group, The parameters are learnable; MLP stands for Multilayer Perceptron.

[0156] go through After iterating through layers (usually 3 to 5 layers), the feature vector of each node... Includes the node and its Structural information of skip neighborhoods. To obtain the global feature vector of the entire subgraph, the features of all nodes are aggregated using summation pooling. ,in Let be the set of nodes in the subgraph. Summation pooling preserves the permutation invariance of the graph, meaning that changes in the order of nodes do not affect the final graph eigenvectors.

[0157] Received This refers to the topological feature vector of a locally connected subgraph, typically with 128 or 256 dimensions. Graph isomorphic network models are pre-trained on large-scale graph datasets to learn general graph structure representation capabilities, and then fine-tuned on subgraph data of medical knowledge graphs to capture graph structure patterns specific to the medical field.

[0158] For each benchmark pathological topology graph in the benchmark pathological topology graph set, graph feature extraction is performed to obtain the benchmark topology feature vector. The same feature extraction method as for locally connected subgraphs is used to ensure the comparability of feature vectors. For each benchmark pathological topology graph in the complication pathological structure library, the same graph isomorphic network model is used for feature extraction, encoding the structural information of the benchmark topology graph into a fixed-dimensional vector representation.

[0159] Since the number of baseline pathological topologies is large and their structure is relatively stable, batch feature extraction can be performed on all baseline topologies during the system initialization phase. The extracted feature vectors are pre-calculated and stored in the database, avoiding repeated calculations during each matching and improving system response speed. The pre-calculated baseline topology feature vectors are organized in the form of a vector database, supporting efficient vector retrieval and similarity calculation.

[0160] For the baseline topology graph newly added to the complication pathology structure database, feature extraction and storage of its feature vectors are performed upon database entry. The feature extraction process is completely consistent with that of the locally connected subgraph, including steps such as initial node feature encoding, multi-layer graph convolution update, and summation pooling aggregation, ensuring that the feature vectors of the locally connected subgraph and the baseline topology feature vectors are in the same feature space and are comparable.

[0161] The baseline topological feature vector not only contains structural information of the graph but can also incorporate metadata information from the baseline topological graph. For example, information such as disease types and clinical departments can be encoded as additional feature dimensions and concatenated into the structural feature vector to form an enhanced feature representation. This enhanced feature can further consider semantic similarity on top of structural matching, improving matching accuracy.

[0162] When calculating the structural similarity between the topological feature vectors of a locally connected subgraph and each baseline topological feature vector, it is necessary to quantify the degree of similarity between the two graph structures. Structural similarity is the core indicator for determining whether a locally connected subgraph is isomorphic to a certain baseline pathological topological graph. The higher the similarity, the closer the structures of the two graphs are, and the more likely they are to represent the same pathological association pattern.

[0163] Similarity calculation employs a combination of multiple metrics, starting with cosine similarity based on feature vectors, calculated using the following formula: ,in The feature vectors of a locally connected subgraph. The base topological feature vectors are used. The cosine similarity value ranges from -1 to 1; the closer the value is to 1, the more consistent the directions of the two vectors, and the more similar the corresponding graph structures. Cosine similarity is simple and efficient to calculate, making it suitable as a preliminary screening metric to quickly filter out obviously dissimilar base topological graphs.

[0164] For candidate benchmark topologies with high cosine similarity, a more precise similarity calculation based on graph edit distance is employed. Graph edit distance is defined as the minimum number of edit operations required to transform one graph into another. Edit operations include node insertion, node deletion, edge insertion, and edge deletion, each with a unit cost. Calculating graph edit distance is an NP-hard problem, and an A* algorithm is used. The search algorithm uses an approximate solution method to obtain an approximate optimal solution within an acceptable time.

[0165] Get the graph edit distance Then, the structural difference degree is calculated. ,in Let be the number of nodes in the two graphs. Let be the number of edges. Normalization is applied to ensure the dissimilarity is between 0 and 1. Considering the impact of graph complexity on matching difficulty, a penalty coefficient is introduced. The difference item was corrected as follows: Structural similarity is calculated as follows: .

[0166] To combine the advantages of the two similarity measures, a weighted average was used for the final structural similarity. Weight Based on experimental optimization, values ​​are typically set to 0.3 and 0.7, with a greater emphasis on accurate graph editing distance metrics. Similarity is calculated between locally connected subgraphs and each baseline topological feature vector, resulting in a set of similarity scores, with each score corresponding to a baseline pathological topological graph.

[0167] When the structural similarity exceeds a preset isomorphic matching threshold, the locally connected subgraph is determined to have an isomorphic relationship with the corresponding baseline pathological topology graph. The isomorphic matching threshold serves as the decision boundary for determining whether two graphs represent the same pathological pattern. The threshold setting needs to balance the sensitivity and specificity of the matching. A threshold that is too low may lead to graphs with significant structural differences being classified as isomorphic, resulting in false positives and potentially mistaking noisy paths for genuine pathological associations. A threshold that is too high may cause truly isomorphic graphs to be missed, resulting in false negatives and failing to identify genuine latent clinical associations.

[0168] The isomorphic matching threshold is typically set between 0.75 and 0.85, with the specific value optimized based on clinical validation data. In practical applications, a dynamic threshold strategy can be adopted, adjusting the threshold according to the complexity of the locally connected subgraph and the convergent semantic feature value. For subgraphs with simple structures and high semantic convergence strength, a higher threshold (e.g., 0.85) is used to improve matching accuracy; for subgraphs with complex structures and moderate semantic convergence strength, a lower threshold (e.g., 0.75) is used to increase the chance of matching.

[0169] The determination process iterates through the similarity scores of all baseline pathological topologies. If the similarity score of a baseline topology is greater than a threshold, the locally connected subgraph is determined to have an isomorphic relationship with that baseline topology. There may be cases where multiple baseline topologies have similarity scores exceeding the threshold. In such cases, the baseline topology with the highest similarity score is selected as the best match, or all baseline topologies exceeding the threshold are retained as a candidate matching set for further analysis.

