A knowledge-based reasoning-based method for real-time intelligent review and error correction of medical record content

Abstract

This invention provides a real-time intelligent review and error correction method for medical record content based on knowledge reasoning, relating to the field of medical information processing. The method includes acquiring a medical knowledge graph and medical record documents, performing medical entity recognition and semantic parsing, constructing a medical event timeline chain for timeline verification, identifying conflicting entities based on counterfactual reasoning, extracting alternative entities, and generating error correction suggestions. This invention enables knowledge-based review of medical record content, improves medical record quality, reduces medical errors, lowers the workload of manual review, and ensures medical safety.

Description

Technical Field

[0001] This invention relates to medical information processing technology, and more particularly to a method for real-time intelligent review and error correction of medical record content based on knowledge reasoning. Background Technology

[0002] Electronic medical records, as a digital representation of medical records, have become an important component of modern healthcare services. With the rapid expansion of medical data, traditional manual review of medical record content can no longer meet practical needs. Therefore, research on automated review and error correction methods for medical records based on artificial intelligence technology has gradually gained attention.

[0003] Existing medical record review technologies lack deep integration of medical knowledge graphs and medical record content, failing to effectively utilize the complex relationships between medical entities for knowledge reasoning. This results in limited ability to identify implicit logical errors in medical records. Current medical record review methods do not adequately consider the chronological logic of disease development, making it difficult to detect chronological inversions in medical records that violate the laws of medical development, thus reducing the accuracy and reliability of the review. After discovering errors in medical records, they often can only point out the location of the error, lacking a mechanism for generating error correction suggestions based on knowledge reasoning. This fails to provide medical staff with targeted modification references, affecting review efficiency and practicality. Summary of the Invention

[0004] This invention provides a real-time intelligent review and error correction method for medical record content based on knowledge reasoning, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for real-time intelligent review and error correction of medical record content based on knowledge reasoning, comprising:

[0006] Obtain a medical knowledge graph containing medical entity nodes, relationship edges between medical entities, and the temporal pattern of disease development, as well as medical record documents awaiting review;

[0007] The medical record documents to be reviewed are subjected to medical entity recognition and semantic parsing. The parsed medical entities are mapped to a medical knowledge graph to generate a set of semantic representations and knowledge constraint relationships of the medical records.

[0008] Extract medical events with time attributes from the semantic representation of medical records, construct a time sequence chain of medical events, match and verify the time sequence of disease development in the medical knowledge graph, and identify time inversion points.

[0009] For counterfactual medical entities corresponding to the time-inverted point, remove them one by one and then perform knowledge reasoning based on the medical knowledge graph to generate counterfactual reasoning results. Compare the original reasoning results with the counterfactual reasoning results to calculate the degree of influence value, and filter out target conflicting entities whose degree of influence value exceeds the preset degree threshold.

[0010] For the target conflicting entity, alternative entities that satisfy the set of temporal and knowledge constraints are extracted from the medical knowledge graph. Contextual evidence of the alternative entities is retrieved from the semantic representation of the medical record to form a reasoning link and generate error correction suggestions. The error correction suggestions are then output to the medical record document to be reviewed.

[0011] Obtaining a medical knowledge graph that includes medical entity nodes, relationships between medical entities, and the temporal patterns of disease development includes:

[0012] Medical entity terms and their relationship descriptions are extracted from medical literature using a natural language processing model, and a set of medical entity terms and a set of medical entity relationship descriptions are constructed.

[0013] Standardized medical terms and relationships between terms are extracted from the medical ontology database. Terms in the medical entity terminology set are mapped to standardized medical terms to generate standardized medical entity nodes.

[0014] Map the set of medical entity relationship descriptions to the terminology relationships to construct standardized relationship edges between medical entities;

[0015] Disease diagnosis and treatment process information is extracted from clinical diagnosis and treatment guidelines. Based on the standardized medical entity nodes, the temporal dependencies in the disease diagnosis and treatment process information are analyzed to generate disease development temporal patterns.

[0016] The standardized medical entity nodes, standardized relationship edges between medical entities, and disease development timeline patterns are integrated to construct a medical knowledge graph.

[0017] The medical record documents to be reviewed undergo medical entity recognition and semantic parsing. The parsed medical entities are mapped to a medical knowledge graph, generating a set of semantic representations and knowledge constraint relationships for the medical records, including:

[0018] Perform entity boundary identification on the medical record documents to be reviewed and extract medical entity terms;

[0019] Semantic matching is performed between medical entity terms and medical entity nodes in the medical knowledge graph. When there are multiple candidate mapping nodes for a medical entity term, the contextual text fragments of the medical entity terms in the medical record document to be reviewed are extracted.

[0020] The context text fragments and the candidate mapping nodes are semantically consistent with each other in the medical knowledge graph. The medical entity nodes that pass the semantic consistency verification are selected as the mapping targets.

[0021] Constructing semantic representations of medical records based on mapping targets;

[0022] The relationship edges between mapping targets are extracted from the medical knowledge graph. When the relationship edge is a mutually exclusive relationship, the mapping targets connected by the relationship edge are identified as mutually exclusive entity pairs. When mutually exclusive entity pairs exist in the semantic representation of medical records, the relationship edge is marked as a conflicting relationship edge that violates the constraints.

[0023] In the medical knowledge graph, multi-hop dependency paths of the mapped target are retrieved. When the starting node and ending node of the multi-hop dependency path exist in the semantic representation of the medical record but the intermediate node is missing, the multi-hop dependency path is marked as a missing dependency path that violates the constraints.

[0024] Conflicting edges and missing dependency paths are treated as a set of knowledge constraint relations.

[0025] Medical events carrying time attributes are extracted from the semantic representation of medical records, and a time sequence chain of medical events is constructed. The time sequence chain of medical events is matched and verified with the disease development time sequence pattern in the medical knowledge graph. The time sequence inversion points are identified, including:

[0026] Extract medical events with time attributes from the semantic representation of medical records, and construct a time sequence chain of medical events by sorting them according to timestamp information;

[0027] The causal transmission path is extracted from the disease development time sequence pattern in the medical knowledge graph, and the starting event and ending event in the causal transmission path are located in the medical event time sequence chain.

[0028] When the timestamp of the termination event is earlier than the timestamp of the start event, the causal transmission path is traced back to extract the intermediate transmission events between the start and termination events, and the start event, termination event, and intermediate transmission events are marked as the causal inversion influence domain.

[0029] Calculate the time interval between adjacent medical events in the medical event time sequence chain, construct the time interval into a time interval vector sequence according to the time sequence of medical events, and extract the standard time evolution pattern from the disease development time sequence pattern in the medical knowledge graph;

[0030] The deviation is calculated by matching the time interval vector sequence with the standard time evolution pattern position by position. When the deviation exceeds the preset deviation threshold, the transition event defined in the corresponding position interval is extracted from the standard time evolution pattern, and the time interval with missing transition event is marked as a time evolution gap.