[0170] The determination of isomorphism is based not only on structural similarity but also on semantic consistency checks. For example, it verifies whether the node types in a locally connected subgraph match the node types in the baseline topology graph, and whether the edge relationship types are consistent. Semantic consistency checks are performed by comparing the attribute labels of nodes and edges. If there is a type mismatch, even if the structural similarity is high, it is not considered isomorphic.

[0171] By extracting specific latent attributes from the baseline pathological topology map that exhibits isomorphic relationships, and confirming the existence of specific latent clinical associations between the target node and the source node set based on these attributes, the transformation from structural matching to clinical semantic confirmation is completed. Specific latent attributes are clinical labels attached to the baseline pathological topology map, describing the special properties and clinical significance of the pathological pattern.

[0172] Common specific latent attributes include: "Rare complication screening," indicating that the pattern involves complications with low incidence but serious clinical consequences, requiring screening through specific examinations; "Cross-departmental pathological association," indicating that the pattern involves disease interaction across multiple clinical departments, requiring interdisciplinary diagnostic considerations; "Multi-system involvement," indicating that the pattern involves simultaneous lesions in multiple organ systems; "Adverse drug reaction monitoring," indicating that the pattern is associated with monitoring the side effects of specific drugs; and "Hereditary disease screening," indicating that the pattern involves screening for family history of hereditary diseases, etc. These attribute tags are annotated by clinical experts when constructing the complication pathology structure database, reflecting the clinical specificity of this pathological pattern.

[0173] The extraction process reads specific latent attribute fields from the metadata of the successfully matched baseline pathological topology map and obtains all labeled attribute tags. Based on these attribute tags, it confirms that there is a specific latent clinical association between the target node (the item to be examined) and the source node set (patient symptoms and diseases). That is, although the examination item does not have a direct first-order association with the patient's chief symptoms or primary diagnosis, it has clinical necessity and rationality based on complex complication mechanisms, rare pathological pathways, or interdisciplinary associations.

[0174] The confirmation results contain two parts of information: first, the association type, i.e., specific latent attribute labels; and second, the source of evidence, i.e., the case sources and literature support for the matched baseline pathological topology. This information provides detailed clinical evidence for the subsequent generation of natural language inference chains, enabling the system to explain to doctors why the examination is appropriate in complex cases.

[0175] This embodiment effectively solves the core technical problem of how to verify whether automatically discovered inference paths correspond to real clinical pathological mechanisms. By introducing a clinically validated pathological pattern library, an authoritative reference standard is provided for inference path verification, enabling the system to distinguish between true implicit associations and graph noise. A unified graph neural network encoding method is used to transform complex graph structures into comparable vector representations, establishing the mathematical foundation for structural matching. A combination of cosine similarity and graph edit distance measures ensures both computational efficiency and matching accuracy. Threshold determination and semantic consistency verification ensure the reliability of matching results and effectively avoid false positives. By extracting specific latent attribute labels, structural matching results are transformed into clinical semantic confirmation, completing the crucial transformation from mathematical calculation to medical judgment and providing clinical evidence for inference chain generation. The entire technical solution, through comparison and verification with real pathological patterns, effectively distinguishes between reasonable latent associations and random noise paths, significantly reducing the probability of misjudging atypical but reasonable examinations as over-examination, and significantly improving the system's accuracy in complex complication scenarios.

[0176] In one embodiment of this invention, calculating the structural similarity between the topological feature vectors of the locally connected subgraphs and each reference topological feature vector includes the following steps: S610. Calculate the minimum number of editing operations between the locally connected subgraph and the baseline pathological topology graph, wherein the editing operations include node insertion, node deletion, edge insertion, and edge deletion; S620. Calculate the structural difference degree based on the minimum number of editing operations; S630. Convert the structural difference degree into the structural similarity degree, where the structural similarity degree and the structural difference degree are negatively correlated.

[0177] In this embodiment, calculating the minimum number of edit operations between the locally connected subgraph and the baseline pathological topology graph includes: Traverse the node space of the locally connected subgraph and the baseline pathological topology graph to identify nodes of mismatched types, and record the nodes of mismatched types as the cost of node replacement operation. Identify missing or redundant edge connections and record them as the cost of topology repair operations. The minimum number of edit operations is obtained by summing the costs of node replacement and topology repair operations.

[0178] The structural difference degree is calculated based on the minimum number of editing operations, including: Obtain the total number of nodes and edges of the locally connected subgraph as the normalization base; The structural difference is obtained by dividing the minimum number of editing operations by the normalized base.

[0179] Calculating the minimum number of edit operations required to transform a locally connected subgraph into a baseline pathological topology graph requires quantifying the topological differences between the two graph structures. Graph edit distance is a classic metric in graph theory for measuring the similarity between two graph structures. It reflects their structural differences by calculating the minimum number of edit operations required to transform one graph into another.

[0180] Editing operations include four basic types: node insertion, node deletion, edge insertion, and edge deletion. A node insertion operation adds a new node to the graph, a node deletion operation removes an existing node, an edge insertion operation adds a new edge between two nodes, and an edge deletion operation removes an existing edge between two nodes. In the context of medical knowledge graphs, these operations have clear clinical significance: node insertion or deletion indicates the addition or removal of a symptom, disease, or examination step in a pathological pathway, while edge insertion or deletion indicates the addition or removal of a clinical association.

[0181] Calculating the minimum number of edit operations requires finding a sequence of edit operations that transforms a locally connected subgraph into a baseline pathological topology with the fewest total operations. Approximate algorithms or heuristic search methods are employed. Commonly used methods include A... Search algorithms, through the design of reasonable heuristic functions and pruning strategies, can obtain near-optimal solutions in polynomial time.

[0182] In practical calculations, the first step is to align the nodes of the two graphs, attempting to find the optimal matching relationship between nodes in the locally connected subgraph and nodes in the baseline pathological topology graph. Node alignment is based on the similarity of node attributes, including a comprehensive matching of node type (symptom, disease, examination, etc.), node label (medical terminology), and the node's structural role in the graph (degree, centrality, etc.).

[0183] For node pairs that can be successfully aligned, no node insertion or deletion operations are needed; only the consistency of their edge connections needs to be checked. For nodes that cannot be aligned, node insertion or deletion operations need to be counted. Edge matching is based on the aligned node pairs, checking whether there is an edge connection between corresponding node pairs in the two graphs. If it exists in one graph but not in the other, edge insertion or deletion operations are required. By systematically traversing all nodes and edges and accumulating the required number of edit operations, an approximate value for the minimum number of edit operations is obtained.

[0184] In practice, a step-by-step strategy is adopted to improve computational efficiency and accuracy when calculating the minimum number of editing operations. First, the node space of the locally connected subgraph and the baseline pathological topology graph is traversed to identify nodes with mismatched types. Node type is a core attribute of a node, and in the medical knowledge graph, it includes categories such as symptom nodes, disease nodes, examination nodes, and pathological mechanism nodes.

[0185] The traversal process employs a bidirectional matching strategy. For each node in the locally connected subgraph, it searches for nodes of the same type and with the most similar attributes in the baseline pathological topology graph for pairing. Attribute similarity is calculated through the semantic similarity of node labels. A medical term vector model is used to encode the medical terminology names of nodes into vectors, and the cosine similarity between vectors is calculated. If the similarity exceeds a threshold (e.g., 0.8), the two nodes are considered to be alignable; if the similarity is below the threshold or no candidate node of the same type can be found, the node cannot be aligned and needs to be recorded as a node replacement operation.

[0186] The cost of a node replacement operation depends on the node's importance in the graph. Core nodes with high degrees have higher replacement costs, while edge nodes with low degrees have lower replacement costs. The total cost of a node replacement operation is obtained by summing the replacement costs of all misaligned nodes.

[0187] Next, we identify missing or redundant edges. After node alignment, we check the edge connections between corresponding node pairs in the locally connected subgraph and the baseline pathology topology graph. Edges that exist in the locally connected subgraph but not in the baseline pathology topology graph are recorded as redundant edges and need to be deleted. Edges that exist in the baseline pathology topology graph but not in the locally connected subgraph are recorded as missing edges and need to be inserted.

[0188] The operation cost of an edge is set according to the relationship type of the edge. For strong relationships (such as diagnostic methods), the operation cost is higher; for weak relationships (such as concurrency relationships), the operation cost is lower. The operation costs of all missing and redundant edges are summed to obtain the topology repair operation cost. Finally, the node replacement operation cost and the topology repair operation cost are summed to obtain the total minimum number of edit operations.

[0189] When calculating structural dissimilarity based on the minimum number of edit operations, the edit distance needs to be normalized to eliminate the influence of graph size differences. Different locally connected subgraphs and baseline pathological topologies may have different numbers of nodes and edges. If the number of edit operations is directly used as the dissimilarity index, even if the graphs are structurally similar, large-scale graphs may produce large edit distances, leading to evaluation bias.

[0190] Normalization is achieved by dividing the number of edit operations by the graph's size index to obtain the relative dissimilarity. First, the total number of nodes in the locally connected subgraph is obtained. Total number of sides and the total number of nodes in the baseline pathology topology graph. Total number of sides .

[0191] There are several options for choosing the normalization cardinality. Commonly used options include: Option 1, using the sum of the number of nodes and edges in the two graphs as the normalization cardinality, i.e. The first approach takes into account the overall size of the graph; the second approach uses the sum of the number of nodes and edges of the larger of the two graphs, i.e. This scheme avoids normalization bias when matching small graphs with large graphs; Scheme 3 uses only the sum of the number of nodes as the normalization base, i.e. This scheme focuses on the differences in node structure.

[0192] In medical knowledge graph applications, Option 1 is more suitable because nodes and edges are equally important for describing pathological paths. The minimum number of edit operations should be considered. Divide by the normalized base To obtain the structural difference degree Theoretically, the structural dissimilarity score ranges from 0 to 1, where 0 indicates that the two graphs are completely identical, and 1 indicates that the two graphs are completely different. In actual calculations, due to the cost settings of editing operations and the sparsity of graphs, the dissimilarity score is usually between 0 and 0.5.

[0193] Converting structural dissimilarity to structural similarity establishes a mapping from a measure of dissimilarity to a measure of similarity. Structural dissimilarity reflects the degree of difference between two graphs; a larger value indicates a greater dissimilarity. Structural similarity, on the other hand, reflects the degree of similarity between two graphs; a larger value indicates a higher similarity. There is a negative correlation between the two: the greater the dissimilarity, the smaller the similarity.

[0194] The simplest transformation method is linear mapping, and the formula is: ,in For structural similarity, The structural dissimilarity level is used. This method is simple and intuitive; when the dissimilarity level is 0, the similarity level is 1, and when the dissimilarity level is 1, the similarity level is 0. However, linear mapping has limitations. It responds uniformly to changes in dissimilarity level, failing to highlight subtle differences in high-similarity intervals or suppress noise in low-similarity intervals.

[0195] To improve the conversion effect, a nonlinear mapping method is adopted, introducing an exponential function for smooth conversion. The exponential mapping formula is as follows: ,in The smoothing parameter controls the steepness of the transformation curve. When When the difference is large, small changes in the degree of difference can lead to significant changes in the degree of similarity, making it suitable for fine-grained differentiation of high similarity intervals; when When the value is small, the conversion curve is relatively flat, making it suitable for robust processing of low similarity intervals.