[0031] The causal inversion influence domain and the time-series evolution gap are taken as the time-series inversion point.

[0032] For counterfactual medical entities corresponding to the time-reversal points, they are removed one by one. Counterfactual reasoning results are then generated based on a medical knowledge graph. The impact value is calculated by comparing the original reasoning results with the counterfactual reasoning results. Target conflicting entities whose impact value exceeds a preset threshold are selected, including:

[0033] Extract the corresponding counterfactual medical entities from the time-reversed points and construct a set of counterfactual medical entities;

[0034] Knowledge reasoning is performed based on medical record semantic representation and medical knowledge graph, generating original reasoning results containing reasoning paths, extracting the medical entity nodes passed through the reasoning paths of the original reasoning results, and constructing the original reasoning dependency graph.

[0035] For each counterfactual medical entity in the counterfactual medical entity set, after removing the counterfactual medical entity from the medical record semantic representation, knowledge reasoning is performed based on the medical knowledge graph to generate a counterfactual reasoning result containing the reasoning path. The medical entity nodes passed through the reasoning path of the counterfactual reasoning result are extracted to construct a counterfactual reasoning dependency graph.

[0036] The original reasoning dependency graph and the counterfactual reasoning dependency graph are compared in terms of topology. The number of broken reasoning paths in the topology comparison is counted as the degree of influence. Counterfactual medical entities whose degree of influence exceeds the preset degree threshold are selected as target conflict entities.

[0037] For each counterfactual medical entity in the counterfactual medical entity set, after removing the counterfactual medical entity from the medical record semantic representation, knowledge reasoning is performed based on the medical knowledge graph to generate counterfactual reasoning results containing the reasoning path, including:

[0038] Select entities to be verified from the set of counterfactual medical entities;

[0039] Locate the entity to be verified in the semantic representation of the medical record, delete the entity to be verified and its associated relationships, and reorganize the remaining medical entities and their associated relationships into the remaining semantic representation.

[0040] Extract the retained medical entities from the remaining semantic representation, map the retained medical entities to medical entity nodes in the medical knowledge graph, obtain the relationship edges between medical entities of the medical entity nodes, and perform multi-hop reasoning along the relationship edges between medical entities starting from the medical entity nodes.

[0041] During multi-hop reasoning, the traversed medical entity nodes and the relationship edges between medical entities are recorded. The traversed medical entity nodes and the relationship edges between medical entities are connected in the reasoning order to form a reasoning path. The reasoning path is then encapsulated as a counterfactual reasoning result.

[0042] For the target conflicting entity, substitute entities that satisfy the set of temporal and knowledge constraint relationships are extracted from the medical knowledge graph. Contextual evidence of the substitute entities is retrieved from the semantic representation of the medical records to form a reasoning chain and generate error correction suggestions. These error correction suggestions are then output to the medical record document to be reviewed, including:

[0043] Candidate entities are extracted from the medical knowledge graph. The temporal consistency between the candidate entities and the medical event timeline is verified, and the logical compatibility between the candidate entities and the knowledge constraint relationship set is verified. Candidate entities that pass both the temporal consistency verification and the logical compatibility verification are selected as replacement entities.

[0044] Locate substitute entities in the semantic representation of medical records, extract the contextual medical entities surrounding the substitute entities, retrieve the association paths between the substitute entities and the contextual medical entities from the medical knowledge graph, and use the association paths as contextual evidence.

[0045] The inference link is constructed by associating and combining the substitute entity with contextual evidence. Replacement instructions are generated from the inference link. The location information of the target conflicting entity in the medical record document to be reviewed, the substitute entity, and the replacement instructions are organized into error correction suggestions. The error correction suggestions are output to the medical record document to be reviewed.

[0046] A second aspect of the present invention provides an electronic device, comprising:

[0047] processor;

[0048] Memory used to store processor-executable instructions;

[0049] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0050] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0051] In this embodiment, by constructing a medical event timeline and matching it with the disease development timeline patterns in the medical knowledge graph, the system can accurately identify timeline inversion points in medical records, effectively solving the technical problem of difficulty in detecting timeline logic errors in traditional medical record review. Counterfactual reasoning is employed for timeline inversion points. By removing suspicious medical entities and calculating their impact value, the system filters out the truly conflicting entities affecting the medical record logic, improving the accuracy and targeting of the review. Based on the medical knowledge graph, substitute entities satisfying the timeline and knowledge constraints are extracted, and combined with contextual evidence to form a complete reasoning chain. This enables the generation of well-reasoned error correction suggestions, overcoming the limitations of traditional error correction methods that lack knowledge support. Real-time intelligent review and correction of medical record content is achieved, significantly improving the efficiency and accuracy of medical record quality management. This helps medical staff understand and adopt the system's proposed modifications, enhancing the system's practical value and credibility. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the real-time intelligent review and error correction method for medical record content based on knowledge reasoning, as described in an embodiment of the present invention.

[0053] Figure 2 This is an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0056] Figure 1 This is a flowchart illustrating the real-time intelligent review and error correction method for medical record content based on knowledge reasoning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes:

[0057] Obtain a medical knowledge graph containing medical entity nodes, relationship edges between medical entities, and the temporal pattern of disease development, as well as medical record documents awaiting review;

[0058] The medical record documents to be reviewed are subjected to medical entity recognition and semantic parsing. The parsed medical entities are mapped to a medical knowledge graph to generate a set of semantic representations and knowledge constraint relationships of the medical records.

[0059] Extract medical events with time attributes from the semantic representation of medical records, construct a time sequence chain of medical events, match and verify the time sequence of disease development in the medical knowledge graph, and identify time inversion points.

[0060] For counterfactual medical entities corresponding to the time-inverted point, remove them one by one and then perform knowledge reasoning based on the medical knowledge graph to generate counterfactual reasoning results. Compare the original reasoning results with the counterfactual reasoning results to calculate the degree of influence value, and filter out target conflicting entities whose degree of influence value exceeds the preset degree threshold.

[0061] For the target conflicting entity, alternative entities that satisfy the set of temporal and knowledge constraints are extracted from the medical knowledge graph. Contextual evidence of the alternative entities is retrieved from the semantic representation of the medical record to form a reasoning link and generate error correction suggestions. The error correction suggestions are then output to the medical record document to be reviewed.

[0062] In one optional implementation, obtaining a medical knowledge graph containing medical entity nodes, relationship edges between medical entities, and the temporal pattern of disease development includes:

[0063] Medical entity terms and their relationship descriptions are extracted from medical literature using a natural language processing model, and a set of medical entity terms and a set of medical entity relationship descriptions are constructed.

[0064] Standardized medical terms and relationships between terms are extracted from the medical ontology database. Terms in the medical entity terminology set are mapped to standardized medical terms to generate standardized medical entity nodes.