[0196] In the application of medical knowledge graphs, Typically, the similarity is set to 3 to 5, ensuring that the similarity remains at a high level above 0.8 when the difference is between 0 and 0.2, while the similarity rapidly decays to below 0.3 when the difference exceeds 0.4, thus prioritizing the identification of high-quality matches. After conversion, the structural similarity value ranges from 0 to 1, and can be directly compared with the preset isomorphic matching threshold to determine whether an isomorphic relationship exists.

[0197] In practical applications, the calculation of structural similarity also needs to consider the impact of graph complexity on matching difficulty. Graph complexity includes multiple dimensions such as the number of nodes, the number of edges, the average degree, and the diameter of the graph. The higher the complexity of the graph, the more difficult the structural matching is, and the tolerance for differences should be increased accordingly.

[0198] To reflect this characteristic, a complexity penalty coefficient is introduced to correct the difference before converting the difference to similarity. Complexity penalty coefficient The formula is calculated based on the scale index of the graph. ,in The penalty strength parameter is typically set between 0.05 and 0.1. The use of a logarithmic function makes the penalty coefficient grow sublinearly with the graph size, avoiding over-penalization of large-scale graphs.

[0199] The corrected degree of difference is Then Substituting the values ​​into the similarity conversion formula yields the final structural similarity. The complexity penalty mechanism appropriately increases the tolerance for structural differences when matching large-scale complex graphs, avoiding matching failures caused by graph complexity and improving the ability to identify complex pathological patterns.

[0200] In addition, a weighting mechanism can be introduced to address the differences in edge attributes. Different types of edges are assigned different weights in editing operations. The editing cost of strongly correlated edges (such as diagnostic methods) is higher than that of weakly correlated edges (such as concurrency relationships), so that the editing distance can more accurately reflect the structural differences in clinical practice.

[0201] This embodiment effectively solves the core technical problem of how to accurately quantify the similarity between two graph structures. By employing a graph edit distance algorithm, abstract graph structural differences are transformed into quantifiable edit operation counts, providing an objective basis for difference measurement. Separating node matching and edge matching and introducing differentiated operation costs ensures that the edit distance accurately reflects the structural differences in the medical field. Normalization eliminates the influence of graph size differences, making the differences between graphs of different sizes comparable and ensuring the fairness of similarity judgment. A nonlinear mapping method transforms the difference into a similarity index that conforms to clinical judgment logic, enhancing the ability to identify high-quality matches. The introduction of a complexity penalty mechanism appropriately increases the system's tolerance for matching large-scale complex graphs, avoiding matching failures due to graph complexity. The entire technical solution forms a complete computational process from topological difference quantification to similarity assessment, ensuring mathematical rigor while incorporating medical expertise. It achieves accurate, efficient, and clinically relevant graph structure matching, providing a reliable technical guarantee for accurately identifying pathological patterns corresponding to locally connected subgraphs.

[0202] In one embodiment of this invention, converting structural dissimilarity into structural similarity includes the following steps: S710, Obtain the penalty coefficient for graph complexity; S720. Multiply the structural difference degree by the penalty coefficient to obtain the corrected difference term; S730. Subtract the correction difference term from the preset theoretical maximum similarity constant to obtain the initial similarity score; S740. Perform exponential smoothing on the initial similarity scores to obtain the structural similarity.

[0203] When obtaining the penalty coefficient for graph complexity, it is necessary to quantify the impact of the graph structure's complexity on matching difficulty. Graph complexity reflects the scale and topological complexity of the graph structure, including multiple dimensions such as the number of nodes, the number of edges, the average node degree, the graph's connectivity, and the distribution of path lengths.

[0204] In the application of medical knowledge graphs, highly complex locally connected subgraphs typically correspond to more complex patient conditions, involving multiple concurrent diseases, involvement of multiple organ systems, or interdisciplinary pathological mechanisms. When matching such complex graphs with baseline pathological topology graphs, the large number of nodes and edges, and the greater diversity of topological structures mean that even if the two graphs represent the same pathological pattern, subtle differences can lead to significant edit distances.

[0205] Without considering complexity, directly using edit distance to calculate dissimilarity would unfairly penalize complex graphs, causing genuine pathological associations to be misjudged as mismatches. The penalty coefficient is designed to moderately amplify the dissimilarity based on graph complexity, allowing complex graphs to have a higher tolerance during matching.

[0206] The penalty coefficient is calculated using a logarithmic function, as shown in the formula: ,in The number of nodes in the locally connected subgraph. The number of nodes in the baseline pathological topology graph. The penalty strength parameter is added by 2 to avoid the logarithmic function becoming unstable when the number of nodes is small.

[0207] The choice of the logarithmic function is based on the following considerations: when the graph size is small (e.g., the number of nodes is less than 10), the difference in complexity is not obvious, the penalty coefficient is close to 1, and the correction for the difference is small; when the graph size increases, the logarithmic function exhibits sublinear growth, the penalty coefficient gradually increases but the growth rate decreases, thus avoiding excessive penalty for large-scale graphs.

[0208] Penalty intensity parameters Optimization is performed based on clinical validation data, typically set between 0.05 and 0.15, with smaller values ​​being preferable. A larger value is suitable for scenarios requiring high matching accuracy. This value is suitable for scenarios where it is necessary to improve the recognition rate of complex cases.

[0209] In practical applications, the contribution of the number of edges to the penalty coefficient can also be introduced, and the modified formula is as follows: ,in Let be the number of edges in the two graphs. The penalty strength parameters are for nodes and edges. This multi-dimensional penalty coefficient design can more comprehensively reflect the complexity characteristics of the graph.

[0210] Multiplying the structural dissimilarity by a penalty coefficient to obtain the corrected dissimilarity term achieves complexity correction for the original dissimilarity. Original structural dissimilarity It is obtained by normalization using the minimum number of edit operations, reflecting the pure difference in topology between the two graphs, but does not consider the impact of graph complexity on matching difficulty.