[0065] Map the set of medical entity relationship descriptions to the terminology relationships to construct standardized relationship edges between medical entities;

[0066] Disease diagnosis and treatment process information is extracted from clinical diagnosis and treatment guidelines. Based on the standardized medical entity nodes, the temporal dependencies in the disease diagnosis and treatment process information are analyzed to generate disease development temporal patterns.

[0067] The standardized medical entity nodes, standardized relationship edges between medical entities, and disease development timeline patterns are integrated to construct a medical knowledge graph.

[0068] First, medical entity and relational information is extracted through literature analysis. In the medical literature processing stage, a named entity recognition model combining a Bidirectional Long Short-Term Memory (BiLSTM) network and a Conditional Random Field (CRF) is employed to identify medical entities such as diseases, symptoms, drugs, and treatments in the literature. This model uses pre-trained word vectors from a medical corpus as input, captures contextual information through BiLSTM, and then optimizes entity boundary annotations using a CRF layer, thereby extracting complete medical entity terms.

[0069] In the relation extraction stage, a remote supervised learning method is used to construct a set of sentences containing medical entity pairs. A convolutional neural network based on an attention mechanism is then applied to identify the relationships between entities. The model first extracts sentence-level features, then aggregates the representations of different sentences through a multi-head attention mechanism, and finally classifies and determines the type of relationship between entities, such as "treatment", "diagnosis", "cause", etc., forming a set of medical entity relationship descriptions.

[0070] To ensure the standardization of medical concepts, standardized terms and relationships are extracted from medical ontology databases such as SNOMED CT and ICD-10. The terminology standardization process employs a vector-based matching algorithm to calculate the semantic similarity between medical entity entries and standardized terms. First, term vector representations are constructed using TF-IDF. Then, a hybrid metric combining edit distance and semantic similarity is used to find the most matching standardized term for each medical entity entry, generating standardized medical entity nodes. For polysemous words, contextual relevance analysis is used to disambiguate and ensure mapping accuracy.

[0071] In constructing standardized relation edges, the extracted relation descriptions are normalized to the standard relation types defined by the medical ontology. A mapping table between relation descriptions and standard relations is established using a combination of template matching and rule transformation. For example, "for treatment" is mapped to the "treats" relation, and "diagnostic basis" is mapped to the "diagnoses" relation. For complex relations, semantic decomposition techniques are used to break them down into basic relation combinations, ensuring the standardization and consistency of relation expression.

[0072] When extracting temporal information from clinical practice guidelines, the process first involves using PDF document structured parsing technology to identify flowcharts, decision trees, and clinical pathways within the guidelines. Regular expressions and syntactic analysis methods are then used to extract key phrases describing temporal relationships, such as "after," "first," and "next." Finally, based on standardized medical entity nodes, these temporal expressions are parsed to construct a time dependency graph of disease development and the treatment process.

[0073] The temporal patterns are generated using a directed acyclic graph (DAG) model to represent the disease progression path, where nodes represent disease states or medical interventions, and edges represent state transitions or causal relationships. Each edge is assigned a temporal attribute, including a relative time window and a transition probability. By analyzing the temporal patterns of the same disease in multiple clinical guidelines, common patterns are extracted to form a standardized temporal model of disease progression.

[0074] Finally, graph database technology is used to integrate the above three parts of information. An attribute graph model is used to store the knowledge graph, which includes standardized medical entity nodes, standardized relationship edges between medical entities, and disease temporal patterns. Nodes store attributes such as standard terminology, synonym sets, and categories of entities; relationship edges store attributes such as relationship type, evidence source, and credibility; and temporal patterns are stored using specific path patterns and time constraints.

[0075] After the knowledge graph is constructed, consistency verification is performed to detect conflicting relationships and circular dependencies. The accuracy of important nodes and relationships in the graph is evaluated through a combination of medical expert review and literature verification. For any issues discovered, a decision is made based on confidence scores to retain, correct, or delete them, ensuring the quality of the knowledge graph.

[0076] In application scenarios, this medical knowledge graph can support intelligent consultation systems in making disease diagnostic inferences, deducing possible diseases based on patient symptoms and following the relational paths within the knowledge graph. Clinical decision support systems can utilize disease temporal patterns to provide doctors with personalized treatment suggestions based on the patient's disease stage. Furthermore, in the field of drug development, analyzing the molecular-disease association network in the graph can help discover potential drug targets and retargeting opportunities.

[0077] The medical knowledge graph constructed through the above steps not only includes static medical entities and relationship knowledge, but also integrates the dynamic temporal patterns of disease development, providing a comprehensive knowledge foundation for intelligent medical applications.

[0078] In one optional implementation, the medical record document to be reviewed undergoes medical entity recognition and semantic parsing. The parsed medical entities are mapped to a medical knowledge graph, generating a set of semantic representations and knowledge constraint relationships for the medical record, including:

[0079] Perform entity boundary identification on the medical record documents to be reviewed and extract medical entity terms;

[0080] Semantic matching is performed between medical entity terms and medical entity nodes in the medical knowledge graph. When there are multiple candidate mapping nodes for a medical entity term, the contextual text fragments of the medical entity terms in the medical record document to be reviewed are extracted.

[0081] The context text fragments and the candidate mapping nodes are semantically consistent with each other in the medical knowledge graph. The medical entity nodes that pass the semantic consistency verification are selected as the mapping targets.

[0082] Constructing semantic representations of medical records based on mapping targets;

[0083] The relationship edges between mapping targets are extracted from the medical knowledge graph. When the relationship edge is a mutually exclusive relationship, the mapping targets connected by the relationship edge are identified as mutually exclusive entity pairs. When mutually exclusive entity pairs exist in the semantic representation of medical records, the relationship edge is marked as a conflicting relationship edge that violates the constraints.

[0084] In the medical knowledge graph, multi-hop dependency paths of the mapped target are retrieved. When the starting node and ending node of the multi-hop dependency path exist in the semantic representation of the medical record but the intermediate node is missing, the multi-hop dependency path is marked as a missing dependency path that violates the constraints.

[0085] Conflicting edges and missing dependency paths are treated as a set of knowledge constraint relations.

[0086] First, entity boundary recognition is performed on the medical records to be reviewed, and medical entity terms are extracted using a medical entity recognition model. This model adopts a pre-trained BERT-BiLSTM-CRF architecture, which is fine-tuned for medical text. The input is a sequence of medical record text, and the output is a label for each word (such as "B-disease", "I-disease", "B-drug", etc.), thereby determining the entity boundaries. For example, in the medical record fragment "The patient was admitted to the hospital due to upper abdominal pain accompanied by nausea and vomiting for three days", three symptom entities, "upper abdominal pain", "nausea", and "vomiting", were identified.

[0087] Semantic matching is performed between medical entity terms and entity nodes in a medical knowledge graph. First, an edit distance algorithm is used for initial screening of entity terms, selecting knowledge graph nodes with an edit distance less than a threshold as candidate mapping nodes. For example, the entity term "diabetes" may match multiple candidate nodes such as "type 1 diabetes," "type 2 diabetes," and "gestational diabetes." When multiple candidate mapping nodes exist, the contextual text fragments of the medical entity term in the medical record document to be reviewed are extracted, with the context window size set to 50 characters before and after the entity.

[0088] Semantic consistency verification is then performed, employing a bidirectional attention mechanism to match the context text fragments with the adjacent nodes of the candidate mapping nodes in the medical knowledge graph. First, the set of first-order adjacent nodes for each candidate node is obtained from the knowledge graph, including related symptoms, examinations, and treatment plans. Then, the semantic similarity between the context text and the set of adjacent nodes is calculated. A pre-trained medical domain BERT model is used to extract text representations, and cosine similarity is used for matching scoring. Finally, the medical entity node with the highest semantic consistency verification score is selected as the mapping target.

[0089] In the above example, if the medical record context mentions information such as "elevated fasting blood glucose" and "excessive thirst," it is more likely to map "diabetes" to the "type 2 diabetes" node, because these symptoms are strongly associated with "type 2 diabetes" in the knowledge graph.

[0090] A semantic representation of medical records is constructed based on the mapping target, and a heterogeneous graph structure is used to store the semantic information of the medical records. In this graph structure, nodes represent medical entities, and node attributes include entity type (such as disease, symptom, drug, etc.), mapping confidence, and position in the original text; edges represent relationships between entities, such as "symptom-disease" and "drug-indication" relationship types. This representation method preserves the semantic associations between medical entities, which facilitates subsequent analysis.

[0091] Relationships between mapping targets are extracted from the medical knowledge graph. When a relationship edge is a mutually exclusive relationship, the mapping targets connected by the edge are identified as mutually exclusive entity pairs. For example, if the semantic representation of a medical record contains both the entities "pregnancy" and "glaucoma," and the drug "trazodone" in the knowledge graph is connected to both "contraindicated during pregnancy" and "used to treat glaucoma," then this set of relationships is marked as a conflicting relationship edge that violates constraints. The identification of mutually exclusive relationships is based on a predefined dictionary of mutually exclusive relationships and a rule base, including drug contraindications, mutually exclusive relationships between diseases, etc.

[0092] This study retrieves multi-hop dependency paths between mapped targets in a medical knowledge graph. A breadth-first search algorithm for directed graphs is employed, with a maximum hop count of three, to retrieve dependency paths between all mapped targets. When a multi-hop dependency path exists in the semantic representation of a medical record but lacks intermediate nodes, the path is marked as a missing dependency path that violates constraints. For example, if a medical record mentions the use of "insulin" treatment but lacks the necessary intermediate step of "blood glucose monitoring," then "insulin use → blood glucose monitoring → dosage adjustment" is marked as a missing dependency path.

[0093] The identified conflicting edges and missing dependency paths are output as a set of knowledge constraints to provide decision support for medical record review. This set is stored in a structured format, with each record containing information such as constraint type (conflicting / missing), related entities, relationship type, and severity of constraint violation, which can be intuitively displayed to medical personnel for review decisions.

[0094] By using the methods described above, we can fully utilize the professional knowledge contained in medical knowledge graphs to automatically discover potential medical logical errors or missing information in medical records, thereby improving the quality of medical records and medical safety.

[0095] In one optional implementation, medical events carrying time attributes are extracted from the semantic representation of medical records, a time-series chain of medical events is constructed, and the time-series chain of medical events is matched and verified with the disease development time-series patterns in a medical knowledge graph. Identifying time-series inversion points includes:

[0096] Extract medical events with time attributes from the semantic representation of medical records, and construct a time sequence chain of medical events by sorting them according to timestamp information;

[0097] The causal transmission path is extracted from the disease development time sequence pattern in the medical knowledge graph, and the starting event and ending event in the causal transmission path are located in the medical event time sequence chain.

[0098] When the timestamp of the termination event is earlier than the timestamp of the start event, the causal transmission path is traced back to extract the intermediate transmission events between the start and termination events, and the start event, termination event, and intermediate transmission events are marked as the causal inversion influence domain.

[0099] Calculate the time interval between adjacent medical events in the medical event time sequence chain, construct the time interval into a time interval vector sequence according to the time sequence of medical events, and extract the standard time evolution pattern from the disease development time sequence pattern in the medical knowledge graph;

[0100] The deviation is calculated by matching the time interval vector sequence with the standard time evolution pattern position by position. When the deviation exceeds the preset deviation threshold, the transition event defined in the corresponding position interval is extracted from the standard time evolution pattern, and the time interval with missing transition event is marked as a time evolution gap.

[0101] The causal inversion influence domain and the time-series evolution gap are taken as the time-series inversion point.

[0102] When extracting medical events from the semantic representation of medical records, the first step is to identify medical event descriptions containing time information using natural language processing techniques. These events typically contain explicit time markers, such as "fever symptoms appeared on March 12, 2022," or relative time descriptions, such as "chest tightness appeared three days before admission." For each medical event, its time attribute and event content are extracted to form a "time-event" pair. For events without explicit time markers, a contextual reasoning method is used to estimate the time based on the time information of nearby events; for example, "followed by dyspnea" can be inferred to have occurred after the aforementioned event.

[0103] Time standardization is a crucial step, converting various time expressions into a standardized timestamp format. For example, "last year's Spring Festival" is converted into a specific date, and "the third day after medication" is converted into a specific date relative to the medication time. Time standardization employs a combination of rule matching and time function transformation to ensure consistency and comparability of time representations.

[0104] After time standardization, medical events are sorted according to timestamp information to construct a time-series chain of medical events. The data structure of the time-series chain is an ordered list, and each node contains attributes such as timestamp, event content, and event type. For example, a patient's course of illness might be represented as: "2025-01-01 Fever 38.5℃ → 2025-01-03 Coughing → 2025-01-05 Diagnosed with pneumonia".

[0105] When extracting the temporal patterns of disease development from a medical knowledge graph, the focus is on the causal transmission paths of the disease. The knowledge graph stores a large number of standard patterns of disease development, including the order of symptom appearance, the transmission relationship from cause to symptom, and the timing of complications. Using a graph traversal algorithm, the causal transmission paths related to patient diagnosis are extracted, such as "viral infection → fever → viral replication → lung infection → pneumonia," where the edges between each node represent causal relationships and standard time intervals.

[0106] Locating the start and end events in a medical event timeline is a crucial step in identifying time-series reversals. First, the standard development path of the target disease is determined within the knowledge graph. Then, matching events are retrieved from the medical event timeline. A semantic similarity calculation method is used to match events in the knowledge graph with events in the medical event timeline, finding the corresponding items with the highest semantic similarity. For example, "viral infection" in the knowledge graph might correspond to "contact with an influenza patient" in the medical record, and "pneumonia" might correspond to "X-ray showing lung infection."