[0211] The correction process involves multiplying by a penalty coefficient. Complexity is incorporated into the variability calculation, and the formula is as follows: Corrected differences Compared to the original difference The difference increases, and the magnitude of the increase is determined by the penalty coefficient. For simple graphs with low complexity, the penalty coefficient is close to 1, and the corrected difference term is similar to the original difference degree, so the correction effect is not obvious. For complex graphs with high complexity, the penalty coefficient is significantly greater than 1, and the corrected difference term is significantly greater than the original difference degree, reflecting the compensation for the difficulty of matching complex graphs.

[0212] The clinical significance of this correction mechanism lies in the fact that when a patient's condition is complex and involves multiple complications, the automatically generated locally connected subgraph has a complex structure. Even if there are certain structural differences when matching it with the baseline pathological topology, it should not be over-penalized, because complex conditions themselves involve individual differences and atypical manifestations. By correcting the penalty coefficient, the tolerance for matching in complex cases is increased, reducing the probability of misjudging true implicit associations as mismatches.

[0213] The numerical range of the correction difference term is typically between 1 and 1.5 times the original difference, with the specific multiple depending on the complexity of the graph. An upper limit truncation mechanism can also be introduced during the correction process to set a maximum value for the correction difference term (e.g., 1.0) to avoid overcorrection caused by extremely complex graphs.

[0214] Subtracting the correction difference term from the preset theoretical maximum similarity constant to obtain the initial similarity score completes the basic conversion from difference to similarity. The theoretical maximum similarity constant represents the upper limit of similarity when two graphs are completely identical, and is usually set to 1.0, representing 100% similarity.

[0215] The conversion formula is ,in This is the initial similarity score. This is the theoretical maximum similarity constant. To correct for the difference term, this transformation establishes a linear negative correlation between the degree of difference and the similarity: when the corrected difference term is 0, the initial similarity score reaches its maximum value of 1.0, indicating that the two images are completely identical; as the corrected difference term increases, the initial similarity score decreases linearly; when the corrected difference term is equal to or exceeds 1.0, the initial similarity score drops to 0 or a negative value, indicating that the two images are completely different.

[0216] The advantages of linear transformation are its simplicity, intuitiveness, and ability to directly reflect the impact of differences on similarity. However, linear transformation also has limitations; its response to changes in differences is uniform, failing to reflect the non-linear characteristics of similarity assessment. In real-world clinical scenarios, subtle differences in high similarity intervals (e.g., similarity above 0.8) are of significant discriminative importance, requiring more sensitive differentiation capabilities; while differences in low similarity intervals (e.g., similarity below 0.3) have a smaller impact on the final judgment and do not require excessive attention.

[0217] Theoretically, the initial similarity score ranges from negative infinity to 1.0. However, in practical applications, since the correction difference term is usually less than 1.0, the initial similarity score is mainly distributed between 0 and 1.0. For negative values, they can be truncated to 0, indicating that the two images are completely dissimilar.

[0218] By performing exponential smoothing on the initial similarity scores, nonlinear optimization of the similarity scores is achieved when obtaining structural similarity scores. The purpose of exponential smoothing is to adjust the distribution characteristics of the similarity scores, enhance the ability to distinguish high similarity intervals, and suppress the noise influence in low similarity intervals.

[0219] Exponential smoothing uses an exponential function for transformation, as shown in the formula: ,in For the final structural similarity, This is the initial similarity score. This is the smoothing intensity parameter. The formula is designed based on the following considerations: when the initial similarity score... When it is large (close to 1.0), the exponent term Approaching 0, structural similarity Approaching 1.0, it maintains the characteristic of high similarity; when the initial similarity score When it is small (close to 0), the exponential term Approaching 1, structural similarity Approaching 0 indicates low similarity; in the intermediate region, the nonlinearity of the exponential function causes the rate of change of similarity to vary with the initial score.

[0220] Smoothness intensity parameter The steepness of the nonlinear transformation was controlled. The larger the value, the stronger the ability to distinguish high-similarity intervals, but the more severe the compression of low-similarity intervals. The smaller the value, the smoother the transformation curve, approaching a linear relationship. In the application of medical knowledge graphs, Typically, the similarity score is set to 2 to 5, so that when the initial similarity score is above 0.7, the structural similarity can be maintained above 0.75, meeting the requirement of the isomorphic matching threshold; while when the initial similarity score is below 0.4, the structural similarity rapidly decays to below 0.3, clearly excluding mismatches.

[0221] Exponential smoothing also has the advantages of good numerical stability and high computational efficiency, making it suitable for application in real-time systems. After exponential smoothing, the resulting structural similarity is between 0 and 1, exhibiting good numerical characteristics and discriminative ability, and can be directly used for comparison with isomorphic matching thresholds.

[0222] This embodiment effectively solves the key technical problem of how to fairly and accurately convert structural dissimilarity into structural similarity. By quantifying graph complexity using a logarithmic function, a scientific basis is provided for fair correction of dissimilarity, ensuring that graphs of different complexities are treated fairly during matching. Complexity correction is achieved by multiplying by a penalty coefficient, effectively incorporating graph complexity factors into dissimilarity calculation, improving the tolerance for matching complex cases, and reducing the misclassification rate. A basic mapping relationship between dissimilarity and similarity is established through linear subtraction, providing input for subsequent nonlinear optimization. Nonlinear transformation optimizes the distribution characteristics of similarity, enhancing the ability to distinguish high-similarity intervals while suppressing noise influence in low-similarity intervals, making similarity judgments more consistent with clinical decision-making logic. The entire technical solution considers the fairness factors of graph complexity and improves discriminative ability through nonlinear optimization, achieving scientific, fair, and accurate similarity calculation, and providing reliable quantitative indicators for accurately identifying pathological patterns corresponding to locally connected subgraphs.