[0107] The determination of time sequence reversal is achieved by comparing the timestamps of the start event and the end event. If the end event time is earlier than the start event time, then time sequence reversal exists. For example, the knowledge graph shows that "contact with influenza patient → fever → pneumonia" is the standard sequence, but the medical record is "diagnosed with pneumonia on 2025-01-05 → recorded history of contact with influenza patient on 2025-01-07", which clearly violates the natural law of disease development and indicates that time sequence reversal exists.

[0108] When a chronological inversion is detected, it is necessary to trace back the causal transmission path and extract all intermediate transmission events between the starting and ending events. For example, in the path "viral infection → fever → viral replication → lung infection → pneumonia," if "viral infection" and "pneumonia" are inverted, then "fever," "viral replication," and "lung infection" are all marked as intermediate transmission events. These events, together with the starting and ending events, constitute the domain of influence of the causal inversion.

[0109] Calculating the time interval between adjacent events in a medical event time series is fundamental to identifying gaps in the temporal evolution. For each pair of adjacent events Ei and Ei+1 in the time series, their time interval Ti = T(Ei+1) - T(Ei) is calculated, forming a sequence of time interval vectors. This vector sequence characterizes the temporal features of each stage in the disease development process.

[0110] Standard time progression patterns are extracted from medical knowledge graphs to obtain standard time intervals for each stage of a specific disease's development. For example, the standard time progression pattern for influenza might be [2 days (from exposure to fever), 1 day (from fever to body aches), 3 days (from systemic symptoms to pronounced respiratory symptoms)]. These patterns are derived from extensive clinical data statistics and reflect the general patterns of disease development.

[0111] The time interval vector sequence is matched position-by-position with the standard time evolution pattern, and the deviation between the actual time interval and the standard time interval is calculated. The deviation is calculated by dividing the difference between the actual time interval and the standard time interval by the standard time interval. If the deviation at a certain position exceeds a preset threshold (e.g., 50%), the corresponding transition event at that position is extracted from the standard pattern, and the time interval with missing transition events is marked as a time evolution gap.

[0112] Finally, the causal inversion domain and the temporal evolution gap are used as temporal inversion points. These inversion points usually indicate potential problems in medical records, such as recording errors, patient recall errors, or atypical disease progression. For example, if the medical record shows only one day between "fever" and "pneumonia," while the standard model shows about seven days with transitional events such as "worsening cough" and "dyspnea," then this is marked as a temporal evolution gap, which may indicate a missed diagnosis or incomplete recording.

[0113] Identifying chronological inversion points can improve the efficiency of medical record quality management, assist physicians in identifying potential diagnostic and treatment problems, and support clinical decision-making. In practical applications, this method can be integrated into electronic medical record systems to automate chronological consistency checks, reduce errors in medical records, and improve the quality of healthcare services.

[0114] In one optional implementation, for the counterfactual medical entities corresponding to the time-reversal points, each is removed one by one, and counterfactual reasoning results are generated based on the medical knowledge graph. The original reasoning results are compared with the counterfactual reasoning results to calculate the degree of influence. Target conflicting entities whose degree of influence exceeds a preset threshold are selected, including:

[0115] Extract the corresponding counterfactual medical entities from the time-reversed points and construct a set of counterfactual medical entities;

[0116] Knowledge reasoning is performed based on medical record semantic representation and medical knowledge graph, generating original reasoning results containing reasoning paths, extracting the medical entity nodes passed through the reasoning paths of the original reasoning results, and constructing the original reasoning dependency graph.

[0117] For each counterfactual medical entity in the counterfactual medical entity set, after removing the counterfactual medical entity from the medical record semantic representation, knowledge reasoning is performed based on the medical knowledge graph to generate a counterfactual reasoning result containing the reasoning path. The medical entity nodes passed through the reasoning path of the counterfactual reasoning result are extracted to construct a counterfactual reasoning dependency graph.

[0118] The original reasoning dependency graph and the counterfactual reasoning dependency graph are compared in terms of topology. The number of broken reasoning paths in the topology comparison is counted as the degree of influence. Counterfactual medical entities whose degree of influence exceeds the preset degree threshold are selected as target conflict entities.

[0119] This process extracts counterfactual medical entities corresponding to chronological inversion points from medical record texts. A chronological inversion point refers to a discrepancy between the chronological order in the medical record narrative and the actual order of events. For example, in the statement "The patient was diagnosed with cirrhosis three years ago and was recently diagnosed with liver cancer after seeking medical attention for fever," the narrative order of cirrhosis and liver cancer differs from their actual developmental order. Using natural language processing (NLP) techniques, chronological relation words such as "once," "before," and "later" are identified. Combined with the temporal attributes of medical entities, chronological inversion points are determined, and corresponding counterfactual medical entities are extracted, forming a set of counterfactual medical entities. For example, the counterfactual medical entities extracted from the above case might include "cirrhosis."

[0120] Knowledge reasoning is performed based on a complete semantic representation of medical records and a medical knowledge graph. The semantic representation of medical records is achieved by converting medical record text into a structured semantic network, containing medical entities, relationships, and attributes. The medical knowledge graph contains entities such as diseases, symptoms, and drugs, and their interrelationships. During the reasoning process, path search algorithms, such as bidirectional breadth-first search, are used to find reasoning paths connecting key medical entities in the medical records within the knowledge graph. For example, the reasoning path from "fever" to "liver cancer" might pass through the "cirrhosis" node. All medical entity nodes traversed during the reasoning process and their relationships are extracted to construct an original reasoning dependency graph, which illustrates the dependencies between the reasoning result and each medical entity.

[0121] For each entity in the counterfactual medical entity set, the entity is temporarily removed from the medical record semantic representation, and then knowledge reasoning is re-executed based on the modified semantic representation and medical knowledge graph. For example, after removing "cirrhosis," it may be impossible to directly infer the association between "fever" and "liver cancer," or find different reasoning paths. The medical entity nodes and their relationships traversed in this reasoning process are extracted to construct a counterfactual reasoning dependency graph.

[0122] The original inference dependency graph and the counterfactual inference dependency graph are compared topologically to calculate the impact value of counterfactual medical entities. The topological comparison focuses on the differences between the two dependency graphs, particularly broken inference paths. A broken inference path is a path that exists in the original inference dependency graph but not in the counterfactual inference dependency graph. For example, removing "cirrhosis" might break the direct inference path from "fever" to "liver cancer." The number of all broken inference paths is counted as the impact value.

[0123] By setting an appropriate preset threshold, counterfactual medical entities whose impact exceeds the threshold are filtered out and marked as target conflict entities. The preset threshold can be flexibly adjusted according to the specific application scenario, and can usually be set as a specific proportion of the number of broken paths to the total number of original inference paths, such as 30%. For example, if removing "cirrhosis" causes 50% of the inference paths to break, and the preset threshold is 30%, then "cirrhosis" is identified as a target conflict entity.

[0124] In practical applications, the accuracy of this method can be improved through iterative optimization. The removal of counterfactual medical entities can employ two strategies: complete removal or weight reduction. Complete removal is suitable for explicit conflict situations, while weight reduction is suitable for situations where partial relevance may exist. Furthermore, the calculation of the impact value can also consider path importance, assigning higher weights to breaks in critical paths.

[0125] The target conflict entities identified through this method can help doctors understand the disease progression more accurately, avoid diagnostic misjudgments caused by time reversal, and improve the accuracy and efficiency of medical diagnosis.

[0126] In one optional implementation, for each counterfactual medical entity in the counterfactual medical entity set, after removing the counterfactual medical entity from the medical record semantic representation, knowledge reasoning is performed based on the medical knowledge graph to generate a counterfactual reasoning result containing the reasoning path, including:

[0127] Select entities to be verified from the set of counterfactual medical entities;

[0128] Locate the entity to be verified in the semantic representation of the medical record, delete the entity to be verified and its associated relationships, and reorganize the remaining medical entities and their associated relationships into the remaining semantic representation.

[0129] Extract the retained medical entities from the remaining semantic representation, map the retained medical entities to medical entity nodes in the medical knowledge graph, obtain the relationship edges between medical entities of the medical entity nodes, and perform multi-hop reasoning along the relationship edges between medical entities starting from the medical entity nodes.

[0130] During multi-hop reasoning, the traversed medical entity nodes and the relationship edges between medical entities are recorded. The traversed medical entity nodes and the relationship edges between medical entities are connected in the reasoning order to form a reasoning path. The reasoning path is then encapsulated as a counterfactual reasoning result.

[0131] First, select one entity to be verified from the counterfactual medical entity set. This set may contain multiple medical entities that may contain errors or require verification, such as symptoms, diseases, and medications in a medical record. For example, in a diabetic patient's medical record, the counterfactual medical entity set may contain entities such as "tuberculosis" and "Parkinson's disease," which are inconsistent with the original condition. In this step, "tuberculosis" is selected as the entity to be verified.

[0132] The selected entity to be verified is located in the medical record semantic representation. The medical record semantic representation is typically a structured graph containing medical entity nodes and relational edges. In this step, by traversing all nodes in the medical record semantic representation, the node matching "tuberculosis" is identified. Then, this node and all its associated relationships are deleted, such as "tuberculosis - manifests as - cough", "tuberculosis - requires - anti-tuberculosis drugs", etc. After deletion, the remaining medical entities and relationships are reorganized to form the remaining semantic representation.

[0133] Retained medical entities, such as "diabetes," "hypertension," "dizziness," and "elevated blood sugar," are extracted from the remaining semantic representation. These retained medical entities are then mapped to corresponding nodes in the medical knowledge graph. The medical knowledge graph is a network structure containing a large number of medical entities and their relationships. Entities include diseases, symptoms, drugs, and examinations, while relationships represent various medical associations between entities, such as "cause," "treatment," and "prevention."

[0134] During the mapping process, entity name matching or synonym expansion techniques are used to ensure accurate location of the corresponding node in the knowledge graph. For example, "diabetes" is mapped to the "diabetes" node in the knowledge graph, and "elevated blood sugar" is mapped to the "high blood sugar" node. For each successfully mapped medical entity node, its relation edges in the knowledge graph are obtained, and these edges connect to other related medical entity nodes.

[0135] Starting from the mapped medical entity nodes, multi-hop reasoning is performed along the relationship edges between medical entities. Multi-hop reasoning refers to the process of reaching other related nodes by starting from the initial node and making multiple jumps (i.e., moving along different relationship edges). During the reasoning process, either breadth-first search or depth-limited depth-first search algorithms are used to control the breadth and depth of the reasoning.

[0136] For example, starting from the "diabetes" node, possible reasoning paths include: "diabetes - leads to - hyperglycemia - manifests as - thirst", "diabetes - complications - diabetic nephropathy - manifests as - proteinuria", etc. In multi-hop reasoning, a maximum number of hops limit (such as 3 hops or 5 hops) is set to control computational complexity, and a relevance scoring mechanism is used to filter out low-relevance paths.

[0137] While performing multi-hop reasoning, all traversed medical entity nodes and the relationship edges between them are recorded. This information is connected in the order of reasoning to form a complete reasoning path. Each reasoning path includes a starting node, intermediate nodes, an ending node, and the relationship edges connecting these nodes.

[0138] Finally, all generated inference paths are organized and filtered, selecting those paths that are highly relevant to the entity to be verified and have high confidence levels. These paths are then packaged into counterfactual inference results. The structure of a counterfactual inference result includes: the entity to be verified, a set of inference paths (each path contains a sequence of nodes and a sequence of relationships), the inference conclusion, and the confidence level.

[0139] For example, for the entity to be verified, "tuberculosis", the possible counterfactual reasoning result is as follows: based on the patient's symptoms of "diabetes" and "dizziness", it is inferred through knowledge graph that these are more likely to be related to "diabetic complications" rather than "tuberculosis"; the reasoning path is "diabetes - leading to - hyperglycemia - affecting - central nervous system - manifesting as - dizziness", and the reasoning conclusion is "dizziness symptoms can be caused by diabetes and have no direct correlation with tuberculosis", with a confidence level of 0.85.

[0140] The above methods can effectively verify potential counterfactual medical entities in medical records, improve the accuracy and reliability of medical records, and provide a more reliable basis for medical decisions.

[0141] In one optional implementation, for the target conflicting entity, alternative entities satisfying the set of temporal and knowledge constraint relationships are extracted from the medical knowledge graph. Contextual evidence of the alternative entities is retrieved from the semantic representation of the medical records, forming a reasoning chain and generating error correction suggestions. The error correction suggestions are then output to the medical record document to be reviewed, including:

[0142] Candidate entities are extracted from the medical knowledge graph. The temporal consistency between the candidate entities and the medical event timeline is verified, and the logical compatibility between the candidate entities and the knowledge constraint relationship set is verified. Candidate entities that pass both the temporal consistency verification and the logical compatibility verification are selected as replacement entities.

[0143] Locate substitute entities in the semantic representation of medical records, extract the contextual medical entities surrounding the substitute entities, retrieve the association paths between the substitute entities and the contextual medical entities from the medical knowledge graph, and use the association paths as contextual evidence.

[0144] The inference link is constructed by associating and combining the substitute entity with contextual evidence. Replacement instructions are generated from the inference link. The location information of the target conflicting entity in the medical record document to be reviewed, the substitute entity, and the replacement instructions are organized into error correction suggestions. The error correction suggestions are output to the medical record document to be reviewed.