[0223] In one embodiment of this example, if there is a specific implicit clinical association between the target node and the source node set, the locally connected subgraph is expanded into a text sequence of node entities and edge relationships, and the text sequence is used as a limiting context to be concatenated into a preset prompt word template to obtain a natural language inference chain, including the following steps: S810. If there is a specific implicit clinical association between the target node and the source node set, perform a depth-first search traversal on the locally connected subgraph. During the traversal, extract the medical standard names of the node entities that have been passed. S820. Extract the attribute description field of the edge connecting each node entity; S830. Following the traversal order, alternately concatenate the medical specification name and attribute description fields to generate a linear text sequence; S840. Load a preset prompt word template containing blank context slots, and concatenate the linear text sequence as a limiting context into the prompt word template to obtain a natural language inference chain. The prompt word template includes task instructions, inference constraints, and output format requirements.

[0224] If a specific implicit clinical association is confirmed between the target node and the source node set, a depth-first search traversal is performed on the locally connected subgraph, extracting the medical standard names of the traversed node entities during the traversal. Depth-first search is a classic graph traversal algorithm, characterized by prioritizing exploration along the depth direction of the graph until it cannot be explored further, at which point it backtracks to the previous level node and continues exploring other branches.

[0225] In traversing a locally connected subgraph, depth-first search can form a complete reasoning path from the source node to the target node. The coherence of this path conforms to the logical order of clinical reasoning, facilitating the subsequent generation of natural language descriptions. The traversal process begins by selecting a starting node from the set of source nodes, typically choosing the source node with the highest clinical importance score as the starting point to ensure that the reasoning chain originates from the most important clinical clues.

[0226] The traversal algorithm maintains an array of visit markers to record whether each node has been visited, avoiding duplicate visits and loops. When visiting each node, its medical standard name is extracted from the graph database. The medical standard name is the formal name of the node in the standard medical terminology system, such as "type 2 diabetes," "coronary atherosclerotic heart disease," and "coronary CT angiography." These standard names have been standardized to conform to international medical coding standards such as ICD-10 or SNOMED CT, and have clear medical meaning and clinical acceptance.

[0227] The extracted medical standard names are stored in an ordered list in traversal order, which records the complete node sequence from the source node to the target node. For locally connected subgraphs containing multiple source nodes, multiple depth-first searches may be required, starting from different source nodes to obtain multiple traversal paths.

[0228] To avoid excessively long text sequences due to too many paths, a maximum number of paths can be set (e.g., a maximum of 3 paths can be retained), or only the shortest path with the fewest nodes can be kept to ensure that the generated reasoning chain is concise and clear. The time complexity of depth-first search is O(log n). ,in For the number of nodes, The number of edges is given. For smaller graph structures like locally connected subgraphs, the traversal efficiency is very high, and can be completed in milliseconds.

[0229] When extracting the attribute description fields of edges connecting various node entities, it is necessary to obtain the semantic information of the clinical relationships represented by the edges. In medical knowledge graphs, edges not only indicate the existence of connections between nodes, but more importantly, they describe the specific nature and clinical significance of the connections. The attribute description fields of edges are stored in natural language and are used to describe the medical relationship between two nodes, such as "may cause", "is a complication of", "requires diagnosis through", "manifests as", "commonly seen in", etc.

[0230] These descriptive fields are annotated by clinical experts or automatically extracted from medical literature during the construction of the medical knowledge graph, and possess clear medical semantics. The extraction process follows a depth-first search traversal path, extracting the attribute description fields of the edges connecting adjacent nodes sequentially according to the order of node visits. For the path... Extracting edges The attribute description fields form an edge description sequence corresponding to the node sequence.

[0231] Edge attribute description fields may contain placeholders used to dynamically insert node names. For example, the placeholder in the description field "[Node A] may cause [Node B]" will be replaced with the actual node name during actual use. Extraction requires parsing these placeholders and replacing them according to the nodes connected by the edge to generate a complete relationship description. For instance, if an edge connects the nodes "Type 2 Diabetes" and "Coronary Heart Disease," and the description field is "may cause," then the complete description is "Type 2 diabetes may cause coronary heart disease."

[0232] The description of the attribute fields differs depending on the type of edge relationship. Triggering relationships typically use verbs such as "triggers," "causes," or "induces," while diagnostic means relationships use expressions like "requires diagnosis through..." or "can be detected through...," and concurrent relationships use expressions like "frequently concurrent" or "may exist simultaneously." The extracted edge description sequence corresponds one-to-one with the node name sequence, providing complete semantic information for subsequent alternating concatenation.

[0233] By alternately concatenating the medical specification name and attribute description fields according to the traversal order, a linear text sequence is generated, completing the conversion from structured graph data to natural language text. The concatenation process follows an alternating pattern of "node-edge-node-edge-...", transforming the traversal path into a fluent natural language description. For the path... The concatenation result is "[name of v1] [description of e1] [name of v2] [description of e2] [name of v3]".

[0234] For example, if the path is "Type 2 diabetes - triggering - coronary heart disease - needs to be diagnosed through... - coronary CTA", then the spliced ​​result is "The patient has a history of type 2 diabetes. Diabetes may trigger coronary atherosclerotic heart disease. Coronary heart disease needs to be diagnosed through coronary CT angiography".

[0235] The assembly process requires grammatical optimization, including adding appropriate conjunctions (such as "therefore," "so," "furthermore"), adjusting word order to conform to Chinese expression habits, and adding subjects and predicates to make sentences complete. For source nodes, prefixes such as "patient's chief complaint" or "patient's previous diagnosis" can be added to clarify the clinical role of the node; for target nodes, guiding words such as "therefore it is necessary to perform" or "it is recommended" can be added to highlight the recommended nature of the examination.