[0145] Candidate entities are extracted from a medical knowledge graph as potential alternatives. The medical knowledge graph contains medical entities such as diseases, symptoms, drugs, and examinations, along with their relationships, forming a rich knowledge network. During extraction, entities of the same category in the knowledge graph are queried based on the attribute category of the target conflicting entity (e.g., disease, symptom, drug). For example, if a conflict is found between "type 2 diabetes" in a medical record and the context description, all diabetes-related diseases are extracted from the knowledge graph as candidates, such as "type 1 diabetes" and "gestational diabetes." Entity similarity calculation methods are used during extraction, including semantic similarity and character similarity. Entities with similarity exceeding a preset threshold are selected as the initial candidate set.

[0146] Verify the temporal consistency between candidate entities and the medical event timeline. Medical events in medical records typically follow strict temporal logic, such as the order of symptom onset, examination, diagnosis, and treatment. Extract time stamps and event sequences from medical records to construct a medical event timeline. For each candidate entity, verify whether it conforms to the established timeline constraints. For example, if the medical record states "the patient's blood sugar rose instead of falling after taking insulin," and the conflicting entity is "type 2 diabetes," then the substitute entity must satisfy the temporal characteristic of "abnormal response to insulin." Substitute the candidate entities into the timeline and judge their rationality through the temporal relationships in the medical knowledge graph, filtering out a subset of candidate entities that conform to the temporal constraints.

[0147] The logical compatibility of candidate entities with the set of knowledge constraints is verified. This set includes association rules between medical concepts, such as the correspondence between diseases and symptoms, indications for treatment plans, and drug interactions. Contextual medical entities related to the conflicting entity in the target are extracted from the medical record to form the knowledge constraint set. For each candidate entity that passes the temporal verification, its compatibility with these knowledge constraints is checked. For example, if the medical record states that the patient has "renal insufficiency," the substitute entity should be compatible with or have a reasonable association with renal insufficiency. Using the relationship paths between entities in the knowledge graph, a compatibility score between the candidate entity and the constraints is calculated, and entities with scores exceeding a threshold are retained.

[0148] After verifying both temporal consistency and logical compatibility, qualified alternative entities are selected. If multiple candidate entities pass the verification, they are ranked according to their similarity to the conflicting entity and their compatibility score with the constraint relationship, and the optimal alternative entity is selected.

[0149] After identifying the substitute entity, it needs to be located in the semantic representation of the medical record. The semantic representation of the medical record transforms the original medical record text into a structured semantic network, where nodes represent medical entities and edges represent relationships between entities. Using a semantic matching algorithm, the position of the substitute entity is found in the semantic representation, determining its role and function within the semantic structure of the medical record.

[0150] Contextual medical entities surrounding the surrogate entity are extracted. Using the surrogate entity as the center, a semantic distance threshold is set, and medical entities in the semantic network whose distance from the surrogate entity is less than the threshold are extracted as contextual entities. These contextual entities constitute the semantic environment of the surrogate entity, providing a foundation for subsequent evidence collection.

[0151] This study retrieves association paths between substitute entities and contextual medical entities from a medical knowledge graph. An association path is a sequence of edges connecting two entities in the knowledge graph, reflecting the semantic relationships between them. A depth-first search algorithm is used to find the shortest or semantically strongest paths from the substitute entity to each context entity in the knowledge graph. For example, if the substitute entity is "type 1 diabetes" and the context entity includes "insulin resistance," the association path might be "type 1 diabetes - leads to - insufficient insulin secretion - distinguishes from - insulin resistance." These association paths serve as contextual evidence supporting the substitution, enhancing the interpretability of the error correction.

[0152] By associating surrogate entities with contextual evidence, a reasoning chain is constructed. A reasoning chain is a structured logical deduction process that connects surrogate entities with contextual evidence through causal, classification, and comparative relationships to form a complete reasoning path. For example, based on contextual evidence such as "the patient is young, has a rapid onset, experiences large fluctuations in blood sugar, and requires frequent insulin injections," the reasoning chain could be "young patient + acute onset + absolute insulin deficiency → type 1 diabetes is more reasonable than type 2 diabetes."

[0153] Generate a replacement description from the constructed reasoning chain. The replacement description should include three parts: a replacement suggestion, supporting evidence, and medical basis. Transform the reasoning chain into a natural language description, using medical terminology to accurately express the reasons for the replacement. For example: "It is recommended to change 'type 2 diabetes' to 'type 1 diabetes' because the patient is young (16 years old), has a rapid onset, has ketosis, requires insulin treatment, and has a negative family history, which meets the clinical characteristics of type 1 diabetes."

[0154] Finally, the location information of the conflicting entity in the medical record document to be reviewed, the alternative entity, and the replacement explanation are organized into a correction suggestion and output to the medical record document to be reviewed. The correction suggestion is presented in a structured format, including location markers, original entity, recommended alternative entity, basis for replacement, and confidence score. Doctors can use this to quickly locate the problem, understand the replacement suggestion and its rationale, and decide whether to adopt the modification suggestion, thereby improving the efficiency and accuracy of medical record review.

[0155] Through the above steps, intelligent error correction of conflicting entities in medical records based on medical knowledge graphs is realized, providing intelligent support for medical document review and effectively improving the quality of medical records and the accuracy of medical decisions.

[0156] A second aspect of the present invention provides an electronic device, comprising:

[0157] processor;

[0158] Memory used to store processor-executable instructions;

[0159] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0160] A third aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0161] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0162] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time intelligent review and error correction method for medical record content based on knowledge reasoning, characterized in that: include: Obtain a medical knowledge graph containing medical entity nodes, relationship edges between medical entities, and the time sequence of disease development, as well as medical record documents awaiting review; The medical record documents to be reviewed are subjected to medical entity recognition and semantic parsing. The parsed medical entities are mapped to a medical knowledge graph to generate a set of semantic representations and knowledge constraint relationships of the medical records. Extract medical events with time attributes from the semantic representation of medical records, construct a time sequence chain of medical events, match and verify the time sequence of disease development in the medical knowledge graph, and identify time inversion points. For counterfactual medical entities corresponding to the time-inverted point, remove them one by one and then perform knowledge reasoning based on the medical knowledge graph to generate counterfactual reasoning results. Compare the original reasoning results with the counterfactual reasoning results to calculate the degree of influence value, and filter out target conflicting entities whose degree of influence value exceeds the preset degree threshold. For the target conflicting entity, alternative entities that satisfy the set of temporal and knowledge constraint relationships are extracted from the medical knowledge graph. Contextual evidence of the alternative entities is retrieved from the semantic representation of the medical record to form a reasoning link and generate error correction suggestions. The error correction suggestions are then output to the medical record document to be reviewed. For counterfactual medical entities corresponding to the time-reversal points, they are removed one by one. Counterfactual reasoning results are then generated based on a medical knowledge graph. The impact value is calculated by comparing the original reasoning results with the counterfactual reasoning results. Target conflicting entities whose impact value exceeds a preset threshold are selected, including: Extract the corresponding counterfactual medical entities from the time-reversed points and construct a set of counterfactual medical entities; Knowledge reasoning is performed based on medical record semantic representation and medical knowledge graph, generating original reasoning results containing reasoning paths, extracting the medical entity nodes passed through the reasoning paths of the original reasoning results, and constructing the original reasoning dependency graph. For each counterfactual medical entity in the counterfactual medical entity set, after removing the counterfactual medical entity from the medical record semantic representation, knowledge reasoning is performed based on the medical knowledge graph to generate a counterfactual reasoning result containing the reasoning path. The medical entity nodes passed through the reasoning path of the counterfactual reasoning result are extracted to construct a counterfactual reasoning dependency graph. The original reasoning dependency graph and the counterfactual reasoning dependency graph are compared in terms of topology. The number of broken reasoning paths in the topology comparison is counted as the degree of influence. Counterfactual medical entities whose degree of influence exceeds the preset degree threshold are selected as target conflict entities.

2. The method according to claim 1, characterized in that, Obtaining a medical knowledge graph that includes medical entity nodes, relationships between medical entities, and the temporal patterns of disease development includes: Medical entity terms and their relationship descriptions are extracted from medical literature using a natural language processing model, and a set of medical entity terms and a set of medical entity relationship descriptions are constructed. Standardized medical terms and relationships between terms are extracted from the medical ontology database. Terms in the medical entity terminology set are mapped to standardized medical terms to generate standardized medical entity nodes. Map the set of medical entity relationship descriptions to the terminology relationships to construct standardized relationship edges between medical entities; Disease diagnosis and treatment process information is extracted from clinical diagnosis and treatment guidelines. Based on the standardized medical entity nodes, the temporal dependencies in the disease diagnosis and treatment process information are analyzed to generate disease development temporal patterns. The standardized medical entity nodes, standardized relationship edges between medical entities, and disease development timeline patterns are integrated to construct a medical knowledge graph.

3. The method according to claim 1, characterized in that, The medical record documents to be reviewed undergo medical entity recognition and semantic parsing. The parsed medical entities are mapped to a medical knowledge graph, generating a set of semantic representations and knowledge constraint relationships for the medical records, including: Perform entity boundary identification on the medical record documents to be reviewed and extract medical entity terms; Semantic matching is performed between medical entity terms and medical entity nodes in the medical knowledge graph. When there are multiple candidate mapping nodes for a medical entity term, the contextual text fragments of the medical entity terms in the medical record document to be reviewed are extracted. The context text fragments and the candidate mapping nodes are semantically consistent with each other in the medical knowledge graph. The medical entity nodes that pass the semantic consistency verification are selected as the mapping targets. Constructing semantic representations of medical records based on mapping targets; The relationship edges between mapping targets are extracted from the medical knowledge graph. When the relationship edge is a mutually exclusive relationship, the mapping targets connected by the relationship edge are identified as mutually exclusive entity pairs. When mutually exclusive entity pairs exist in the semantic representation of medical records, the relationship edge is marked as a conflicting relationship edge that violates the constraints. In the medical knowledge graph, multi-hop dependency paths of the mapped target are retrieved. When the starting node and ending node of the multi-hop dependency path exist in the semantic representation of the medical record but the intermediate node is missing, the multi-hop dependency path is marked as a missing dependency path that violates the constraints. Conflicting edges and missing dependency paths are treated as a set of knowledge constraint relations.

4. The method according to claim 1, characterized in that, Medical events carrying time attributes are extracted from the semantic representation of medical records, and a time sequence chain of medical events is constructed. The time sequence chain of medical events is matched and verified with the disease development time sequence pattern in the medical knowledge graph. The time sequence inversion points are identified, including: Extract medical events with time attributes from the semantic representation of medical records, and construct a time sequence chain of medical events by sorting them according to timestamp information; The causal transmission path is extracted from the disease development time sequence pattern in the medical knowledge graph, and the starting event and ending event in the causal transmission path are located in the medical event time sequence chain. When the timestamp of the termination event is earlier than the timestamp of the start event, the causal transmission path is traced back to extract the intermediate transmission events between the start and termination events, and the start event, termination event, and intermediate transmission events are marked as the causal inversion influence domain. Calculate the time interval between adjacent medical events in the medical event time sequence chain, construct the time interval into a time interval vector sequence according to the time sequence of medical events, and extract the standard time evolution pattern from the disease development time sequence pattern in the medical knowledge graph; The deviation is calculated by matching the time interval vector sequence with the standard time evolution pattern position by position. When the deviation exceeds the preset deviation threshold, the transition event defined in the corresponding position interval is extracted from the standard time evolution pattern, and the time interval with missing transition event is marked as a time evolution gap. The causal inversion influence domain and the time-series evolution gap are taken as the time-series inversion point.

5. The method according to claim 1, characterized in that, For each counterfactual medical entity in the counterfactual medical entity set, after removing the counterfactual medical entity from the medical record semantic representation, knowledge reasoning is performed based on the medical knowledge graph to generate counterfactual reasoning results containing the reasoning path, including: Select entities to be verified from the set of counterfactual medical entities; Locate the entity to be verified in the semantic representation of medical records, delete the entity to be verified and its associated relationships, and reorganize the remaining medical entities and their associated relationships into the remaining semantic representation. Extract the retained medical entities from the remaining semantic representation, map the retained medical entities to medical entity nodes in the medical knowledge graph, obtain the relationship edges between medical entities of the medical entity nodes, and perform multi-hop reasoning along the relationship edges between medical entities starting from the medical entity nodes. During multi-hop reasoning, the traversed medical entity nodes and the relationship edges between medical entities are recorded. The traversed medical entity nodes and the relationship edges between medical entities are connected in the reasoning order to form a reasoning path. The reasoning path is then encapsulated as a counterfactual reasoning result.

6. The method according to claim 1, characterized in that, For the target conflicting entity, substitute entities that satisfy the set of temporal and knowledge constraint relationships are extracted from the medical knowledge graph. Contextual evidence of the substitute entities is retrieved from the semantic representation of the medical records to form a reasoning chain and generate error correction suggestions. These error correction suggestions are then output to the medical record document to be reviewed, including: Candidate entities are extracted from the medical knowledge graph. The temporal consistency between the candidate entities and the medical event timeline is verified, and the logical compatibility between the candidate entities and the knowledge constraint relationship set is verified. Candidate entities that pass both the temporal consistency verification and the logical compatibility verification are selected as replacement entities. Locate substitute entities in the semantic representation of medical records, extract the contextual medical entities surrounding the substitute entities, retrieve the association paths between the substitute entities and the contextual medical entities from the medical knowledge graph, and use the association paths as contextual evidence. The inference link is constructed by associating and combining the substitute entity with contextual evidence. Replacement instructions are generated from the inference link. The location information of the target conflicting entity in the medical record document to be reviewed, the substitute entity, and the replacement instructions are organized into error correction suggestions. The error correction suggestions are output to the medical record document to be reviewed.

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

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

Images