[0236] For a locally connected subgraph containing multiple paths, the text sequences of each path can be separated by semicolons or newlines to form a multi-segment inference description. For example, "Path 1: The patient complains of chest tightness and shortness of breath, which may indicate heart failure and requires evaluation by echocardiography; Path 2: The patient has a history of diabetes, which is often complicated by coronary heart disease, which needs to be diagnosed by coronary CTA."

[0237] The generated linear text sequences are typically between 100 and 500 characters in length, containing complete reasoning logic while maintaining conciseness, making them easy for doctors to read and understand quickly. The generation of text sequences can also incorporate template technology, pre-defining description templates for different types of paths. Appropriate templates are selected based on the node type and edge relationship type of the path, and specific node names and edge descriptions are filled in, improving generation efficiency and consistency.

[0238] By loading a pre-defined cue word template containing blank context slots and concatenating a linear text sequence as the limiting context into the cue word template, the transformation from text sequence to structured inference output is completed when a natural language inference chain is obtained. The cue word template is a pre-designed text framework used to guide large language models in inference and output generation.

[0239] The template comprises three core parts: task instructions, inference constraints, and output format requirements. The task instructions section clearly defines the objective of the inference task, such as "Based on the following pathological association paths, determine the medical rationality of the examination items and provide a detailed reasoning process." This section uses explicit instructional language to ensure that the model understands the task requirements.

[0240] The inference constraints section specifies the boundary conditions and rules for inference, such as "inference is based solely on the provided pathological pathways, without introducing external medical knowledge," "one-way causal logic must be followed, and reverse reasoning is prohibited," and "the probability of disease occurrence and clinical commonality must be considered." These constraints ensure the rigor and clinical relevance of the inference process, preventing the model from producing unreasonable inference results.

[0241] The output format requirements section specifies the structure of the model's output, for example, "Output format: Reasoning process: [detailed description of reasoning logic], Conclusion: [reasonable / excessive], Confidence level: [a value between 0 and 1]". Formatted output facilitates the system's parsing of the model's reasoning results and extracts key information for subsequent decision-making.

[0242] The prompt template contains blank context slots, usually marked with special tags (such as "[CONTEXT]"), used to insert specific pathological path text. The concatenation process replaces the context slots in the template with the generated linear text sequence to form the complete prompt text. For example, the complete prompt text might be: "Please determine the medical rationale for the examination based on the following pathological association path: The patient has a history of type 2 diabetes, diabetes may cause coronary heart disease, and coronary heart disease needs to be diagnosed by coronary CTA. Inference constraints: Based only on the provided path, follow causal logic. Output format: Inference process: [...], Conclusion: [Reasonable / Excessive]".

[0243] The complete prompt text is input into a pre-trained large language model (such as GPT-4 or a model fine-tuned for the medical field). The model reasons based on the prompt, generating a natural language reasoning chain that includes the reasoning process and conclusion. The reasoning chain elaborates on the logical connection from the patient's symptoms and disease to the items to be examined. For example, "The patient has a history of diabetes, which is an important risk factor for coronary heart disease. When the patient experiences chest tightness, it is necessary to investigate coronary artery lesions. Coronary CTA is the standard examination method for diagnosing coronary heart disease, therefore this examination is medically justified."

[0244] This embodiment effectively solves the key technical problem of how to transform structured graph data into clinically understandable natural language reasoning chains. Through a systematic graph traversal algorithm, the structural information of locally connected subgraphs is transformed into an ordered sequence of nodes, ensuring the coherence and completeness of the reasoning path. By acquiring the semantic information of edges, graph relationships are transformed into natural language expressions, providing crucial semantic connections for generating coherent reasoning text. Through the ordered combination of node names and edge descriptions, and grammatical optimization, discrete graph elements are transformed into fluent natural language text, achieving a complete conversion from structured data to readable text. Through structured template design and the reasoning capabilities of a large language model, text sequences are transformed into complete reasoning chains containing reasoning processes and conclusions, enhancing the interpretability and credibility of the system's judgments. The entire technical solution not only achieves technical data transformation but, more importantly, establishes an effective communication bridge between the system and doctors, enabling complex multi-hop reasoning logic to be presented in a natural language form familiar to doctors. This significantly improves the interpretability and clinical acceptability of the system, effectively solving the trust issue caused by the lack of explanation in intelligent review systems.

[0245] This application also provides an electronic device, including: The memory is configured to store instructions; and The processor is configured to retrieve the instructions from the memory and, when executing the instructions, to implement the over-checking detection method described above.

[0246] In this embodiment, the electronic device can be a tablet computer, desktop computer, laptop computer, handheld computer, wearable device, laptop computer, ultra-mobile personal computer (UMPC), netbook, or other device with a processor. Of course, the electronic device can also be a server. This application embodiment does not impose any special limitations on the specific form of the electronic device.

[0247] This application also provides a machine-readable storage medium storing instructions for causing a machine to perform the above-described over-checking detection method.

[0248] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0249] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0250] 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.

[0251] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0252] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0253] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0254] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0255] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0256] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for detecting over-checking, characterized in that, The method includes: In response to receiving the doctor's order for examination items, the system retrieves the patient's electronic medical record data, which includes a set of entities for the chief complaint symptoms and a set of entities for previously diagnosed diseases. The entity set of chief complaint symptoms and the entity set of previously diagnosed diseases are aligned with nodes in a pre-built medical knowledge graph to obtain a source node set, and the items to be examined are mapped to target nodes. The source node set is used as the initial semantic scalar. The graph convolutional network is iteratively computed based on the initial semantic scalar until the semantic distribution reaches a steady state. The convergent semantic features of the target node under the steady state are then extracted. Based on the extraction of convergent semantic features, the main transmission links from the source node set to the target node are obtained, and a locally connected subgraph is generated based on the main transmission links; The local connected subgraphs are subjected to subgraph isomorphic matching with a pre-defined complication pathology structure library to determine whether there is a specific implicit clinical association between the target node and the source node set. If there is a specific implicit clinical association between the target node and the source node set, the local connected subgraph is expanded into a text sequence of node entities and edge relationships, and the text sequence is used as a limiting context to be concatenated into a preset prompt word template to obtain a natural language inference chain. The medical rationality of the item to be examined under the current symptoms is determined based on the text sequence, and the reasonable or excessive classification label and the corresponding natural language inference chain are output. If the category label is appropriate, allow the order to be issued for the item to be inspected. When the category label is "excessive", an interception pop-up is triggered and the natural language inference chain is displayed.

2. The method according to claim 1, characterized in that, The set of entities representing chief complaints and symptoms and the set of entities representing previously diagnosed diseases are aligned with nodes in a pre-constructed medical knowledge graph to obtain a source node set, including: The entity set of chief complaint symptoms and the entity set of previously diagnosed diseases are standardized to obtain a standardized entity set. Retrieve node identifiers that match the standardized entity set in the medical knowledge graph; Activate the corresponding graph node based on the node identifier, and mark the activated graph node as the source node set.

3. The method according to claim 1, characterized in that, The source node set is used as the initial semantic scalar. A graph convolutional network is iteratively computed based on this initial semantic scalar until the semantic distribution reaches a steady state. The convergent semantic features of the target nodes in the steady state are then extracted, including: Obtain the attribute information of edges in the medical knowledge graph, including the triggering relationship, the diagnostic method relationship, and the concurrency relationship; Based on the edge attribute information, initialize the semantic transmission damping coefficient between each adjacent node; Extract the clinical importance score of each source node in the source node set; The clinical importance score is mapped to the initial semantic scalar magnitude of each source node; Based on the semantic transmission damping coefficient, a semantic diffusion matrix of a graph structure is constructed. The initial semantic scalar magnitude is iteratively diffused along the semantic diffusion matrix, and the mutation rate of the global node scalar is calculated in two adjacent iterations. When the mutation rate of the global node scalar is lower than the preset stability convergence factor, the semantic distribution is determined to have reached a steady state, and the accumulated scalar value of the target node is extracted as the convergent semantic feature under the steady state.

4. The method according to claim 1, characterized in that, Based on convergent semantic feature extraction, the main propagation links from the source node set to the target node are obtained, and a locally connected subgraph is generated based on the main propagation links, including: Tracing back the scalar source path of the convergent semantic features of the target node in steady state; Calculate the path decay integral for each scalar source path; All scalar source paths are sorted in ascending order based on the path attenuation integral, and the paths ranked in the preset first percentage interval are extracted as candidate transmission links. Cross-reference frequency statistics are performed on the nodes in the candidate propagation links, and isolated paths with zero cross-reference frequency are removed. The remaining paths are used as the main propagation links. Merge the node and edge relationships in all major transmission links to construct a locally connected subgraph.

5. The method according to claim 1, characterized in that, The locally connected subgraphs are subjected to subgraph isomorphic matching with a pre-defined complication pathology structure library to determine whether there are specific implicit clinical associations between the target node and the source node set, including: Obtain a set of baseline pathological topologies for confirmed complex complication cases from the complication pathology structure library; Graph feature extraction is performed on locally connected subgraphs to obtain topological feature vectors of the locally connected subgraphs; Graph feature extraction is performed on each benchmark pathological topology graph in the benchmark pathological topology graph set to obtain the benchmark topology feature vector; Calculate the structural similarity between the topological feature vectors of the locally connected subgraphs and each baseline topological feature vector; When the structural similarity is greater than the preset isomorphic matching threshold, it is determined that the local connected subgraph and the corresponding baseline pathological topology graph have an isomorphic relationship. Specific latent attributes marked in the baseline pathological topology map with isomorphic relationships are extracted, and specific latent clinical associations between the target node and the source node set are confirmed based on the specific latent attributes.

6. The method according to claim 1, characterized in that, Calculate the structural similarity between the topological feature vectors of locally connected subgraphs and each baseline topological feature vector, including: Calculate the minimum number of edit operations between the locally connected subgraph and the baseline pathological topology, where the edit operations include node insertion, node deletion, edge insertion, and edge deletion; Calculate the structural difference degree based on the minimum number of editing operations; The structural dissimilarity is converted into structural similarity, where structural similarity and structural dissimilarity are negatively correlated.

7. The method according to claim 6, characterized in that, Converting structural dissimilarity to structural similarity includes: Obtain the penalty coefficient for graph complexity; Multiply the structural difference by the penalty coefficient to obtain the corrected difference term; The initial similarity score is obtained by subtracting the correction difference term from the preset theoretical maximum similarity constant. The initial similarity scores are subjected to exponential smoothing to obtain the structural similarity.

8. The method according to claim 1, characterized in that, If a specific implicit clinical association exists between the target node and the source node set, the locally connected subgraph is expanded into a text sequence of node entities and edge relationships. This text sequence is then used as a limiting context and concatenated into a preset prompt word template to obtain a natural language inference chain, including: If there is a specific implicit clinical association between the target node and the source node set, perform a depth-first search traversal on the locally connected subgraph, and extract the medical standard names of the node entities passed through during the traversal. Extract the attribute description fields of the edges connecting each node entity; Following the traversal order, the medical specification name and attribute description fields are alternately concatenated to generate a linear text sequence; Load a preset prompt word template containing blank context slots, and concatenate a linear text sequence as a limiting context into the prompt word template to obtain a natural language inference chain. The prompt word template includes task instructions, inference constraints, and output format requirements.

9. An electronic device, characterized in that, include: The memory is configured to store instructions; as well as A processor configured to retrieve the instructions from the memory and, when executing the instructions, to implement the over-checking detection method according to any one of claims 1 to 8.

10. A machine-readable storage medium, characterized in that, The machine-readable storage medium stores instructions for causing the machine to perform the over-checking detection method according to any one of claims 1 to 8.

Images